path
stringlengths
7
265
concatenated_notebook
stringlengths
46
17M
notebooks/MLencoding.ipynb
###Markdown How to use the MLencoding classThis is a tutorial of how to use our MLencoding package to build encoding models a predict spikes. ###Code import warnings import numpy as np import pandas as pd import scipy.io ###Output _____no_output_____ ###Markdown Load encoding package ###Code from mlencoding import * ###Output Using Theano backend. ###Markdown 1. DataBelow we load a dataset available on CRCNS: a [Macaque M1](http://crcns.org/data-sets/movements/dream/downloading-dream) (from [Stevenston et al. 2011](http://jn.physiology.org/content/106/2/764.short)).The data has been organized in Matlab into neat arrays for easy loading here.We will soon want a single numpy array representing the external covariates, and a single numpy vector representing the neural response. The data array X will be of dimensions (n, p), where n is the number of time bins and p is the number of covariates, and the response y will be of dimensions (n, ) . We use pandas as an intermediate tool for data organizing, but it's really not necessary - if using your own data just wrangle it into numpy arrays of proper dimension. Load data ###Code m1_imported = scipy.io.loadmat('../data/m1_stevenson_2011.mat') ###Output _____no_output_____ ###Markdown 1.1 CovariatesPull into pandas dataframe. This allows us to easily access covariates by name. ###Code data = pd.DataFrame() data['time'] = m1_imported['time'][0] data['handPos_x'] = m1_imported['handPos'][0] data['handPos_y'] = m1_imported['handPos'][1] data['handVel_x'] = m1_imported['handVel'][0] data['handVel_y'] = m1_imported['handVel'][1] #### Compute more covariates/features #These will be used as the 'engineered' features for improving the GLM's performance. data['velDir'] = np.arctan2(data['handVel_y'], data['handVel_x']) data['cos_velDir'] = np.cos(data['velDir']) data['sin_velDir'] = np.sin(data['velDir']) data['speed'] = np.sqrt(data['handVel_x'].values**2+data['handVel_y'].values**2) r = np.arctan2(data['handPos_y'], data['handPos_x']) data['cos_PosDir'] = np.cos(r) data['sin_PosDir'] = np.sin(r) data['radial_Pos'] = np.sqrt(data['handPos_x'].values**2+data['handPos_y'].values**2) data.head() ###Output _____no_output_____ ###Markdown 2. Making an encoding modelWe instantiate the object like this: ###Code glm_model = MLencoding(tunemodel = 'glm') ###Output _____no_output_____ ###Markdown We can then train it on some data. Let's go for 3/4 of the data we have for some neuron. ###Code neuron_n = 1 X = data[['handPos_x','handPos_y','handVel_x','handVel_y']].values y = m1_imported['spikes'][neuron_n] n_samples = X.shape[0] threefourths = int(n_samples*3/4) X_train = X[:threefourths,:] y_train = y[:threefourths] # Now we train the model glm_model.fit(X_train,y_train) ###Output _____no_output_____ ###Markdown Let's predict the neural response on the training set. ###Code X_test = X[threefourths:,:] y_test = y[threefourths:] y_hat = glm_model.predict(X_test) ###Output _____no_output_____ ###Markdown How did we do? We can score this prediction with the class's internal function 'poisson_pseudoR2'. ###Code # The 'null model' we measure against is the mean of the train dataset. y_null = np.mean(y_train) pr2_glm = glm_model.poisson_pseudoR2(y_test, y_hat, y_null) print(pr2_glm) ###Output 0.0625913964434 ###Markdown Cross-validationLet's now obtain the predictions and scores of 10-fold cross-validation for a GLM. ###Code Y_hat, PR2s = glm_model.fit_cv(X,y, n_cv = 10, verbose = 2) ###Output ...runnning cv-fold 1 of 10 pR2: 0.0488023178838 ...runnning cv-fold 2 of 10 pR2: 0.0434830590622 ...runnning cv-fold 3 of 10 pR2: 0.0513488923378 ...runnning cv-fold 4 of 10 pR2: 0.0521074580784 ...runnning cv-fold 5 of 10 pR2: 0.0449312912574 ...runnning cv-fold 6 of 10 pR2: 0.062685886475 ...runnning cv-fold 7 of 10 pR2: 0.0459586387009 ...runnning cv-fold 8 of 10 pR2: 0.0578141187789 ...runnning cv-fold 9 of 10 pR2: 0.0523027349251 ...runnning cv-fold 10 of 10 pR2: 0.0496125678667 pR2_cv: 0.050905 (+/- 0.001765) ###Markdown Other methods: neural networks, random forest, XGBoostUsing other encoding models is as simple as this: ###Code nn_model = MLencoding(tunemodel='feedforward_nn') Y_hat, PR2s = nn_model.fit_cv(X,y, n_cv = 10, verbose = 2) ###Output ...runnning cv-fold 1 of 10 ###Markdown Predicting spikes using spike or covariate historyMLencoding supports models that also use previous covariate values to predict the current spike rate. Spike history is also supported.When you instantiate a model with the `spike_history=True` or `cov_history=True` keywords, all future calls to `fit`, `predict`, and `fit_cv` will automatically construct a new covariate matrix with additional columns. These columns represent the covariate history. This matrix is then used for fitting.Currently, covariate history columns are raised cosine basis functions. You can define how many temporal basis you want with `n_filters`, which will span the interval [0, `max_time`]. Times are measured in milliseconds. In order to perform this calculation, the model needs to know how many milliseconds are in each time bin. (Set this with `window`). ###Code xgb_history = MLencoding(tunemodel = 'xgboost', cov_history = False, spike_history=True, # We can choose! window = 50, #this dataset has 50ms time bins n_filters = 2, max_time = 250 ) xgb_history.fit_cv(X,y, verbose = 2, continuous_folds = True); ###Output ...runnning cv-fold 0 of 10 pR2: 0.172896381616 ...runnning cv-fold 1 of 10 pR2: 0.151629755677 ...runnning cv-fold 2 of 10 pR2: 0.183958679349 ...runnning cv-fold 3 of 10 pR2: 0.149697611433 ...runnning cv-fold 4 of 10 pR2: 0.127944114605 ...runnning cv-fold 5 of 10 pR2: 0.146583568384 ...runnning cv-fold 6 of 10 pR2: 0.227747587776 ...runnning cv-fold 7 of 10 pR2: 0.265500709309 ...runnning cv-fold 8 of 10 pR2: 0.275622248323 ...runnning cv-fold 9 of 10 pR2: 0.266005528721 pR2_cv: 0.196759 (+/- 0.017011) ###Markdown Here is a version that uses spike history with random folds. ###Code # First we need to set n_every > max_time/window. xgb_history_rand = MLencoding(tunemodel = 'xgboost', cov_history = False, spike_history=True, window = 50, n_filters = 2, max_time = 250, n_every = 6 ) xgb_history_rand.fit_cv(X,y, verbose = 2, continuous_folds = False); ###Output ...runnning cv-fold 1 of 10 pR2: 0.172660664684 ...runnning cv-fold 2 of 10 pR2: 0.201177093824 ...runnning cv-fold 3 of 10 pR2: 0.181089793866 ...runnning cv-fold 4 of 10 pR2: 0.148885335305 ...runnning cv-fold 5 of 10 pR2: 0.183087263289 ...runnning cv-fold 6 of 10 pR2: 0.17288721494 ...runnning cv-fold 7 of 10 pR2: 0.130874947193 ...runnning cv-fold 8 of 10 pR2: 0.175744079298 ...runnning cv-fold 9 of 10 pR2: 0.149755921527 ...runnning cv-fold 10 of 10 pR2: 0.0676141844825 pR2_cv: 0.158378 (+/- 0.011328) ###Markdown Fitting an LSTMThere's nothing special about fitting an LSTM in our implementation. Just be sure to set `spike_history=True` and `cov_history = True`, and to use continuous CV folds. ###Code lstm = MLencoding(tunemodel = 'lstm', cov_history = True, spike_history=True, # We can choose! window = 50, #this dataset has 50ms time bins n_filters = 4, max_time = 250 ) lstm.fit_cv(X,y, verbose = 2, continuous_folds = True); ###Output ...runnning cv-fold 0 of 10 pR2: 0.178190232035 ...runnning cv-fold 1 of 10 pR2: 0.169885240103 ...runnning cv-fold 2 of 10 pR2: 0.176461553019 ...runnning cv-fold 3 of 10 pR2: 0.161520848555 ...runnning cv-fold 4 of 10 pR2: 0.132098223238 ...runnning cv-fold 5 of 10 pR2: 0.149307415463 ...runnning cv-fold 6 of 10 pR2: 0.246421820011 ...runnning cv-fold 7 of 10 pR2: 0.269467384959 ...runnning cv-fold 8 of 10 pR2: 0.276576452182 ...runnning cv-fold 9 of 10 pR2: 0.281433609311 pR2_cv: 0.204136 (+/- 0.017292) ###Markdown Getting and setting model parametersTo get the current set of parameters, we can either run: ###Code nn_model.params # or nn_model.get_params() ###Output _____no_output_____ ###Markdown We can set the parameters with the `set_params` method. This method takes a dictionary, which update the current set of parameters used. ###Code nn_model.set_params({'dropout':0.3}) nn_model.params ###Output _____no_output_____ ###Markdown Hyperparameter optimization using hyperoptWe might not want the default parameters. Here's how to set some better ones ###Code from hyperopt import fmin, hp, Trials, tpe, STATUS_OK # Makes sure these are in nn_models.params, otherwise you'll get a key error space4rf = { 'dropout': hp.uniform('dropout', 0., 0.6), 'n1': hp.uniform('n1', 2,128), 'n2': hp.uniform('n2', 1,15), } #object that holds iteration results trials = Trials() #define model nn_model = MLencoding(tunemodel='feedforward_nn') #function to minimize def fnc(params): # make sure parameters are integers that need to be. params['n1'] = int(params['n1']) params['n2'] = int(params['n2']) nn_model.set_params(params) # Remember that X and y have been defined above. Y_hat, PR2s = nn_model.fit_cv(X,y, n_cv = 5, verbose = 0) # return negative since hyperopt always minimizes the function return -np.mean(pseudo_R2) ###Output _____no_output_____ ###Markdown Let's assume that our neuron 1 is a held-out neuron for parameter optimization. Let's optimize: ###Code hyperoptBest = fmin(fnc, space4rf, algo=tpe.suggest, max_evals=50, trials=trials) ###Output _____no_output_____ ###Markdown Defining your own modelsThe `MLencoding` class is flexible and can be used with predefined models as long as they have `fit` and `predict` methods.Let's build a different type of neural network, for example. ###Code my_model = Sequential() my_model.add(Dense(100, input_dim=np.shape(X)[1], init='glorot_normal', activation='relu',)) my_model.add(Dense(1,activation='softplus')) optim = Nadam() my_model.compile(loss='poisson', optimizer=optim,) my_enc = MLencoding(tunemodel = my_model) my_enc.fit_cv(X,y,n_cv=5,verbose=2); ###Output ...runnning cv-fold 1 of 5 pR2: -0.00401729001754 ...runnning cv-fold 2 of 5 pR2: -0.00440856722819 ...runnning cv-fold 3 of 5 pR2: -0.00344133554292 ...runnning cv-fold 4 of 5 pR2: -0.000698628352245 ...runnning cv-fold 5 of 5 pR2: -0.00209311949187 pR2_cv: -0.002932 (+/- 0.000610) ###Markdown How to use the MLencoding classThis is a tutorial of how to use our MLencoding package to build encoding models a predict spikes. ###Code import warnings import numpy as np import pandas as pd import scipy.io ###Output _____no_output_____ ###Markdown Load encoding package ###Code from mlencoding import * ###Output Using Theano backend. ###Markdown 1. DataBelow we load a dataset available on CRCNS: a [Macaque M1](http://crcns.org/data-sets/movements/dream/downloading-dream) (from [Stevenston et al. 2011](http://jn.physiology.org/content/106/2/764.short)).The data has been organized in Matlab into neat arrays for easy loading here.We will soon want a single numpy array representing the external covariates, and a single numpy vector representing the neural response. The data array X will be of dimensions (n, p), where n is the number of time bins and p is the number of covariates, and the response y will be of dimensions (n, ) . We use pandas as an intermediate tool for data organizing, but it's really not necessary - if using your own data just wrangle it into numpy arrays of proper dimension. Load data ###Code m1_imported = scipy.io.loadmat('../data/m1_stevenson_2011.mat') ###Output _____no_output_____ ###Markdown 1.1 CovariatesPull into pandas dataframe. This allows us to easily access covariates by name. ###Code data = pd.DataFrame() data['time'] = m1_imported['time'][0] data['handPos_x'] = m1_imported['handPos'][0] data['handPos_y'] = m1_imported['handPos'][1] data['handVel_x'] = m1_imported['handVel'][0] data['handVel_y'] = m1_imported['handVel'][1] #### Compute more covariates/features #These will be used as the 'engineered' features for improving the GLM's performance. data['velDir'] = np.arctan2(data['handVel_y'], data['handVel_x']) data['cos_velDir'] = np.cos(data['velDir']) data['sin_velDir'] = np.sin(data['velDir']) data['speed'] = np.sqrt(data['handVel_x'].values**2+data['handVel_y'].values**2) r = np.arctan2(data['handPos_y'], data['handPos_x']) data['cos_PosDir'] = np.cos(r) data['sin_PosDir'] = np.sin(r) data['radial_Pos'] = np.sqrt(data['handPos_x'].values**2+data['handPos_y'].values**2) data.head() ###Output _____no_output_____ ###Markdown 2. Making an encoding modelWe instantiate the object like this: ###Code glm_model = MLencoding(tunemodel = 'glm') ###Output _____no_output_____ ###Markdown We can then train it on some data. Let's go for 3/4 of the data we have for some neuron. ###Code neuron_n = 1 X = data[['handPos_x','handPos_y','handVel_x','handVel_y']].values y = m1_imported['spikes'][neuron_n] n_samples = X.shape[0] threefourths = int(n_samples*3/4) X_train = X[:threefourths,:] y_train = y[:threefourths] # Now we train the model glm_model.fit(X_train,y_train) ###Output _____no_output_____ ###Markdown Let's predict the neural response on the training set. ###Code X_test = X[threefourths:,:] y_test = y[threefourths:] y_hat = glm_model.predict(X_test) ###Output _____no_output_____ ###Markdown How did we do? We can score this prediction with the class's internal function 'poisson_pseudoR2'. ###Code # The 'null model' we measure against is the mean of the train dataset. y_null = np.mean(y_train) pr2_glm = glm_model.poisson_pseudoR2(y_test, y_hat, y_null) print(pr2_glm) ###Output 0.0625913964434 ###Markdown Cross-validationLet's now obtain the predictions and scores of 10-fold cross-validation for a GLM. ###Code Y_hat, PR2s = glm_model.fit_cv(X,y, n_cv = 10, verbose = 2) ###Output ...runnning cv-fold 1 of 10 pR2: 0.0488023178838 ...runnning cv-fold 2 of 10 pR2: 0.0434830590622 ...runnning cv-fold 3 of 10 pR2: 0.0513488923378 ...runnning cv-fold 4 of 10 pR2: 0.0521074580784 ...runnning cv-fold 5 of 10 pR2: 0.0449312912574 ...runnning cv-fold 6 of 10 pR2: 0.062685886475 ...runnning cv-fold 7 of 10 pR2: 0.0459586387009 ...runnning cv-fold 8 of 10 pR2: 0.0578141187789 ...runnning cv-fold 9 of 10 pR2: 0.0523027349251 ...runnning cv-fold 10 of 10 pR2: 0.0496125678667 pR2_cv: 0.050905 (+/- 0.001765) ###Markdown Other methods: neural networks, random forest, XGBoostUsing other encoding models is as simple as this: ###Code nn_model = MLencoding(tunemodel='feedforward_nn') Y_hat, PR2s = nn_model.fit_cv(X,y, n_cv = 10, verbose = 2) ###Output ...runnning cv-fold 1 of 10 ###Markdown Predicting spikes using spike or covariate historyMLencoding supports models that also use previous covariate values to predict the current spike rate. Spike history is also supported.When you instantiate a model with the `spike_history=True` or `cov_history=True` keywords, all future calls to `fit`, `predict`, and `fit_cv` will automatically construct a new covariate matrix with additional columns. These columns represent the covariate history. This matrix is then used for fitting.Currently, covariate history columns are raised cosine basis functions. You can define how many temporal basis you want with `n_filters`, which will span the interval [0, `max_time`]. Times are measured in milliseconds. In order to perform this calculation, the model needs to know how many milliseconds are in each time bin. (Set this with `window`). ###Code xgb_history = MLencoding(tunemodel = 'xgboost', cov_history = False, spike_history=True, # We can choose! window = 50, #this dataset has 50ms time bins n_filters = 2, max_time = 250 ) xgb_history.fit_cv(X,y, verbose = 2, continuous_folds = True); ###Output ...runnning cv-fold 0 of 10 pR2: 0.172896381616 ...runnning cv-fold 1 of 10 pR2: 0.151629755677 ...runnning cv-fold 2 of 10 pR2: 0.183958679349 ...runnning cv-fold 3 of 10 pR2: 0.149697611433 ...runnning cv-fold 4 of 10 pR2: 0.127944114605 ...runnning cv-fold 5 of 10 pR2: 0.146583568384 ...runnning cv-fold 6 of 10 pR2: 0.227747587776 ...runnning cv-fold 7 of 10 pR2: 0.265500709309 ...runnning cv-fold 8 of 10 pR2: 0.275622248323 ...runnning cv-fold 9 of 10 pR2: 0.266005528721 pR2_cv: 0.196759 (+/- 0.017011) ###Markdown Here is a version that uses spike history with random folds. ###Code # First we need to set n_every > max_time/window. xgb_history_rand = MLencoding(tunemodel = 'xgboost', cov_history = False, spike_history=True, window = 50, n_filters = 2, max_time = 250, n_every = 6 ) xgb_history_rand.fit_cv(X,y, verbose = 2, continuous_folds = False); ###Output ...runnning cv-fold 1 of 10 pR2: 0.172660664684 ...runnning cv-fold 2 of 10 pR2: 0.201177093824 ...runnning cv-fold 3 of 10 pR2: 0.181089793866 ...runnning cv-fold 4 of 10 pR2: 0.148885335305 ...runnning cv-fold 5 of 10 pR2: 0.183087263289 ...runnning cv-fold 6 of 10 pR2: 0.17288721494 ...runnning cv-fold 7 of 10 pR2: 0.130874947193 ...runnning cv-fold 8 of 10 pR2: 0.175744079298 ...runnning cv-fold 9 of 10 pR2: 0.149755921527 ...runnning cv-fold 10 of 10 pR2: 0.0676141844825 pR2_cv: 0.158378 (+/- 0.011328) ###Markdown Fitting an LSTMThere's nothing special about fitting an LSTM in our implementation. Just be sure to set `spike_history=True` and `cov_history = True`, and to use continuous CV folds. ###Code lstm = MLencoding(tunemodel = 'lstm', cov_history = True, spike_history=True, # We can choose! window = 50, #this dataset has 50ms time bins n_filters = 4, max_time = 250 ) lstm.fit_cv(X,y, verbose = 2, continuous_folds = True); ###Output ...runnning cv-fold 0 of 10 pR2: 0.178190232035 ...runnning cv-fold 1 of 10 pR2: 0.169885240103 ...runnning cv-fold 2 of 10 pR2: 0.176461553019 ...runnning cv-fold 3 of 10 pR2: 0.161520848555 ...runnning cv-fold 4 of 10 pR2: 0.132098223238 ...runnning cv-fold 5 of 10 pR2: 0.149307415463 ...runnning cv-fold 6 of 10 pR2: 0.246421820011 ...runnning cv-fold 7 of 10 pR2: 0.269467384959 ...runnning cv-fold 8 of 10 pR2: 0.276576452182 ...runnning cv-fold 9 of 10 pR2: 0.281433609311 pR2_cv: 0.204136 (+/- 0.017292) ###Markdown Getting and setting model parametersTo get the current set of parameters, we can either run: ###Code nn_model.params # or nn_model.get_params() ###Output _____no_output_____ ###Markdown We can set the parameters with the `set_params` method. This method takes a dictionary, which update the current set of parameters used. ###Code nn_model.set_params({'dropout':0.3}) nn_model.params ###Output _____no_output_____ ###Markdown Hyperparameter optimization using hyperoptWe might not want the default parameters. Here's how to set some better ones ###Code from hyperopt import fmin, hp, Trials, tpe, STATUS_OK # Makes sure these are in nn_models.params, otherwise you'll get a key error space4rf = { 'dropout': hp.uniform('dropout', 0., 0.6), 'n1': hp.uniform('n1', 2,128), 'n2': hp.uniform('n2', 1,15), } #object that holds iteration results trials = Trials() #define model nn_model = MLencoding(tunemodel='feedforward_nn') #function to minimize def fnc(params): # make sure parameters are integers that need to be. params['n1'] = int(params['n1']) params['n2'] = int(params['n2']) nn_model.set_params(params) # Remember that X and y have been defined above. Y_hat, PR2s = nn_model.fit_cv(X,y, n_cv = 5, verbose = 0) # return negative since hyperopt always minimizes the function return -np.mean(pseudo_R2) ###Output _____no_output_____ ###Markdown Let's assume that our neuron 1 is a held-out neuron for parameter optimization. Let's optimize: ###Code hyperoptBest = fmin(fnc, space4rf, algo=tpe.suggest, max_evals=50, trials=trials) ###Output _____no_output_____ ###Markdown Defining your own modelsThe `MLencoding` class is flexible and can be used with predefined models as long as they have `fit` and `predict` methods.Let's build a different type of neural network, for example. ###Code my_model = Sequential() my_model.add(Dense(100, input_dim=np.shape(X)[1], init='glorot_normal', activation='relu',)) my_model.add(Dense(1,activation='softplus')) optim = Nadam() my_model.compile(loss='poisson', optimizer=optim,) my_enc = MLencoding(tunemodel = my_model) my_enc.fit_cv(X,y,n_cv=5,verbose=2); ###Output ...runnning cv-fold 1 of 5 pR2: -0.00401729001754 ...runnning cv-fold 2 of 5 pR2: -0.00440856722819 ...runnning cv-fold 3 of 5 pR2: -0.00344133554292 ...runnning cv-fold 4 of 5 pR2: -0.000698628352245 ...runnning cv-fold 5 of 5 pR2: -0.00209311949187 pR2_cv: -0.002932 (+/- 0.000610)
starter_notebook_reverse_training_Setswana2.ipynb
###Markdown Masakhane - Reverse Machine Translation for African Languages (Using JoeyNMT) > NB> - The purpose of this Notebook is to build models that translate African languages(target language) *into* English(source language). This will allow us to in future be able to make translations from one African language to the other. If you'd like to translate *from* English, please use [this](https://github.com/masakhane-io/masakhane-mt/blob/master/starter_notebook.ipynb) starter notebook instead.> - We call this reverse training because normally we build models that make translations from the source language(English) to the target language. But in this case we are doing the reverse; building models that make translations from the target language to the source(English) Note before beginning: - The idea is that you should be able to make minimal changes to this in order to get SOME result for your own translation corpus. - The tl;dr: Go to the **"TODO"** comments which will tell you what to update to get up and running - If you actually want to have a clue what you're doing, read the text and peek at the links - With 100 epochs, it should take around 7 hours to run in Google Colab - Once you've gotten a result for your language, please attach and email your notebook that generated it to [email protected] - If you care enough and get a chance, doing a brief background on your language would be amazing. See examples in [(Martinus, 2019)](https://arxiv.org/abs/1906.05685) Retrieve your data & make a parallel corpusIf you are wanting to use the JW300 data referenced on the Masakhane website or in our GitHub repo, you can use `opus-tools` to convert the data into a convenient format. `opus_read` from that package provides a convenient tool for reading the native aligned XML files and to convert them to TMX format. The tool can also be used to fetch relevant files from OPUS on the fly and to filter the data as necessary. [Read the documentation](https://pypi.org/project/opustools-pkg/) for more details.Once you have your corpus files in TMX format (an xml structure which will include the sentences in your target language and your source language in a single file), we recommend reading them into a pandas dataframe. Thankfully, Jade wrote a silly `tmx2dataframe` package which converts your tmx file to a pandas dataframe. Submitted by Tebello Lebesa 2388016Sumbitted by Korstiaan Wapenaar 1492459 ###Code from google.colab import drive drive.mount('/content/drive') # TODO: Set your source and target languages. Keep in mind, these traditionally use language codes as found here: # These will also become the suffix's of all vocab and corpus files used throughout import os source_language = "en" target_language = "tn" lc = False # If True, lowercase the data. seed = 42 # Random seed for shuffling. tag = "baseline" # Give a unique name to your folder - this is to ensure you don't rewrite any models you've already submitted os.environ["src"] = source_language # Sets them in bash as well, since we often use bash scripts os.environ["tgt"] = target_language os.environ["tag"] = tag # This will save it to a folder in our gdrive instead! !mkdir -p "/content/drive/My Drive/masakhane/$tgt-$src-$tag" os.environ["gdrive_path"] = "/content/drive/My Drive/masakhane/%s-%s-%s" % (target_language, source_language, tag) !echo $gdrive_path # Install opus-tools ! pip install opustools-pkg # Downloading our corpus ! opus_read -d JW300 -s $src -t $tgt -wm moses -w jw300.$src jw300.$tgt -q # extract the corpus file ! gunzip JW300_latest_xml_$src-$tgt.xml.gz # Download the global test set. ! wget https://raw.githubusercontent.com/juliakreutzer/masakhane/master/jw300_utils/test/test.en-any.en # And the specific test set for this language pair. os.environ["trg"] = target_language os.environ["src"] = source_language ! wget https://raw.githubusercontent.com/juliakreutzer/masakhane/master/jw300_utils/test/test.en-$trg.en ! mv test.en-$trg.en test.en ! wget https://raw.githubusercontent.com/juliakreutzer/masakhane/master/jw300_utils/test/test.en-$trg.$trg ! mv test.en-$trg.$trg test.$trg # Read the test data to filter from train and dev splits. # Store english portion in set for quick filtering checks. en_test_sents = set() filter_test_sents = "test.en-any.en" j = 0 with open(filter_test_sents) as f: for line in f: en_test_sents.add(line.strip()) j += 1 print('Loaded {} global test sentences to filter from the training/dev data.'.format(j)) import pandas as pd # TMX file to dataframe source_file = 'jw300.' + source_language target_file = 'jw300.' + target_language source = [] target = [] skip_lines = [] # Collect the line numbers of the source portion to skip the same lines for the target portion. with open(source_file) as f: for i, line in enumerate(f): # Skip sentences that are contained in the test set. if line.strip() not in en_test_sents: source.append(line.strip()) else: skip_lines.append(i) with open(target_file) as f: for j, line in enumerate(f): # Only add to corpus if corresponding source was not skipped. if j not in skip_lines: target.append(line.strip()) print('Loaded data and skipped {}/{} lines since contained in test set.'.format(len(skip_lines), i)) df = pd.DataFrame(zip(source, target), columns=['source_sentence', 'target_sentence']) # if you get TypeError: data argument can't be an iterator is because of your zip version run this below #df = pd.DataFrame(list(zip(source, target)), columns=['source_sentence', 'target_sentence']) df.head(3) ###Output Loaded data and skipped 5241/909627 lines since contained in test set. ###Markdown Pre-processing and exportIt is generally a good idea to remove duplicate translations and conflicting translations from the corpus. In practice, these public corpora include some number of these that need to be cleaned.In addition we will split our data into dev/test/train and export to the filesystem. ###Code # drop duplicate translations df_pp = df.drop_duplicates() # drop conflicting translations # (this is optional and something that you might want to comment out # depending on the size of your corpus) df_pp.drop_duplicates(subset='source_sentence', inplace=True) df_pp.drop_duplicates(subset='target_sentence', inplace=True) # Shuffle the data to remove bias in dev set selection. df_pp = df_pp.sample(frac=1, random_state=seed).reset_index(drop=True) # Install fuzzy wuzzy to remove "almost duplicate" sentences in the # test and training sets. ! pip install fuzzywuzzy ! pip install python-Levenshtein import time from fuzzywuzzy import process import numpy as np from os import cpu_count from functools import partial from multiprocessing import Pool # reset the index of the training set after previous filtering df_pp.reset_index(drop=False, inplace=True) # Remove samples from the training data set if they "almost overlap" with the # samples in the test set. # Filtering function. Adjust pad to narrow down the candidate matches to # within a certain length of characters of the given sample. def fuzzfilter(sample, candidates, pad): candidates = [x for x in candidates if len(x) <= len(sample)+pad and len(x) >= len(sample)-pad] if len(candidates) > 0: return process.extractOne(sample, candidates)[1] else: return np.nan # start_time = time.time() # ### iterating over pandas dataframe rows is not recomended, let use multi processing to apply the function # with Pool(cpu_count()-1) as pool: # scores = pool.map(partial(fuzzfilter, candidates=list(en_test_sents), pad=5), df_pp['source_sentence']) # hours, rem = divmod(time.time() - start_time, 3600) # minutes, seconds = divmod(rem, 60) # print("done in {}h:{}min:{}seconds".format(hours, minutes, seconds)) # # Filter out "almost overlapping samples" # df_pp = df_pp.assign(scores=scores) # df_pp = df_pp[df_pp['scores'] < 95] # This section does the split between train/dev for the parallel corpora then saves them as separate files # We use 1000 dev test and the given test set. import csv # Do the split between dev/train and create parallel corpora num_dev_patterns = 1000 # Optional: lower case the corpora - this will make it easier to generalize, but without proper casing. if lc: # Julia: making lowercasing optional df_pp["source_sentence"] = df_pp["source_sentence"].str.lower() df_pp["target_sentence"] = df_pp["target_sentence"].str.lower() # Julia: test sets are already generated dev = df_pp.tail(num_dev_patterns) # Herman: Error in original stripped = df_pp.drop(df_pp.tail(num_dev_patterns).index) with open("train."+source_language, "w") as src_file, open("train."+target_language, "w") as trg_file: for index, row in stripped.iterrows(): src_file.write(row["source_sentence"]+"\n") trg_file.write(row["target_sentence"]+"\n") with open("dev."+source_language, "w") as src_file, open("dev."+target_language, "w") as trg_file: for index, row in dev.iterrows(): src_file.write(row["source_sentence"]+"\n") trg_file.write(row["target_sentence"]+"\n") #stripped[["source_sentence"]].to_csv("train."+source_language, header=False, index=False) # Herman: Added `header=False` everywhere #stripped[["target_sentence"]].to_csv("train."+target_language, header=False, index=False) # Julia: Problematic handling of quotation marks. #dev[["source_sentence"]].to_csv("dev."+source_language, header=False, index=False) #dev[["target_sentence"]].to_csv("dev."+target_language, header=False, index=False) # Doublecheck the format below. There should be no extra quotation marks or weird characters. ! head train.* ! head dev.* ###Output ==> train.en <== When a Christian dies , his material wealth is of no more value to him than Jesus ’ garment was to him when he died . “ When I think of the five years I spent in the bush , killing people and being shot at , I feel pretty stupid , ” said the fighter . The two females , named Owalla and Durga , were introduced to the Pilanesberg Reserve of Bophuthatswana in 1982 . We deliberately allowed our children to see that just as they were struggling with the anxieties of youth , we were struggling with the anxieties of adults . ” ( Matthew 24 : 3-14 ; Luke 21 : 11 ) — 12 / 15 , page 11 . She was baptized in 1982 . A romantic attraction to someone outside the marriage could be an indication that a husband and a wife are not attentive to each other ’ s needs . Jesus established the limit to the honor that ought to be rendered to others when he told his disciples : “ Do not you be called Rabbi , for one is your teacher , whereas all you are brothers . “ If sin , sickness , and death were understood as nothingness , they would disappear . ” — Science and Health With Key to the Scriptures . Create an atmosphere that will allow your child to talk about death and its meaning . ==> train.st <== Ha Mokreste a e - shoa , leruo la hae ha e be la bohlokoa joalokaha seaparo sa Jesu e sa ka ea e - ba sa bohlokoa ho eena ha a e - shoa . Lesole leo le re : “ Ha ke nahana ka lilemo tse hlano tseo ke li qetileng morung , ke bolaea batho ’ me le ’ na ke thunngoa , ke ikutloa ke le sethoto sa lithoto . Litlou tse peli tse tšehali tse bitsoang Owalla le Durga li ile tsa isoa Pilanesberg Reserve ea Bophuthatswana ka 1982 . Ka morero re ile ra lumella hore bana ba rōna ba bone hore joalokaha ba ne ba loana le matšoenyeho a bocha , le rōna re ne re loana le matšoenyeho a batho ba baholo . ” ( Matheu 24 : 3 - 14 ; Luka 21 : 11 ) — 12 / 15 , leqepheng la 11 . O ile a kolobetsoa ka 1982 . Ho rata motho e mong ka ntle ho molekane oa hao e ka ’ na ea e - ba pontšo ea hore monna le mosali ha ba hlokomelane . Jesu o ile a bontša hore na tlhompho e fuoang ba bang e lokela ho fella hokae ha a re ho barutuoa ba hae : “ Le se ke la bitsoa Rabi , kaha mosuoe oa lōna o mong , athe lōna bohle le bara ba motho . Haeba sebe , ho kula le lefu li ne li nkoa e se letho , li ne li tla fela . ” — Science and Health With Key to the Scriptures . Etsang hore bana ba lōna ba phutholohe ha ba bua ka lefu . ==> dev.en <== ( Luke 23 : 43 ) And they will never need to die at all ! In the future , after the war of Armageddon , anointed Christians will become Christ ’ s bride . As with most aspects of child training , example is an effective teacher . I was pioneering in Melbourne at the time and living at the Society ’ s literature depot . A Miracle in New York ? Thus , they may expect to survive the most calamitous time of distress ever to strike the nations. — Daniel 12 : 1 ; Matthew 24 : 13 , 21 , 22 . They were not to look back , but Lot ’ s wife did so , perhaps longing for the material things left behind . Since ancient times , people observed these changes and attributed great meaning to them . Then , discuss what your teenager would do . 9 Whatever the topics , our conversations will build others up if they adhere to the apostle Paul ’ s admonition to the congregation in Philippi . ==> dev.st <== ( Luka 23 : 43 ) ’ Me ba ke ke ba hlola ba e - shoa le ka mohla ! Nakong e tlang , ka mor’a ntoa ea Armagedone , Bakreste ba tlotsitsoeng e tla ba monyaluoa oa Kreste . Tsela e atlehang ka ho fetisisa ea ho koetlisa bana ke ha batsoali ba ba behela mohlala o motle . Ke ne ke bula maliboho Melbourne ka nako eo ’ me ke phela motebong oa libuka oa Mokhatlo . Ho Etsahetse Mohlolo New York ? Kahoo , li ka lebella ho pholoha nako e mahlonoko ka ho fetisisa ea tlokotsi e kileng ea oela lichaba . — Daniele 12 : 1 ; Mattheu 24 : 13 , 21 , 22 . Ba ne ba sa tlameha ho hetla , empa mosali oa Lota o ile a hetla , mohlomong a laba - labela lintho tse bonahalang tse setseng morao . Ho tloha mehleng ea boholo - holo , batho ba ne ba ithuta liphetoho tsena ebe ba re ho na le ntho e khōlō eo li e bolelang . Mo botse hore na eena o ne a ka etsa’ng . 9 Ho sa tsotellehe litaba tseo re buang ka tsona , meqoqo ea rōna e tla haha ba bang haeba e lumellana le keletso ea moapostola Pauluse e eang phuthehong ea Filipi . ###Markdown --- Installation of JoeyNMTJoeyNMT is a simple, minimalist NMT package which is useful for learning and teaching. Check out the documentation for JoeyNMT [here](https://joeynmt.readthedocs.io) ###Code # Install JoeyNMT ! git clone https://github.com/joeynmt/joeynmt.git ! cd joeynmt; pip3 install . # Install Pytorch with GPU support v1.7.1. ! pip install torch==1.9.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html ###Output Cloning into 'joeynmt'... remote: Enumerating objects: 3224, done. remote: Counting objects: 100% (273/273), done. remote: Compressing objects: 100% (139/139), done. remote: Total 3224 (delta 157), reused 206 (delta 134), pack-reused 2951 Receiving objects: 100% (3224/3224), 8.17 MiB | 15.02 MiB/s, done. Resolving deltas: 100% (2186/2186), done. Processing /content/joeynmt  DEPRECATION: A future pip version will change local packages to be built in-place without first copying to a temporary directory. We recommend you use --use-feature=in-tree-build to test your packages with this new behavior before it becomes the default. pip 21.3 will remove support for this functionality. You can find discussion regarding this at https://github.com/pypa/pip/issues/7555. Requirement already satisfied: future in /usr/local/lib/python3.7/dist-packages (from joeynmt==1.3) (0.16.0) Requirement already satisfied: pillow in /usr/local/lib/python3.7/dist-packages (from joeynmt==1.3) (7.1.2) Requirement already satisfied: numpy>=1.19.5 in /usr/local/lib/python3.7/dist-packages (from joeynmt==1.3) (1.19.5) Requirement already satisfied: setuptools>=41.0.0 in /usr/local/lib/python3.7/dist-packages (from joeynmt==1.3) (57.4.0) Requirement already satisfied: torch>=1.9.0 in /usr/local/lib/python3.7/dist-packages (from joeynmt==1.3) (1.9.0+cu111) Requirement already satisfied: tensorboard>=1.15 in /usr/local/lib/python3.7/dist-packages (from joeynmt==1.3) (2.6.0) Requirement already satisfied: torchtext>=0.10.0 in /usr/local/lib/python3.7/dist-packages (from joeynmt==1.3) (0.10.0) Collecting sacrebleu>=2.0.0 Downloading sacrebleu-2.0.0-py3-none-any.whl (90 kB)  |████████████████████████████████| 90 kB 3.8 MB/s [?25hCollecting subword-nmt Downloading subword_nmt-0.3.7-py2.py3-none-any.whl (26 kB) Requirement already satisfied: matplotlib in /usr/local/lib/python3.7/dist-packages (from joeynmt==1.3) (3.2.2) Requirement already satisfied: seaborn in /usr/local/lib/python3.7/dist-packages (from joeynmt==1.3) (0.11.2) Collecting pyyaml>=5.1 Downloading PyYAML-5.4.1-cp37-cp37m-manylinux1_x86_64.whl (636 kB)  |████████████████████████████████| 636 kB 19.6 MB/s [?25hCollecting pylint>=2.9.6 Downloading pylint-2.11.1-py3-none-any.whl (392 kB)  |████████████████████████████████| 392 kB 38.0 MB/s [?25hCollecting six==1.12 Downloading six-1.12.0-py2.py3-none-any.whl (10 kB) Collecting wrapt==1.11.1 Downloading wrapt-1.11.1.tar.gz (27 kB) Collecting typing-extensions>=3.10.0 Downloading typing_extensions-3.10.0.2-py3-none-any.whl (26 kB) Collecting astroid<2.9,>=2.8.0 Downloading astroid-2.8.2-py3-none-any.whl (246 kB)  |████████████████████████████████| 246 kB 42.7 MB/s [?25hRequirement already satisfied: toml>=0.7.1 in /usr/local/lib/python3.7/dist-packages (from pylint>=2.9.6->joeynmt==1.3) (0.10.2) Collecting mccabe<0.7,>=0.6 Downloading mccabe-0.6.1-py2.py3-none-any.whl (8.6 kB) Collecting platformdirs>=2.2.0 Downloading platformdirs-2.4.0-py3-none-any.whl (14 kB) Collecting isort<6,>=4.2.5 Downloading isort-5.9.3-py3-none-any.whl (106 kB)  |████████████████████████████████| 106 kB 51.6 MB/s [?25hCollecting typed-ast<1.5,>=1.4.0 Downloading typed_ast-1.4.3-cp37-cp37m-manylinux1_x86_64.whl (743 kB)  |████████████████████████████████| 743 kB 39.4 MB/s [?25hCollecting lazy-object-proxy>=1.4.0 Downloading lazy_object_proxy-1.6.0-cp37-cp37m-manylinux1_x86_64.whl (55 kB)  |████████████████████████████████| 55 kB 3.0 MB/s [?25hCollecting portalocker Downloading portalocker-2.3.2-py2.py3-none-any.whl (15 kB) Requirement already satisfied: regex in /usr/local/lib/python3.7/dist-packages (from sacrebleu>=2.0.0->joeynmt==1.3) (2019.12.20) Collecting colorama Downloading colorama-0.4.4-py2.py3-none-any.whl (16 kB) Requirement already satisfied: tabulate>=0.8.9 in /usr/local/lib/python3.7/dist-packages (from sacrebleu>=2.0.0->joeynmt==1.3) (0.8.9) Requirement already satisfied: requests<3,>=2.21.0 in /usr/local/lib/python3.7/dist-packages (from tensorboard>=1.15->joeynmt==1.3) (2.23.0) Requirement already satisfied: grpcio>=1.24.3 in /usr/local/lib/python3.7/dist-packages (from tensorboard>=1.15->joeynmt==1.3) (1.41.0) Requirement already satisfied: tensorboard-plugin-wit>=1.6.0 in /usr/local/lib/python3.7/dist-packages (from tensorboard>=1.15->joeynmt==1.3) (1.8.0) Requirement already satisfied: google-auth<2,>=1.6.3 in /usr/local/lib/python3.7/dist-packages (from tensorboard>=1.15->joeynmt==1.3) (1.35.0) Requirement already satisfied: wheel>=0.26 in /usr/local/lib/python3.7/dist-packages (from tensorboard>=1.15->joeynmt==1.3) (0.37.0) Requirement already satisfied: google-auth-oauthlib<0.5,>=0.4.1 in /usr/local/lib/python3.7/dist-packages (from tensorboard>=1.15->joeynmt==1.3) (0.4.6) Requirement already satisfied: werkzeug>=0.11.15 in /usr/local/lib/python3.7/dist-packages (from tensorboard>=1.15->joeynmt==1.3) (1.0.1) Requirement already satisfied: absl-py>=0.4 in /usr/local/lib/python3.7/dist-packages (from tensorboard>=1.15->joeynmt==1.3) (0.12.0) Requirement already satisfied: tensorboard-data-server<0.7.0,>=0.6.0 in /usr/local/lib/python3.7/dist-packages (from tensorboard>=1.15->joeynmt==1.3) (0.6.1) Requirement already satisfied: markdown>=2.6.8 in /usr/local/lib/python3.7/dist-packages (from tensorboard>=1.15->joeynmt==1.3) (3.3.4) Requirement already satisfied: protobuf>=3.6.0 in /usr/local/lib/python3.7/dist-packages (from tensorboard>=1.15->joeynmt==1.3) (3.17.3) Requirement already satisfied: rsa<5,>=3.1.4 in /usr/local/lib/python3.7/dist-packages (from google-auth<2,>=1.6.3->tensorboard>=1.15->joeynmt==1.3) (4.7.2) Requirement already satisfied: cachetools<5.0,>=2.0.0 in /usr/local/lib/python3.7/dist-packages (from google-auth<2,>=1.6.3->tensorboard>=1.15->joeynmt==1.3) (4.2.4) Requirement already satisfied: pyasn1-modules>=0.2.1 in /usr/local/lib/python3.7/dist-packages (from google-auth<2,>=1.6.3->tensorboard>=1.15->joeynmt==1.3) (0.2.8) Requirement already satisfied: requests-oauthlib>=0.7.0 in /usr/local/lib/python3.7/dist-packages (from google-auth-oauthlib<0.5,>=0.4.1->tensorboard>=1.15->joeynmt==1.3) (1.3.0) Requirement already satisfied: importlib-metadata in /usr/local/lib/python3.7/dist-packages (from markdown>=2.6.8->tensorboard>=1.15->joeynmt==1.3) (4.8.1) Requirement already satisfied: pyasn1<0.5.0,>=0.4.6 in /usr/local/lib/python3.7/dist-packages (from pyasn1-modules>=0.2.1->google-auth<2,>=1.6.3->tensorboard>=1.15->joeynmt==1.3) (0.4.8) Requirement already satisfied: urllib3!=1.25.0,!=1.25.1,<1.26,>=1.21.1 in /usr/local/lib/python3.7/dist-packages (from requests<3,>=2.21.0->tensorboard>=1.15->joeynmt==1.3) (1.24.3) Requirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.7/dist-packages (from requests<3,>=2.21.0->tensorboard>=1.15->joeynmt==1.3) (2021.5.30) Requirement already satisfied: chardet<4,>=3.0.2 in /usr/local/lib/python3.7/dist-packages (from requests<3,>=2.21.0->tensorboard>=1.15->joeynmt==1.3) (3.0.4) Requirement already satisfied: idna<3,>=2.5 in /usr/local/lib/python3.7/dist-packages (from requests<3,>=2.21.0->tensorboard>=1.15->joeynmt==1.3) (2.10) Requirement already satisfied: oauthlib>=3.0.0 in /usr/local/lib/python3.7/dist-packages (from requests-oauthlib>=0.7.0->google-auth-oauthlib<0.5,>=0.4.1->tensorboard>=1.15->joeynmt==1.3) (3.1.1) Requirement already satisfied: tqdm in /usr/local/lib/python3.7/dist-packages (from torchtext>=0.10.0->joeynmt==1.3) (4.62.3) Requirement already satisfied: zipp>=0.5 in /usr/local/lib/python3.7/dist-packages (from importlib-metadata->markdown>=2.6.8->tensorboard>=1.15->joeynmt==1.3) (3.6.0) Requirement already satisfied: cycler>=0.10 in /usr/local/lib/python3.7/dist-packages (from matplotlib->joeynmt==1.3) (0.10.0) Requirement already satisfied: python-dateutil>=2.1 in /usr/local/lib/python3.7/dist-packages (from matplotlib->joeynmt==1.3) (2.8.2) Requirement already satisfied: pyparsing!=2.0.4,!=2.1.2,!=2.1.6,>=2.0.1 in /usr/local/lib/python3.7/dist-packages (from matplotlib->joeynmt==1.3) (2.4.7) Requirement already satisfied: kiwisolver>=1.0.1 in /usr/local/lib/python3.7/dist-packages (from matplotlib->joeynmt==1.3) (1.3.2) Requirement already satisfied: scipy>=1.0 in /usr/local/lib/python3.7/dist-packages (from seaborn->joeynmt==1.3) (1.4.1) Requirement already satisfied: pandas>=0.23 in /usr/local/lib/python3.7/dist-packages (from seaborn->joeynmt==1.3) (1.1.5) Requirement already satisfied: pytz>=2017.2 in /usr/local/lib/python3.7/dist-packages (from pandas>=0.23->seaborn->joeynmt==1.3) (2018.9) Building wheels for collected packages: joeynmt, wrapt Building wheel for joeynmt (setup.py) ... [?25l[?25hdone Created wheel for joeynmt: filename=joeynmt-1.3-py3-none-any.whl size=86029 sha256=9cf2ca4c26274d4054c39fdc30dfaf0ba310c6316021ec2f7424258690d7aec8 Stored in directory: /tmp/pip-ephem-wheel-cache-gciht89z/wheels/0a/f4/bf/6c9d3b8efbfece6cd209f865be37382b02e7c3584df2e28ca4 Building wheel for wrapt (setup.py) ... [?25l[?25hdone Created wheel for wrapt: filename=wrapt-1.11.1-cp37-cp37m-linux_x86_64.whl size=68437 sha256=2abf7cfa0208c10d77e60b78de85b364d802f8b1038841523467211057b0ac37 Stored in directory: /root/.cache/pip/wheels/4e/58/9d/da8bad4545585ca52311498ff677647c95c7b690b3040171f8 Successfully built joeynmt wrapt Installing collected packages: typing-extensions, six, wrapt, typed-ast, lazy-object-proxy, portalocker, platformdirs, mccabe, isort, colorama, astroid, subword-nmt, sacrebleu, pyyaml, pylint, joeynmt Attempting uninstall: typing-extensions Found existing installation: typing-extensions 3.7.4.3 Uninstalling typing-extensions-3.7.4.3: Successfully uninstalled typing-extensions-3.7.4.3 Attempting uninstall: six Found existing installation: six 1.15.0 Uninstalling six-1.15.0: Successfully uninstalled six-1.15.0 Attempting uninstall: wrapt Found existing installation: wrapt 1.12.1 Uninstalling wrapt-1.12.1: Successfully uninstalled wrapt-1.12.1 Attempting uninstall: pyyaml Found existing installation: PyYAML 3.13 Uninstalling PyYAML-3.13: Successfully uninstalled PyYAML-3.13 ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. tensorflow 2.6.0 requires six~=1.15.0, but you have six 1.12.0 which is incompatible. tensorflow 2.6.0 requires typing-extensions~=3.7.4, but you have typing-extensions 3.10.0.2 which is incompatible. tensorflow 2.6.0 requires wrapt~=1.12.1, but you have wrapt 1.11.1 which is incompatible. google-colab 1.0.0 requires six~=1.15.0, but you have six 1.12.0 which is incompatible. google-api-python-client 1.12.8 requires six<2dev,>=1.13.0, but you have six 1.12.0 which is incompatible. google-api-core 1.26.3 requires six>=1.13.0, but you have six 1.12.0 which is incompatible. datascience 0.10.6 requires folium==0.2.1, but you have folium 0.8.3 which is incompatible. albumentations 0.1.12 requires imgaug<0.2.7,>=0.2.5, but you have imgaug 0.2.9 which is incompatible. Successfully installed astroid-2.8.2 colorama-0.4.4 isort-5.9.3 joeynmt-1.3 lazy-object-proxy-1.6.0 mccabe-0.6.1 platformdirs-2.4.0 portalocker-2.3.2 pylint-2.11.1 pyyaml-5.4.1 sacrebleu-2.0.0 six-1.12.0 subword-nmt-0.3.7 typed-ast-1.4.3 typing-extensions-3.10.0.2 wrapt-1.11.1 Looking in links: https://download.pytorch.org/whl/torch_stable.html Collecting torch==1.8.0+cu101 Downloading https://download.pytorch.org/whl/cu101/torch-1.8.0%2Bcu101-cp37-cp37m-linux_x86_64.whl (763.5 MB)  |████████████████████████████████| 763.5 MB 14 kB/s [?25hRequirement already satisfied: typing-extensions in /usr/local/lib/python3.7/dist-packages (from torch==1.8.0+cu101) (3.10.0.2) Requirement already satisfied: numpy in /usr/local/lib/python3.7/dist-packages (from torch==1.8.0+cu101) (1.19.5) Installing collected packages: torch Attempting uninstall: torch Found existing installation: torch 1.9.0+cu111 Uninstalling torch-1.9.0+cu111: Successfully uninstalled torch-1.9.0+cu111 ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. torchvision 0.10.0+cu111 requires torch==1.9.0, but you have torch 1.8.0+cu101 which is incompatible. torchtext 0.10.0 requires torch==1.9.0, but you have torch 1.8.0+cu101 which is incompatible. joeynmt 1.3 requires torch>=1.9.0, but you have torch 1.8.0+cu101 which is incompatible. Successfully installed torch-1.8.0+cu101 ###Markdown Preprocessing the Data into Subword BPE Tokens- One of the most powerful improvements for agglutinative languages (a feature of most Bantu languages) is using BPE tokenization [ (Sennrich, 2015) ](https://arxiv.org/abs/1508.07909).- It was also shown that by optimizing the umber of BPE codes we significantly improve results for low-resourced languages [(Sennrich, 2019)](https://www.aclweb.org/anthology/P19-1021) [(Martinus, 2019)](https://arxiv.org/abs/1906.05685)- Below we have the scripts for doing BPE tokenization of our data. We use 4000 tokens as recommended by [(Sennrich, 2019)](https://www.aclweb.org/anthology/P19-1021). You do not need to change anything. Simply running the below will be suitable. ###Code # One of the huge boosts in NMT performance was to use a different method of tokenizing. # Usually, NMT would tokenize by words. However, using a method called BPE gave amazing boosts to performance # Do subword NMT from os import path os.environ["src"] = source_language # Sets them in bash as well, since we often use bash scripts os.environ["tgt"] = target_language # Learn BPEs on the training data. os.environ["data_path"] = path.join("joeynmt", "data",target_language + source_language ) # Herman! ! subword-nmt learn-joint-bpe-and-vocab --input train.$src train.$tgt -s 4000 -o bpe.codes.4000 --write-vocabulary vocab.$src vocab.$tgt # Apply BPE splits to the development and test data. ! subword-nmt apply-bpe -c bpe.codes.4000 --vocabulary vocab.$src < train.$src > train.bpe.$src ! subword-nmt apply-bpe -c bpe.codes.4000 --vocabulary vocab.$tgt < train.$tgt > train.bpe.$tgt ! subword-nmt apply-bpe -c bpe.codes.4000 --vocabulary vocab.$src < dev.$src > dev.bpe.$src ! subword-nmt apply-bpe -c bpe.codes.4000 --vocabulary vocab.$tgt < dev.$tgt > dev.bpe.$tgt ! subword-nmt apply-bpe -c bpe.codes.4000 --vocabulary vocab.$src < test.$src > test.bpe.$src ! subword-nmt apply-bpe -c bpe.codes.4000 --vocabulary vocab.$tgt < test.$tgt > test.bpe.$tgt # Create directory, move everyone we care about to the correct location ! mkdir -p $data_path ! cp train.* $data_path ! cp test.* $data_path ! cp dev.* $data_path ! cp bpe.codes.4000 $data_path ! ls $data_path # Also move everything we care about to a mounted location in google drive (relevant if running in colab) at gdrive_path ! cp train.* "$gdrive_path" ! cp test.* "$gdrive_path" ! cp dev.* "$gdrive_path" ! cp bpe.codes.4000 "$gdrive_path" ! ls "$gdrive_path" # Create that vocab using build_vocab ! sudo chmod 777 joeynmt/scripts/build_vocab.py ! joeynmt/scripts/build_vocab.py joeynmt/data/$tgt$src/train.bpe.$src joeynmt/data/$tgt$src/train.bpe.$tgt --output_path joeynmt/data/$tgt$src/vocab.txt # Some output ! echo "BPE Sesotho Sentences" ! tail -n 5 test.bpe.$tgt ! echo "Combined BPE Vocab" ! tail -n 10 joeynmt/data/$tgt$src/vocab.txt # Herman # Also move everything we care about to a mounted location in google drive (relevant if running in colab) at gdrive_path ! cp train.* "$gdrive_path" ! cp test.* "$gdrive_path" ! cp dev.* "$gdrive_path" ! cp bpe.codes.4000 "$gdrive_path" ! ls "$gdrive_path" ###Output bpe.codes.4000 dev.en test.bpe.st test.st train.en dev.bpe.en dev.st test.en train.bpe.en train.st dev.bpe.st test.bpe.en test.en-any.en train.bpe.st ###Markdown Creating the JoeyNMT ConfigJoeyNMT requires a yaml config. We provide a template below. We've also set a number of defaults with it, that you may play with!- We used Transformer architecture - We set our dropout to reasonably high: 0.3 (recommended in [(Sennrich, 2019)](https://www.aclweb.org/anthology/P19-1021))Things worth playing with:- The batch size (also recommended to change for low-resourced languages)- The number of epochs (we've set it at 30 just so it runs in about an hour, for testing purposes)- The decoder options (beam_size, alpha)- Evaluation metrics (BLEU versus Crhf4) ###Code # This creates the config file for our JoeyNMT system. It might seem overwhelming so we've provided a couple of useful parameters you'll need to update # (You can of course play with all the parameters if you'd like!) name = '%s%s' % (target_language, source_language) # gdrive_path = os.environ["gdrive_path"] # Create the config config = """ name: "{target_language}{source_language}_reverse_transformer" data: src: "{target_language}" trg: "{source_language}" train: "data/{name}/train.bpe" dev: "data/{name}/dev.bpe" test: "data/{name}/test.bpe" level: "bpe" lowercase: False max_sent_length: 100 src_vocab: "data/{name}/vocab.txt" trg_vocab: "data/{name}/vocab.txt" testing: beam_size: 5 alpha: 1.0 training: #load_model: "{gdrive_path}/models/{name}_transformer/1.ckpt" # if uncommented, load a pre-trained model from this checkpoint random_seed: 42 optimizer: "adam" normalization: "tokens" adam_betas: [0.9, 0.999] scheduling: "Noam scheduling" # TODO: try switching from plateau to Noam scheduling. plateau to Noam scheduling patience: 5 # For plateau: decrease learning rate by decrease_factor if validation score has not improved for this many validation rounds. learning_rate_factor: 0.5 # factor for Noam scheduler (used with Transformer) learning_rate_warmup: 1000 # warmup steps for Noam scheduler (used with Transformer) decrease_factor: 0.7 loss: "crossentropy" learning_rate: 0.0003 learning_rate_min: 0.00000001 weight_decay: 0.0 label_smoothing: 0.1 batch_size: 4096 batch_type: "token" eval_batch_size: 3600 eval_batch_type: "token" batch_multiplier: 1 early_stopping_metric: "ppl" epochs: 3 # TODO: Decrease for when playing around and checking of working. Around 30 is sufficient to check if its working at all. 5 - 3 validation_freq: 1000 # TODO: Set to at least once per epoch. logging_freq: 100 eval_metric: "bleu" model_dir: "models/{name}_reverse_transformer" overwrite: True # TODO: Set to True if you want to overwrite possibly existing models. shuffle: True use_cuda: True max_output_length: 100 print_valid_sents: [0, 1, 2, 3] keep_last_ckpts: 3 model: initializer: "xavier" bias_initializer: "zeros" init_gain: 1.0 embed_initializer: "xavier" embed_init_gain: 1.0 tied_embeddings: True tied_softmax: True encoder: type: "transformer" num_layers: 6 num_heads: 8 # TODO: Increase to 8 for larger data. 4 - 8 embeddings: embedding_dim: 512 # TODO: Increase to 512 for larger data. 256 -512 scale: True dropout: 0.2 # typically ff_size = 4 x hidden_size hidden_size: 512 # TODO: Increase to 512 for larger data. 256 - 512 ff_size: 2048 # TODO: Increase to 2048 for larger data. 1024 - 2048 dropout: 0.3 decoder: type: "transformer" num_layers: 6 num_heads: 8 # TODO: Increase to 8 for larger data. 4 - 8 embeddings: embedding_dim: 512 # TODO: Increase to 512 for larger data. 256 - 512 scale: True dropout: 0.2 # typically ff_size = 4 x hidden_size hidden_size: 512 # TODO: Increase to 512 for larger data. 256 - 512 ff_size: 2048 # TODO: Increase to 2048 for larger data. 1024 - 2048 dropout: 0.3 """.format(name=name, gdrive_path=os.environ["gdrive_path"], source_language=source_language, target_language=target_language) with open("joeynmt/configs/transformer_reverse_{name}.yaml".format(name=name),'w') as f: f.write(config) ###Output _____no_output_____ ###Markdown Train the ModelThis single line of joeynmt runs the training using the config we made above ###Code # Train the model # You can press Ctrl-C to stop. And then run the next cell to save your checkpoints! !cd joeynmt; python3 -m joeynmt train configs/transformer_reverse_$tgt$src.yaml # Copy the created models from the notebook storage to google drive for persistant storage !cp -r joeynmt/models/${tgt}${src}_reverse_transformer/* "$gdrive_path/models/${src}${tgt}_reverse_transformer/" # Output our validation accuracy ! cat "$gdrive_path/models/${tgt}${src}_reverse_transformer/validations.txt" # Test our model ! cd joeynmt; python3 -m joeynmt test "$gdrive_path/models/${tgt}${src}_reverse_transformer/config.yaml" ###Output _____no_output_____
nbs/55_cosine_search.ipynb
###Markdown Cosine ###Code # default_exp cosine # export import numpy as np import pandas as pd from forgebox.category import Category ###Output _____no_output_____ ###Markdown Cosine Similarity ###Code # export class CosineSearch: """ Build a index search on cosine distance cos = CosineSearch(base_array) idx_order = cos(vec) """ def __init__(self, base): assert len(base.shape) == 2,\ f"Base array has to be 2 dimentional, input is {len(base.shape)}" self.base = base self.base_norm = self.calc_base_norm(self.base) self.normed_base = self.base/self.base_norm[:, None] self.dim = self.base.shape[1] def __len__(self): return base.shape[0] @staticmethod def calc_base_norm(base: np.ndarray) -> np.ndarray: return np.sqrt(np.power(base, 2).sum(1)) def search(self, vec: np.ndarray, return_similarity: bool = False): if return_similarity: similarity = (vec * self.normed_base / (np.power(vec, 2).sum())).sum(1) order = similarity.argsort()[::-1] return order, similarity[order] return self(vec) def __call__(self, vec: np.ndarray) -> np.ndarray: """ Return the order index of the closest vector to the furthest vec: an 1 dimentional vector """ return (vec * self.normed_base).sum(1).argsort()[::-1] class CosineSearchWithCategory(CosineSearch): """ Combine with the category manager The class can return a dataframe with category information search_dataframe """ def __init__(self, base: np.ndarray, category: np.ndarray): super().__init__(base) self.category = category assert len(self.category) >= len(self), "category number too small" def search_dataframe( self, vec, return_similarity=True ) -> pd.DataFrame: """ return a dataframe from the closest category to the furthest """ if return_similarity: idx, similarity = self.search(vec, return_similarity) return pd.DataFrame({ "category": self.category.i2c[idx], "idx": idx, "similarity": similarity}) idx = self.search(vec, return_similarity) return pd.DataFrame({ "category": self.category.i2c[idx], "idx": idx}) ###Output _____no_output_____ ###Markdown Test search ###Code base = np.random.rand(50000,100)-.2 vec = base[200] cosine = CosineSearch(base) cosine(vec) cosine.search(vec, return_similarity=True) # cos_cat = CosineSearchWithCategory(base, Category(list(f"c{i}" for i in range(len(base))))) %%time for i in range(100): cosine(vec) ###Output CPU times: user 1.21 s, sys: 147 ms, total: 1.36 s Wall time: 1.37 s
DeepFake-Xception.ipynb
###Markdown Dir ###Code train_dir = '/mnt/a/fakedata/deepfake/train' validation_dir = '/mnt/a/fakedata/deepfake/val' test50_dir = '/mnt/a/fakedata/deepfake/test' ###Output _____no_output_____ ###Markdown Xception ###Code img_input = Input(shape=(img_height, img_width, 3)) # layer 1 # x = Conv2D(filters=32, kernel_size=(3, 3), strides=2, padding='valid', use_bias=False)(img_input) x = BatchNormalization()(x) x = Activation('relu')(x) # layer 2 # x = Conv2D(filters=64, kernel_size=(3, 3), padding='valid', use_bias=False)(x) x = BatchNormalization()(x) x = Activation('relu')(x) # skip layer 1 # res = Conv2D(filters=128, kernel_size=(1, 1), strides=2, padding='same', use_bias=False)(x) res = BatchNormalization()(res) # layer 3 # x = SeparableConv2D(filters=128, kernel_size=(3, 3), strides=1, padding='same', use_bias=False)(x) x = BatchNormalization()(x) # layer 4 # x = Activation('relu')(x) x = SeparableConv2D(filters=128, kernel_size=(3,3), strides=1, padding='same', use_bias=False)(x) x = MaxPooling2D(pool_size=(3, 3), strides=2, padding='same')(x) x = Add()([x, res]) # skip layer 2 # res = Conv2D(filters=256, kernel_size=(1, 1), strides=2, padding='same', use_bias=False)(x) res = BatchNormalization()(res) # layer 5 # x = Activation('relu')(x) x = SeparableConv2D(filters=256, kernel_size=(3, 3), strides=1, padding='same', use_bias=False)(x) x = BatchNormalization()(x) # layer 6 # x = Activation('relu')(x) x = SeparableConv2D(filters=256, kernel_size=(3,3), strides=1, padding='same', use_bias=False)(x) x = MaxPooling2D(pool_size=(3, 3), strides=2, padding='same')(x) x = Add()([x, res]) # skip layer 3 # res = Conv2D(filters=728, kernel_size=(1, 1), strides=2, padding='same', use_bias=False)(x) res = BatchNormalization()(res) # layer 7 # x = Activation('relu')(x) x = SeparableConv2D(filters=728, kernel_size=(3, 3), strides=1, padding='same', use_bias=False)(x) x = BatchNormalization()(x) # layer 8 # x = Activation('relu')(x) x = SeparableConv2D(filters=728, kernel_size=(3,3), strides=1, padding='same', use_bias=False)(x) x = MaxPooling2D(pool_size=(3, 3), strides=2, padding='same')(x) x = Add()([x, res]) # ======== middle flow ========= # for i in range(8): # layer 9, 10, 11, 12, 13, 14, 15, 16, 17 # res = x x = Activation('relu')(x) x = SeparableConv2D(filters=728, kernel_size=(3, 3), strides=1, padding='same', use_bias=False)(x) x = BatchNormalization()(x) x = Activation('relu')(x) x = SeparableConv2D(filters=728, kernel_size=(3, 3), strides=1, padding='same', use_bias=False)(x) x = BatchNormalization()(x) x = Activation('relu')(x) x = SeparableConv2D(filters=728, kernel_size=(3, 3), strides=1, padding='same', use_bias=False)(x) x = BatchNormalization()(x) x = Add()([x, res]) # ======== exit flow ========== # # skip layer 4 # res = Conv2D(filters=1024, kernel_size=(1, 1), strides=2, padding='same', use_bias=False)(x) res = BatchNormalization()(res) # layer 18 # x = Activation('relu')(x) x = SeparableConv2D(filters=728, kernel_size=(3, 3), strides=1, padding='same', use_bias=False)(x) x = BatchNormalization()(x) # layer 19 # x = Activation('relu')(x) x = SeparableConv2D(filters=1024, kernel_size=(3, 3), strides=1, padding='same', use_bias=False)(x) x = BatchNormalization()(x) x = MaxPooling2D(pool_size=(3, 3), strides=2, padding='same')(x) x = Add()([x, res]) # layer 20 # x = SeparableConv2D(filters=1536, kernel_size=(3, 3), strides=1, padding='same', use_bias=False)(x) x = BatchNormalization()(x) x = Activation('relu')(x) # layer 21 # x = SeparableConv2D(filters=2048, kernel_size=(3, 3), strides=1, padding='same', use_bias=False)(x) x = BatchNormalization()(x) x = Activation('relu')(x) x_gap = GlobalAveragePooling2D()(x) output = Dense(units=2, activation='softmax')(x_gap) model = Model(img_input, output) model.summary() model.compile(optimizer=Adam(), loss='categorical_crossentropy', metrics=['accuracy']) print(len(model.trainable_weights)) def bgr(img): return cv2.cvtColor(img, cv2.COLOR_BGR2RGB) ###Output _____no_output_____ ###Markdown Data generator ###Code train_datagen = ImageDataGenerator(rescale=1./255, preprocessing_function=bgr) test_datagen = ImageDataGenerator(rescale=1./255, preprocessing_function=bgr) train_generator = train_datagen.flow_from_directory(train_dir, target_size=(img_height, img_width), batch_size=batch_size, shuffle=True, class_mode='categorical') validation_generator = train_datagen.flow_from_directory(validation_dir, target_size=(img_height, img_width), batch_size=batch_size, shuffle=False, class_mode='categorical') test50_generator = test_datagen.flow_from_directory(test50_dir, target_size=(img_height, img_width), batch_size=batch_size, shuffle=False, class_mode='categorical') # callback_list = [EarlyStopping(monitor='val_accuracy', patience=10), # ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=3)] # history = model.fit_generator(train_generator, # steps_per_epoch=200, # epochs=100, # validation_data=validation_generator, # validation_steps=len(validation_generator), # callbacks=callback_list) # model.save('/home/www/fake_detection/model/deepfake_xception.h5') # model = load_model('/home/www/fake_detection/model/deepfake_xception.h5') # output = model.predict_generator(test50_generator, steps=len(test50_generator), verbose=1) # np.set_printoptions(formatter={'float': lambda x: "{0:0.3f}".format(x)}) # print(test50_generator.class_indices) # print(output) # output_score50 = [] # output_class50 = [] # answer_class50 = [] # answer_class50_1 =[] # for i in trange(len(test50_generator)): # output50 = model.predict_on_batch(test50_generator[i][0]) # output_score50.append(output50) # answer_class50.append(test50_generator[i][1]) # output_score50 = np.concatenate(output_score50) # answer_class50 = np.concatenate(answer_class50) # output_class50 = np.argmax(output_score50, axis=1) # answer_class50_1 = np.argmax(answer_class50, axis=1) # print(output_class50) # print(answer_class50_1) # cm50 = confusion_matrix(answer_class50_1, output_class50) # report50 = classification_report(answer_class50_1, output_class50) # recall50 = cm50[0][0] / (cm50[0][0] + cm50[0][1]) # fallout50 = cm50[1][0] / (cm50[1][0] + cm50[1][1]) # fpr50, tpr50, thresholds50 = roc_curve(answer_class50_1, output_score50[:, 1], pos_label=1.) # eer50 = brentq(lambda x : 1. - x - interp1d(fpr50, tpr50)(x), 0., 1.) # thresh50 = interp1d(fpr50, thresholds50)(eer50) # print(report50) # print(cm50) # print("AUROC: %f" %(roc_auc_score(answer_class50_1, output_score50[:, 1]))) # print(thresh50) # print('test_acc: ', len(output_class50[np.equal(output_class50, answer_class50_1)]) / len(output_class50)) def cutout(img): """ # Function: RandomCrop (ZeroPadded (4, 4)) + random occulusion image # Arguments: img: image # Returns: img """ img = bgr(img) height = img.shape[0] width = img.shape[1] channels = img.shape[2] MAX_CUTS = 3 # chance to get more cuts MAX_LENGTH_MUTIPLIER = 10 # chance to get larger cuts # 16 for cifar10, 8 for cifar100 # Zero-padded (4, 4) # img = np.pad(img, ((4,4),(4,4),(0,0)), mode='constant', constant_values=(0)) # # random-crop 64x64 # dy, dx = height, width # x = np.random.randint(0, width - dx + 1) # y = np.random.randint(0, height - dy + 1) # img = img[y:(y+dy), x:(x+dx)] # mean norm # mean = img.mean(keepdims=True) # img -= mean img *= 1./255 mask = np.ones((height, width, channels), dtype=np.float32) nb_cuts = np.random.randint(0, MAX_CUTS + 1) # cutout for i in range(nb_cuts): y = np.random.randint(height) x = np.random.randint(width) length = 4 * np.random.randint(1, MAX_LENGTH_MUTIPLIER+1) y1 = np.clip(y-length//2, 0, height) y2 = np.clip(y+length//2, 0, height) x1 = np.clip(x-length//2, 0, width) x2 = np.clip(x+length//2, 0, width) mask[y1:y2, x1:x2, :] = 0. img = img * mask return img class ReLU6(Layer): def __init__(self): super().__init__(name="ReLU6") self.relu6 = ReLU(max_value=6, name="ReLU6") def call(self, input): return self.relu6(input) class HardSigmoid(Layer): def __init__(self): super().__init__() self.relu6 = ReLU6() def call(self, input): return self.relu6(input + 3.0) / 6.0 class HardSwish(Layer): def __init__(self): super().__init__() self.hard_sigmoid = HardSigmoid() def call(self, input): return input * self.hard_sigmoid(input) class Attention(Layer): def __init__(self, ch, **kwargs): super(Attention, self).__init__(**kwargs) self.channels = ch self.filters_f_g = self.channels // 8 self.filters_h = self.channels def build(self, input_shape): kernel_shape_f_g = (1, 1) + (self.channels, self.filters_f_g) print(kernel_shape_f_g) kernel_shape_h = (1, 1) + (self.channels, self.filters_h) # Create a trainable weight variable for this layer: self.gamma = self.add_weight(name='gamma', shape=[1], initializer='zeros', trainable=True) self.kernel_f = self.add_weight(shape=kernel_shape_f_g, initializer='glorot_uniform', name='kernel_f') self.kernel_g = self.add_weight(shape=kernel_shape_f_g, initializer='glorot_uniform', name='kernel_g') self.kernel_h = self.add_weight(shape=kernel_shape_h, initializer='glorot_uniform', name='kernel_h') self.bias_f = self.add_weight(shape=(self.filters_f_g,), initializer='zeros', name='bias_F') self.bias_g = self.add_weight(shape=(self.filters_f_g,), initializer='zeros', name='bias_g') self.bias_h = self.add_weight(shape=(self.filters_h,), initializer='zeros', name='bias_h') super(Attention, self).build(input_shape) # Set input spec. self.input_spec = InputSpec(ndim=4, axes={3: input_shape[-1]}) self.built = True def call(self, x): def hw_flatten(x): return K.reshape(x, shape=[K.shape(x)[0], K.shape(x)[1]*K.shape(x)[2], K.shape(x)[-1]]) f = K.conv2d(x, kernel=self.kernel_f, strides=(1, 1), padding='same') # [bs, h, w, c'] f = K.bias_add(f, self.bias_f) g = K.conv2d(x, kernel=self.kernel_g, strides=(1, 1), padding='same') # [bs, h, w, c'] g = K.bias_add(g, self.bias_g) h = K.conv2d(x, kernel=self.kernel_h, strides=(1, 1), padding='same') # [bs, h, w, c] h = K.bias_add(h, self.bias_h) s = tf.matmul(hw_flatten(g), hw_flatten(f), transpose_b=True) # # [bs, N, N] beta = K.softmax(s, axis=-1) # attention map o = K.batch_dot(beta, hw_flatten(h)) # [bs, N, C] o = K.reshape(o, shape=K.shape(x)) # [bs, h, w, C] x = self.gamma * o + x return x def compute_output_shape(self, input_shape): return input_shape ft_dir = '/mnt/a/fakedata/deepfake/finetune' train_gen_aug = ImageDataGenerator(shear_range=0, zoom_range=0, rotation_range=0.2, width_shift_range=2., height_shift_range=2., horizontal_flip=True, zca_whitening=False, fill_mode='nearest', preprocessing_function=cutout) test_datagen = ImageDataGenerator(rescale=1./255, preprocessing_function=bgr) ft_gen = train_gen_aug.flow_from_directory(ft_dir, target_size=(img_height, img_width), batch_size=batch_size, shuffle=True, class_mode='categorical') validation_generator = test_datagen.flow_from_directory(validation_dir, target_size=(img_height, img_width), batch_size=batch_size, shuffle=False, class_mode='categorical') test50_generator = test_datagen.flow_from_directory(test50_dir, target_size=(img_height, img_width), batch_size=batch_size, shuffle=False, class_mode='categorical') model_ft = load_model('/home/www/fake_detection/model/deepfake_xception.h5') for i in range(2): model_ft.layers.pop() im_in = Input(shape=(img_width, img_height, 3)) base_model = Model(img_input, x) base_model.set_weights(model_ft.get_weights()) # for i in range(len(base_model.layers) - 0): # base_model.layers[i].trainable = False x1 = base_model(im_in) # (12, 12, 32) ########### Mobilenet block bneck 3x3 (32 --> 128) ################# expand1 = Conv2D(576, kernel_size=1, strides=1, kernel_regularizer=l2(1e-5), use_bias=False)(x1) expand1 = BatchNormalization()(expand1) expand1 = HardSwish()(expand1) dw1 = DepthwiseConv2D(kernel_size=(3,3), strides=(2,2), padding='same', depthwise_regularizer=l2(1e-5), use_bias=False)(expand1) dw1 = BatchNormalization()(dw1) se_gap1 = GlobalAveragePooling2D()(dw1) se_gap1 = Reshape([1, 1, -1])(se_gap1) se1 = Conv2D(144, kernel_size=1, strides=1, padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(se_gap1) se1 = Activation('relu')(se1) se1 = Conv2D(576, kernel_size=1, strides=1, padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(se1) se1 = HardSigmoid()(se1) se1 = Multiply()([expand1, se1]) project1 = HardSwish()(se1) project1 = Conv2D(128, kernel_size=(1, 1), padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(project1) project1 = BatchNormalization()(project1) ########### Mobilenet block bneck 5x5 (128 --> 128) ################# expand2 = Conv2D(576, kernel_size=1, strides=1, kernel_regularizer=l2(1e-5), use_bias=False)(project1) expand2 = BatchNormalization()(expand2) expand2 = HardSwish()(expand2) dw2 = DepthwiseConv2D(kernel_size=(5,5), strides=(1,1), padding='same', depthwise_regularizer=l2(1e-5), use_bias=False)(expand2) dw2 = BatchNormalization()(dw2) se_gap2 = GlobalAveragePooling2D()(dw2) se_gap2 = Reshape([1, 1, -1])(se_gap2) se2 = Conv2D(144, kernel_size=1, strides=1, padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(se_gap2) se2 = Activation('relu')(se2) se2 = Conv2D(576, kernel_size=1, strides=1, padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(se2) se2 = HardSigmoid()(se2) se2 = Multiply()([expand2, se2]) project2 = HardSwish()(se2) project2 = Conv2D(128, kernel_size=(1, 1), padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(project2) project2 = BatchNormalization()(project2) project2 = Add()([project1, project2]) ########### Mobilenet block bneck 5x5 (128 --> 128) ################# expand3 = Conv2D(576, kernel_size=1, strides=1, kernel_regularizer=l2(1e-5), use_bias=False)(project2) expand3 = BatchNormalization()(expand3) expand3 = HardSwish()(expand3) dw3 = DepthwiseConv2D(kernel_size=(5,5), strides=(1,1), padding='same', depthwise_regularizer=l2(1e-5), use_bias=False)(expand3) dw3 = BatchNormalization()(dw3) se_gap3 = GlobalAveragePooling2D()(dw3) se_gap3 = Reshape([1, 1, -1])(se_gap3) se3 = Conv2D(144, kernel_size=1, strides=1, padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(se_gap3) se3 = Activation('relu')(se3) se3 = Conv2D(576, kernel_size=1, strides=1, padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(se3) se3 = HardSigmoid()(se3) se3 = Multiply()([expand3, se3]) project3 = HardSwish()(se3) project3 = Conv2D(128, kernel_size=(1, 1), padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(project3) project3 = BatchNormalization()(project3) project3 = Add()([project2, project3]) expand4 = Conv2D(576, kernel_size=1, strides=1, kernel_regularizer=l2(1e-5), use_bias=False)(project3) expand4 = BatchNormalization()(expand4) expand4 = HardSwish()(expand4) dw4 = DepthwiseConv2D(kernel_size=(5,5), strides=(1,1), padding='same', depthwise_regularizer=l2(1e-5), use_bias=False)(expand4) dw4 = BatchNormalization()(dw4) se_gap4 = GlobalAveragePooling2D()(dw4) se_gap4 = Reshape([1, 1, -1])(se_gap4) se4 = Conv2D(144, kernel_size=1, strides=1, padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(se_gap4) se4 = Activation('relu')(se4) se4 = Conv2D(576, kernel_size=1, strides=1, padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(se4) se4 = HardSigmoid()(se4) se4 = Multiply()([expand4, se4]) project4 = HardSwish()(se4) project4 = Conv2D(128, kernel_size=(1, 1), padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(project4) project4 = BatchNormalization()(project4) project4 = Add()([project3, project4]) ########## Classification ########## x2 = Conv2D(576, kernel_size=1, strides=1, padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(project4) x2 = BatchNormalization()(x2) x2 = HardSwish()(x2) x2 = GlobalAveragePooling2D()(x2) ######### Image Attention Model ######### ### Block 1 ### x3 = SeparableConv2D(32, kernel_size=(3, 3), strides=(2,2), padding='same', depthwise_regularizer=l2(1e-5), pointwise_regularizer=l2(1e-5), use_bias=False)(im_in) x3 = BatchNormalization()(x3) x3 = Activation('relu')(x3) x3 = Attention(32)(x3) ### Block 2 ### x4 = SeparableConv2D(64, kernel_size=(3, 3), strides=(2,2), padding='same', depthwise_regularizer=l2(1e-5), pointwise_regularizer=l2(1e-5), use_bias=False)(x3) x4 = BatchNormalization()(x4) x4 = Activation('relu')(x4) x4 = Attention(64)(x4) ### Block 3 ### x5 = SeparableConv2D(128, kernel_size=(3, 3), strides=(2,2), padding='same', depthwise_regularizer=l2(1e-5), pointwise_regularizer=l2(1e-5), use_bias=False)(x4) x5 = BatchNormalization()(x5) x5 = Activation('relu')(x5) x5 = Attention(128)(x5) ### final stage ### x6 = Conv2D(576, kernel_size=1, strides=1, padding='valid', kernel_regularizer=l2(1e-5), use_bias=False)(x5) x6 = BatchNormalization()(x6) x6 = Activation('relu')(x6) x6 = GlobalAveragePooling2D()(x6) ######## final addition ######### x2 = Add()([x2, x6]) x2 = Dense(2, kernel_regularizer=l2(1e-5))(x2) x2 = Activation('softmax')(x2) model_top = Model(inputs=im_in, outputs=x2) model_top.summary() # optimizer = SGD(lr=1e-3, momentum=0.9, nesterov=True) optimizer = Adam() model_top.compile(optimizer=optimizer, loss='categorical_crossentropy', metrics=['acc']) callback_list = [EarlyStopping(monitor='val_acc', patience=30), ReduceLROnPlateau(monitor='loss', factor=np.sqrt(0.5), cooldown=0, patience=5, min_lr=0.5e-5)] output = model_top.fit_generator(ft_gen, steps_per_epoch=200, epochs=300, validation_data=validation_generator, validation_steps=len(validation_generator), callbacks=callback_list) output_score50 = [] output_class50 = [] answer_class50 = [] answer_class50_1 =[] for i in trange(len(test50_generator)): output50 = model_top.predict_on_batch(test50_generator[i][0]) output_score50.append(output50) answer_class50.append(test50_generator[i][1]) output_score50 = np.concatenate(output_score50) answer_class50 = np.concatenate(answer_class50) output_class50 = np.argmax(output_score50, axis=1) answer_class50_1 = np.argmax(answer_class50, axis=1) print(output_class50) print(answer_class50_1) cm50 = confusion_matrix(answer_class50_1, output_class50) report50 = classification_report(answer_class50_1, output_class50) recall50 = cm50[0][0] / (cm50[0][0] + cm50[0][1]) fallout50 = cm50[1][0] / (cm50[1][0] + cm50[1][1]) fpr50, tpr50, thresholds50 = roc_curve(answer_class50_1, output_score50[:, 1], pos_label=1.) eer50 = brentq(lambda x : 1. - x - interp1d(fpr50, tpr50)(x), 0., 1.) thresh50 = interp1d(fpr50, thresholds50)(eer50) print(report50) print(cm50) print("AUROC: %f" %(roc_auc_score(answer_class50_1, output_score50[:, 1]))) print(thresh50) print('test_acc: ', len(output_class50[np.equal(output_class50, answer_class50_1)]) / len(output_class50)) model_top.save("/home/www/fake_detection/model/deepfake_xception_ft1.h5") ###Output _____no_output_____
Magnetic Pickup/ReadScopeData.ipynb
###Markdown Read data from Rigol DS1054Z scope https://readthedocs.org/projects/ds1054z/downloads/pdf/stable/ Import the libraries ###Code from ds1054z import DS1054Z import matplotlib.pyplot as plt import numpy as np import math import scipy.io as sio import scipy.signal as sig from scipy.fft import rfft, rfftfreq import pyvisa as visa import time import os import shutil ###Output _____no_output_____ ###Markdown Define plot mode. Interactive mode is helpful for visuallizing the program execution ###Code #%matplotlib widget ###Output _____no_output_____ ###Markdown Verify scope connection ###Code scope = DS1054Z('192.168.1.206') print(scope.idn) ###Output RIGOL TECHNOLOGIES,DS1054Z,DS1ZA200902668,00.04.04.SP3 ###Markdown Test Description This sheet is designed to take data on a Rigol DS1054z oscilloscope with channel 1 of the scope connected to an MSP6729 magnetic pick up viewing the chuck of a 7" x 10" mini lathe. Channel 1 common of the oscilloscope is connected to the magnetic pick up shield wire and the black wire. The signal of channel 1 is connected to the red wire of the mag pick up. ![alt text](IMG_8440_Scope2MagPickup_Labels.png "Oscilloscope Connections") Define functions used in the test This is the function that sets the trigger level ###Code def b_set_trigger(d_trigger_level = 1e-01): """Set the trigger configuration Keyword arguments: d_trigger_level -- Voltage level to trigger scope (default: 0.1 volts) Return values: [None] """ scope.write(':trigger:edge:source CHAN1') scope.write(':trigger:edge:level ' + format(d_trigger_level)) scope.single() ###Output _____no_output_____ ###Markdown Function that contains the commands that setup the scope ###Code def b_setup_scope(scope, d_ch1_scale=5.e-1, timebase_scale=5e-2, d_trigger_level = 1e-01, b_single = True): """Setup Rigol ds1054z to read a 3/8-24 magnetic pickup Keyword arguments: scope -- Connection to scope d_ch1_scale -- Channel 1 scale (default: 0.5 volts) timebase_scale -- Time scale for data (default: 0.005 seconds) d_trigger_level -- Voltage level to trigger scope (default: 0.1 volts) b_trigger -- If true, then use trigger levels (default: True) Return values: d_ch1_scale_actual -- The closest value chosen by the scope """ scope.timebase_scale = timebase_scale scope.run() scope.display_channel(1,enable=True) scope.set_probe_ratio(1,1) scope.set_channel_scale(1,"{:e}".format(d_ch1_scale) +'V') scope.write(':CHANnel1:COUPling AC') scope.display_channel(2,enable=False) scope.display_channel(3,enable=False) scope.display_channel(4,enable=False) # Do we need a trigger? if b_single: # Set the scope to capture after trigger b_set_trigger(d_trigger_level) else: # No trigger, useful for seeing the scope data when you aren't sure # what the signal looks like scope.write(":TRIGger:SWEep AUTO") return scope.get_channel_scale(1) ###Output _____no_output_____ ###Markdown Verify the help comments are at least somewhat on point ###Code help(b_setup_scope) ###Output Help on function b_setup_scope in module __main__: b_setup_scope(scope, d_ch1_scale=0.5, timebase_scale=0.05, d_trigger_level=0.1, b_single=True) Setup Rigol ds1054z to read a 3/8-24 magnetic pickup Keyword arguments: scope -- Connection to scope d_ch1_scale -- Channel 1 scale (default: 0.5 volts) timebase_scale -- Time scale for data (default: 0.005 seconds) d_trigger_level -- Voltage level to trigger scope (default: 0.1 volts) b_trigger -- If true, then use trigger levels (default: True) Return values: d_ch1_scale_actual -- The closest value chosen by the scope ###Markdown Define the function that acquires data from scopeThis one is a little tricky because it can take time to acquire the signal so there are pause statements to allow data to accumulate at the scope. If the acquisition terminates before the sampling is complete there will be NaN's in the list. In this case the NaN's are converte zeros to allow processing to continue. It can be helpful to see a partial waveform to troubleshoot timing at the scope. ###Code def d_get_data(i_ch=1, timebase_scale=5e-2): """Get data from the scope Keyword arguments: i_ch -- 1-based index of channel to sample (default: 1) Return values: np_d_ch1 -- numpy array of values from the scope """ # Calculate the delay time d_time_delay = timebase_scale*32 + 1. # Acquire the data time.sleep(d_time_delay) d_ch1 = scope.get_waveform_samples(i_ch, mode='NORM') time.sleep(d_time_delay) scope.run() # Convert the list to a numpy array and replace NaN's with zeros np_d_ch1 = np.array(d_ch1) np_d_ch1 = np.nan_to_num(np_d_ch1) return np.array(np_d_ch1) ###Output _____no_output_____ ###Markdown Verify the help text ###Code help(d_get_data) ###Output Help on function d_get_data in module __main__: d_get_data(i_ch=1, timebase_scale=0.05) Get data from the scope Keyword arguments: i_ch -- 1-based index of channel to sample (default: 1) Return values: np_d_ch1 -- numpy array of values from the scope ###Markdown Define the function that extracts features from the data ###Code class cl_sig_features: """Class to manage signal features on scope data Example usage: cl_test = cl_sig_features(np.array([1.,2., 3.]),1.1) Should produce: print('np_d_ch1: '+ np.array2string(cl_test.np_d_ch1)) print('timebase_scale: ' + '%0.3f' % cl_test.timebase_scale) print('i_ns: ' + '%3.f' % cl_test.i_ns) print('d_t_del: ' + '%0.3f' % cl_test.d_t_del) print('d_time' + np.array2string(cl_test.d_time)) np_d_ch1: [1. 2. 3.] timebase_scale: 1.000 i_ns: 3 d_t_del: 4.000 d_time[0. 4. 8.] """ def __init__(self, np_d_ch1, timebase_scale): self.__np_d_ch1 = np_d_ch1 self.__timebase_scale = float(timebase_scale) self.__np_d_rpm = np.zeros_like(self.np_d_ch1) self.__d_thresh = np.NaN self.__d_events_per_rev = np.NaN @property def np_d_ch1(self): """Numpy array containing the scope data""" return self.__np_d_ch1 @property def timebase_scale(self): """Scope time scale""" return self.__timebase_scale @property def i_ns(self): """Number of samples in the scope data""" self.__i_ns = len(self.__np_d_ch1) return self.__i_ns @property def d_t_del(self): """Delta time between each sample""" self.__d_t_del = (12.*float(self.timebase_scale))/float(self.i_ns) return self.__d_t_del @property def d_time(self): """Numpy array with time values, in seconds""" self.__d_time = np.linspace(0,(self.i_ns-1),self.i_ns)*self.d_t_del return self.__d_time @property def d_fs(self): """Sampling frequeny in hertz""" self.__d_fs = 1.0/(self.__d_time[1]-self.__d_time[0]) return self.__d_fs @property def np_d_ch1_filt(self): """ Return the signal, filtered with Savitsky-Golay""" self.__i_win_len = 31; self.__i_poly_order = 1; self.__np_d_ch1_filt = sig.savgol_filter(self.np_d_ch1, self.__i_win_len, self.__i_poly_order); self.__str_filt_desc = ('Savitsky-Golay | Window Length: ' + '%3.f' % self.__i_win_len + ' | Polynomial Order: ' + '%2.f' % self.__i_poly_order) self.__str_filt_desc_short = 'SGolay' return self.__np_d_ch1_filt @property def str_filt_desc(self): "Complete Filt description of the Savitsky-Golay filter design" return self.__str_filt_desc @property def str_filt_desc_short(self): """Short Filt description, useful for plot legend labels""" return self.__str_filt_desc_short @property def np_d_ch1_filt1(self): """ Return the signal, filtered with butter FIR filter""" self.__i_poles = 1 if self.d_fs < 300: self.__d_wn = fs/8 else: self.__d_wn = 100 self.__sos = sig.butter(self.__i_poles, self.__d_wn, btype='low', fs=self.d_fs, output = 'sos') self.__np_d_ch1_filt1 = sig.sosfilt(self.__sos, self.np_d_ch1) self.__str_filt1_desc = ('Butterworth | Poles: ' + '%2.f' % self.__i_poles + ' | Lowpass corner (Hz): ' + '%0.2f' % self.__d_wn) self.__str_filt1_desc_short = 'Butter' return self.__np_d_ch1_filt1 @property def str_filt1_desc(self): "Complete Filt1 description of the Butterworth filter design" return self.__str_filt1_desc @property def str_filt1_desc_short(self): """Short Filt1 description, useful for plot legend labels""" return self.__str_filt1_desc_short @property def np_d_eventtimes(self): """Numpy array of trigger event times""" return self.__np_d_eventtimes @property def d_thresh(self): """Trigger threshold value""" return self.__d_thresh @property def np_d_rpm(self): """Estimated RPM values""" return self.__np_d_rpm @property def d_events_per_rev(self): """Events per revolution""" return self.__d_events_per_rev @np_d_ch1.setter def np_d_ch1(self, np_d_ch1): self.__np_d_ch1 = np_d_ch1 @timebase_scale.setter def timebase_scale(self, timebase_scale): self.__timebase_scale = timebase_scale # Method for calculating the spectrum for a real signal def d_fft_real(self): """Calculate the half spectrum since this is a real-valued signal""" d_y = rfft(np_d_ch1) d_ws = rfftfreq(self.i_ns, 1./self.d_fs) return([d_ws, d_y]) # Plotting method, time domain signals. def plt_sigs(self): """Plot out the data in this signal feature class in the time domain Return values: handle to the plot """ plt.figure() plt.plot(self.d_time, self.np_d_ch1) plt.plot(self.d_time, self.np_d_ch1_filt) plt.plot(self.d_time, self.np_d_ch1_filt1) plt.grid() plt.xlabel("Time, seconds") plt.ylabel("Channel output, volts") plt.legend(['as-aquired', self.str_filt_desc_short, self.str_filt1_desc_short]) plt.show() self.__plot_handle = plt.gcf() return self.__plot_handle # Plotting method for single-sided (real signal) spectrum def plt_spec(self): """Plot data in frequency domain. This method assumes a real signal Return values: handle to the plot """ self.__spec = self.d_fft_real() plt.figure() plt.plot(self.__spec[0], np.abs(self.__spec[1])) plt.grid() plt.xlabel("Frequency, hertz") plt.ylabel("Channel amplitude, volts") plt.show() self.__plot_handle = plt.gcf() return [self.__plot_handle, self.__spec[0], self.__spec[1]] # Plotting method for the eventtimes def plt_eventtimes(self): """Plot event data in time. Return values: list: [handle to the plot, np array of eventtimes] """ # The eventtimes all should have threshold value for voltage self.__np_d_eventvalue = np.ones_like(self.__np_d_eventtimes)*self.d_thresh # Put up the the plot time plt.figure() plt.plot(self.__d_time, self.np_d_ch1) plt.plot(self.np_d_eventtimes, self.__np_d_eventvalue, "ok") plt.xlabel('Time, seconds') plt.ylabel('Amplitude, volts') plt.legend(['as-aquired', 'eventtimes']) plt.title('Amplitude and eventtimes vs. time') self.__plot_handle = plt.gcf() return [self.__plot_handle, self.__np_d_eventtimes] # Plotting method for the eventtimes def plt_rpm(self): """Plot rpm data in time. Return values: list: [handle to the plot, np array of RPM values] """ # Put up the the plot time fig,ax1 = plt.subplots() ax2 = ax1.twinx() ax1.plot(self.__d_time, self.np_d_ch1) ax2.plot(self.np_d_eventtimes, self.__np_d_rpm, "ok") ax1.set_xlabel('Time, seconds') ax1.set_ylabel('Amplitude, volts') ax2.set_ylabel('Event speed, RPM') plt.legend(['as-aquired', 'RPM']) plt.title('Amplitude and eventtimes vs. time') plt.show() self.__plot_handle = plt.gcf() return [self.__plot_handle, self.__np_d_rpm] # Estimate triggers for speed, public method def np_d_est_triggers(self, i_direction=0, d_thresh=0, d_hyst=0.1, i_kernel=5, b_verbose=False): """ This method estimates speed by identifying trigger points in time, a given threshold and hysteresis. When the signal level crosses the threshold, the trigger holds off. The trigger holds off until the signal crosses hysteresis levels. Hysteresis is defined relative to the threshold voltage. The trigger times can be used to estimate the rotating speed. Keyword arguments: i_direction -- 0 to search for threshold on rising signal, 1 to search on a falling signal. d_thresh -- Threshold value (default: 0.0 volts for zero crossings) d_hyst -- Hysteresis value (default: 0.1 volts) i_kernel -- Number of samples to consider in estimating slope, must be an odd number (default: 5) b_verbose -- Print the intermediate steps (default: False). Useful for stepping through the method to troubleshoot or understand it better. Return values: np_d_eventtimes -- numpy array with list of trigger event times """ # Store to local private member, it gets used in other places in the class self.__d_thresh = d_thresh # Initialize trigger state to hold off: the trigger will be active # once the signal crosses the hysteresis b_trigger_hold = True # half kernel get used a lot self.__i_half_kernel = int((i_kernel - 1)/2.) # Use smoothing and derivative functions of S-G filter for estimating rise/fall self.__np_d_ch1_dir = sig.savgol_filter(self.np_d_ch1, i_kernel, 1, deriv=1); # Initiate state machine: one state for rising signal, 'up', (i_direction = 0) # and another for falling signal, 'down', (i_direction = 1) self.__d_hyst_abs = 0 idx_event = 0 self.__np_d_eventtimes = np.zeros_like(self.np_d_ch1) if i_direction == 0: # Define the absolute hysteretic value, rising self.__d_hyst_ab = self.__d_thresh - d_hyst # Loop through the signal for idx,x in enumerate(self.np_d_ch1): # Intermediate results if b_verbose: print('idx: ' + '%2.f' % idx + ' | x: ' + '%0.5f' % x + ' | s-g: ' + '%0.4f' % self.__np_d_ch1_dir[idx]) # The trigger leaves 'hold-off' state if the slope is # negative and we fall below the threshold if (x <= self.__d_hyst_ab and self.__np_d_ch1_dir[idx] < 0 and b_trigger_hold == True): # Next time the signal rises above the threshold, trigger # will be set to hold-off state b_trigger_hold = False # If we are on the rising portion of the signal and there is no hold off # state on the trigger, trigger, and change state if (x >= self.__d_thresh and self.__np_d_ch1_dir[idx] > 0 and b_trigger_hold == False): # Change state to hold off b_trigger_hold = True # Estimate time of crossing with interpolation if idx>0: # Interpolate to estimate the actual crossing from the 2 nearest points xp = np.array([self.np_d_ch1[idx-1], self.np_d_ch1[idx]]) fp = np.array([self.__d_time[idx-1], self.__d_time[idx]]) self.__np_d_eventtimes[idx_event] = np.interp(d_thresh, xp, fp) # More intermediate results if b_verbose: print('xp: ' + np.array2string(xp) + ' | fp: ' + np.array2string(fp) + ' | d_thresh: ' + '%0.4f' % d_thresh + ' | eventtimes: ' + '%0.4f' % self.__np_d_eventtimes[idx_event]) # Increment the eventtimes index idx_event += 1 else: # Define the absolute hysteretic value, falling self.__d_hyst_ab = self.__d_thresh + d_hyst # Loop through the signal for idx,x in enumerate(self.np_d_ch1): # Intermediate results if b_verbose: print('idx: ' + '%2.f' % idx + ' | x: ' + '%0.5f' % x + ' | s-g: ' + '%0.4f' % self.__np_d_ch1_dir[idx]) # The trigger leaves 'hold-off' state if the slope is # positive and we rise above the threshold if (x >= self.__d_hyst_ab and self.__np_d_ch1_dir[idx] > 0 and b_trigger_hold == True): # Next time the signal rises above the threshold, trigger # will be set to hold-off state b_trigger_hold = False # If we are on the falling portion of the signal and there is no hold off # state on the trigger, trigger, and change state if (x <= self.__d_thresh and self.__np_d_ch1_dir[idx] < 0 and b_trigger_hold == False): # Change state to hold off b_trigger_hold = True # Estimate time of crossing with interpolation if idx>0: # Interpolate to estimate the actual crossing from the 2 nearest points xp = np.array([self.np_d_ch1[idx-1], self.np_d_ch1[idx]]) fp = np.array([self.__d_time[idx-1], self.__d_time[idx]]) self.__np_d_eventtimes[idx_event] = np.interp(d_thresh, xp, fp) # More intermediate results if b_verbose: print('xp: ' + np.array2string(xp) + ' | fp: ' + np.array2string(fp) + ' | d_thresh: ' + '%0.4f' % d_thresh + ' | eventtimes: ' + '%0.4f' % self.__np_d_eventtimes[idx_event]) # Increment the eventtimes index idx_event += 1 # Remove zero-valued element self.__np_d_eventtimes = np.delete(self.__np_d_eventtimes, np.where(self.__np_d_eventtimes == 0)) return self.__np_d_eventtimes # Method to estimate the RPM values def d_est_rpm(self, d_events_per_rev=1): """ Estimate the RPM from the signal using eventtimes which must have calculate with a previous call to the method np_d_est_triggers. """ # Store the new value in the object self.__d_events_per_rev = d_events_per_rev # Calculate the RPM using the difference in event times self.__np_d_rpm = 60./(np.diff(self.np_d_eventtimes)*float(d_events_per_rev)) # To keep the lengths the same, append the last sample self.__np_d_rpm = np.append(self.__np_d_rpm, self.__np_d_rpm[len(self.__np_d_rpm)-1]) return self.__np_d_rpm # Save the data def b_save_data(self, str_data_prefix = 'testclass', idx_data=1): """ Save the data in the object to a .csv file Keyword arguments: str_data_prefix -- String with file prefix (defaults to 'testclass') idx_data -- File index (defaults to 1) Return values: True if write succeeds """ str_file = str_data_prefix + '_' '%03.0f' % idx_data + '.csv' file_data = open(str_file,'w+') file_data.write('X,CH1,Start,Increment,\n') str_line = 'Sequence,Volt,Volt,0.000000e-03,' + str(self.d_t_del) file_data.write(str_line+'\n') for idx_line in range(0, self.i_ns): str_line = str(idx_line) + ',' + '%0.5f' % self.np_d_ch1[idx_line] + ',' + ',' file_data.write(str_line+'\n') file_data.close() return True ###Output _____no_output_____ ###Markdown Verify help and class structure ###Code help(cl_sig_features) ###Output Help on class cl_sig_features in module __main__: class cl_sig_features(builtins.object) | cl_sig_features(np_d_ch1, timebase_scale) | | Class to manage signal features on scope data | | Example usage: | cl_test = cl_sig_features(np.array([1.,2., 3.]),1.1) | | Should produce: | | print('np_d_ch1: '+ np.array2string(cl_test.np_d_ch1)) | print('timebase_scale: ' + '%0.3f' % cl_test.timebase_scale) | print('i_ns: ' + '%3.f' % cl_test.i_ns) | print('d_t_del: ' + '%0.3f' % cl_test.d_t_del) | print('d_time' + np.array2string(cl_test.d_time)) | | np_d_ch1: [1. 2. 3.] | timebase_scale: 1.000 | i_ns: 3 | d_t_del: 4.000 | d_time[0. 4. 8.] | | Methods defined here: | | __init__(self, np_d_ch1, timebase_scale) | Initialize self. See help(type(self)) for accurate signature. | | b_save_data(self, str_data_prefix='testclass', idx_data=1) | Save the data in the object to a .csv file | | Keyword arguments: | str_data_prefix -- String with file prefix (defaults to 'testclass') | idx_data -- File index (defaults to 1) | | Return values: | True if write succeeds | | d_est_rpm(self, d_events_per_rev=1) | Estimate the RPM from the signal using eventtimes which must have calculate | with a previous call to the method np_d_est_triggers. | | d_fft_real(self) | Calculate the half spectrum since this is a real-valued signal | | np_d_est_triggers(self, i_direction=0, d_thresh=0, d_hyst=0.1, i_kernel=5, b_verbose=False) | This method estimates speed by identifying trigger points in time, | a given threshold and hysteresis. When the signal level crosses the threshold, | the trigger holds off. The trigger holds off until the signal crosses | hysteresis levels. Hysteresis is defined relative to the threshold voltage. | | The trigger times can be used to estimate the rotating speed. | | Keyword arguments: | i_direction -- 0 to search for threshold on rising signal, 1 to search | on a falling signal. | d_thresh -- Threshold value (default: 0.0 volts for zero crossings) | d_hyst -- Hysteresis value (default: 0.1 volts) | i_kernel -- Number of samples to consider in estimating slope, | must be an odd number (default: 5) | b_verbose -- Print the intermediate steps (default: False). Useful | for stepping through the method to troubleshoot or | understand it better. | | Return values: | np_d_eventtimes -- numpy array with list of trigger event times | | plt_eventtimes(self) | Plot event data in time. | | Return values: | list: [handle to the plot, np array of eventtimes] | | plt_rpm(self) | Plot rpm data in time. | | Return values: | list: [handle to the plot, np array of RPM values] | | plt_sigs(self) | Plot out the data in this signal feature class in the time domain | | Return values: | handle to the plot | | plt_spec(self) | Plot data in frequency domain. This method assumes a real signal | | Return values: | handle to the plot | | ---------------------------------------------------------------------- | Readonly properties defined here: | | d_events_per_rev | Events per revolution | | d_fs | Sampling frequeny in hertz | | d_t_del | Delta time between each sample | | d_thresh | Trigger threshold value | | d_time | Numpy array with time values, in seconds | | i_ns | Number of samples in the scope data | | np_d_ch1_filt | Return the signal, filtered with Savitsky-Golay | | np_d_ch1_filt1 | Return the signal, filtered with butter FIR filter | | np_d_eventtimes | Numpy array of trigger event times | | np_d_rpm | Estimated RPM values | | str_filt1_desc | Complete Filt1 description of the Butterworth filter design | | str_filt1_desc_short | Short Filt1 description, useful for plot legend labels | | str_filt_desc | Complete Filt description of the Savitsky-Golay filter design | | str_filt_desc_short | Short Filt description, useful for plot legend labels | | ---------------------------------------------------------------------- | Data descriptors defined here: | | __dict__ | dictionary for instance variables (if defined) | | __weakref__ | list of weak references to the object (if defined) | | np_d_ch1 | Numpy array containing the scope data | | timebase_scale | Scope time scale ###Markdown Setup the test sequence Seed the sequence, these typically work well for 200-500 RPM and the mag pick-up gapped at 10 mils ###Code str_data_prefix = 'test010' idx_data = 0 d_timebase_scale = 1e-1 d_ch1_scale = 5.e-1 ###Output _____no_output_____ ###Markdown Acquisition loopThis loop has several steps: Acquire discovery signalThe code does not assume an RPM so it derives it from signal features. Scope setupSetup the vertical and horizontal scales on the scope. For the first pass, no trigger is used and time scale is set so that at least 2 revolutions of the lathe should be seen in the signal. This lets us see if the signal is valid and how we want to configure the trigger. Initial acquisitionOnce the setup is complete acquire the data and push the information into the signal feature class. The speed will be used to set the timescale so that we get about 5 events in the scope window Visualize the dataA few plots are presented of the scope data. Useful for troubleshooting Estimate the speedWith the event frequency known, the scope timebase_scale can be calculated ###Code while True: # Setup the scope for the trial sample acquisition d_ch1_scale = b_setup_scope(scope, d_ch1_scale=d_ch1_scale, timebase_scale=d_timebase_scale, d_trigger_level = 1e-01, b_single = False) # Acquire the test sample np_d_ch1 = d_get_data(i_ch=1, timebase_scale=d_timebase_scale) # Instatiate the class, send the waveform samples and scales cl_sig_no_trigger = cl_sig_features(np_d_ch1, d_timebase_scale) # Plot out the signal hp = cl_sig_no_trigger.plt_sigs() # The shape of the response is similar as speed increased, but the # triggering threshold has to increase to accomodate the higher # amplitudes d_thresh_est = 0.2 * (d_ch1_scale/0.5) # Calculate the trigger event times np_d_eventtimes = cl_sig_no_trigger.np_d_est_triggers(i_direction=0, d_thresh=d_thresh_est, d_hyst=0.2, b_verbose=False) np_d_eventtimes # Visualize the eventtimes hp = cl_sig_no_trigger.plt_eventtimes() # Calculated the desired timebase_scale print("d_timebase_scale (prior to adjustment): " '%0.3f' % d_timebase_scale) d_timebase_scale = (6./12.)*(np.mean(np.diff(np_d_eventtimes))) print("d_timebase_scale (after adjustment): " '%0.3f' % d_timebase_scale) # Check for clipping and correct scaling. The scope has 8 vertical division so the # total voltage range on the screen is 8 * d_ch1_scale d_pkpk = np.max(np_d_ch1) - np.min(np_d_ch1) print("d_pkpk: " + "%0.4f" % d_pkpk) d_volts_scale = (8*d_ch1_scale) print("d_volts_scale: " + "%0.4f" % d_volts_scale) if ( d_pkpk > d_volts_scale ): print("Voltage scale reduced") d_ch1_scale = d_ch1_scale*2. # Could be the vertical scale is too small, check for that if ( abs( d_volts_scale/d_pkpk > 2. )): print("Voltage scale increased") d_ch1_scale = d_ch1_scale/2. # The scope trigger setting scales with the overall amplitude since the # shape of the response is similar d_trigger_level_est = 0.2 * (d_ch1_scale/0.5) # Reset the scope with the adjusted features, set to trigger on single sample b_setup_scope(scope, d_ch1_scale=d_ch1_scale, timebase_scale=d_timebase_scale, d_trigger_level = 1e-01, b_single = True) # Acquire the sample np_d_ch1 = d_get_data(i_ch=1, timebase_scale=d_timebase_scale) # Reset back to free-run b_setup_scope(scope, d_ch1_scale=d_ch1_scale, timebase_scale=d_timebase_scale, d_trigger_level = 1e-01, b_single = False) # Instatiate the class, send the waveform samples and scales cl_sig_no_trigger = cl_sig_features(np_d_ch1, d_timebase_scale) # Visualize the data hp = cl_sig_no_trigger.plt_sigs() # Save it off to a file b_file_save = cl_sig_no_trigger.b_save_data(str_data_prefix = str_data_prefix, idx_data = idx_data) # Wait for the next speed adjustment and continue idx_data += 1 time.sleep(2) input("Press Enter to continue...") ###Output _____no_output_____
Course2/week2/week-2-multiple-regression-assignment-2-blank.ipynb
###Markdown Regression Week 2: Multiple Regression (gradient descent) In the first notebook we explored multiple regression using graphlab create. Now we will use graphlab along with numpy to solve for the regression weights with gradient descent.In this notebook we will cover estimating multiple regression weights via gradient descent. You will:* Add a constant column of 1's to a graphlab SFrame to account for the intercept* Convert an SFrame into a Numpy array* Write a predict_output() function using Numpy* Write a numpy function to compute the derivative of the regression weights with respect to a single feature* Write gradient descent function to compute the regression weights given an initial weight vector, step size and tolerance.* Use the gradient descent function to estimate regression weights for multiple features Fire up graphlab create Make sure you have the latest version of graphlab (>= 1.7) ###Code import graphlab ###Output _____no_output_____ ###Markdown Load in house sales dataDataset is from house sales in King County, the region where the city of Seattle, WA is located. ###Code sales = graphlab.SFrame('kc_house_data.gl/') ###Output [INFO] graphlab.cython.cy_server: GraphLab Create v2.1 started. Logging: /tmp/graphlab_server_1477228368.log ###Markdown If we want to do any "feature engineering" like creating new features or adjusting existing ones we should do this directly using the SFrames as seen in the other Week 2 notebook. For this notebook, however, we will work with the existing features. Convert to Numpy Array Although SFrames offer a number of benefits to users (especially when using Big Data and built-in graphlab functions) in order to understand the details of the implementation of algorithms it's important to work with a library that allows for direct (and optimized) matrix operations. Numpy is a Python solution to work with matrices (or any multi-dimensional "array").Recall that the predicted value given the weights and the features is just the dot product between the feature and weight vector. Similarly, if we put all of the features row-by-row in a matrix then the predicted value for *all* the observations can be computed by right multiplying the "feature matrix" by the "weight vector". First we need to take the SFrame of our data and convert it into a 2D numpy array (also called a matrix). To do this we use graphlab's built in .to_dataframe() which converts the SFrame into a Pandas (another python library) dataframe. We can then use Panda's .as_matrix() to convert the dataframe into a numpy matrix. ###Code import numpy as np # note this allows us to refer to numpy as np instead ###Output _____no_output_____ ###Markdown Now we will write a function that will accept an SFrame, a list of feature names (e.g. ['sqft_living', 'bedrooms']) and an target feature e.g. ('price') and will return two things:* A numpy matrix whose columns are the desired features plus a constant column (this is how we create an 'intercept')* A numpy array containing the values of the outputWith this in mind, complete the following function (where there's an empty line you should write a line of code that does what the comment above indicates)**Please note you will need GraphLab Create version at least 1.7.1 in order for .to_numpy() to work!** ###Code def get_numpy_data(data_sframe, features, output): data_sframe['constant'] = 1 # this is how you add a constant column to an SFrame # add the column 'constant' to the front of the features list so that we can extract it along with the others: features = ['constant'] + features # this is how you combine two lists # select the columns of data_SFrame given by the features list into the SFrame features_sframe (now including constant): features_sframe = graphlab.SFrame() for feature in features: features_sframe[feature] = data_sframe[feature] # the following line will convert the features_SFrame into a numpy matrix: feature_matrix = features_sframe.to_numpy() # assign the column of data_sframe associated with the output to the SArray output_sarray output_sarray = data_sframe[output] # the following will convert the SArray into a numpy array by first converting it to a list output_array = output_sarray.to_numpy() return(feature_matrix, output_array) ###Output _____no_output_____ ###Markdown For testing let's use the 'sqft_living' feature and a constant as our features and price as our output: ###Code (example_features, example_output) = get_numpy_data(sales, ['sqft_living'], 'price') # the [] around 'sqft_living' makes it a list print example_features[0,:] # this accesses the first row of the data the ':' indicates 'all columns' print example_output[0] # and the corresponding output ###Output [ 1.00000000e+00 1.18000000e+03] 221900.0 ###Markdown Predicting output given regression weights Suppose we had the weights [1.0, 1.0] and the features [1.0, 1180.0] and we wanted to compute the predicted output 1.0\*1.0 + 1.0\*1180.0 = 1181.0 this is the dot product between these two arrays. If they're numpy arrayws we can use np.dot() to compute this: ###Code my_weights = np.array([1., 1.]) # the example weights my_features = example_features[0,] # we'll use the first data point predicted_value = np.dot(my_features, my_weights) print predicted_value ###Output 1181.0 ###Markdown np.dot() also works when dealing with a matrix and a vector. Recall that the predictions from all the observations is just the RIGHT (as in weights on the right) dot product between the features *matrix* and the weights *vector*. With this in mind finish the following predict_output function to compute the predictions for an entire matrix of features given the matrix and the weights: ###Code def predict_output(feature_matrix, weights): # assume feature_matrix is a numpy matrix containing the features as columns and weights is a corresponding numpy array # create the predictions vector by using np.dot() predictions = np.dot(feature_matrix, weights) return(predictions) ###Output _____no_output_____ ###Markdown If you want to test your code run the following cell: ###Code test_predictions = predict_output(example_features, my_weights) print test_predictions[0] # should be 1181.0 print test_predictions[1] # should be 2571.0 ###Output 1181.0 2571.0 ###Markdown Computing the Derivative We are now going to move to computing the derivative of the regression cost function. Recall that the cost function is the sum over the data points of the squared difference between an observed output and a predicted output.Since the derivative of a sum is the sum of the derivatives we can compute the derivative for a single data point and then sum over data points. We can write the squared difference between the observed output and predicted output for a single point as follows:(w[0]\*[CONSTANT] + w[1]\*[feature_1] + ... + w[i] \*[feature_i] + ... + w[k]\*[feature_k] - output)^2Where we have k features and a constant. So the derivative with respect to weight w[i] by the chain rule is:2\*(w[0]\*[CONSTANT] + w[1]\*[feature_1] + ... + w[i] \*[feature_i] + ... + w[k]\*[feature_k] - output)\* [feature_i]The term inside the paranethesis is just the error (difference between prediction and output). So we can re-write this as:2\*error\*[feature_i]That is, the derivative for the weight for feature i is the sum (over data points) of 2 times the product of the error and the feature itself. In the case of the constant then this is just twice the sum of the errors!Recall that twice the sum of the product of two vectors is just twice the dot product of the two vectors. Therefore the derivative for the weight for feature_i is just two times the dot product between the values of feature_i and the current errors. With this in mind complete the following derivative function which computes the derivative of the weight given the value of the feature (over all data points) and the errors (over all data points). ###Code def feature_derivative(errors, feature): # Assume that errors and feature are both numpy arrays of the same length (number of data points) # compute twice the dot product of these vectors as 'derivative' and return the value derivative = 2 * np.dot(errors, feature) return(derivative) ###Output _____no_output_____ ###Markdown To test your feature derivartive run the following: ###Code (example_features, example_output) = get_numpy_data(sales, ['sqft_living'], 'price') my_weights = np.array([0., 0.]) # this makes all the predictions 0 test_predictions = predict_output(example_features, my_weights) # just like SFrames 2 numpy arrays can be elementwise subtracted with '-': errors = test_predictions - example_output # prediction errors in this case is just the -example_output feature = example_features[:,0] # let's compute the derivative with respect to 'constant', the ":" indicates "all rows" derivative = feature_derivative(errors, feature) print derivative print -np.sum(example_output)*2 # should be the same as derivative ###Output -23345850022.0 -23345850022.0 ###Markdown Gradient Descent Now we will write a function that performs a gradient descent. The basic premise is simple. Given a starting point we update the current weights by moving in the negative gradient direction. Recall that the gradient is the direction of *increase* and therefore the negative gradient is the direction of *decrease* and we're trying to *minimize* a cost function. The amount by which we move in the negative gradient *direction* is called the 'step size'. We stop when we are 'sufficiently close' to the optimum. We define this by requiring that the magnitude (length) of the gradient vector to be smaller than a fixed 'tolerance'.With this in mind, complete the following gradient descent function below using your derivative function above. For each step in the gradient descent we update the weight for each feature befofe computing our stopping criteria ###Code from math import sqrt # recall that the magnitude/length of a vector [g[0], g[1], g[2]] is sqrt(g[0]^2 + g[1]^2 + g[2]^2) def regression_gradient_descent(feature_matrix, output, initial_weights, step_size, tolerance): converged = False weights = np.array(initial_weights) # make sure it's a numpy array while not converged: # compute the predictions based on feature_matrix and weights using your predict_output() function predictions = predict_output(feature_matrix, weights) # compute the errors as predictions - output errors = predictions - output gradient_sum_squares = 0 # initialize the gradient sum of squares # while we haven't reached the tolerance yet, update each feature's weight for i in range(len(weights)): # loop over each weight # Recall that feature_matrix[:, i] is the feature column associated with weights[i] # compute the derivative for weight[i]: _derivative = feature_derivative(errors, feature_matrix[:, i]) # add the squared value of the derivative to the gradient sum of squares (for assessing convergence) gradient_sum_squares += _derivative * _derivative # subtract the step size times the derivative from the current weight weights[i] -= _derivative * step_size # compute the square-root of the gradient sum of squares to get the gradient magnitude: gradient_magnitude = sqrt(gradient_sum_squares) if gradient_magnitude < tolerance: converged = True return(weights) ###Output ###Markdown A few things to note before we run the gradient descent. Since the gradient is a sum over all the data points and involves a product of an error and a feature the gradient itself will be very large since the features are large (squarefeet) and the output is large (prices). So while you might expect "tolerance" to be small, small is only relative to the size of the features. For similar reasons the step size will be much smaller than you might expect but this is because the gradient has such large values. Running the Gradient Descent as Simple Regression First let's split the data into training and test data. ###Code train_data,test_data = sales.random_split(.8,seed=0) ###Output _____no_output_____ ###Markdown Although the gradient descent is designed for multiple regression since the constant is now a feature we can use the gradient descent function to estimat the parameters in the simple regression on squarefeet. The folowing cell sets up the feature_matrix, output, initial weights and step size for the first model: ###Code # let's test out the gradient descent simple_features = ['sqft_living'] my_output = 'price' (simple_feature_matrix, output) = get_numpy_data(train_data, simple_features, my_output) initial_weights = np.array([-47000., 1.]) step_size = 7e-12 tolerance = 2.5e7 ###Output _____no_output_____ ###Markdown Next run your gradient descent with the above parameters. ###Code myGrad1 = regression_gradient_descent(simple_feature_matrix, output, initial_weights, step_size, tolerance) print myGrad1 ###Output [-46999.88716555 281.91211912] ###Markdown How do your weights compare to those achieved in week 1 (don't expect them to be exactly the same)? **Quiz Question: What is the value of the weight for sqft_living -- the second element of ‘simple_weights’ (rounded to 1 decimal place)?** Use your newly estimated weights and your predict_output() function to compute the predictions on all the TEST data (you will need to create a numpy array of the test feature_matrix and test output first: ###Code (test_simple_feature_matrix, test_output) = get_numpy_data(test_data, simple_features, my_output) ###Output _____no_output_____ ###Markdown Now compute your predictions using test_simple_feature_matrix and your weights from above. ###Code test_predictions = predict_output(test_simple_feature_matrix, myGrad1) ###Output _____no_output_____ ###Markdown **Quiz Question: What is the predicted price for the 1st house in the TEST data set for model 1 (round to nearest dollar)?** ###Code print test_predictions[0] ###Output 356134.443171 ###Markdown Now that you have the predictions on test data, compute the RSS on the test data set. Save this value for comparison later. Recall that RSS is the sum of the squared errors (difference between prediction and output). ###Code i=0 RSS = 0 while i < len(test_data): error = test_predictions[i] - test_data["price"][i] error = error*error RSS += error i += 1 print RSS ###Output 2.75400047593e+14 ###Markdown Running a multiple regression Now we will use more than one actual feature. Use the following code to produce the weights for a second model with the following parameters: ###Code model_features = ['sqft_living', 'sqft_living15'] # sqft_living15 is the average squarefeet for the nearest 15 neighbors. my_output = 'price' (feature_matrix, output) = get_numpy_data(train_data, model_features, my_output) initial_weights = np.array([-100000., 1., 1.]) step_size = 4e-12 tolerance = 1e9 ###Output _____no_output_____ ###Markdown Use the above parameters to estimate the model weights. Record these values for your quiz. ###Code myGrad2 = regression_gradient_descent(feature_matrix, output, initial_weights, step_size, tolerance) print myGrad2 ###Output [ -9.99999688e+04 2.45072603e+02 6.52795277e+01] ###Markdown Use your newly estimated weights and the predict_output function to compute the predictions on the TEST data. Don't forget to create a numpy array for these features from the test set first! ###Code (test_simple_feature_matrix2, test_output) = get_numpy_data(test_data, model_features, my_output) test_predictions2 = predict_output(test_simple_feature_matrix2, myGrad2) ###Output _____no_output_____ ###Markdown **Quiz Question: What is the predicted price for the 1st house in the TEST data set for model 2 (round to nearest dollar)?** ###Code test_predictions2[0] ###Output _____no_output_____ ###Markdown What is the actual price for the 1st house in the test data set? ###Code test_data[0]["price"] ###Output _____no_output_____ ###Markdown **Quiz Question: Which estimate was closer to the true price for the 1st house on the TEST data set, model 1 or model 2?**model2 Now use your predictions and the output to compute the RSS for model 2 on TEST data. ###Code i=0 RSS = 0 while i < len(test_data): error = test_predictions2[i] - test_data["price"][i] error = error*error RSS += error i += 1 print RSS ###Output 2.70263446465e+14
Python/SpaCy/spaCy.ipynb
###Markdown spaCy experiments Imports & initialization Import the required modules. ###Code import collections import itertools import matplotlib.pyplot as plt import numpy as np import spacy ###Output _____no_output_____ ###Markdown Create a language model, English in this case. ###Code en_nlp = spacy.load('en_core_web_sm') ###Output _____no_output_____ ###Markdown Part of speech tagging (POS) Read a text file into a string variable. ###Code with open('Data/frost.txt') as file: text = ''.join(file.readlines()) ###Output _____no_output_____ ###Markdown Parse the text using the language model. ###Code doc = en_nlp(text) ###Output _____no_output_____ ###Markdown Show the part of speech tags, as well as the context of the words. ###Code for word in doc: print(f'{word.text!r}: {word.pos_}, ' f'{word.left_edge.text!r} <- {word.head.text!r} -> {word.right_edge.text!r}') ###Output _____no_output_____ ###Markdown Since we can't use backslashes in f-strings, we define a constant to represent it. ###Code newline = '\n' ###Output _____no_output_____ ###Markdown To split a text in sentences, a statistical model is used that was obtained from the training corpus. ###Code for i, sentence in enumerate(doc.sents): print(f'{i:3d} {sentence.text.replace(newline, " ")}') ###Output _____no_output_____ ###Markdown For poetry, sentences seem somewhat hard to detect. However, it is possible to define a language model for English and add a rule-based sentencizer to it. ###Code en_nlp_alt = spacy.lang.en.English() sentencizer = en_nlp_alt.create_pipe('sentencizer') en_nlp_alt.add_pipe(sentencizer) doc = en_nlp_alt(text) for i, sentence in enumerate(doc.sents): print(f'{i:3d} {sentence.text.replace(newline, " ").strip()}') ###Output _____no_output_____ ###Markdown Lemmatization By way of example, consider Plato's *Republic*. This is a fairly long text. ###Code !wc -l -w Data/republic.mb.txt with open('Data/republic.mb.txt') as file: text = ''.join(file.readlines()) ###Output _____no_output_____ ###Markdown The full result of the language model parsing this text would be rather large, but for our purposes, we require only tokenization, not POS or NER, hence we disable these features. ###Code doc = en_nlp(text, disable=['parser', 'ner']) ###Output _____no_output_____ ###Markdown We can now perform lemmatization on all words that are not stop words, and we also eliminate named entities (`-PROP-` as value for lemma) and punctuation. On the resulting list, a word count is performed. ###Code stopwords = en_nlp.Defaults.stop_words | {'\n', '\n\n', '-PRON-'} punctuation = ',.;?!:-' counts = collections.Counter([token.lemma_.lower() for token in doc if token.lemma_ not in stopwords and token.lemma_ not in punctuation]) ###Output _____no_output_____ ###Markdown The top-20 words are given below. ###Code counts.most_common(20) def plot_distr(counts, nr_words): words = list() numbers = list() for word, number in counts.most_common(nr_words): words.append(word) numbers.append(number) figure, axes = plt.subplots(1, 1, figsize=(15, 5)) axes.bar(words, numbers) axes.set_xticklabels(words, rotation=45) plot_distr(counts, 30) ###Output _____no_output_____ ###Markdown Named entiry recognition (NER) Named entity recognition is supported as well. ###Code sentence = 'Music by Johann Sebastian Bach is better than that by Friederich Buxtehude. Both lived in Germany' doc = en_nlp(sentence) for i, word in enumerate(doc): print(f'{i:3d} {word.text!r}: {word.pos_}, {word.ent_type_}') ###Output _____no_output_____ ###Markdown It is also possible to retrieve named entities from the document explicitly. ###Code for entity in doc.ents: print(f'{entity} ({entity.label_}): {entity.start} -> {entity.end}') ###Output _____no_output_____ ###Markdown Note that the first name of Buxtehude is in fact Dietrich, not Friederich. Nevertheless, the NER marks `Friederich Buxtehude` as a person. This can also be visualized as markup in the sentence. ###Code spacy.displacy.render(doc, style='ent', jupyter=True) spacy.displacy.render(doc, style='dep', jupyter=True, options={'distance': 140, 'compact': True}) ###Output _____no_output_____ ###Markdown Similarity Document similarity can also be computed conveniently. ###Code doc1 = en_nlp('The book is nice') doc2 = en_nlp('The novel is beautiful') doc1.similarity(doc2) doc1 = en_nlp('The book is nice') doc2 = en_nlp('The house is on fire') doc1.similarity(doc2) words = ['queen', 'lady', 'girl', 'king', 'lord', 'boy', 'cat', 'dog', 'lion'] similarity = np.empty((len(words), len(words))) for i, word1 in enumerate(words): for j, word2 in enumerate(words): similarity[i, j] = en_nlp(word1).similarity(en_nlp(word2)) ###Output _____no_output_____ ###Markdown The similarity matrix can be visualized as a heat map using the following function: ###Code def plot_similarity_matrix(sim, words, cmap=plt.cm.Blues): figure, axes = plt.subplots(figsize=(6, 6)) axes.imshow(sim, interpolation='nearest', cmap=cmap) axes.set_xticks(range(len(words))) axes.set_xticklabels(words, rotation=45) axes.set_yticks(range(len(words))) axes.set_yticklabels(words) fmt = '{0:.2f}' thresh = 0.5*(sim.max() + sim.min()) for i, j in itertools.product(range(sim.shape[0]), range(sim.shape[1])): axes.text(j, i, fmt.format(sim[i, j]), horizontalalignment="center", color="white" if sim[i, j] > thresh else "black", fontsize=8) figure.tight_layout() axes.set_xlabel('word 1') axes.set_ylabel('word 2') plot_similarity_matrix(similarity, words) ###Output _____no_output_____ ###Markdown However, small language models don't contain real word vectors, only context sensitive tensors. We can repeat the computation above with a medium sized language model. ###Code en_nlp_md = spacy.load('en_core_web_md') similarity = np.empty((len(words), len(words))) for i, word1 in enumerate(words): for j, word2 in enumerate(words): similarity[i, j] = en_nlp_md(word1).similarity(en_nlp_md(word2)) plot_similarity_matrix(similarity, words) doc1 = en_nlp_md('The book is nice') doc2 = en_nlp_md('The novel is beautiful') doc1.similarity(doc2) doc1 = en_nlp_md('The book is nice') doc2 = en_nlp_md('The house is on fire') doc1.similarity(doc2) doc1 = en_nlp_md('Stock prices for Intel are on the rise.') doc2 = en_nlp_md('The value of NVIDIA shares is increasing.') doc1.similarity(doc2) doc1 = en_nlp_md('Stock prices for Intel are on the rise.') doc2 = en_nlp_md('The economy of Denmark is flourishing.') doc1.similarity(doc2) doc1 = en_nlp_md('Stock prices for Intel are on the rise.') doc2 = en_nlp_md('The value of NVIDIA shares is plumetting.') doc1.similarity(doc2) ###Output _____no_output_____
Matplotlib_multivariate/Encodings_Practice.ipynb
###Markdown In this notebook, you'll be working with the Pokémon dataset from the univariate plots lesson. ###Code pokemon = pd.read_csv('./data/pokemon.csv') pokemon.head() ###Output _____no_output_____ ###Markdown **Task 1**: To start, let's look at the relationship between the Pokémon combat statistics of Speed, Defense, and Special-Defense. If a Pokémon has higher defensive statistics, does it necessarily sacrifice speed? Create a single plot to depict this relationship. ###Code # YOUR CODE HERE # run this cell to check your work against ours encodings_solution_1() ###Output When creating the plot, I made the figure size bigger and set axis limits to zoom into the majority of data points. I might want to apply some manual jitter to the data since I suspect there to be a lot of overlapping points. From the plot as given, I see a slight increase in speed as both defense and special defense increase. However, the brightest points seem to be clumped up in the center in the 60-80 defense and special defense ranges with the two brightest points on the lower left of the diagonal. ###Markdown To complete the second task, we need to first reshape the dataset so that all Pokémon types are recorded in a single column. This will add duplicates of Pokémon with two types, which is fine for the task to be performed. ###Code type_cols = ['type_1','type_2'] non_type_cols = pokemon.columns.difference(type_cols) pkmn_types = pokemon.melt(id_vars = non_type_cols, value_vars = type_cols, var_name = 'type_level', value_name = 'type').dropna() pkmn_types.head() ###Output _____no_output_____ ###Markdown **Task 2**: How do weights and heights compare between Fairy type Pokémon and Dragon type Pokémon? You may want to subset your dataframe before proceeding with the plotting code. **Hint**: If you remember from the univariate plots lesson, one of your axis variables may need to be transformed. If you plan on using FacetGrid, its `.set()` method will be vital for adjusting the axis scaling and tick marks. Check the [last example in the Seaborn documentation](https://seaborn.pydata.org/generated/seaborn.FacetGrid.html) for an example of how the `.set()` method is used, and the [matplotlib documentation of the Axes class](https://matplotlib.org/api/axes_api.html) for properties that you might want to set. ###Code # YOUR CODE HERE # run this cell to check your work against ours encodings_solution_2() ###Output After subsetting the data, I used FacetGrid to set up and generate the plot. I used the .set() method for FacetGrid objects to set the x-scaling and tick marks. The plot shows the drastic difference in sizes and weights for the Fairy and Dragon Pokemon types.
Frequentist_vs_Bayesian_Regression.ipynb
###Markdown 1. Simple Linear Regression Load the data ###Code datafolder = "/content/drive/My Drive/NUS/BT4012/" file_name = "student-mat.csv" df_data = pd.read_csv(datafolder + file_name, sep=';', index_col=None) df_data.rename(columns={'G3': 'Grade'}, inplace=True) df_data = df_data[~df_data['Grade'].isin([0, 1])] df_used = df_data[['studytime', 'Medu', 'Grade']] df_used.head(2) df_X = df_used[['studytime', 'Medu']] #store features df_y = df_used[['Grade']] # Split into training/testing sets with 25% split X_train, X_test, y_train, y_test = train_test_split(df_X, df_y, test_size = 0.25, random_state=123) ###Output _____no_output_____ ###Markdown *Train linear regression model on X_train and y_train** Adopt the default hyperparamter setting ###Code ## write your code lr = LinearRegression() lr.fit(X_train, y_train) ###Output _____no_output_____ ###Markdown *Check MAE and RMSE on testing data* ###Code ## Write your code predictions = lr.predict(X_test) mae = np.mean(abs(predictions - y_test)) rmse = np.sqrt(np.mean((predictions - y_test)**2)) print('MAE: %0.2f' % mae) print('RMSE: %0.2f' % rmse) ###Output MAE: 2.84 RMSE: 3.45 ###Markdown *Check the predicted grade of one student*The 10th student in the testing data ###Code ## write your code predictions[9] ###Output _____no_output_____ ###Markdown *Print the learned model parameters of linear regression* ###Code ## Write your code intercept = lr.intercept_[0] coef = lr.coef_ formula = 'Grade = %0.2f +' % intercept for i, col in enumerate(X_train.columns): formula += ' %0.2f * %s +' % (coef[0][i], col) print(formula[:-2]) print() print("For an unit increase in study time, expected grade increases by 0.68, independent of all other variables.\ Likewise, for an unit increase in Medu, expected grade increases by 0.47, independent of all other variables.") ###Output Grade = 8.75 + 0.68 * studytime + 0.47 * Medu For an unit increase in study time, expected grade increases by 0.68, independent of all other variables. Likewise, for an unit increase in Medu, expected grade increases by 0.47, independent of all other variables. ###Markdown 2. Bayesian Linear RegressionHere, two bayesian models will be implemented with **two different sets of prior functions**. The first bayeisan model is given as:$u_i = \beta_0 + \beta_1*{studytime}_i + \beta_2*{medu}_i$$grade_i \sim Norm(u_i, \sigma^2_\epsilon)$$\beta_0 \sim Norm(0, 1)$$\beta_1 \sim Norm(0, 1)$$\beta_2 \sim Norm(0, 100)$$\sigma_\epsilon \sim {Uniform}(0, 10)$Here, $\beta_0$ is the intercept. Then, $\beta_1$ and $\beta_2$ are the coefficients for features: studytime and medu. For the i-th datasample, a mean $u_i$ can be computed linearly from two features. Then, the target grade $y_i$ is assumed to be normally distributed around this $u_i$. Make sure the version of pymc3 is 3.8 ###Code ! pip install pymc3==3.8 import pymc3 as pm ## Define your model here def model_build(df_train, df_label=None): """ build genearlized linear model """ with pm.Model() as model: ## write your code here num_fea = df_train.shape[1] #error term sigma = pm.Uniform('sigma', 0, 10) #intercept mu_infe = pm.Normal('intercept', mu=0, sigma=1) #beta1 mu_infe = mu_infe + pm.Normal('beta_1_coeff_for_{}'.format(df_train.columns[0]), mu=0, sigma=1)*df_train.loc[:, df_train.columns[0]] #beta2 mu_infe = mu_infe + pm.Normal('beta_2_coeff_for_{}'.format(df_train.columns[1]), mu=0, sigma=10)*df_train.loc[:, df_train.columns[1]] if df_label is None: # inference likelihood = pm.Normal('y', mu=mu_infe, sigma=sigma, observed = False) else: # training likelihood = pm.Normal('y', mu=mu_infe, sigma=sigma, observed = df_label['Grade'].values) return model # Use MCMC algorithm to draw samples to approximate the posterior for model parameters (error term, bias term and all coefficients) with model_build(X_train, y_train): trace = pm.sample(draws=2000, chains = 2, tune = 500) # sample the posterior predictive distribution for the 10th student in testing data # 4000 samples (2 chains and each chain has 2000 samples) will be sampled for this student. with model_build(X_test.iloc[9:10,:]): ppc = pm.sample_posterior_predictive(trace) ###Output 100%|██████████| 4000/4000 [00:07<00:00, 521.53it/s] ###Markdown *Compute the mean and standard deviation of your prediction.* ###Code ## write your code print("The mean of prediction is %0.3f and the standard deviation is %0.3f " %(np.mean(ppc['y']),np.std(ppc['y']))) ###Output The mean of prediction is 12.353 and the standard deviation is 3.163 ###Markdown *Check the posterior distribution for the model parameters*$p(w|D)$ ###Code ## write your code here print(pm.summary(trace).round(5)) pm.plot_posterior(trace, figsize = (12, 3)) ###Output /usr/local/lib/python3.6/dist-packages/arviz/data/io_pymc3.py:89: FutureWarning: Using `from_pymc3` without the model will be deprecated in a future release. Not using the model will return less accurate and less useful results. Make sure you use the model argument or call from_pymc3 within a model context. FutureWarning, ###Markdown The other bayeisan model is given as:$u_i = \beta_0 + \beta_1*{studytime}_i + \beta_2*{medu}_i$$grade_i \sim Norm(u_i, \sigma^2_\epsilon)$$\beta_0 \sim Norm(0, 100)$$\beta_1 \sim Norm(0, 100)$$\beta_2 \sim Norm(0, 100)$$\sigma_\epsilon \sim {Uniform}(0, 10)$ ###Code ## Define your model here def model_build(df_train, df_label=None): """ build genearlized linear model """ with pm.Model() as model: ## write your code here num_fea = df_train.shape[1] #error term sigma = pm.Uniform('sigma', 0, 10) #intercept mu_infe = pm.Normal('intercept', mu=0, sigma=10) #beta1 mu_infe = mu_infe + pm.Normal('beta_1_coeff_for_{}'.format(df_train.columns[0]), mu=0, sigma=10)*df_train.loc[:, df_train.columns[0]] #beta2 mu_infe = mu_infe + pm.Normal('beta_2_coeff_for_{}'.format(df_train.columns[1]), mu=0, sigma=10)*df_train.loc[:, df_train.columns[1]] if df_label is None: # inference likelihood = pm.Normal('y', mu=mu_infe, sigma=sigma, observed = False) else: # training likelihood = pm.Normal('y', mu=mu_infe, sigma=sigma, observed = df_label['Grade'].values) return model # Use MCMC algorithm to draw samples to approximate the posterior for model parameters (error term, bias term and all coefficients) with model_build(X_train, y_train): trace = pm.sample(draws=2000, chains = 2, tune = 500) ###Output Auto-assigning NUTS sampler... Initializing NUTS using jitter+adapt_diag... Sequential sampling (2 chains in 1 job) NUTS: [beta_2_coeff_for_Medu, beta_1_coeff_for_studytime, intercept, sigma] Sampling chain 0, 0 divergences: 100%|██████████| 2500/2500 [00:06<00:00, 367.14it/s] Sampling chain 1, 0 divergences: 100%|██████████| 2500/2500 [00:06<00:00, 394.84it/s] The acceptance probability does not match the target. It is 0.881162011783612, but should be close to 0.8. Try to increase the number of tuning steps. ###Markdown *Similar to the first bayesian linear regression model, check the distribution for model parameters and the prediction distribution of the previous chosen data sample* ###Code ## write your code # prediction distribution with model_build(X_test.iloc[9:10,:]): # 9th student ppc = pm.sample_posterior_predictive(trace) print("The mean of prediction is %0.3f and the standard deviation is %0.3f " %(np.mean(ppc['y']),np.std(ppc['y']))) # distribution of model parameters print(pm.summary(trace).round(5)) pm.plot_posterior(trace, figsize = (12, 3)) ###Output 100%|██████████| 4000/4000 [00:05<00:00, 688.23it/s] /usr/local/lib/python3.6/dist-packages/arviz/data/io_pymc3.py:89: FutureWarning: Using `from_pymc3` without the model will be deprecated in a future release. Not using the model will return less accurate and less useful results. Make sure you use the model argument or call from_pymc3 within a model context. FutureWarning,
notebooks/10-data-prep/01.20-ol-entry.ipynb
###Markdown Code for cleanning and preparing the entry surveys data ###Code from __future__ import absolute_import, division, print_function import datetime import time import os import pandas as pd import sys sys.path.insert(0, '../../src/data/') from config import * ###Output _____no_output_____ ###Markdown Config ###Code # Matplotlib for additional customization from matplotlib import pyplot as plt %matplotlib inline # Seaborn for plotting and styling import seaborn as sns ###Output 2019-01-18 13:20:44,777 - DEBUG - backend module://ipykernel.pylab.backend_inline version unknown ###Markdown read ###Code participants_entry_survey_data_anon = pd.read_hdf(surveys_anon_store_path, 'entry/participants_entry_survey_data_anon') participants_entry_survey_data_anon.head() ###Output _____no_output_____ ###Markdown Rename columns ###Code columns_dict = { 'member':'member', 'Q3.3':'race', 'Q3.3_7_TEXT':'race_other', 'Q3.2':'age', 'Q3.1':'gender', 'Q3.4':'citizneships', 'Q5.3':'is_cofounder', 'Q5.5_1_TEXT':'title', 'Q5.6':'time_in_startup', 'Q6.1':'HHH_type', 'Q6.2':'experience', 'Q7.2_1':'TIPI_1', 'Q7.2_2':'TIPI_2', 'Q7.2_3':'TIPI_3', 'Q7.2_4':'TIPI_4', 'Q7.2_5':'TIPI_5', 'Q7.2_6':'TIPI_6', 'Q7.2_7':'TIPI_7', 'Q7.2_8':'TIPI_8', 'Q7.2_9':'TIPI_9', 'Q7.2_10':'TIPI_10', } participants_entry_survey_data_clean = participants_entry_survey_data_anon.rename(columns=columns_dict) participants_entry_survey_data_clean.head(5) ###Output _____no_output_____ ###Markdown Translate values Big 5 TIPI Transalte text to numbers ###Code participants_entry_survey_data_clean.TIPI_1.unique() def TIPI_Translation(answer): """ A function to translate the string answers from the TIPI survey to numbers """ if answer == 'Agree strongly': return(7.0) if answer == 'Agree moderately': return(6.0) if answer == 'Agree a little': return(5.0) if answer == 'Neither agree nor disagree': return(4.0) if answer == 'Disagree a little': return(3.0) if answer == 'Disagree moderately': return(2.0) if answer == 'Disagree strongly ': #extra space.... return(1.0) if answer == 'Disagree strongly': return(1.0) return answer #Apply translation functions for i in range(1,11): participants_entry_survey_data_clean['TIPI_{}'.format(i)] = participants_entry_survey_data_clean['TIPI_{}'.format(i)].apply(TIPI_Translation) ###Output _____no_output_____ ###Markdown Transalte answers to personalityTIPI scale scoring (“R” denotes reverse-scored items):* Extraversion: 1, 6R* Agreeableness: 2R, 7* Conscientiousness; 3, 8R* Emotional Stability: 4R, 9* Openness to Experiences: 5, 10RScoring the TIPI:1. Recode the reverse-scored items (i.e., recode a 7 with a 1, a 6 with a 2, a 5 with a 3, etc.). The reverse scored items are 2, 4, 6, 8, & 10.2. Take the AVERAGE of the two items (the standard item and the recoded reverse-scored item) that make up each scale.Example using the Extraversion scale: A participant has scores of 5 on item 1 (Extraverted, enthusiastic) and and 2 on item 6 (Reserved, quiet). First, recode the reverse-scored item (i.e., item 6), replacing the 2 with a 6. Second, take the average of the score for item 1 and the (recoded) score for item 6. So the TIPI Extraversion scale score would be: (5 + 6)/2 = 5.5 ###Code for i in range(2,11,2): participants_entry_survey_data_clean['TIPI_{}R'.format(i)] = 8 - participants_entry_survey_data_clean['TIPI_{}'.format(i)] df = participants_entry_survey_data_clean participants_entry_survey_data_clean['TIPI_extraversion'] = (df['TIPI_1']+df['TIPI_6R'])/2 participants_entry_survey_data_clean['TIPI_agreeableness'] = (df['TIPI_2R']+df['TIPI_7'])/2 participants_entry_survey_data_clean['TIPI_conscientiousness'] = (df['TIPI_3']+df['TIPI_8R'])/2 participants_entry_survey_data_clean['TIPI_emotional_stability'] = (df['TIPI_4R']+df['TIPI_9'])/2 participants_entry_survey_data_clean['TIPI_openness'] = (df['TIPI_5']+df['TIPI_10R'])/2 for i in range(1,11): del participants_entry_survey_data_clean['TIPI_{}'.format(i)] for i in range(2,11,2): del participants_entry_survey_data_clean['TIPI_{}R'.format(i)] participants_entry_survey_data_clean.head() ###Output _____no_output_____ ###Markdown is CEO? ###Code participants_entry_survey_data_clean.title.value_counts() #CEO #Ceo #participants_entry_survey_data_clean['is_ceo'] = \ # participants_entry_survey_data_clean.title.str.contains('ceo',case=False)\ # .fillna(False) participants_entry_survey_data_clean['is_ceo'] = 0 cond = (participants_entry_survey_data_clean.title.fillna("").str.contains("ceo",case=False)) participants_entry_survey_data_clean.loc[cond,'is_ceo'] = 1 participants_entry_survey_data_clean.query('is_ceo == 1')[['title','is_ceo']] participants_entry_survey_data_clean[['member','title','is_ceo']].query('is_ceo') participants_entry_survey_data_clean.head() ###Output _____no_output_____ ###Markdown Is co-founder ###Code # Set for 'Yes and maybe' participants_entry_survey_data_clean['is_cofounder_temp'] = 0 cond = (participants_entry_survey_data_clean.is_cofounder != 'No') participants_entry_survey_data_clean.loc[cond,'is_cofounder_temp'] = 1 print(len(participants_entry_survey_data_clean.query('is_cofounder_temp == 1'))) # Manually fix cond = (participants_entry_survey_data_clean.member == 'XLIPIHEOIT') participants_entry_survey_data_clean.loc[cond,'is_cofounder_temp'] = 0 print(len(participants_entry_survey_data_clean.query('is_cofounder_temp == 1'))) participants_entry_survey_data_clean['is_cofounder'] = participants_entry_survey_data_clean['is_cofounder_temp'] del participants_entry_survey_data_clean['is_cofounder_temp'] ###Output _____no_output_____ ###Markdown HHH ###Code participants_entry_survey_data_clean.HHH_type.value_counts() def HHH_Translation(answer): """ """ if answer == 'Hacker (you can solve any technical problem and make anything work)': return('Hacker') if answer == 'Hustler (you are the one who closes deals and brings back the money)': return('Hustler') if answer == 'Hipster (you are a creative design genius who makes the user experience awesome)': return('Hispster') return answer participants_entry_survey_data_clean['HHH_type'] = participants_entry_survey_data_clean['HHH_type'].apply(HHH_Translation) participants_entry_survey_data_clean.HHH_type.value_counts() participants_entry_survey_data_clean.head() ###Output _____no_output_____ ###Markdown Gender ###Code def gender_Translation(answer): """ """ if answer == 'Male': return('M') if answer == 'Female': return('F') else: return('U') return answer participants_entry_survey_data_clean['gender'] = participants_entry_survey_data_clean['gender'].apply(gender_Translation) #participants_entry_survey_data_clean.query('gender == "U"') #participants_entry_survey_data_clean.loc['UAXR5EMOI2','gender']='M' participants_entry_survey_data_clean.loc[participants_entry_survey_data_clean.member =='UAXR5EMOI2','gender']='M' participants_entry_survey_data_clean.gender.value_counts() ###Output _____no_output_____ ###Markdown tests ###Code participants_entry_survey_data_clean.dtypes participants_entry_survey_data_clean.HHH_type.value_counts() participants_entry_survey_data_clean.TIPI_extraversion.hist() participants_entry_survey_data_clean.TIPI_agreeableness.hist() participants_entry_survey_data_clean.TIPI_openness.hist() ###Output _____no_output_____ ###Markdown Store ###Code participants_entry_survey_data_clean.set_index('member', inplace=True) with pd.HDFStore(surveys_clean_store_path) as store: store.put('entry/participants_entry_survey_data_clean', participants_entry_survey_data_clean, format='table') ###Output _____no_output_____ ###Markdown Sainty check ###Code members = pd.read_hdf(analysis_store_path, 'metadata/members') members.head() a = participants_entry_survey_data_clean.join(members) a.query('is_ceo == 1')[['company']].sort_values('company') ###Output _____no_output_____
python/d2l-en/pytorch/chapter_recurrent-modern/gru.ipynb
###Markdown Gated Recurrent Units (GRU):label:`sec_gru`In :numref:`sec_bptt`,we discussed how gradients are calculatedin RNNs.In particular we found that long products of matrices can leadto vanishing or exploding gradients.Let us briefly think about what suchgradient anomalies mean in practice:* We might encounter a situation where an early observation is highly significant for predicting all future observations. Consider the somewhat contrived case where the first observation contains a checksum and the goal is to discern whether the checksum is correct at the end of the sequence. In this case, the influence of the first token is vital. We would like to have some mechanisms for storing vital early information in a *memory cell*. Without such a mechanism, we will have to assign a very large gradient to this observation, since it affects all the subsequent observations.* We might encounter situations where some tokens carry no pertinent observation. For instance, when parsing a web page there might be auxiliary HTML code that is irrelevant for the purpose of assessing the sentiment conveyed on the page. We would like to have some mechanism for *skipping* such tokens in the latent state representation.* We might encounter situations where there is a logical break between parts of a sequence. For instance, there might be a transition between chapters in a book, or a transition between a bear and a bull market for securities. In this case it would be nice to have a means of *resetting* our internal state representation.A number of methods have been proposed to address this. One of the earliest is long short-term memory :cite:`Hochreiter.Schmidhuber.1997` which wewill discuss in :numref:`sec_lstm`. The gated recurrent unit (GRU):cite:`Cho.Van-Merrienboer.Bahdanau.ea.2014` is a slightly more streamlinedvariant that often offers comparable performance and is significantly faster tocompute :cite:`Chung.Gulcehre.Cho.ea.2014`.Due to its simplicity, let us start with the GRU. Gated Hidden StateThe key distinction between vanilla RNNs and GRUsis that the latter support gating of the hidden state.This means that we have dedicated mechanisms forwhen a hidden state should be *updated* andalso when it should be *reset*.These mechanisms are learned and they address the concerns listed above.For instance, if the first token is of great importancewe will learn not to update the hidden state after the first observation.Likewise, we will learn to skip irrelevant temporary observations.Last, we will learn to reset the latent state whenever needed.We discuss this in detail below. Reset Gate and Update GateThe first thing we need to introduce arethe *reset gate* and the *update gate*.We engineer them to be vectors with entries in $(0, 1)$such that we can perform convex combinations.For instance,a reset gate would allow us to control how much of the previous state we might still want to remember.Likewise, an update gate would allow us to control how much of the new state is just a copy of the old state.We begin by engineering these gates.:numref:`fig_gru_1` illustrates the inputs for boththe reset and update gates in a GRU, given the inputof the current time stepand the hidden state of the previous time step.The outputs of two gatesare given by two fully-connected layerswith a sigmoid activation function.![Computing the reset gate and the update gate in a GRU model.](../img/gru-1.svg):label:`fig_gru_1`Mathematically,for a given time step $t$,suppose that the input isa minibatch$\mathbf{X}_t \in \mathbb{R}^{n \times d}$ (number of examples: $n$, number of inputs: $d$) and the hidden state of the previous time step is $\mathbf{H}_{t-1} \in \mathbb{R}^{n \times h}$ (number of hidden units: $h$). Then, the reset gate $\mathbf{R}_t \in \mathbb{R}^{n \times h}$ and update gate $\mathbf{Z}_t \in \mathbb{R}^{n \times h}$ are computed as follows:$$\begin{aligned}\mathbf{R}_t = \sigma(\mathbf{X}_t \mathbf{W}_{xr} + \mathbf{H}_{t-1} \mathbf{W}_{hr} + \mathbf{b}_r),\\\mathbf{Z}_t = \sigma(\mathbf{X}_t \mathbf{W}_{xz} + \mathbf{H}_{t-1} \mathbf{W}_{hz} + \mathbf{b}_z),\end{aligned}$$where $\mathbf{W}_{xr}, \mathbf{W}_{xz} \in \mathbb{R}^{d \times h}$ and$\mathbf{W}_{hr}, \mathbf{W}_{hz} \in \mathbb{R}^{h \times h}$ are weightparameters and $\mathbf{b}_r, \mathbf{b}_z \in \mathbb{R}^{1 \times h}$ arebiases.Note that broadcasting (see :numref:`subsec_broadcasting`) is triggered during the summation.We use sigmoid functions (as introduced in :numref:`sec_mlp`) to transform input values to the interval $(0, 1)$. Candidate Hidden StateNext, let usintegrate the reset gate $\mathbf{R}_t$ withthe regular latent state updating mechanismin :eqref:`rnn_h_with_state`.It leads to the following*candidate hidden state*$\tilde{\mathbf{H}}_t \in \mathbb{R}^{n \times h}$ at time step $t$:$$\tilde{\mathbf{H}}_t = \tanh(\mathbf{X}_t \mathbf{W}_{xh} + \left(\mathbf{R}_t \odot \mathbf{H}_{t-1}\right) \mathbf{W}_{hh} + \mathbf{b}_h),$$:eqlabel:`gru_tilde_H`where $\mathbf{W}_{xh} \in \mathbb{R}^{d \times h}$ and $\mathbf{W}_{hh} \in \mathbb{R}^{h \times h}$are weight parameters,$\mathbf{b}_h \in \mathbb{R}^{1 \times h}$is the bias,and the symbol $\odot$ is the Hadamard (elementwise) product operator.Here we use a nonlinearity in the form of tanh to ensure that the values in the candidate hidden state remain in the interval $(-1, 1)$.The result is a *candidate* since we still need to incorporate the action of the update gate.Comparing with :eqref:`rnn_h_with_state`,now the influence of the previous statescan be reduced with theelementwise multiplication of$\mathbf{R}_t$ and $\mathbf{H}_{t-1}$in :eqref:`gru_tilde_H`.Whenever the entries in the reset gate $\mathbf{R}_t$ are close to 1, we recover a vanilla RNN such as in :eqref:`rnn_h_with_state`.For all entries of the reset gate $\mathbf{R}_t$ that are close to 0, the candidate hidden state is the result of an MLP with $\mathbf{X}_t$ as the input. Any pre-existing hidden state is thus *reset* to defaults.:numref:`fig_gru_2` illustrates the computational flow after applying the reset gate.![Computing the candidate hidden state in a GRU model.](../img/gru-2.svg):label:`fig_gru_2` Hidden StateFinally, we need to incorporate the effect of the update gate $\mathbf{Z}_t$. This determines the extent to which the new hidden state $\mathbf{H}_t \in \mathbb{R}^{n \times h}$ is just the old state $\mathbf{H}_{t-1}$ and by how much the new candidate state $\tilde{\mathbf{H}}_t$ is used.The update gate $\mathbf{Z}_t$ can be used for this purpose, simply by taking elementwise convex combinations between both $\mathbf{H}_{t-1}$ and $\tilde{\mathbf{H}}_t$.This leads to the final update equation for the GRU:$$\mathbf{H}_t = \mathbf{Z}_t \odot \mathbf{H}_{t-1} + (1 - \mathbf{Z}_t) \odot \tilde{\mathbf{H}}_t.$$Whenever the update gate $\mathbf{Z}_t$ is close to 1, we simply retain the old state. In this case the information from $\mathbf{X}_t$ is essentially ignored, effectively skipping time step $t$ in the dependency chain. In contrast, whenever $\mathbf{Z}_t$ is close to 0, the new latent state $\mathbf{H}_t$ approaches the candidate latent state $\tilde{\mathbf{H}}_t$. These designs can help us cope with the vanishing gradient problem in RNNs and better capture dependencies for sequences with large time step distances.For instance,if the update gate has been close to 1for all the time steps of an entire subsequence,the old hidden state at the time step of its beginningwill be easily retained and passedto its end,regardless of the length of the subsequence.:numref:`fig_gru_3` illustrates the computational flow after the update gate is in action.![Computing the hidden state in a GRU model.](../img/gru-3.svg):label:`fig_gru_3`In summary, GRUs have the following two distinguishing features:* Reset gates help capture short-term dependencies in sequences.* Update gates help capture long-term dependencies in sequences. Implementation from ScratchTo gain a better understanding of the GRU model, let us implement it from scratch. We begin by reading the time machine dataset that we used in :numref:`sec_rnn_scratch`. The code for reading the dataset is given below. ###Code import torch from torch import nn from d2l import torch as d2l batch_size, num_steps = 32, 35 train_iter, vocab = d2l.load_data_time_machine(batch_size, num_steps) ###Output _____no_output_____ ###Markdown (**Initializing Model Parameters**)The next step is to initialize the model parameters.We draw the weights from a Gaussian distributionwith standard deviation to be 0.01 and set the bias to 0. The hyperparameter `num_hiddens` defines the number of hidden units.We instantiate all weights and biases relating to the update gate, the reset gate, the candidate hidden state,and the output layer. ###Code def get_params(vocab_size, num_hiddens, device): num_inputs = num_outputs = vocab_size def normal(shape): return torch.randn(size=shape, device=device)*0.01 def three(): return (normal((num_inputs, num_hiddens)), normal((num_hiddens, num_hiddens)), torch.zeros(num_hiddens, device=device)) W_xz, W_hz, b_z = three() # Update gate parameters W_xr, W_hr, b_r = three() # Reset gate parameters W_xh, W_hh, b_h = three() # Candidate hidden state parameters # Output layer parameters W_hq = normal((num_hiddens, num_outputs)) b_q = torch.zeros(num_outputs, device=device) # Attach gradients params = [W_xz, W_hz, b_z, W_xr, W_hr, b_r, W_xh, W_hh, b_h, W_hq, b_q] for param in params: param.requires_grad_(True) return params ###Output _____no_output_____ ###Markdown Defining the ModelNow we will define [**the hidden state initialization function**] `init_gru_state`. Just like the `init_rnn_state` function defined in :numref:`sec_rnn_scratch`, this function returns a tensor with a shape (batch size, number of hidden units) whose values are all zeros. ###Code def init_gru_state(batch_size, num_hiddens, device): return (torch.zeros((batch_size, num_hiddens), device=device), ) ###Output _____no_output_____ ###Markdown Now we are ready to [**define the GRU model**].Its structure is the same as that of the basic RNN cell, except that the update equations are more complex. ###Code def gru(inputs, state, params): W_xz, W_hz, b_z, W_xr, W_hr, b_r, W_xh, W_hh, b_h, W_hq, b_q = params H, = state outputs = [] for X in inputs: Z = torch.sigmoid((X @ W_xz) + (H @ W_hz) + b_z) R = torch.sigmoid((X @ W_xr) + (H @ W_hr) + b_r) H_tilda = torch.tanh((X @ W_xh) + ((R * H) @ W_hh) + b_h) H = Z * H + (1 - Z) * H_tilda Y = H @ W_hq + b_q outputs.append(Y) return torch.cat(outputs, dim=0), (H,) ###Output _____no_output_____ ###Markdown Training and Predicting[**Training**] and prediction work in exactly the same manner as in :numref:`sec_rnn_scratch`.After training,we print out the perplexity on the training setand the predicted sequence followingthe provided prefixes "time traveller" and "traveller", respectively. ###Code vocab_size, num_hiddens, device = len(vocab), 256, d2l.try_gpu() num_epochs, lr = 500, 1 model = d2l.RNNModelScratch(len(vocab), num_hiddens, device, get_params, init_gru_state, gru) d2l.train_ch8(model, train_iter, vocab, lr, num_epochs, device) ###Output perplexity 1.1, 23227.8 tokens/sec on cuda:0 time traveller with a slight accession ofcheerfulness really thi ###Markdown [**Concise Implementation**]In high-level APIs,we can directlyinstantiate a GPU model.This encapsulates all the configuration detail that we made explicit above.The code is significantly faster as it uses compiled operators rather than Python for many details that we spelled out before. ###Code num_inputs = vocab_size gru_layer = nn.GRU(num_inputs, num_hiddens) model = d2l.RNNModel(gru_layer, len(vocab)) model = model.to(device) d2l.train_ch8(model, train_iter, vocab, lr, num_epochs, device) ###Output perplexity 1.0, 292351.9 tokens/sec on cuda:0 time travelleryou can show black is white by argument said filby
exercises/E06Solutions_Analog-signals.ipynb
###Markdown E06 Analog signals This weeks homework asks you to perform a fast Fourier transform (FFT) on a given signal. The aim is to find out which frequency components are contained in the noisy but stationary signal. The signal is stored in an numpy array saved in the file `analag-signal1.npy`. The array contains two rows, the first row is the time axis and the second row is the signal. You first need to download the file `analag-signal1.npy` from the moodle or the github repository. Next you are asked to perform the discrete fast Fourier transform (numpy function `numpy.fft.fft`) on the signal and determine which frequencies are contained in the signal. Note that the signal is stationary, in other words, the frequency content does not change over time and you can use the entire signal to compute the FFT.Adapt the 'Fast Fourier transform of a constructed signal' code from the in-class tutorial in order to do implement this exercise.Here are the specific questions. 1. What is the sampling rate of the signal? In other words, at which frequency was the signal acquired? (Note that you first need to load the array from the `analag-signal1.npy` file by using the numpy `np.load()` function). ###Code import numpy as np import matplotlib.pyplot as plt %matplotlib inline data = np.load('analog-signal1b.npy') dt = np.diff(data[0]) print(dt) print('The sampling rate of the signal is : ',1/dt[0],' Hz') ###Output [0.0004 0.0004 0.0004 ... 0.0004 0.0004 0.0004] The sampling rate of the signal is : 2500.0 Hz ###Markdown 2. Perform the fast Fourier transform (FFT, with numpy function `numpy.fft.fft`) on the signal. Determine through plotting the FFT of the signal (power on y-axis over frequency on x-axis) which frequencies are contained in the signal (determine the frequencies through visual inspection of the plot). Note that you need the sampling rate obtained above to get the correct scaling of the frequency axis. ###Code # performing FFT on signal ###################### fs = 1./dt[0] nyquist = fs/2. fSpaceSignal = np.fft.fft(data[1]) fBase = np.linspace(0,nyquist,np.floor(len(data[1])/2)+1) halfTheSignal = fSpaceSignal[:len(fBase)] complexConjugate = np.conj(halfTheSignal) powe = halfTheSignal*complexConjugate # plotting results ############################## fig = plt.figure(figsize=(10,10)) ax0 = fig.add_subplot(2,1,1) ax0.plot(data[0],data[1]) ax2 = fig.add_subplot(2,1,2) ax2.plot(fBase,powe/max(powe)) ax2.set_xlim([0,15]) plt.show() ###Output /home/mgraupe/.virtualenvs/locorungs/lib/python3.6/site-packages/ipykernel_launcher.py:5: DeprecationWarning: object of type <class 'numpy.float64'> cannot be safely interpreted as an integer. """
GEE/jrc_gsw/NormalizedDifference.ipynb
###Markdown View source on GitHub Notebook Viewer Run in Google Colab Install Earth Engine API and geemapInstall the [Earth Engine Python API](https://developers.google.com/earth-engine/python_install) and [geemap](https://geemap.org). The **geemap** Python package is built upon the [ipyleaflet](https://github.com/jupyter-widgets/ipyleaflet) and [folium](https://github.com/python-visualization/folium) packages and implements several methods for interacting with Earth Engine data layers, such as `Map.addLayer()`, `Map.setCenter()`, and `Map.centerObject()`.The following script checks if the geemap package has been installed. If not, it will install geemap, which automatically installs its [dependencies](https://github.com/giswqs/geemapdependencies), including earthengine-api, folium, and ipyleaflet. ###Code # Installs geemap package import subprocess try: import geemap except ImportError: print("Installing geemap ...") subprocess.check_call(["python", "-m", "pip", "install", "geemap"]) import ee import geemap ###Output _____no_output_____ ###Markdown Create an interactive map The default basemap is `Google Maps`. [Additional basemaps](https://github.com/giswqs/geemap/blob/master/geemap/basemaps.py) can be added using the `Map.add_basemap()` function. ###Code Map = geemap.Map(center=[40, -100], zoom=4) Map ###Output _____no_output_____ ###Markdown Add Earth Engine Python script ###Code # Add Earth Engine dataset # NormalizedDifference example. # # Compute Normalized Difference Vegetation Index over MOD09GA product. # NDVI = (NIR - RED) / (NIR + RED), where # RED is sur_refl_b01, 620-670nm # NIR is sur_refl_b02, 841-876nm # Load a MODIS image. img = ee.Image('MODIS/006/MOD09GA/2012_03_09') # Use the normalizedDifference(A, B) to compute (A - B) / (A + B) ndvi = img.normalizedDifference(['sur_refl_b02', 'sur_refl_b01']) # Make a palette: a list of hex strings. palette = ['FFFFFF', 'CE7E45', 'DF923D', 'F1B555', 'FCD163', '99B718', '74A901', '66A000', '529400', '3E8601', '207401', '056201', '004C00', '023B01', '012E01', '011D01', '011301'] # Center the map Map.setCenter(-94.84497, 39.01918, 8) # Display the input image and the NDVI derived from it. Map.addLayer(img.select(['sur_refl_b01', 'sur_refl_b04', 'sur_refl_b03']), {'gain': [0.1, 0.1, 0.1]}, 'MODIS bands 1/4/3') Map.addLayer(ndvi, {'min': 0, 'max': 1, 'palette': palette}, 'NDVI') ###Output _____no_output_____ ###Markdown Display Earth Engine data layers ###Code Map.addLayerControl() # This line is not needed for ipyleaflet-based Map. Map ###Output _____no_output_____
Exercise 2 - CNN.ipynb
###Markdown Brief IntroWith no doubt, Convolutional Neural Networks (CNN) has been the most successful model for computer vision tasks. Convolutional operation has similar mechanism to the way that human eyes work in visual perception. When humans explore the visual world, the eyes behave in a pattern of alternative fixations and saccades. The saccadic eye movements bring the visual target to the fovea abruptly (about 20ms), and the target information is then processed during eye fixations when the eyes stay relatively stable (e.g 200ms). We are usually un-aware of the eye movements, as they are programed and executed automatically by cognitive brain process. Our brain then aggregate all these local information to a global decision, based on previous knowledge/experience. The visual field is not explored as a whole. Only a selective set of local positions are viewed, and that turns out to be enough to serve the perception needs in our daily lives (It means images are extremely redundant to serve the recognition/classification popurse. Duplicated and irrelavant information should be effectively discarded to gain efficiency, e.g through weighting and local operator (local operators also can be considered as weighting by penalizing weights of the positions outside the receptive field to 0). Images are too rich and also too costy.).From this perspective, CNN is very much bio-inspired methodology: local-to-global, like divide-and-conquer (e.g to sort a list, you can sort the sublists (local) then merge to have the global solution). It acts like information selector and aggragator, grab the needed and throw away the rest. OK, too much talking, stop brain storming and code it. Let code say ###Code ## load libs %matplotlib inline import time import warnings warnings.filterwarnings('ignore') import numpy as np import matplotlib.pyplot as plt from sklearn.datasets import fetch_mldata from sklearn.metrics import confusion_matrix, classification_report, accuracy_score from sklearn.preprocessing import OneHotEncoder ###Output _____no_output_____ ###Markdown Load MNIST ###Code mnist = fetch_mldata('mnist original', data_home = 'datasets/') X, y = mnist['data'], mnist['target'] X.shape, y.shape ## shape check ###Output C:\ProgramData\Anaconda3\lib\site-packages\sklearn\utils\deprecation.py:77: DeprecationWarning: Function fetch_mldata is deprecated; fetch_mldata was deprecated in version 0.20 and will be removed in version 0.22 warnings.warn(msg, category=DeprecationWarning) C:\ProgramData\Anaconda3\lib\site-packages\sklearn\utils\deprecation.py:77: DeprecationWarning: Function mldata_filename is deprecated; mldata_filename was deprecated in version 0.20 and will be removed in version 0.22 warnings.warn(msg, category=DeprecationWarning) ###Markdown Preprocess MNIST ###Code X = X.T X = X / 255.0 Y = OneHotEncoder().fit_transform(y.reshape(-1,1).astype('int32')).toarray().T X.shape, Y.shape ###Output _____no_output_____ ###Markdown Make Train/Test Splits ###Code m = 60000 X_train, X_test = X[:,:m].reshape(1,28,28,-1), X[:,m:].reshape(1,28,28,-1) Y_train, Y_test = Y[:,:m], Y[:,m:] X_train.shape, Y_train.shape, X_test.shape, Y_test.shape ###Output _____no_output_____ ###Markdown Shuffle Train set ###Code np.random.seed(54321) shuffle = np.random.permutation(m) X_train, Y_train = X_train[:,:,:,shuffle], Y_train[:,shuffle] X_train.shape, Y_train.shape ###Output _____no_output_____ ###Markdown Visual check ###Code idx = 134 plt.imshow(X_train[:,:,:,idx].squeeze(), cmap = 'binary_r') plt.title(np.argmax(Y_train[:,idx])) plt.axis('off') plt.show() ###Output _____no_output_____ ###Markdown Define network ###Code ## input layer input_depth = 1 input_height = 28 input_width = 28 ## convolution layer conv_depth = 2 conv_height = 3 conv_width = 3 ## trainable parameters connecting input & convolution layers W1 = np.random.randn(conv_depth, input_depth, conv_height, conv_width) b1 = np.zeros((conv_depth, 1)) ## densely connected (fc) layer fc_dims = 32 flatten_dims = conv_depth * (input_height - conv_height + 1) * (input_width - conv_width + 1) ## trainable parameters connecting convolution & dense layers W2 = np.random.randn(fc_dims, flatten_dims) b2 = np.zeros((fc_dims, 1)) ## output layer output_dims = 10 ## trainable parameters connecting dense & output layers W3 = np.random.randn(output_dims, fc_dims) b3 = np.zeros((output_dims, 1)) ###Output _____no_output_____ ###Markdown Training CNN ###Code ## prepare inputs Input = X_train.copy() Target = Y_train.copy() Input.shape, Target.shape ## initialize convolution output conv_output_height = input_height - conv_height + 1 conv_output_width = input_width - conv_width + 1 conv_output = np.zeros((conv_depth, conv_output_height, conv_output_width, Input.shape[-1])) for epoch in range(20): #------------------------------------------------------------------FORWARD BLOCK ## feed forward: convolution operation for f in range(conv_depth): for r in range(conv_output_height): for c in range(conv_output_width): current_patch = Input[:, r : r + conv_height, c : c + conv_width] current_filter = np.expand_dims(W1[f,:,:,:], axis = 3) ## to match shape for broadcasting conv_output[f, r, c] = (current_patch * current_filter + b1[f]).reshape(-1, Input.shape[-1]).sum(axis = 0) ## reshape 2X faster # conv_output[f, r, c] += (current_patch * current_filter + b1[f]).sum(axis = 0).sum(axis = 0).sum(axis = 0) ## feed forward: flatten the convolution output conv_output_flatten = conv_output.reshape(-1, Input.shape[-1]) A1 = 1 / (1 + np.exp(-conv_output_flatten)) ## sigmoid ## feed forward: affine operation Z2 = W2 @ A1 + b2 A2 = 1/(1 + np.exp(-Z2)) ## geed forward: affine + softmax operation Z3 = W3 @ A2 + b3 Z3 = Z3 - np.max(Z3, axis = 0) A3 = np.exp(Z3)/np.exp(Z3).sum(axis = 0) #------------------------------------------------------------------BACKWARD BLOCK ## backpropagation: softmax layer dZ3 = A3 - Y_train dW3 = dZ3 @ A2.T / Input.shape[-1] db3 = dZ3.mean(axis = 1, keepdims = True) ## backpropagation: dense layer dA2 = W3.T @ dZ3 dZ2 = dA2 * A2 * (1 - A2) dW2 = dZ2 @ A1.T / Input.shape[-1] db2 = dZ2.mean(axis = 1, keepdims = True) ## backpropagation: convolution layer dA1 = W2.T @ dZ2 d_conv_flatten = dA1 * A1 * (1 - A1) d_conv_matrix = d_conv_flatten.reshape(conv_output.shape) ## backpropagation: convolution layer --> weight dW1 = np.zeros(W1.shape) for in_c in range(Input.shape[0]): for out_c in range(conv_output.shape[0]): for r in range(conv_height): for c in range(conv_width): conv_input_patch = Input[in_c, r : r + conv_output_height, c : c + conv_output_width, :] ## conv input conv_output_vals = d_conv_matrix[out_c] ## conv results dW1[out_c, in_c, r, c] = np.sum(conv_input_patch * conv_output_vals)/Input.shape[-1] ## backpropagation: convolution layer --> bias db1 = d_conv_matrix.sum(axis = 1).sum(axis = 1).mean(axis = 1, keepdims = True) # equivalent # db1 = np.zeros((b1.shape)) # for out_c in range(d_conv_matrix.shape[0]): # db1[out_c] += d_conv_matrix[out_c].sum()/Input.shape[-1] ## backpropagation: convolution layer --> Input dInput = np.zeros_like(Input) for in_c in range(Input.shape[0]): for out_c in range(conv_output.shape[0]): current_filter = np.expand_dims(W1[out_c, in_c], axis = 2) for r in range(conv_output_height): for c in range(conv_output_width): d_conv_val = d_conv_matrix[out_c, in_c, r, c] dInput[in_c, r : r+conv_height, c : c + conv_width, :] += d_conv_val * current_filter #------------------------------------------------------------------ UPDATE PARAMETERS ## update model lr = 1 W3 -= dW3 * lr W2 -= dW2 * lr W1 -= dW1 * lr b3 -= db3 * lr b2 -= db2 * lr b1 -= db1 * lr ## compute loss Loss = -np.mean(Y_train * np.log(A3), axis = 1) print('epoch:', epoch, ', loss:', Loss.sum()) ###Output epoch: 0 , loss: [1.28140686 0.20822068 0.2338827 0.46949416 0.60362163 0.17409459 0.35062963 1.02211311 1.32937071 0.04804767] epoch: 1 , loss: [0.38369461 0.06354422 0.77907208 0.20330568 0.38941393 0.18225178 0.46335044 0.22145518 0.68721877 1.18083523] epoch: 2 , loss: [0.34118477 0.98553669 0.58973446 0.09616901 0.11638067 0.16605581 0.25018618 0.26129972 0.40057913 0.76627435] epoch: 3 , loss: [0.13881807 0.66915701 0.38335981 0.55071828 0.3880977 0.23016209 0.24817241 0.08562567 0.22102089 0.52127314] epoch: 4 , loss: [0.55372952 0.38240065 0.14696803 0.54680683 0.17156462 0.17519124 0.13035284 0.70305301 0.20451274 0.36796369] epoch: 5 , loss: [0.34151146 0.16762261 0.26628403 0.32332559 0.20978761 0.18224469 0.3076453 0.46574045 0.17083644 0.18504516] epoch: 6 , loss: [0.21586048 0.32785749 0.19406651 0.20420862 0.21118384 0.21637319 0.20095451 0.29251818 0.2560406 0.24154615] epoch: 7 , loss: [0.23242883 0.22555531 0.25360815 0.25162868 0.23244112 0.21426809 0.24437044 0.2203629 0.21580981 0.22066201] epoch: 8 , loss: [0.22568329 0.2612961 0.22039339 0.22442451 0.22329748 0.2172761 0.22149135 0.24540321 0.23199796 0.23268553] epoch: 9 , loss: [0.23015672 0.23679541 0.23447607 0.23837881 0.22862081 0.21729422 0.23231189 0.23134363 0.22498903 0.22764145] epoch: 10 , loss: [0.22758303 0.25200254 0.22677901 0.23018803 0.22573063 0.21695097 0.22648256 0.23839846 0.22770123 0.22964755] epoch: 11 , loss: [0.22917324 0.24169142 0.23076751 0.23478301 0.22740242 0.21740039 0.22956132 0.23468208 0.22678078 0.22901753] epoch: 12 , loss: [0.22819881 0.24834426 0.22857501 0.23208175 0.22642583 0.21705683 0.22785971 0.23653148 0.2270004 0.2291191 ] epoch: 13 , loss: [0.2288109 0.24388698 0.22978328 0.23365703 0.22701621 0.21730915 0.22881446 0.2356316 0.22704205 0.2292138 ] epoch: 14 , loss: [0.22842485 0.24680646 0.22909761 0.23271827 0.22665451 0.21713782 0.22826391 0.23604126 0.22693296 0.22907742] epoch: 15 , loss: [0.22867101 0.24486233 0.22949335 0.23328011 0.22688052 0.21725326 0.2285881 0.23587407 0.22704354 0.22920425] epoch: 16 , loss: [0.22851321 0.24614424 0.22926003 0.23293913 0.2267379 0.21717704 0.22839324 0.23592658 0.22695367 0.22910357] epoch: 17 , loss: [0.22861485 0.24529276 0.22940014 0.23314757 0.22682893 0.21722734 0.22851247 0.23592458 0.2270208 0.22917828] epoch: 18 , loss: [0.22854897 0.24585603 0.22931462 0.23301865 0.22677045 0.21719432 0.22843837 0.23590762 0.22697309 0.22912515] epoch: 19 , loss: [0.22859166 0.24548223 0.22936765 0.23309886 0.22680828 0.21721603 0.22848499 0.23592912 0.22700614 0.22916205] ###Markdown Test ###Code ## initialize convolution output Input = X_test.copy() conv_output_height = input_height - conv_height + 1 conv_output_width = input_width - conv_width + 1 conv_output = np.zeros((conv_depth, conv_output_height, conv_output_width, Input.shape[-1])) ## feed forward: convolution operation for f in range(conv_depth): for r in range(conv_output_height): for c in range(conv_output_width): current_patch = Input[:, r : r + conv_height, c : c + conv_width] current_filter = np.expand_dims(W1[f,:,:,:], axis = 3) ## to match shape for broadcasting conv_output[f, r, c] = (current_patch * current_filter + b1[f]).reshape(-1, Input.shape[-1]).sum(axis = 0) ## reshape 2X faster # conv_output[f, r, c] += (current_patch * current_filter + b1[f]).sum(axis = 0).sum(axis = 0).sum(axis = 0) ## feed forward: flatten the convolution output conv_output_flatten = conv_output.reshape(-1, Input.shape[-1]) A1 = 1 / (1 + np.exp(-conv_output_flatten)) ## sigmoid ## feed forward: affine operation Z2 = W2 @ A1 + b2 A2 = 1/(1 + np.exp(-Z2)) ## geed forward: affine + softmax operation Z3 = W3 @ A2 + b3 Z3 = Z3 - np.max(Z3, axis = 0) A3 = np.exp(Z3)/np.exp(Z3).sum(axis = 0) preds = np.argmax(A3, axis = 0) truth = np.argmax(Y_test, axis = 0) ###Output _____no_output_____ ###Markdown Results Report ###Code print(accuracy_score(truth, preds)) print(confusion_matrix(truth, preds)) print(classification_report(truth, preds)) ## something wrong inside, to be corrected ###Output _____no_output_____
Session_13_logistic_regression.ipynb
###Markdown Logistic regressionI think it most simple to first demonstrate what type of problem that we are interested in. So here lets compare displacement (disp) for US made and non-US made cars.Let's first assign mtcars to a new dataframe and then identify the US (1) vs. non_US (0) under the column label "origin." ###Code cars <- mtcars non_US <- c(1:3, 8:14, 18:21, 26:28, 30:32) origin <- rep(1, nrow(cars)) origin[non_US] <- 0 cars$origin <- origin head(cars) ###Output _____no_output_____ ###Markdown Now that we have an origin column in our dataframe, let's plot it compared to engine displacement (disp): ###Code par(pin=c(3,3)) plot(cars$disp, cars$origin) ###Output _____no_output_____ ###Markdown You can see from the figure that there is an obvious bias to the left for non_US (origin = 0) cars versus US (origin = 1). So clearly, there is a relationship between origin and displacement---US cars have larger engine displacements.The question that we may then ask is, "Given the engine displacement of a car, is that car likely to be made in the US or not?" Logistic regression is one approach to dealing with this classification problem. Definition: Binary logistic regressionFor a binary categorical (this would be the dependant variable) versus a continuous (independent) variable we can describe this simple system with the following equations: $$\begin{align}t =& \beta_0 + \beta_1 x \\Pr(y|x) = p(x) =& \frac{1}{1 + e^{-t}}\end{align}$$The function t is called the linear predictor function which includes an intercept and coefficient for the independent variable x. The probability of a positive classification y given x is Pr(y|x). The logit function is the natural logarithm of the odds ratio of the outcome:$$\begin{align}logit (p) =& \ln \left(\frac{p(x)}{1-p(x)}\right) = \beta_0 + \beta_1 x \end{align}$$To determine if the value of a given variable should be classified as being of y or not y, the following holds:$$y =\left\{ \begin{array}{ll} 1 \mbox{ if: } \beta_0 + \beta_1 x + \epsilon > 0\\ 0 \mbox{ else:} \end{array} \right. $$Unlike linear regression, we cannot fit this equation to our data with a simple least-squares fit. Logistic regression requires the use of a different approach called maximum likelihood estimation (MLE). The computational overhead of finding the best fitting solution means that this approach only became viable with the ubiquity of computers.Let's plot this function and play with the parameters. ###Code #Definition of the logistic function x <- seq(-20,20, length=100) par(mfrow=c(2,2)) beta0 <- 0 beta1 <- 1 pr <- 1/(1+exp(- (beta1*x + beta0) )) plot(x,pr, xlim=c(-20,20), type="l", xlab='Cont. var.', main='beta(0,1)') beta0 <- -5 beta1 <- 1 t <- beta0 + beta1*x pr <- 1/(1+exp(- (t) )) plot(x,pr, xlim=c(-20,20), type="l", xlab='Cont. var.', main='beta(-5,1)') beta0 <- 0 beta1 <- 0.2 t <- beta0 + beta1*x pr <- 1/(1+exp(- (t) )) plot(x,pr, xlim=c(-20,20), type="l", xlab='Cont. var.', main='beta(0,0.2)') beta0 <- -5 beta1 <- 0.2 t <- beta0 + beta1*x pr <- 1/(1+exp(- (t) )) plot(x,pr, xlim=c(-20,20), type="l", xlab='Cont. var.', main='beta(-5,0.2)') ###Output _____no_output_____ ###Markdown Note that since the linear predictor function (t) is inside an exponential term, the offset resulting from the \beta_0 term is scaled non-linearly with the other terms.Going back to our data for car origin versus engine displacement, we can use the general linear model to calculate the \beta parameters: ###Code model <- glm(origin ~ disp, family=binomial(link='logit'), data=cars) summary(model) ###Output _____no_output_____ ###Markdown From the model coefficients, we can plot the probability density function against the real values: ###Code x <- seq(50,500, length=100) beta0 <- -8.73 beta1 <- 0.0315 t <- beta0 + beta1*x pr <- 1/(1+exp(- (t) )) par(pin=c(3,3)) plot(x,pr, type="l", xlab="disp") points(cars$disp,cars$origin,col="red") ###Output _____no_output_____ ###Markdown Odds ratioIn order to interpret the coefficients given by the summary of the general linear model, we need to extract them from the logit equation. Here we get an expression that tells us "for every one unit increase in x, the odds multiply by e^{\beta_1}":$$\begin{align}odds =& e^{\beta_1}\end{align}$$ ###Code or = exp(model$coefficient[2]) #We can extract the \beta_1 from the model directly print(or) ###Output disp 1.032036 ###Markdown We can then interpret this as, "for every increase in engine displacement by 1 cu-in, the likely odds that the car has a US origin increases by 3%." Multiple logistic regressionThere is no restriction on the number of terms present in the linear predictor function that we used in establishing the model. In the general case we may write the function as:$$\begin{align}t =& \beta_0 + \beta_1 x_1 + \beta_2 x_2 + \ldots \\=& \sum_{i=1}^m \beta_i x_i\end{align}$$Note:- It is generally recommended that all predictors used in a model have at least ten corresponding events, although this is a theoretical argument. https://bmcmedresmethodol.biomedcentral.com/articles/10.1186/s12874-016-0267-3 Classification based on a predictorThe logistic regression model may then be used as a binary classifier for a general set of predictor variables. So given a set of data to "train" on, we can calculate the coefficients and use the predictor function to determine if a case should be classified as yes/no.We can do this for the existing data by spliting it into a training and query set and calculating the model using only the training set. We can then use the predictor function to decide which way a case from the query set should be classified. ###Code #Take a random sample of the data set.seed(42) #This fixes the random number generator so that you get the same "random" sample every time trainingIndex <- sample(1:nrow(cars), 0.8*nrow(cars)) trainingData <- cars[trainingIndex,] queryData <- cars[-trainingIndex,] print(queryData) logmod <- glm(origin ~ disp, family=binomial(link='logit'), data=trainingData) #Create the model print(logmod) #\beta_0 + \beta_1 x dispP <- predict(logmod, queryData) print(dispP) ###Output mpg cyl disp hp drat wt qsec vs am gear carb origin Datsun 710 22.8 4 108.0 93 3.85 2.320 18.61 1 1 4 1 0 Valiant 18.1 6 225.0 105 2.76 3.460 20.22 1 0 3 1 1 Merc 450SE 16.4 8 275.8 180 3.07 4.070 17.40 0 0 3 3 0 Merc 450SL 17.3 8 275.8 180 3.07 3.730 17.60 0 0 3 3 0 Merc 450SLC 15.2 8 275.8 180 3.07 3.780 18.00 0 0 3 3 0 Fiat X1-9 27.3 4 79.0 66 4.08 1.935 18.90 1 1 4 1 0 Ford Pantera L 15.8 8 351.0 264 4.22 3.170 14.50 0 1 5 4 1 Call: glm(formula = origin ~ disp, family = binomial(link = "logit"), data = trainingData) Coefficients: (Intercept) disp -10.43294 0.03948 Degrees of Freedom: 24 Total (i.e. Null); 23 Residual Null Deviance: 33.65 Residual Deviance: 5.976 AIC: 9.976 Datsun 710 Valiant Merc 450SE Merc 450SL Merc 450SLC -6.169078 -1.549898 0.455695 0.455695 0.455695 Fiat X1-9 Ford Pantera L -7.314003 3.424604 ###Markdown From the model of the training data, we can use the \beta coefficients in the linear predictor function (t) to determine whether a classification of US or non-US is more likely. Recall that if the t value is greater than zero, then the probability of positive classification is greater than 0.5. Reciprocally, if the t value is less than zero, we can negatively classify the input.Finally, we can compare the results of origin with predicted origin in a table. ###Code dispP[dispP>0] <- 1 dispP[dispP<0] <- 0 queryData['predicted'] <- dispP print(queryData) xtabs( ~ predicted + origin, queryData) #Contingency table or confusion matrix ###Output mpg cyl disp hp drat wt qsec vs am gear carb origin Datsun 710 22.8 4 108.0 93 3.85 2.320 18.61 1 1 4 1 0 Valiant 18.1 6 225.0 105 2.76 3.460 20.22 1 0 3 1 1 Merc 450SE 16.4 8 275.8 180 3.07 4.070 17.40 0 0 3 3 0 Merc 450SL 17.3 8 275.8 180 3.07 3.730 17.60 0 0 3 3 0 Merc 450SLC 15.2 8 275.8 180 3.07 3.780 18.00 0 0 3 3 0 Fiat X1-9 27.3 4 79.0 66 4.08 1.935 18.90 1 1 4 1 0 Ford Pantera L 15.8 8 351.0 264 4.22 3.170 14.50 0 1 5 4 1 predicted Datsun 710 0 Valiant 0 Merc 450SE 1 Merc 450SL 1 Merc 450SLC 1 Fiat X1-9 0 Ford Pantera L 1 ###Markdown Contingency tablesWe can look at this table in terms of a "contingency table" or in the general classification case "confusion matrix." Here the "truth" of car origin is represented by the columns, where zero is non-US and unity is US. The "predicted" classification is presented by row.This resembles the false positive (type I) false negative (type II) table for determining the validity of a statistical hypothesis.For a perfect model the diagonal would constitute the sum of all cases and the off-diagonal would sum to zero. In our case the classifier is fairly poor: Of the two US cars in the sample, one was classified as US (true positive) and one as non-US (false negative). Of the five non-US cars, two were classified as non-US (true negative) and three as US (false positive). Precision & recallPrecision is the ratio of true positives to true positives plus true negatives. Recall is the true positives to true positives plus false positives.$$\begin{align}P = \frac{tp}{tp + fp} \\R = \frac{tp}{tp + fn}\end{align}$$In the case of predicting origin based on engine displacement alone:$$\begin{align}P = \frac{1}{1 + 3}= 1/3 \\R = \frac{1}{1 + 1} = 1/2\end{align}$$ F_1 scoreThe F_1 score is a measure of a test's accuracy that considers both precision and recall.$$\begin{align}F_1 =& 2 \frac{P R}{P+R}\end{align}$$Calculated for the prior case:$$\begin{align}F_1 =& 2 \frac{1/6}{5/6} \\=& 1/5\end{align}$$ Expanding the modelLet's train a new model that includes the cars weight to determine if our contingency table scores improve. ###Code logmod <- glm(origin ~ disp + wt, family=binomial(link='logit'), data=trainingData) #Create the model print(logmod) #\beta_0 + \beta_1 x dispP <- predict(logmod, queryData) dispP[dispP>0] <- 1 dispP[dispP<0] <- 0 queryData['predicted'] <- dispP print(queryData) xtabs( ~ predicted + origin, queryData) ###Output Call: glm(formula = origin ~ disp + wt, family = binomial(link = "logit"), data = trainingData) Coefficients: (Intercept) disp wt -3.75075 0.06992 -4.49680 Degrees of Freedom: 24 Total (i.e. Null); 22 Residual Null Deviance: 33.65 Residual Deviance: 5.259 AIC: 11.26 mpg cyl disp hp drat wt qsec vs am gear carb origin Datsun 710 22.8 4 108.0 93 3.85 2.320 18.61 1 1 4 1 0 Valiant 18.1 6 225.0 105 2.76 3.460 20.22 1 0 3 1 1 Merc 450SE 16.4 8 275.8 180 3.07 4.070 17.40 0 0 3 3 0 Merc 450SL 17.3 8 275.8 180 3.07 3.730 17.60 0 0 3 3 0 Merc 450SLC 15.2 8 275.8 180 3.07 3.780 18.00 0 0 3 3 0 Fiat X1-9 27.3 4 79.0 66 4.08 1.935 18.90 1 1 4 1 0 Ford Pantera L 15.8 8 351.0 264 4.22 3.170 14.50 0 1 5 4 1 predicted Datsun 710 0 Valiant 0 Merc 450SE 0 Merc 450SL 0 Merc 450SLC 0 Fiat X1-9 0 Ford Pantera L 1
tikz_jupyter_starter_demo.ipynb
###Markdown Tikz Diagrams in Jupyter Setup Instructions Install Dependencies ```!apt-get -qq install -y texlive-xetex!apt-get -qq install -y imagemagick``` Install Python Interface Libraries```!pip install git+git://github.com/mkrphys/ipython-tikzmagic.git``` Supress Code OutputPut the magic ```%%capture``` at the start of the code cell to discard the output printouts. Import Modulestandard way```import tikz```or via Python Magic```%load_ext tikzmagic``` Demo Install and import ###Code %%capture # dependency installs !apt-get -qq install -y texlive-xetex !apt-get -qq install -y imagemagick # package install !pip install git+git://github.com/mkrphys/ipython-tikzmagic.git # module import %load_ext tikzmagic ###Output _____no_output_____ ###Markdown Draw Something ###Code %%tikz \draw[thick] (0cm,0cm) circle(1cm); \draw[thick] (-0.35,0.35) ellipse (0.1cm and 0.1cm); %draw left eye \draw[thick] (0.35,0.35) ellipse (0.1cm and 0.1cm); %draw right eye \draw[thick] plot [smooth,tension=1.5] coordinates{(-0.5,-0.5) (0,-0.8) (0.5,-0.5)};%draw smile ###Output _____no_output_____
Data Science/Python/.ipynb_checkpoints/03_Practice+Exercise+1-checkpoint.ipynb
###Markdown Practice Exercise 1 You are provided with 2 lists that contain the data of an ecommerce website. The first list contains the data for the number of items sold for a particular product and the second list contains the price of the product sold. As a part of this exercise, solve the questions that are provided below. ###Code number = [8, 9, 9, 1, 6, 9, 5, 7, 3, 9, 7, 3, 4, 8, 3, 5, 8, 4, 8, 7, 5, 7, 3, 6, 1, 2, 7, 4, 7, 7, 8, 4, 3, 4, 2, 2, 2, 7, 3, 5, 6, 1, 1, 3, 2, 1, 1, 7, 7, 1, 4, 4, 5, 6, 1, 2, 7, 4, 5, 8, 1, 4, 8, 6, 2, 4, 3, 7, 3, 6, 2, 3, 3, 3, 2, 4, 6, 8, 9, 3, 9, 3, 1, 8, 6, 6, 3, 3, 9, 4, 6, 4, 9, 6, 7, 1, 2, 8, 7, 8, 1, 4] price = [195, 225, 150, 150, 90, 60, 75, 255, 270, 225, 135, 195, 30, 15, 210, 105, 15, 30, 180, 60, 165, 60, 45, 225, 180, 90, 30, 210, 150, 15, 270, 60, 210, 180, 60, 225, 150, 150, 120, 195, 75, 240, 60, 45, 30, 180, 240, 285, 135, 165, 180, 240, 60, 105, 165, 240, 120, 45, 120, 165, 285, 225, 90, 105, 225, 45, 45, 45, 75, 180, 90, 240, 30, 30, 60, 135, 180, 15, 255, 180, 270, 135, 105, 135, 210, 180, 135, 195, 225, 75, 225, 15, 240, 60, 15, 180, 255, 90, 15, 150, 230, 150] ###Output _____no_output_____ ###Markdown How many different products are sold by the company in total?- 99- 100- 101- 102 ###Code # Type your code here ###Output _____no_output_____ ###Markdown How many items were sold in total?- 460- 490- 500- 520 ###Code # Type your code here ###Output _____no_output_____ ###Markdown What is the average price of the products sold by the ecommerce company?- 139- 151- 142- 128 ###Code # Type your code here ###Output _____no_output_____ ###Markdown What is the price of the costliest item sold?- 225- 310- 280- 285 ###Code # Type your code here ###Output _____no_output_____ ###Markdown What is the total revenue of the company? [Revenue = Price\*Quantity]- 67100- 53900- 45300- 71200 ###Code # Type your code here ###Output _____no_output_____ ###Markdown Demand for the 20th product in the list is more than the 50th product. [True/False]- True- False- Can't be calculated ###Code # Type your code here ###Output _____no_output_____ ###Markdown How many products fall under the category of expensive goods? An expensive good is that good whose price is more than the average price of the products sold by the company.- 48- 50- 52- 54 ###Code # Type your code here ###Output _____no_output_____
Code/Data Preprocessing/Data Resizing.ipynb
###Markdown Importing Libraries ###Code # importing libraries import os import sys import os.path import cv2 from threading import Thread ###Output _____no_output_____ ###Markdown Global Variables for Setting I/O Directories ###Code #############################################Global Variables################################################## General_Directory = 'D:\\Food Datasets\\' #location of full resolution Datasets General_Out_Directory_144p = 'D:\\Food Datasets\\144p_16-9\\' #16:9 images are downscaled to 144p (144x256) General_Out_Directory_224p = 'D:\\Food Datasets\\224p_1-1_4-3\\' #only 1:1 and 4:3 images are downscaled to 224p (224x224) General_Out_Directory_224x224 = 'D:\\Food Datasets\\224p_ALL\\' #all images are downscaled to 224p (224x224) General_Out_Directory_224x224x224 = 'D:\\Food Datasets\\224p_Mixed\\' #Mixed images are downscaled to 224p (224x224) General_Out_Directory_224x224x224x224 = 'D:\\Food Datasets\\224p_4-3_16-9\\' #only 16:9 and 4:3 images are downscaled to 224p (224x224) General_Out_Directory_240p = 'D:\\Food Datasets\\240p_4-3_16-9\\' #16:9 and 4:3 images are dowscaled to 240p (240x360) General_Out_Directory_360p = 'D:\\Food Datasets\\360p_ALL\\' #all images are dowscaled to 360p (360x480) General_Out_Directory_480p = 'D:\\Food Datasets\\480p_4-3\\' #4:3 images are dowscaled to 360p (360x480) General_Out_Directory_640p = 'D:\\Food Datasets\\640p_4-3_16-9\\' #16:9 and 4:3 images are downscaled to 360p (360x640) General_Out_Directory_640x480 = 'D:\\Food Datasets\\640x480_Mixed\\' #Mixed images are downscaled to 480p(480x640) ############################################################################################################## Very_Tiny_General_Directory = General_Directory + 'Full Training Dataset\\' Tiny_General_Directory = General_Directory + 'Large Training Dataset\\' Small_General_Directory = General_Directory + 'Small Training Dataset\\' Balanced_General_Directory = General_Directory + 'Balanced Training Dataset\\' Large_General_Directory = General_Directory + 'Large Training Dataset\\' Full_General_Directory = General_Directory + 'Full Training Dataset\\' Very_Tiny_Out_Directory_144p = General_Out_Directory_144p + 'Very Tiny Training Dataset\\' Tiny_Out_Directory_144p = General_Out_Directory_144p + 'Tiny Training Dataset\\' Small_Out_Directory_144p = General_Out_Directory_144p + 'Small Training Dataset\\' Balanced_Out_Directory_144p = General_Out_Directory_144p + 'Balanced Training Dataset\\' Large_Out_Directory_144p = General_Out_Directory_144p + 'Large Training Dataset\\' Full_Out_Directory_144p = General_Out_Directory_144p + 'Full Training Dataset\\' Very_Tiny_Out_Directory_224p = General_Out_Directory_224p + 'Very Tiny Training Dataset\\' Tiny_Out_Directory_224p = General_Out_Directory_224p + 'Tiny Training Dataset\\' Small_Out_Directory_224p = General_Out_Directory_224p + 'Small Training Dataset\\' Balanced_Out_Directory_224p = General_Out_Directory_224p + 'Balanced Training Dataset\\' Large_Out_Directory_224p = General_Out_Directory_224p + 'Large Training Dataset\\' Full_Out_Directory_224p = General_Out_Directory_224p + 'Full Training Dataset\\' Very_Tiny_Out_Directory_224x224 = General_Out_Directory_224x224 + 'Very Tiny Training Dataset\\' Tiny_Out_Directory_224x224 = General_Out_Directory_224x224 + 'Tiny Training Dataset\\' Small_Out_Directory_224x224 = General_Out_Directory_224x224 + 'Small Training Dataset\\' Balanced_Out_Directory_224x224 = General_Out_Directory_224x224 + 'Balanced Training Dataset\\' Large_Out_Directory_224x224 = General_Out_Directory_224x224 + 'Large Training Dataset\\' Full_Out_Directory_224x224 = General_Out_Directory_224x224 + 'Full Training Dataset\\' Very_Tiny_Out_Directory_224x224x224 = General_Out_Directory_224x224x224 + 'Very Tiny Training Dataset\\' Tiny_Out_Directory_224x224x224 = General_Out_Directory_224x224x224 + 'Tiny Training Dataset\\' Small_Out_Directory_224x224x224 = General_Out_Directory_224x224x224 + 'Small Training Dataset\\' Balanced_Out_Directory_224x224x224 = General_Out_Directory_224x224x224 + 'Balanced Training Dataset\\' Large_Out_Directory_224x224x224 = General_Out_Directory_224x224x224 + 'Large Training Dataset\\' Full_Out_Directory_224x224x224 = General_Out_Directory_224x224x224 + 'Full Training Dataset\\' Very_Tiny_Out_Directory_224x224x224x224 = General_Out_Directory_224x224x224x224 + 'Very Tiny Training Dataset\\' Tiny_Out_Directory_224x224x224x224 = General_Out_Directory_224x224x224x224 + 'Tiny Training Dataset\\' Small_Out_Directory_224x224x224x224 = General_Out_Directory_224x224x224x224 + 'Small Training Dataset\\' Balanced_Out_Directory_224x224x224x224 = General_Out_Directory_224x224x224x224 + 'Balanced Training Dataset\\' Large_Out_Directory_224x224x224x224 = General_Out_Directory_224x224x224x224 + 'Large Training Dataset\\' Full_Out_Directory_224x224x224x224 = General_Out_Directory_224x224x224x224 + 'Full Training Dataset\\' Very_Tiny_Out_Directory_240p = General_Out_Directory_240p + 'Very Tiny Training Dataset\\' Tiny_Out_Directory_240p = General_Out_Directory_240p + 'Tiny Training Dataset\\' Small_Out_Directory_240p = General_Out_Directory_240p + 'Small Training Dataset\\' Balanced_Out_Directory_240p = General_Out_Directory_240p + 'Balanced Training Dataset\\' Large_Out_Directory_240p = General_Out_Directory_240p + 'Large Training Dataset\\' Full_Out_Directory_240p = General_Out_Directory_240p + 'Full Training Dataset\\' Very_Tiny_Out_Directory_360p = General_Out_Directory_360p + 'Very Tiny Training Dataset\\' Tiny_Out_Directory_360p = General_Out_Directory_360p + 'Tiny Training Dataset\\' Small_Out_Directory_360p = General_Out_Directory_360p + 'Small Training Dataset\\' Balanced_Out_Directory_360p = General_Out_Directory_360p + 'Balanced Training Dataset\\' Large_Out_Directory_360p = General_Out_Directory_360p + 'Large Training Dataset\\' Full_Out_Directory_360p = General_Out_Directory_360p + 'Full Training Dataset\\' Very_Tiny_Out_Directory_480p = General_Out_Directory_480p + 'Very Tiny Training Dataset\\' Tiny_Out_Directory_480p = General_Out_Directory_480p + 'Tiny Training Dataset\\' Small_Out_Directory_480p = General_Out_Directory_480p + 'Small Training Dataset\\' Balanced_Out_Directory_480p = General_Out_Directory_480p + 'Balanced Training Dataset\\' Large_Out_Directory_480p = General_Out_Directory_480p + 'Large Training Dataset\\' Full_Out_Directory_480p = General_Out_Directory_480p + 'Full Training Dataset\\' Very_Tiny_Out_Directory_640p = General_Out_Directory_640p + 'Very Tiny Training Dataset\\' Tiny_Out_Directory_640p = General_Out_Directory_640p + 'Tiny Training Dataset\\' Small_Out_Directory_640p = General_Out_Directory_640p + 'Small Training Dataset\\' Balanced_Out_Directory_640p = General_Out_Directory_640p + 'Balanced Training Dataset\\' Large_Out_Directory_640p = General_Out_Directory_640p + 'Large Training Dataset\\' Full_Out_Directory_640p = General_Out_Directory_640p + 'Full Training Dataset\\' Very_Tiny_Out_Directory_640x480 = General_Out_Directory_640x480 + 'Very Tiny Training Dataset\\' Tiny_Out_Directory_640x480 = General_Out_Directory_640x480 + 'Tiny Training Dataset\\' Small_Out_Directory_640x480 = General_Out_Directory_640x480 + 'Small Training Dataset\\' Balanced_Out_Directory_640x480 = General_Out_Directory_640x480 + 'Balanced Training Dataset\\' Large_Out_Directory_640x480 = General_Out_Directory_640x480 + 'Large Training Dataset\\' Full_Out_Directory_640x480 = General_Out_Directory_640x480 + 'Full Training Dataset\\' ########################################################################################################### VTFD_144p = Very_Tiny_Out_Directory_144p + 'Food\\' TFD_144p = Tiny_Out_Directory_144p + 'Food\\' SFD_144p = Small_Out_Directory_144p + 'Food\\' BFD_144p = Balanced_Out_Directory_144p + 'Food\\' LFD_144p = Large_Out_Directory_144p + 'Food\\' FFD_144p = Full_Out_Directory_144p + 'Food\\' VTNFD_144p = Very_Tiny_Out_Directory_144p + 'No Food\\' TNFD_144p = Tiny_Out_Directory_144p + 'No Food\\' SNFD_144p = Small_Out_Directory_144p + 'No Food\\' BNFD_144p = Balanced_Out_Directory_144p + 'No Food\\' LNFD_144p = Large_Out_Directory_144p + 'No Food\\' FNFD_144p = Full_Out_Directory_144p + 'No Food\\' ########################################################################################################### VTFD_224p = Very_Tiny_Out_Directory_224p + 'Food\\' TFD_224p = Tiny_Out_Directory_224p + 'Food\\' SFD_224p = Small_Out_Directory_224p + 'Food\\' BFD_224p = Balanced_Out_Directory_224p + 'Food\\' LFD_224p = Large_Out_Directory_224p + 'Food\\' FFD_224p = Full_Out_Directory_224p + 'Food\\' VTNFD_224p = Very_Tiny_Out_Directory_224p + 'No Food\\' TNFD_224p = Tiny_Out_Directory_224p + 'No Food\\' SNFD_224p = Small_Out_Directory_224p + 'No Food\\' BNFD_224p = Balanced_Out_Directory_224p + 'No Food\\' LNFD_224p = Large_Out_Directory_224p + 'No Food\\' FNFD_224p = Full_Out_Directory_224p + 'No Food\\' ########################################################################################################### VTFD_240p = Very_Tiny_Out_Directory_240p + 'Food\\' TFD_240p = Tiny_Out_Directory_240p + 'Food\\' SFD_240p = Small_Out_Directory_240p + 'Food\\' BFD_240p = Balanced_Out_Directory_240p + 'Food\\' LFD_240p = Large_Out_Directory_240p + 'Food\\' FFD_240p = Full_Out_Directory_240p + 'Food\\' VTNFD_240p = Very_Tiny_Out_Directory_240p + 'No Food\\' TNFD_240p = Tiny_Out_Directory_240p + 'No Food\\' SNFD_240p = Small_Out_Directory_240p + 'No Food\\' BNFD_240p = Balanced_Out_Directory_240p + 'No Food\\' LNFD_240p = Large_Out_Directory_240p + 'No Food\\' FNFD_240p = Full_Out_Directory_240p + 'No Food\\' ########################################################################################################### VTFD_360p = Very_Tiny_Out_Directory_360p + 'Food\\' TFD_360p = Tiny_Out_Directory_360p + 'Food\\' SFD_360p = Small_Out_Directory_360p + 'Food\\' BFD_360p = Balanced_Out_Directory_360p + 'Food\\' LFD_360p = Large_Out_Directory_360p + 'Food\\' FFD_360p = Full_Out_Directory_360p + 'Food\\' VTNFD_360p = Very_Tiny_Out_Directory_360p + 'No Food\\' TNFD_360p = Tiny_Out_Directory_360p + 'No Food\\' SNFD_360p = Small_Out_Directory_360p + 'No Food\\' BNFD_360p = Balanced_Out_Directory_360p + 'No Food\\' LNFD_360p = Large_Out_Directory_360p + 'No Food\\' FNFD_360p = Full_Out_Directory_360p + 'No Food\\' ########################################################################################################### VTFD_480p = Very_Tiny_Out_Directory_480p + 'Food\\' TFD_480p = Tiny_Out_Directory_480p + 'Food\\' SFD_480p = Small_Out_Directory_480p + 'Food\\' BFD_480p = Balanced_Out_Directory_480p + 'Food\\' LFD_480p = Large_Out_Directory_480p + 'Food\\' FFD_480p = Full_Out_Directory_480p + 'Food\\' VTNFD_480p = Very_Tiny_Out_Directory_480p + 'No Food\\' TNFD_480p = Tiny_Out_Directory_480p + 'No Food\\' SNFD_480p = Small_Out_Directory_480p + 'No Food\\' BNFD_480p = Balanced_Out_Directory_480p + 'No Food\\' LNFD_480p = Large_Out_Directory_480p + 'No Food\\' FNFD_480p = Full_Out_Directory_480p + 'No Food\\' ########################################################################################################### VTFD_640p = Very_Tiny_Out_Directory_640p + 'Food\\' TFD_640p = Tiny_Out_Directory_640p + 'Food\\' SFD_640p = Small_Out_Directory_640p + 'Food\\' BFD_640p = Balanced_Out_Directory_640p + 'Food\\' LFD_640p = Large_Out_Directory_640p + 'Food\\' FFD_640p = Full_Out_Directory_640p + 'Food\\' VTNFD_640p = Very_Tiny_Out_Directory_640p + 'No Food\\' TNFD_640p = Tiny_Out_Directory_640p + 'No Food\\' SNFD_640p = Small_Out_Directory_640p + 'No Food\\' BNFD_640p = Balanced_Out_Directory_640p + 'No Food\\' LNFD_640p = Large_Out_Directory_640p + 'No Food\\' FNFD_640p = Full_Out_Directory_640p + 'No Food\\' ########################################################################################################### VTFD_224x224 = Very_Tiny_Out_Directory_224x224 + 'Food\\' TFD_224x224 = Tiny_Out_Directory_224x224 + 'Food\\' SFD_224x224 = Small_Out_Directory_224x224 + 'Food\\' BFD_224x224 = Balanced_Out_Directory_224x224 + 'Food\\' LFD_224x224 = Large_Out_Directory_224x224 + 'Food\\' FFD_224x224 = Full_Out_Directory_224x224 + 'Food\\' VTNFD_224x224 = Very_Tiny_Out_Directory_224x224 + 'No Food\\' TNFD_224x224 = Tiny_Out_Directory_224x224 + 'No Food\\' SNFD_224x224 = Small_Out_Directory_224x224 + 'No Food\\' BNFD_224x224 = Balanced_Out_Directory_224x224 + 'No Food\\' LNFD_224x224 = Large_Out_Directory_224x224 + 'No Food\\' FNFD_224x224 = Full_Out_Directory_224x224 + 'No Food\\' ########################################################################################################### VTFD_224x224x224 = Very_Tiny_Out_Directory_224x224x224 + 'Food\\' TFD_224x224x224 = Tiny_Out_Directory_224x224x224 + 'Food\\' SFD_224x224x224 = Small_Out_Directory_224x224x224 + 'Food\\' BFD_224x224x224 = Balanced_Out_Directory_224x224x224 + 'Food\\' LFD_224x224x224 = Large_Out_Directory_224x224x224 + 'Food\\' FFD_224x224x224 = Full_Out_Directory_224x224x224 + 'Food\\' VTNFD_224x224x224 = Very_Tiny_Out_Directory_224x224x224 + 'No Food\\' TNFD_224x224x224 = Tiny_Out_Directory_224x224x224 + 'No Food\\' SNFD_224x224x224 = Small_Out_Directory_224x224x224 + 'No Food\\' BNFD_224x224x224 = Balanced_Out_Directory_224x224x224 + 'No Food\\' LNFD_224x224x224 = Large_Out_Directory_224x224x224 + 'No Food\\' FNFD_224x224x224 = Full_Out_Directory_224x224x224 + 'No Food\\' ########################################################################################################### VTFD_224x224x224x224 = Very_Tiny_Out_Directory_224x224x224x224 + 'Food\\' TFD_224x224x224x224 = Tiny_Out_Directory_224x224x224x224 + 'Food\\' SFD_224x224x224x224 = Small_Out_Directory_224x224x224x224 + 'Food\\' BFD_224x224x224x224 = Balanced_Out_Directory_224x224x224x224 + 'Food\\' LFD_224x224x224x224 = Large_Out_Directory_224x224x224x224 + 'Food\\' FFD_224x224x224x224 = Full_Out_Directory_224x224x224x224 + 'Food\\' VTNFD_224x224x224x224 = Very_Tiny_Out_Directory_224x224x224x224 + 'No Food\\' TNFD_224x224x224x224 = Tiny_Out_Directory_224x224x224x224 + 'No Food\\' SNFD_224x224x224x224 = Small_Out_Directory_224x224x224x224 + 'No Food\\' BNFD_224x224x224x224 = Balanced_Out_Directory_224x224x224x224 + 'No Food\\' LNFD_224x224x224x224 = Large_Out_Directory_224x224x224x224 + 'No Food\\' FNFD_224x224x224x224 = Full_Out_Directory_224x224x224x224 + 'No Food\\' ########################################################################################################### VTFD_640x480 = Very_Tiny_Out_Directory_640x480 + 'Food\\' TFD_640x480 = Tiny_Out_Directory_640x480 + 'Food\\' SFD_640x480 = Small_Out_Directory_640x480 + 'Food\\' BFD_640x480 = Balanced_Out_Directory_640x480 + 'Food\\' LFD_640x480 = Large_Out_Directory_640x480 + 'Food\\' FFD_640x480 = Full_Out_Directory_640x480 + 'Food\\' VTNFD_640x480 = Very_Tiny_Out_Directory_640x480 + 'No Food\\' TNFD_640x480 = Tiny_Out_Directory_640x480 + 'No Food\\' SNFD_640x480 = Small_Out_Directory_640x480 + 'No Food\\' BNFD_640x480 = Balanced_Out_Directory_640x480 + 'No Food\\' LNFD_640x480 = Large_Out_Directory_640x480 + 'No Food\\' FNFD_640x480 = Full_Out_Directory_640x480 + 'No Food\\' ########################################################################################################### VTFD = Very_Tiny_General_Directory + 'Food\\' VTNFD = Very_Tiny_General_Directory + 'Non Food\\' TFD = Tiny_General_Directory + 'Food\\' TNFD = Tiny_General_Directory + 'Non Food\\' SFD = Small_General_Directory + 'Food\\' SNFD = Small_General_Directory + 'Non Food\\' BFD = Balanced_General_Directory + 'Food\\' BNFD = Balanced_General_Directory + 'Non Food\\' LFD = Large_General_Directory + 'Food\\' LNFD = Large_General_Directory + 'Non Food\\' FFD = Full_General_Directory + 'Food\\' FNFD =Full_General_Directory + 'Non Food\\' ######################################################################################################## #VTFD_G256p = VTFD + "G-256p\\" VTFD_G360p = VTFD + "G-360p\\" VTFD_G480p = VTFD + "G-480p\\" VTFD_G640p = VTFD + "G-640p\\" VTFD_G720p = VTFD + "G-720p\\" VTFD_G1024p = VTFD + "G-1024p\\" VTFD_G1080p = VTFD + "G-1080p\\" #VTFD_GMedium = VTFD + "G-Medium\\" #VTFD_GLarge = VTFD + "G-Large\\" #VTNFD_G256p = VTNFD + "G-256p\\" VTNFD_G360p = VTNFD + "G-360p\\" VTNFD_G480p = VTNFD + "G-480p\\" VTNFD_G640p = VTNFD + "G-640p\\" VTNFD_G720p = VTNFD + "G-720p\\" VTNFD_G1024p = VTNFD + "G-1024p\\" VTNFD_G1080p = VTNFD + "G-1080p\\" #VTNFD_GMedium = VTNFD + "G-Medium\\" #VTNFD_GLarge = VTNFD + "G-Large\\" #VTFD_BSmall = VTFD + "B-Small\\" #VTFD_B640p = VTFD + "B-640p\\" #VTFD_B1000p = VTFD + "B-1000p\\" #VTFD_BMedium = VTFD + "B-Medium\\" #VTFD_BLarge = VTFD + "B-Large\\" #VTFD_B1280p = VTFD + "B-1280p\\" #VTFD_B1920p = VTFD + "B-1920p\\" #VTNFD_BSmall = VTNFD + "B-Small\\" #VTNFD_B640p = VTNFD + "B-640p\\" #VTNFD_B1000p = VTNFD + "B-1000p\\" #VTNFD_BMedium = VTNFD + "B-Medium\\" #VTNFD_BLarge = VTNFD + "B-Large\\" #VTNFD_B1280p = VTNFD + "B-1280p\\" #VTNFD_B1920p = VTNFD + "B-1920p\\" ############################################CONSTANTS################################################### #TFD_G256p = TFD + "G-256p\\" #TFD_G360p = TFD + "G-360p\\" #TFD_G480p = TFD + "G-480p\\" #TFD_G640p = TFD + "G-640p\\" #TFD_G720p = TFD + "G-720p\\" #TFD_G1024p = TFD + "G-1024p\\" #TFD_G1080p = TFD + "G-1080p\\" #TFD_GMedium = TFD + "G-Medium\\" #TFD_GLarge = TFD + "G-Large\\" #TNFD_G256p = TNFD + "G-256p\\" #TNFD_G360p = TNFD + "G-360p\\" #TNFD_G480p = TNFD + "G-480p\\" #TNFD_G640p = TNFD + "G-640p\\" #TNFD_G720p = TNFD + "G-720p\\" #TNFD_G1024p = TNFD + "G-1024p\\" #TNFD_G1080p = TNFD + "G-1080p\\" #TNFD_GMedium = TNFD + "G-Medium\\" #TNFD_GLarge = TNFD + "G-Large\\" TFD_BSmall = TFD + "B-Small\\" TFD_B640p = TFD + "B-640p\\" TFD_B1000p = TFD + "B-1000p\\" TFD_BMedium = TFD + "B-Medium\\" TFD_BLarge = TFD + "B-Large\\" TFD_B1280p = TFD + "B-1280p\\" TFD_B1920p = TFD + "B-1920p\\" TNFD_BSmall = TNFD + "B-Small\\" TNFD_B640p = TNFD + "B-640p\\" TNFD_B1000p = TNFD + "B-1000p\\" TNFD_BMedium = TNFD + "B-Medium\\" TNFD_BLarge = TNFD + "B-Large\\" TNFD_B1280p = TNFD + "B-1280p\\" TNFD_B1920p = TNFD + "B-1920p\\" ######################################################################################################## #SFD_G256p = SFD + "G-256p\\" #SFD_G360p = SFD + "G-360p\\" #SFD_G480p = SFD + "G-480p\\" #SFD_G640p = SFD + "G-640p\\" #SFD_G720p = SFD + "G-720p\\" #SFD_G1024p = SFD + "G-1024p\\" #SFD_G1080p = SFD + "G-1080p\\" #SFD_GMedium = SFD + "G-Medium\\" #SFD_GLarge = SFD + "G-Large\\" #SNFD_G256p = SNFD + "G-256p\\" #SNFD_G360p = SNFD + "G-360p\\" #SNFD_G480p = SNFD + "G-480p\\" #SNFD_G640p = SNFD + "G-640p\\" #SNFD_G720p = SNFD + "G-720p\\" #SNFD_G1024p = SNFD + "G-1024p\\" #SNFD_G1080p = SNFD + "G-1080p\\" #SNFD_GMedium = SNFD + "G-Medium\\" #SNFD_GLarge = SNFD + "G-Large\\" #SFD_BSmall = SFD + "B-Small\\" SFD_B640p = SFD + "B-640p\\" SFD_B1000p = SFD + "B-1000p\\" SFD_BMedium = SFD + "B-Medium\\" SFD_BLarge = SFD + "B-Large\\" SFD_B1280p = SFD + "B-1280p\\" SFD_B1920p = SFD + "B-1920p\\" #SNFD_BSmall = SNFD + "B-Small\\" SNFD_B640p = SNFD + "B-640p\\" SNFD_B1000p = SNFD + "B-1000p\\" SNFD_BMedium = SNFD + "B-Medium\\" SNFD_BLarge = SNFD + "B-Large\\" SNFD_B1280p = SNFD + "B-1280p\\" SNFD_B1920p = SNFD + "B-1920p\\" ######################################################################################################## ###Output _____no_output_____ ###Markdown Datasets Clusters Based on Aspect Ratio ###Code ############################################CLUSTER DIRECTORIES########################################### VTFCALL = [] VTNFCALL = [] VTFCALL.append(VTFD_G360p) VTFCALL.append(VTFD_G480p) VTFCALL.append(VTFD_G640p) VTFCALL.append(VTFD_G720p) VTFCALL.append(VTFD_G1024p) VTFCALL.append(VTFD_G1080p) #VTFCALL.append(VTFD_GLarge) #VTFCALL.append(VTFD_GMedium) #VTFCALL.append(VTFD_BLarge) #VTFCALL.append(VTFD_BMedium) #VTFCALL.append(VTFD_B1280p) #VTFCALL.append(VTFD_B1920p) #VTFCALL.append(VTFD_B1000p) #VTFCALL.append(VTFD_B640p) VTNFCALL.append(VTNFD_G360p) VTNFCALL.append(VTNFD_G480p) VTNFCALL.append(VTNFD_G640p) VTNFCALL.append(VTNFD_G720p) VTNFCALL.append(VTNFD_G1024p) VTNFCALL.append(VTNFD_G1080p) #VTNFCALL.append(VTNFD_GLarge) #VTNFCALL.append(VTNFD_GMedium) #VTNFCALL.append(VTNFD_BLarge) #VTNFCALL.append(VTNFD_BMedium) #VTNFCALL.append(VTNFD_B1280p) #VTNFCALL.append(VTNFD_B1920p) #VTNFCALL.append(VTNFD_B1000p) #VTNFCALL.append(VTNFD_B640p) ########################################################################################################## TFCALL = [] TNFCALL = [] TFCALL.append(TFD_BSmall) TFCALL.append(TFD_BLarge) TFCALL.append(TFD_BMedium) TFCALL.append(TFD_B1280p) TFCALL.append(TFD_B1920p) TFCALL.append(TFD_B1000p) TFCALL.append(TFD_B640p) TNFCALL.append(TNFD_BSmall) TNFCALL.append(TNFD_BLarge) TNFCALL.append(TNFD_BMedium) TNFCALL.append(TNFD_B1280p) TNFCALL.append(TNFD_B1920p) TNFCALL.append(TNFD_B1000p) TNFCALL.append(TNFD_B640p) ############################################################################################################# SFCALL = [] SNFCALL = [] SFCALL.append(SFD_BLarge) SFCALL.append(SFD_BMedium) SFCALL.append(SFD_B1280p) SFCALL.append(SFD_B1920p) SFCALL.append(SFD_B1000p) SFCALL.append(SFD_B640p) SNFCALL.append(SNFD_BLarge) SNFCALL.append(SNFD_BMedium) SNFCALL.append(SNFD_B1280p) SNFCALL.append(SNFD_B1920p) SNFCALL.append(SNFD_B1000p) SNFCALL.append(SNFD_B640p) ############################################################################################################# ###Output _____no_output_____ ###Markdown Functions ###Code def resize144p(cluster, destination): counter = 0 skipped = 0 dim = (256, 144) if not os.path.exists(destination): os.makedirs(destination) for i in range (len(cluster)): directory = cluster[i] print("Processing Directory %d of %d..." %(i+1, len(cluster))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) print('%d images were resized. %d images were skipped' %(counter, skipped)) def resize224p(cluster1, cluster2, destination): counter = 0 skipped = 0 dim = (224, 224) if not os.path.exists(destination): os.makedirs(destination) for i in range (len(cluster1)): directory = cluster1[i] print("Processing Directory %d of %d..." %(i+1, len(cluster1))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) for i in range (len(cluster2)): directory = cluster2[i] print("Processing Directory %d of %d..." %(i+1, len(cluster2))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) print('%d images were resized. %d images were skipped' %(counter, skipped)) def resize224x224(cluster, destination): counter = 0 skipped = 0 dim = (224, 224) if not os.path.exists(destination): os.makedirs(destination) for i in range (len(cluster)): directory = cluster[i] print("Processing Directory %d of %d..." %(i+1, len(cluster))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) print('%d images were resized. %d images were skipped' %(counter, skipped)) def resize224x224x224(cluster, destination): counter = 0 skipped = 0 dim = (224, 224) if not os.path.exists(destination): os.makedirs(destination) for i in range (len(cluster)): directory = cluster[i] print("Processing Directory %d of %d..." %(i+1, len(cluster))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) print('%d images were resized. %d images were skipped' %(counter, skipped)) def resize224x224x224x224(cluster1, cluster2, destination): counter = 0 skipped = 0 dim = (224, 224) if not os.path.exists(destination): os.makedirs(destination) for i in range (len(cluster1)): directory = cluster1[i] print("Processing Directory %d of %d..." %(i+1, len(cluster1))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) for i in range (len(cluster2)): directory = cluster2[i] print("Processing Directory %d of %d..." %(i+1, len(cluster2))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) print('%d images were resized. %d images were skipped' %(counter, skipped)) def resize240p(cluster1, cluster2, destination): counter = 0 skipped = 0 dim = (360, 240) if not os.path.exists(destination): os.makedirs(destination) for i in range (len(cluster1)): directory = cluster1[i] print("Processing Directory %d of %d..." %(i+1, len(cluster1))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) for i in range (len(cluster2)): directory = cluster2[i] print("Processing Directory %d of %d..." %(i+1, len(cluster2))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) print('%d images were resized. %d images were skipped' %(counter, skipped)) def resize360p(cluster, destination): counter = 0 skipped = 0 dim = (480, 360) if not os.path.exists(destination): os.makedirs(destination) for i in range (len(cluster)): directory = cluster[i] print("Processing Directory %d of %d..." %(i+1, len(cluster))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) print('%d images were resized. %d images were skipped' %(counter, skipped)) def resize480p(cluster, destination): counter = 0 skipped = 0 dim = (480, 360) if not os.path.exists(destination): os.makedirs(destination) for i in range (len(cluster)): directory = cluster[i] print("Processing Directory %d of %d..." %(i+1, len(cluster))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) print('%d images were resized. %d images were skipped' %(counter, skipped)) def resize640p(cluster1, cluster2, destination): counter = 0 skipped = 0 dim = (640, 360) if not os.path.exists(destination): os.makedirs(destination) for i in range (len(cluster1)): directory = cluster1[i] print("Processing Directory %d of %d..." %(i+1, len(cluster1))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) for i in range (len(cluster2)): directory = cluster2[i] print("Processing Directory %d of %d..." %(i+1, len(cluster2))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) print('%d images were resized. %d images were skipped' %(counter, skipped)) def resize640x480(cluster, destination): counter = 0 skipped = 0 dim = (640, 480) if not os.path.exists(destination): os.makedirs(destination) for i in range (len(cluster)): directory = cluster[i] print("Processing Directory %d of %d..." %(i+1, len(cluster))) print("Processing Directory: %s" % (directory)) for filename in os.listdir(directory): image_path = os.path.join(directory, filename) img = cv2.imread('%s' % (image_path)) if img is not None: resized = cv2.resize(img, dim, interpolation = cv2.INTER_AREA) cv2.imwrite(destination + '%s.png' %(counter),resized) counter += 1 else: skipped += 1 if(counter % 100 == 0): print('%d images have been resized so far...' %(counter)) print('%d images were resized. %d images were skipped' %(counter, skipped)) ###Output _____no_output_____ ###Markdown Very Tiny Dataset Resizing 144p: (240x144) ###Code VT1 = Thread(target = resize144p, args = (VTFC169, VTFD_144p,)) VT2 = Thread(target = resize144p, args = (VTNFC169, VTNFD_144p,)) VT1.start() VT2.start() VT1.join() VT2.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code VT3 = Thread(target = resize224p, args = (VTFC11, VTFC43, VTFD_224p,)) VT4 = Thread(target = resize224p, args = (VTNFC11, VTNFC43, VTNFD_224p,)) VT3.start() VT4.start() VT3.join() VT4.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code VT5 = Thread(target = resize224x224, args = (VTFCALL, BFD_224x224,)) VT6 = Thread(target = resize224x224, args = (VTNFCALL, BNFD_224x224,)) VT5.start() VT6.start() VT5.join() VT6.join() ###Output Processing Directory 1 of 6... Processing Directory: D:\Food Datasets\Full Training Dataset\Food\G-360p\ Processing Directory 1 of 6... Processing Directory: D:\Food Datasets\Full Training Dataset\Non Food\G-360p\ 100 images have been resized so far... 100 images have been resized so far... 200 images have been resized so far... 200 images have been resized so far... 300 images have been resized so far... 300 images have been resized so far... 400 images have been resized so far... 400 images have been resized so far... 500 images have been resized so far... 500 images have been resized so far... 600 images have been resized so far... 600 images have been resized so far... 700 images have been resized so far... 700 images have been resized so far... 800 images have been resized so far... 800 images have been resized so far... 900 images have been resized so far... 900 images have been resized so far... 1000 images have been resized so far... 1000 images have been resized so far... 1100 images have been resized so far... 1100 images have been resized so far... 1200 images have been resized so far... 1200 images have been resized so far... 1300 images have been resized so far... 1300 images have been resized so far... 1400 images have been resized so far... 1400 images have been resized so far... 1500 images have been resized so far... 1500 images have been resized so far... 1600 images have been resized so far... 1600 images have been resized so far... 1700 images have been resized so far... 1700 images have been resized so far... 1800 images have been resized so far... 1800 images have been resized so far... 1900 images have been resized so far... 1900 images have been resized so far... 2000 images have been resized so far... 2000 images have been resized so far... 2100 images have been resized so far... 2100 images have been resized so far... 2200 images have been resized so far... 2200 images have been resized so far... 2300 images have been resized so far... 2300 images have been resized so far... 2400 images have been resized so far... 2400 images have been resized so far... 2500 images have been resized so far... 2500 images have been resized so far... 2600 images have been resized so far... 2600 images have been resized so far... 2700 images have been resized so far... 2700 images have been resized so far... 2800 images have been resized so far... 2800 images have been resized so far... 2900 images have been resized so far... 2900 images have been resized so far... 3000 images have been resized so far... 3000 images have been resized so far... 3100 images have been resized so far... 3200 images have been resized so far... 3100 images have been resized so far... 3300 images have been resized so far... 3200 images have been resized so far... 3400 images have been resized so far... 3300 images have been resized so far... 3500 images have been resized so far... 3400 images have been resized so far... 3600 images have been resized so far... 3500 images have been resized so far... 3700 images have been resized so far... 3600 images have been resized so far... 3800 images have been resized so far... 3700 images have been resized so far... 3900 images have been resized so far... 3800 images have been resized so far... 4000 images have been resized so far... 3900 images have been resized so far... 4100 images have been resized so far... 4000 images have been resized so far... 4200 images have been resized so far... 4100 images have been resized so far... 4300 images have been resized so far... 4200 images have been resized so far... 4400 images have been resized so far... 4300 images have been resized so far... 4500 images have been resized so far... 4400 images have been resized so far... 4600 images have been resized so far... 4500 images have been resized so far... 4700 images have been resized so far... 4600 images have been resized so far... 4800 images have been resized so far... 4700 images have been resized so far... 4900 images have been resized so far... 4800 images have been resized so far... 5000 images have been resized so far... 4900 images have been resized so far... 5100 images have been resized so far... 5000 images have been resized so far... 5200 images have been resized so far... 5100 images have been resized so far... 5300 images have been resized so far... 5200 images have been resized so far... 5400 images have been resized so far... 5300 images have been resized so far... 5400 images have been resized so far... 5500 images have been resized so far... 5500 images have been resized so far... 5600 images have been resized so far... 5600 images have been resized so far... 5700 images have been resized so far... 5800 images have been resized so far... 5700 images have been resized so far... 5800 images have been resized so far... 5900 images have been resized so far... 6000 images have been resized so far... 5900 images have been resized so far... 6000 images have been resized so far... 6100 images have been resized so far... 6200 images have been resized so far... 6100 images have been resized so far... 6300 images have been resized so far... 6200 images have been resized so far... 6400 images have been resized so far... 6300 images have been resized so far... 6500 images have been resized so far... 6400 images have been resized so far... 6600 images have been resized so far... 6500 images have been resized so far... 6700 images have been resized so far... 6600 images have been resized so far... 6800 images have been resized so far... 6700 images have been resized so far... 6800 images have been resized so far... 6900 images have been resized so far... 6900 images have been resized so far... 7000 images have been resized so far... 7100 images have been resized so far... 7000 images have been resized so far... 7200 images have been resized so far... 7100 images have been resized so far... 7300 images have been resized so far... 7200 images have been resized so far... 7400 images have been resized so far... 7300 images have been resized so far... 7500 images have been resized so far... 7400 images have been resized so far... 7600 images have been resized so far... 7500 images have been resized so far... 7700 images have been resized so far... 7600 images have been resized so far... 7800 images have been resized so far... 7700 images have been resized so far... 7900 images have been resized so far... 7800 images have been resized so far... 8000 images have been resized so far... 7900 images have been resized so far... 8100 images have been resized so far... 8000 images have been resized so far... 8200 images have been resized so far... 8100 images have been resized so far... 8300 images have been resized so far... Processing Directory 2 of 6... Processing Directory: D:\Food Datasets\Full Training Dataset\Food\G-480p\ 8200 images have been resized so far... 8400 images have been resized so far... 8500 images have been resized so far... 8300 images have been resized so far... 8600 images have been resized so far... 8400 images have been resized so far... 8700 images have been resized so far... 8500 images have been resized so far... 8800 images have been resized so far... 8900 images have been resized so far... 8600 images have been resized so far... 9000 images have been resized so far... 8700 images have been resized so far... 9100 images have been resized so far... 9200 images have been resized so far... 8800 images have been resized so far... 9300 images have been resized so far... 8900 images have been resized so far... 9400 images have been resized so far... 9000 images have been resized so far... 9500 images have been resized so far... 9600 images have been resized so far... Processing Directory 2 of 6... Processing Directory: D:\Food Datasets\Full Training Dataset\Non Food\G-480p\ 9100 images have been resized so far... 9700 images have been resized so far... 9200 images have been resized so far... 9800 images have been resized so far... 9300 images have been resized so far... 9900 images have been resized so far... 9400 images have been resized so far... 10000 images have been resized so far... 9500 images have been resized so far... ###Markdown 224p: (224x224) ###Code VT7 = Thread(target = resize224x224x224, args = (VTFCMIX, VTFD_224x224x224,)) VT8 = Thread(target = resize224x224x224, args = (VTNFCMIX, VTNFD_224x224x224,)) VT7.start() VT8.start() VT7.join() VT8.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code VT9 = Thread(target = resize224x224x224x224, args = (VTFC43, VTFC169, VTFD_224x224x224x224,)) VT10 = Thread(target = resize224x224x224x224, args = (VTNFC43, VTNFC169, VTNFD_224x224x224x224,)) VT9.start() VT10.start() VT9.join() VT10.join() ###Output _____no_output_____ ###Markdown 240p: (360x240) ###Code VT11 = Thread(target = resize240p, args = (VTFC43, VTFC169, VTFD_240p,)) VT12 = Thread(target = resize240p, args = (VTNFC43, VTNFC169, VTNFD_240p,)) VT11.start() VT12.start() VT11.join() VT12.join() ###Output _____no_output_____ ###Markdown 360p: (480x360) ###Code VT13 = Thread(target = resize360p, args = (VTFCALL, VTFD_360p,)) VT14 = Thread(target = resize360p, args = (VTNFCALL, VTNFD_360p,)) VT13.start() VT14.start() VT13.join() VT14.join() ###Output _____no_output_____ ###Markdown 360p: (480x360) ###Code VT15 = Thread(target = resize480p, args = (VTFC43, VTFD_480p,)) VT16 = Thread(target = resize480p, args = (VTNFC43, VTNFD_480p,)) VT15.start() VT16.start() VT15.join() VT16.join() ###Output _____no_output_____ ###Markdown 360p: (640x360) ###Code VT17 = Thread(target = resize640p, args = (VTFC43, VTFC169, VTFD_640p,)) VT18 = Thread(target = resize640p, args = (VTNFC43, VTNFC169, VTNFD_640p,)) VT17.start() VT18.start() VT17.join() VT18.join() ###Output _____no_output_____ ###Markdown 480p: (640x480) ###Code VT19 = Thread(target = resize640x480, args = (VTFCMIX, VTFD_640x480,)) VT20 = Thread(target = resize640x480, args = (VTNFCMIX, VTNFD_640x480,)) VT19.start() VT20.start() VT19.join() VT20.join() ###Output _____no_output_____ ###Markdown Tiny Dataset Resizing 144p: (240x144) ###Code T1 = Thread(target = resize144p, args = (TFC169, TFD_144p,)) T2 = Thread(target = resize144p, args = (TNFC169, TNFD_144p,)) T1.start() T2.start() T1.join() T2.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code T3 = Thread(target = resize224p, args = (TFC11, TFC43, TFD_224p,)) T4 = Thread(target = resize224p, args = (TNFC11, TNFC43, TNFD_224p,)) T3.start() T4.start() T3.join() T4.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code T5 = Thread(target = resize224x224, args = (TFCALL, TFD_224x224,)) T6 = Thread(target = resize224x224, args = (TNFCALL, TNFD_224x224,)) T5.start() T6.start() T5.join() T6.join() ###Output Processing Directory 1 of 7...Processing Directory 1 of 7... Processing Directory: D:\Food Datasets\Large Training Dataset\Non Food\B-Small\ Processing Directory: D:\Food Datasets\Large Training Dataset\Food\B-Small\ 100 images have been resized so far... 100 images have been resized so far... 200 images have been resized so far... 200 images have been resized so far... 300 images have been resized so far... 300 images have been resized so far... 400 images have been resized so far... 400 images have been resized so far... 500 images have been resized so far... 500 images have been resized so far... 600 images have been resized so far... 600 images have been resized so far... 700 images have been resized so far... 700 images have been resized so far... 800 images have been resized so far... 800 images have been resized so far... 900 images have been resized so far... 900 images have been resized so far... 1000 images have been resized so far... 1000 images have been resized so far... 1100 images have been resized so far... 1100 images have been resized so far... 1200 images have been resized so far... 1200 images have been resized so far... 1300 images have been resized so far... 1300 images have been resized so far... 1400 images have been resized so far... 1400 images have been resized so far... 1500 images have been resized so far... 1500 images have been resized so far... 1600 images have been resized so far... 1600 images have been resized so far... 1700 images have been resized so far... 1700 images have been resized so far... 1800 images have been resized so far... 1800 images have been resized so far... 1900 images have been resized so far... 1900 images have been resized so far... 2000 images have been resized so far... 2000 images have been resized so far... 2100 images have been resized so far... 2100 images have been resized so far... 2200 images have been resized so far... 2200 images have been resized so far... 2300 images have been resized so far... 2300 images have been resized so far... 2400 images have been resized so far... 2400 images have been resized so far... 2500 images have been resized so far... 2500 images have been resized so far... 2600 images have been resized so far... 2600 images have been resized so far... 2700 images have been resized so far... 2700 images have been resized so far... 2800 images have been resized so far... 2800 images have been resized so far... 2900 images have been resized so far... 2900 images have been resized so far... 3000 images have been resized so far... 3100 images have been resized so far... 3200 images have been resized so far... 3000 images have been resized so far... 3300 images have been resized so far... 3100 images have been resized so far... 3400 images have been resized so far... 3200 images have been resized so far... 3500 images have been resized so far... 3300 images have been resized so far... 3600 images have been resized so far... 3400 images have been resized so far... 3700 images have been resized so far... 3500 images have been resized so far... 3800 images have been resized so far... 3600 images have been resized so far... 3900 images have been resized so far... 3700 images have been resized so far... 4000 images have been resized so far... 3800 images have been resized so far... 4100 images have been resized so far... 4200 images have been resized so far... 3900 images have been resized so far... 4300 images have been resized so far... 4000 images have been resized so far... 4400 images have been resized so far... 4100 images have been resized so far... 4500 images have been resized so far... 4200 images have been resized so far... 4600 images have been resized so far... 4300 images have been resized so far... 4700 images have been resized so far... 4400 images have been resized so far... 4800 images have been resized so far... 4500 images have been resized so far... 4900 images have been resized so far... 5000 images have been resized so far... 4600 images have been resized so far... 5100 images have been resized so far... 4700 images have been resized so far... 5200 images have been resized so far... 4800 images have been resized so far... 5300 images have been resized so far... 4900 images have been resized so far... 5400 images have been resized so far... 5000 images have been resized so far... 5500 images have been resized so far... 5100 images have been resized so far... 5600 images have been resized so far... 5200 images have been resized so far... 5700 images have been resized so far... 5300 images have been resized so far... 5800 images have been resized so far... 5900 images have been resized so far... 5400 images have been resized so far... 6000 images have been resized so far... 5500 images have been resized so far... 5600 images have been resized so far... 6100 images have been resized so far... 5700 images have been resized so far... 5800 images have been resized so far... 6200 images have been resized so far... 5900 images have been resized so far... 6000 images have been resized so far... 6100 images have been resized so far... 6300 images have been resized so far... Processing Directory 2 of 7... Processing Directory: D:\Food Datasets\Large Training Dataset\Food\B-Large\ 6400 images have been resized so far... 6200 images have been resized so far... 6500 images have been resized so far... 6600 images have been resized so far... 6700 images have been resized so far... 6800 images have been resized so far... 6900 images have been resized so far... 7000 images have been resized so far... 6300 images have been resized so far... 7100 images have been resized so far... 7200 images have been resized so far... 7300 images have been resized so far... 7400 images have been resized so far... 7500 images have been resized so far... 7600 images have been resized so far... 7700 images have been resized so far... 6400 images have been resized so far... 7800 images have been resized so far... 7900 images have been resized so far... 8000 images have been resized so far... 8100 images have been resized so far... 8200 images have been resized so far... 8300 images have been resized so far... 8400 images have been resized so far... 6500 images have been resized so far... 8500 images have been resized so far... 8600 images have been resized so far... 8700 images have been resized so far... 8800 images have been resized so far... 8900 images have been resized so far... 9000 images have been resized so far... 6600 images have been resized so far... 9100 images have been resized so far... 9200 images have been resized so far... 9300 images have been resized so far... 9400 images have been resized so far... 9500 images have been resized so far... 6700 images have been resized so far... 9600 images have been resized so far... 9700 images have been resized so far... 9800 images have been resized so far... 6800 images have been resized so far... 9900 images have been resized so far... 10000 images have been resized so far... 10100 images have been resized so far... 10200 images have been resized so far... 10300 images have been resized so far... 6900 images have been resized so far... 10400 images have been resized so far... 10500 images have been resized so far... 10600 images have been resized so far... 10700 images have been resized so far... 7000 images have been resized so far... 10800 images have been resized so far... 10900 images have been resized so far... 11000 images have been resized so far... 7100 images have been resized so far... 11100 images have been resized so far... 11200 images have been resized so far... 11300 images have been resized so far... 11400 images have been resized so far... 11500 images have been resized so far... 7200 images have been resized so far... 11600 images have been resized so far... 11700 images have been resized so far... 11800 images have been resized so far... 11900 images have been resized so far... 12000 images have been resized so far... 7300 images have been resized so far... 12100 images have been resized so far... 12200 images have been resized so far... 12300 images have been resized so far... 12400 images have been resized so far... ###Markdown 224p: (224x224) ###Code T7 = Thread(target = resize224x224x224, args = (TFCMIX, TFD_224x224x224,)) T8 = Thread(target = resize224x224x224, args = (TNFCMIX, TNFD_224x224x224,)) T7.start() T8.start() T7.join() T8.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code T9 = Thread(target = resize224x224x224x224, args = (TFC43, TFC169, TFD_224x224x224x224,)) T10 = Thread(target = resize224x224x224x224, args = (TNFC43, TNFC169, TNFD_224x224x224x224,)) T9.start() T10.start() T9.join() T10.join() ###Output _____no_output_____ ###Markdown 240p: (360x240) ###Code T11 = Thread(target = resize240p, args = (TFC43, TFC169, TFD_240p,)) T12 = Thread(target = resize240p, args = (TNFC43, TNFC169, TNFD_240p,)) T11.start() T12.start() T11.join() T12.join() ###Output _____no_output_____ ###Markdown 360p: (480x360) ###Code T13 = Thread(target = resize360p, args = (TFCALL, TFD_360p,)) T14 = Thread(target = resize360p, args = (TNFCALL, TNFD_360p,)) T13.start() T14.start() T13.join() T14.join() ###Output _____no_output_____ ###Markdown 360p: (480x360) ###Code T15 = Thread(target = resize480p, args = (TFC43, TFD_480p,)) T16 = Thread(target = resize480p, args = (TNFC43, TNFD_480p,)) T15.start() T16.start() T15.join() T16.join() ###Output _____no_output_____ ###Markdown 360p: (640x360) ###Code T17 = Thread(target = resize640p, args = (TFC43, TFC169, TFD_640p,)) T18 = Thread(target = resize640p, args = (TNFC43, TNFC169, TNFD_640p,)) T17.start() T18.start() T17.join() T18.join() ###Output _____no_output_____ ###Markdown 480p: (640x480) ###Code T19 = Thread(target = resize640x480, args = (TFCMIX, TFD_640x480,)) T20 = Thread(target = resize640x480, args = (TNFCMIX, TNFD_640x480,)) T19.start() T20.start() T19.join() T20.join() ###Output _____no_output_____ ###Markdown Small Dataset Resizing 144p: (240x144) ###Code S1 = Thread(target = resize144p, args = (SFC169, SFD_144p,)) S2 = Thread(target = resize144p, args = (SNFC169, SNFD_144p,)) S1.start() S2.start() S1.join() S2.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code S3 = Thread(target = resize224p, args = (SFC11, SFC43, SFD_224p,)) S4 = Thread(target = resize224p, args = (SNFC11, SNFC43, SNFD_224p,)) S3.start() S4.start() S3.join() S4.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code S5 = Thread(target = resize224x224, args = (SFCALL, SFD_224x224,)) S6 = Thread(target = resize224x224, args = (SNFCALL, SNFD_224x224,)) S5.start() S6.start() S5.join() S6.join() ###Output Processing Directory 1 of 6... Processing Directory: D:\Food Datasets\Small Training Dataset\Non Food\B-Large\ Processing Directory 1 of 6... Processing Directory: D:\Food Datasets\Small Training Dataset\Food\B-Large\ 100 images have been resized so far... 100 images have been resized so far... 200 images have been resized so far... 200 images have been resized so far... 300 images have been resized so far... 400 images have been resized so far... 300 images have been resized so far... 500 images have been resized so far... 400 images have been resized so far... 600 images have been resized so far... 500 images have been resized so far... 700 images have been resized so far... 600 images have been resized so far... 800 images have been resized so far... 700 images have been resized so far... 900 images have been resized so far... 800 images have been resized so far... 1000 images have been resized so far... 900 images have been resized so far... 1100 images have been resized so far... 1000 images have been resized so far... 1200 images have been resized so far... 1100 images have been resized so far... 1200 images have been resized so far... 1300 images have been resized so far... 1300 images have been resized so far... 1400 images have been resized so far... 1400 images have been resized so far... 1500 images have been resized so far... 1500 images have been resized so far... 1600 images have been resized so far... 1600 images have been resized so far... 1700 images have been resized so far... 1700 images have been resized so far... 1800 images have been resized so far... 1800 images have been resized so far... 1900 images have been resized so far... 1900 images have been resized so far... 2000 images have been resized so far... 2000 images have been resized so far... 2100 images have been resized so far... 2100 images have been resized so far... 2200 images have been resized so far... 2200 images have been resized so far... 2300 images have been resized so far... 2300 images have been resized so far... 2400 images have been resized so far... 2400 images have been resized so far... 2500 images have been resized so far... 2500 images have been resized so far... 2600 images have been resized so far... 2600 images have been resized so far... 2700 images have been resized so far... 2700 images have been resized so far... 2800 images have been resized so far... 2900 images have been resized so far... 2800 images have been resized so far... 3000 images have been resized so far... 2900 images have been resized so far... 3100 images have been resized so far... 3000 images have been resized so far... 3200 images have been resized so far... 3300 images have been resized so far... 3100 images have been resized so far... 3400 images have been resized so far... 3200 images have been resized so far... 3500 images have been resized so far... 3600 images have been resized so far... 3300 images have been resized so far... 3400 images have been resized so far... 3700 images have been resized so far... 3800 images have been resized so far... 3500 images have been resized so far... 3900 images have been resized so far... 4000 images have been resized so far... 3600 images have been resized so far... 4100 images have been resized so far... 3700 images have been resized so far... 4200 images have been resized so far... 4300 images have been resized so far... 3800 images have been resized so far... 4400 images have been resized so far... 3900 images have been resized so far... 4500 images have been resized so far... 4000 images have been resized so far... 4600 images have been resized so far... 4100 images have been resized so far... 4700 images have been resized so far... 4200 images have been resized so far... 4800 images have been resized so far... 4900 images have been resized so far... 4300 images have been resized so far... 5000 images have been resized so far... 5100 images have been resized so far... 4400 images have been resized so far... 5200 images have been resized so far... 4500 images have been resized so far... 5300 images have been resized so far... 5400 images have been resized so far... 4600 images have been resized so far... 5500 images have been resized so far... 4700 images have been resized so far... 5600 images have been resized so far... 5700 images have been resized so far... 4800 images have been resized so far... 5800 images have been resized so far... 4900 images have been resized so far... 5900 images have been resized so far... 6000 images have been resized so far... 5000 images have been resized so far... 6100 images have been resized so far... 5100 images have been resized so far... 6200 images have been resized so far... 6300 images have been resized so far... 5200 images have been resized so far... 6400 images have been resized so far... 5300 images have been resized so far... 6500 images have been resized so far... 6600 images have been resized so far... 5400 images have been resized so far... 6700 images have been resized so far... 5500 images have been resized so far... 6800 images have been resized so far... 6900 images have been resized so far... 5600 images have been resized so far... 7000 images have been resized so far... 7100 images have been resized so far... 5700 images have been resized so far... 7200 images have been resized so far... 5800 images have been resized so far... 7300 images have been resized so far... 7400 images have been resized so far... 5900 images have been resized so far... 7500 images have been resized so far... 6000 images have been resized so far... 7600 images have been resized so far... 6100 images have been resized so far... Processing Directory 2 of 6... Processing Directory: D:\Food Datasets\Small Training Dataset\Food\B-Medium\ 7700 images have been resized so far... 7800 images have been resized so far... 6200 images have been resized so far... 7900 images have been resized so far... 8000 images have been resized so far... 8100 images have been resized so far... 8200 images have been resized so far... 8300 images have been resized so far... 6300 images have been resized so far... 8400 images have been resized so far... 8500 images have been resized so far... 8600 images have been resized so far... 8700 images have been resized so far... 6400 images have been resized so far... 8800 images have been resized so far... 8900 images have been resized so far... 9000 images have been resized so far... 6500 images have been resized so far... 9100 images have been resized so far... 9200 images have been resized so far... 9300 images have been resized so far... 9400 images have been resized so far... 9500 images have been resized so far... 6600 images have been resized so far... 9600 images have been resized so far... 9700 images have been resized so far... 9800 images have been resized so far... 6700 images have been resized so far... 9900 images have been resized so far... 10000 images have been resized so far... 10100 images have been resized so far... 10200 images have been resized so far... 6800 images have been resized so far... 10300 images have been resized so far... 10400 images have been resized so far... 10500 images have been resized so far... 10600 images have been resized so far... 6900 images have been resized so far... 10700 images have been resized so far... 10800 images have been resized so far... 10900 images have been resized so far... 7000 images have been resized so far... 11000 images have been resized so far... 11100 images have been resized so far... 11200 images have been resized so far... 11300 images have been resized so far... 11400 images have been resized so far... 7100 images have been resized so far... 11500 images have been resized so far... 11600 images have been resized so far... 11700 images have been resized so far... 11800 images have been resized so far... 11900 images have been resized so far... 12000 images have been resized so far... 7200 images have been resized so far... 12100 images have been resized so far... 12200 images have been resized so far... 12300 images have been resized so far... 12400 images have been resized so far... 12500 images have been resized so far... ###Markdown 224p: (224x224) ###Code S7 = Thread(target = resize224x224x224, args = (SFCMIX, SFD_224x224x224,)) S8 = Thread(target = resize224x224x224, args = (SNFCMIX, SNFD_224x224x224,)) S7.start() S8.start() S7.join() S8.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code S9 = Thread(target = resize224x224x224x224, args = (SFC43, SFC169, SFD_224x224x224x224,)) S10 = Thread(target = resize224x224x224x224, args = (SNFC43, SNFC169, SNFD_224x224x224x224,)) S9.start() S10.start() S9.join() S10.join() ###Output _____no_output_____ ###Markdown 240p: (360x240) ###Code S11 = Thread(target = resize240p, args = (SFC43, SFC169, SFD_240p,)) S12 = Thread(target = resize240p, args = (SFC43, SFC169, SFD_240p,)) S11.start() S12.start() S11.join() S12.join() ###Output _____no_output_____ ###Markdown 360p: (480x360) ###Code S13 = Thread(target = resize360p, args = (SFCALL, SFD_360p,)) S14 = Thread(target = resize360p, args = (SNFCALL, SNFD_360p,)) S13.start() S14.start() S13.join() S14.join() ###Output _____no_output_____ ###Markdown 360p: (480x360) ###Code S15 = Thread(target = resize480p, args = (SFC43, SFD_480p,)) S16 = Thread(target = resize480p, args = (SNFC43, SNFD_480p,)) S15.start() S16.start() S15.join() S16.join() ###Output _____no_output_____ ###Markdown 360p: (640x360) ###Code S17 = Thread(target = resize640p, args = (SFC43, SFC169, SFD_640p,)) S18 = Thread(target = resize640p, args = (SNFC43, SNFC169, SNFD_640p,)) S17.start() S18.start() S17.join() S18.join() ###Output _____no_output_____ ###Markdown 480p: (640x480) ###Code S19 = Thread(target = resize640x480, args = (SFCMIX, SFD_640x480,)) S20 = Thread(target = resize640x480, args = (SNFCMIX, SNFD_640x480,)) S19.start() S20.start() S19.join() S20.join() ###Output _____no_output_____ ###Markdown Balanced Dataset Resizing 144p: (240x144) ###Code B1 = Thread(target=resize144p, args = (BFC169, BFD_144p,)) B2 = Thread(target=resize144p, args = (BNFC169, BNFD_144p,)) B1.start() B2.start() B1.join() B2.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code B3 = Thread(target=resize224p, args = (BFC11, BFC43, BFD_224p,)) B4 = Thread(target=resize224p, args = (BNFC11, BNFC43, BNFD_224p,)) B3.start() B4.start() B3.join() B4.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code B5 = Thread(target=resize224x224, args = (BFCALL, BFD_224x224,)) B6 = Thread(target=resize224x224, args = (BNFCALL, BNFD_224x224,)) B5.start() B6.start() B5.join() B6.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code B7 = Thread(target = resize224x224x224, args = (BFCMIX, BFD_224x224x224,)) B8 = Thread(target = resize224x224x224, args = (BNFCMIX, BNFD_224x224x224,)) B7.start() B8.start() B7.join() B8.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code B9 = Thread(target = resize224x224x224x224, args = (BFC43, BFC169, BFD_224x224x224x224,)) B10 = Thread(target = resize224x224x224x224, args = (BNFC43, BNFC169, BNFD_224x224x224x224,)) B9.start() B10.start() B9.join() B10.join() ###Output _____no_output_____ ###Markdown 240p: (360x240) ###Code B11 = Thread(target = resize240p, args = (BFC43, BFC169, BFD_240p,)) B12 = Thread(target = resize240p, args = (BNFC43, BNFC169, BNFD_240p,)) B11.start() B12.start() B11.join() B12.join() ###Output _____no_output_____ ###Markdown 360p: (480x360) ###Code B13 = Thread(target = resize360p, args = (BFCALL, BFD_360p,)) B14 = Thread(target = resize360p, args = (BNFCALL, BNFD_360p,)) B13.start() B14.start() B13.join() B14.join() ###Output _____no_output_____ ###Markdown 360p: (480x360) ###Code B15 = Thread(target = resize480p, args = (BFC43, BFD_480p,)) B16 = Thread(target = resize480p, args = (BNFC43, BNFD_480p,)) B15.start() B16.start() B15.join() B16.join() ###Output _____no_output_____ ###Markdown 360p: (640x360) ###Code B17 = Thread(target = resize640p, args = (BFC43, BFC169, BFD_640p,)) B18 = Thread(target = resize640p, args = (BNFC43, BNFC169, BNFD_640p,)) B17.start() B18.start() B17.join() B18.join() ###Output _____no_output_____ ###Markdown 480p: (640x480) ###Code B19 = Thread(target = resize640x480, args = (BFCMIX, BFD_640x480,)) B20 = Thread(target = resize640x480, args = (BNFCMIX, BNFD_640x480,)) B19.start() B20.start() B19.join() B20.join() ###Output _____no_output_____ ###Markdown Large Dataset Resizing 144p: (240x144) ###Code L1 = Thread(target = resize144p, args = (LFC169, LFD_144p,)) L2 = Thread(target = resize144p, args = (LNFC169, LNFD_144p,)) L1.start() L2.start() L1.join() L2.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code L3 = Thread(target = resize224p, args = (LFC11, LFC43, LFD_224p,)) L4 = Thread(target = resize224p, args = (LNFC11, LNFC43, LNFD_224p,)) L3.start() L4.start() L3.join() L4.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code L5 = Thread(target = resize224x224, args = (LFCALL, LFD_224x224,)) L6 = Thread(target = resize224x224, args = (LNFCALL, LNFD_224x224,)) L5.start() L6.start() L5.join() L6.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code L7 = Thread(target = resize224x224x224, args = (LFCMIX, LFD_224x224x224,)) L8 = Thread(target = resize224x224x224, args = (LNFCMIX, LNFD_224x224x224,)) L7.start() L8.start() L7.join() L8.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code L9 = Thread(target = resize224x224x224x224, args = (LFC43, LFC169, LFD_224x224x224x224,)) L10 = Thread(target = resize224x224x224x224, args = (LNFC43, LNFC169, LNFD_224x224x224x224,)) L9.start() L10.start() L9.join() L10.join() ###Output _____no_output_____ ###Markdown 240p: (360x240) ###Code L11 = Thread(target = resize240p, args = (LFC43, LFC169, LFD_240p,)) L12 = Thread(target = resize240p, args = (LNFC43, LNFC169, LNFD_240p,)) L11.start() L12.start() L11.join() L12.join() ###Output _____no_output_____ ###Markdown 360p: (480x360) ###Code L13 = Thread(target = resize360p, args = (LFCALL, LFD_360p,)) L14 = Thread(target = resize360p, args = (LNFCALL, LNFD_360p,)) L13.start() L14.start() L13.join() L14.join() ###Output _____no_output_____ ###Markdown 360p: (480x360) ###Code L15 = Thread(target = resize480p, args = (LFC43, LFD_480p,)) L16 = Thread(target = resize480p, args = (LNFC43, LNFD_480p,)) L15.start() L16.start() L15.join() L16.join() ###Output _____no_output_____ ###Markdown 360p: (640x360) ###Code L17 = Thread(target = resize640p, args = (LFC43, LFC169, LFD_640p,)) L18 = Thread(target = resize640p, args = (LNFC43, LNFC169, LNFD_640p,)) L17.start() L18.start() L17.join() L18.join() ###Output _____no_output_____ ###Markdown 480p: (640x480) ###Code L19 = Thread(target = resize640x480, args = (LFCMIX, LFD_640x480,)) L20 = Thread(target = resize640x480, args = (LNFCMIX, LNFD_640x480,)) L19.start() L20.start() L19.join() L20.join() ###Output _____no_output_____ ###Markdown Full Dataset Resizing 144p: (240x144) ###Code F1 = Thread(target = resize144p, args = (FFC169, FFD_144p,)) F2 = Thread(target = resize144p, args = (FFC169, FFD_144p,)) F1.start() F2.start() F1.join() F2.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code F3 = Thread(target = resize224p, args = (FFC11, FFC43, FFD_224p,)) F4 = Thread(target = resize224p, args = (FFC11, FFC43, FFD_224p,)) F3.start() F4.start() F3.join() F4.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code F5 = Thread(target = resize224x224, args = (FFCALL, FFD_224x224,)) F6 = Thread(target = resize224x224, args = (FNFCALL, FNFD_224x224,)) F5.start() F6.start() F5.join() F6.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code F7 = Thread(target = resize224x224x224, args = (FFCMIX, FFD_224x224x224,)) F8 = Thread(target = resize224x224x224, args = (FNFCMIX, FNFD_224x224x224,)) F7.start() F8.start() F7.join() F8.join() ###Output _____no_output_____ ###Markdown 224p: (224x224) ###Code F9 = Thread(target = resize224x224x224x224, args = (FFC43, FFC169, FFD_224x224x224x224,)) F10 = Thread(target = resize224x224x224x224, args = (FNFC43, FNFC169, FNFD_224x224x224x224,)) F9.start() F10.start() F9.join() F10.join() ###Output _____no_output_____ ###Markdown 240p: (360x240) ###Code F11 = Thread(target = resize240p, args = (FFC43, FFC169, FFD_240p,)) F12 = Thread(target = resize240p, args = (FNFC43, FNFC169, FNFD_240p,)) F11.start() F12.start() F11.join() F12.join() ###Output _____no_output_____ ###Markdown 360p: (480x360) ###Code F13 = Thread(target = resize360p, args = (FFCALL, FFD_360p,)) F14 = Thread(target = resize360p, args = (FNFCALL, FNFD_360p,)) F13.start() F14.start() F13.join() F14.join() ###Output _____no_output_____ ###Markdown 360p: (480x360) ###Code F15 = Thread(target = resize480p, args = (FFC43, FFD_480p,)) F16 = Thread(target = resize480p, args = (FNFC43, FNFD_480p,)) F15.start() F16.start() F15.join() F16.join() ###Output _____no_output_____ ###Markdown 360p: (640x360) ###Code F17 = Thread(target = resize640p, args = (FFC43, FFC169, FFD_640p,)) F18 = Thread(target = resize640p, args = (FNFC43, FNFC169, FNFD_640p,)) F17.start() F18.start() F17.join() F18.join() ###Output _____no_output_____ ###Markdown 480p: (640x480) ###Code F19 = Thread(target = resize640x480, args = (FFCMIX, FFD_640x480,)) F20 = Thread(target = resize640x480, args = (FNFCMIX, FNFD_640x480,)) F19.start() F20.start() F19.join() F20.join() L5 = Thread(target = resize224x224, args = (LFCALL, LFD_224x224,)) L6 = Thread(target = resize224x224, args = (LNFCALL, LNFD_224x224,)) F5 = Thread(target = resize224x224, args = (FFCALL, FFD_224x224,)) F6 = Thread(target = resize224x224, args = (FNFCALL, FNFD_224x224,)) F7 = Thread(target = resize224x224x224, args = (FFCMIX, FFD_224x224x224,)) F8 = Thread(target = resize224x224x224, args = (FNFCMIX, FNFD_224x224x224,)) F7.start() F8.start() L5.start() L6.start() F5.start() F6.start() F5.join() F6.join() F7.join() F8.join() L5.join() L6.join() ###Output _____no_output_____
nmt/Basque-English_eus-eng.ipynb
###Markdown 基于注意力的神经机器翻译 此笔记本训练一个将巴斯克语翻译为英语的序列到序列(sequence to sequence,简写为 seq2seq)模型。此例子难度较高,需要对序列到序列模型的知识有一定了解。训练完此笔记本中的模型后,你将能够输入一个巴斯克句子,例如 *"Golfak liluratu egiten nau."*,并返回其英语翻译 *"I love golf."*对于一个简单的例子来说,翻译质量令人满意。但是更有趣的可能是生成的注意力图:它显示在翻译过程中,输入句子的哪些部分受到了模型的注意。请注意:运行这个例子用一个 P100 GPU 需要花大约 10 分钟。 ###Code import tensorflow as tf import matplotlib.pyplot as plt import matplotlib.ticker as ticker from sklearn.model_selection import train_test_split import unicodedata import re import numpy as np import os import io import time ###Output _____no_output_____ ###Markdown 下载和准备数据集我们将使用 http://www.manythings.org/anki/ 提供的一个语言数据集。这个数据集包含如下格式的语言翻译对:```May I borrow this book? ¿Puedo tomar prestado este libro?```这个数据集中有很多种语言可供选择。我们将使用英语 - 巴斯克语数据集。为方便使用,我们在谷歌云上提供了此数据集的一份副本。但是你也可以自己下载副本。下载完数据集后,我们将采取下列步骤准备数据:1. 给每个句子添加一个 *开始* 和一个 *结束* 标记(token)。2. 删除特殊字符以清理句子。3. 创建一个单词索引和一个反向单词索引(即一个从单词映射至 id 的词典和一个从 id 映射至单词的词典)。4. 将每个句子填充(pad)到最大长度。 ###Code ''' # 下载文件 path_to_zip = tf.keras.utils.get_file( 'spa-eng.zip', origin='http://storage.googleapis.com/download.tensorflow.org/data/spa-eng.zip', extract=True) path_to_file = os.path.dirname(path_to_zip)+"/spa-eng/spa.txt" ''' path_to_file = "./lan/eus.txt" # 将 unicode 文件转换为 ascii def unicode_to_ascii(s): return ''.join(c for c in unicodedata.normalize('NFD', s) if unicodedata.category(c) != 'Mn') def preprocess_sentence(w): w = unicode_to_ascii(w.lower().strip()) # 在单词与跟在其后的标点符号之间插入一个空格 # 例如: "he is a boy." => "he is a boy ." # 参考:https://stackoverflow.com/questions/3645931/python-padding-punctuation-with-white-spaces-keeping-punctuation w = re.sub(r"([?.!,¿])", r" \1 ", w) w = re.sub(r'[" "]+', " ", w) # 除了 (a-z, A-Z, ".", "?", "!", ","),将所有字符替换为空格 w = re.sub(r"[^a-zA-Z?.!,¿]+", " ", w) w = w.rstrip().strip() # 给句子加上开始和结束标记 # 以便模型知道何时开始和结束预测 w = '<start> ' + w + ' <end>' return w en_sentence = u"May I borrow this book?" sp_sentence = u"¿Puedo tomar prestado este libro?" print(preprocess_sentence(en_sentence)) print(preprocess_sentence(sp_sentence).encode('utf-8')) # 1. 去除重音符号 # 2. 清理句子 # 3. 返回这样格式的单词对:[ENGLISH, SPANISH] def create_dataset(path, num_examples): lines = io.open(path, encoding='UTF-8').read().strip().split('\n') word_pairs = [[preprocess_sentence(w) for w in l.split('\t')] for l in lines[:num_examples]] return zip(*word_pairs) en, sp = create_dataset(path_to_file, None) print(en[-1]) print(sp[-1]) def max_length(tensor): return max(len(t) for t in tensor) def tokenize(lang): lang_tokenizer = tf.keras.preprocessing.text.Tokenizer( filters='') lang_tokenizer.fit_on_texts(lang) tensor = lang_tokenizer.texts_to_sequences(lang) tensor = tf.keras.preprocessing.sequence.pad_sequences(tensor, padding='post') return tensor, lang_tokenizer def load_dataset(path, num_examples=None): # 创建清理过的输入输出对 targ_lang, inp_lang = create_dataset(path, num_examples) input_tensor, inp_lang_tokenizer = tokenize(inp_lang) target_tensor, targ_lang_tokenizer = tokenize(targ_lang) return input_tensor, target_tensor, inp_lang_tokenizer, targ_lang_tokenizer ###Output _____no_output_____ ###Markdown 限制数据集的大小以加快实验速度(可选)在超过 10 万个句子的完整数据集上训练需要很长时间。为了更快地训练,我们可以将数据集的大小限制为 3 万个句子(当然,翻译质量也会随着数据的减少而降低): ###Code # 尝试实验不同大小的数据集 num_examples = 30000 input_tensor, target_tensor, inp_lang, targ_lang = load_dataset(path_to_file, num_examples) # 计算目标张量的最大长度 (max_length) max_length_targ, max_length_inp = max_length(target_tensor), max_length(input_tensor) # 采用 80 - 20 的比例切分训练集和验证集 input_tensor_train, input_tensor_val, target_tensor_train, target_tensor_val = train_test_split(input_tensor, target_tensor, test_size=0.2) # 显示长度 print(len(input_tensor_train), len(target_tensor_train), len(input_tensor_val), len(target_tensor_val)) def convert(lang, tensor): for t in tensor: if t!=0: print ("%d ----> %s" % (t, lang.index_word[t])) print ("Input Language; index to word mapping") convert(inp_lang, input_tensor_train[0]) print () print ("Target Language; index to word mapping") convert(targ_lang, target_tensor_train[0]) ###Output _____no_output_____ ###Markdown 创建一个 tf.data 数据集 ###Code BUFFER_SIZE = len(input_tensor_train) BATCH_SIZE = 64 steps_per_epoch = len(input_tensor_train)//BATCH_SIZE embedding_dim = 256 units = 1024 vocab_inp_size = len(inp_lang.word_index)+1 vocab_tar_size = len(targ_lang.word_index)+1 dataset = tf.data.Dataset.from_tensor_slices((input_tensor_train, target_tensor_train)).shuffle(BUFFER_SIZE) dataset = dataset.batch(BATCH_SIZE, drop_remainder=True) example_input_batch, example_target_batch = next(iter(dataset)) example_input_batch.shape, example_target_batch.shape ###Output _____no_output_____ ###Markdown 编写编码器 (encoder) 和解码器 (decoder) 模型实现一个基于注意力的编码器 - 解码器模型。关于这种模型,你可以阅读 TensorFlow 的 [神经机器翻译 (序列到序列) 教程](https://github.com/tensorflow/nmt)。本示例采用一组更新的 API。此笔记本实现了上述序列到序列教程中的 [注意力方程式](https://github.com/tensorflow/nmtbackground-on-the-attention-mechanism)。下图显示了注意力机制为每个输入单词分配一个权重,然后解码器将这个权重用于预测句子中的下一个单词。下图和公式是 [Luong 的论文](https://arxiv.org/abs/1508.04025v5)中注意力机制的一个例子。输入经过编码器模型,编码器模型为我们提供形状为 *(批大小,最大长度,隐藏层大小)* 的编码器输出和形状为 *(批大小,隐藏层大小)* 的编码器隐藏层状态。下面是所实现的方程式:本教程的编码器采用 [Bahdanau 注意力](https://arxiv.org/pdf/1409.0473.pdf)。在用简化形式编写之前,让我们先决定符号:* FC = 完全连接(密集)层* EO = 编码器输出* H = 隐藏层状态* X = 解码器输入以及伪代码:* `score = FC(tanh(FC(EO) + FC(H)))`* `attention weights = softmax(score, axis = 1)`。 Softmax 默认被应用于最后一个轴,但是这里我们想将它应用于 *第一个轴*, 因为分数 (score) 的形状是 *(批大小,最大长度,隐藏层大小)*。最大长度 (`max_length`) 是我们的输入的长度。因为我们想为每个输入分配一个权重,所以 softmax 应该用在这个轴上。* `context vector = sum(attention weights * EO, axis = 1)`。选择第一个轴的原因同上。* `embedding output` = 解码器输入 X 通过一个嵌入层。* `merged vector = concat(embedding output, context vector)`* 此合并后的向量随后被传送到 GRU每个步骤中所有向量的形状已在代码的注释中阐明: ###Code class Encoder(tf.keras.Model): def __init__(self, vocab_size, embedding_dim, enc_units, batch_sz): super(Encoder, self).__init__() self.batch_sz = batch_sz self.enc_units = enc_units self.embedding = tf.keras.layers.Embedding(vocab_size, embedding_dim) self.gru = tf.keras.layers.GRU(self.enc_units, return_sequences=True, return_state=True, recurrent_initializer='glorot_uniform') def call(self, x, hidden): x = self.embedding(x) output, state = self.gru(x, initial_state = hidden) return output, state def initialize_hidden_state(self): return tf.zeros((self.batch_sz, self.enc_units)) encoder = Encoder(vocab_inp_size, embedding_dim, units, BATCH_SIZE) # 样本输入 sample_hidden = encoder.initialize_hidden_state() sample_output, sample_hidden = encoder(example_input_batch, sample_hidden) print ('Encoder output shape: (batch size, sequence length, units) {}'.format(sample_output.shape)) print ('Encoder Hidden state shape: (batch size, units) {}'.format(sample_hidden.shape)) class BahdanauAttention(tf.keras.layers.Layer): def __init__(self, units): super(BahdanauAttention, self).__init__() self.W1 = tf.keras.layers.Dense(units) self.W2 = tf.keras.layers.Dense(units) self.V = tf.keras.layers.Dense(1) def call(self, query, values): # 隐藏层的形状 == (批大小,隐藏层大小) # hidden_with_time_axis 的形状 == (批大小,1,隐藏层大小) # 这样做是为了执行加法以计算分数 hidden_with_time_axis = tf.expand_dims(query, 1) # 分数的形状 == (批大小,最大长度,1) # 我们在最后一个轴上得到 1, 因为我们把分数应用于 self.V # 在应用 self.V 之前,张量的形状是(批大小,最大长度,单位) score = self.V(tf.nn.tanh( self.W1(values) + self.W2(hidden_with_time_axis))) # 注意力权重 (attention_weights) 的形状 == (批大小,最大长度,1) attention_weights = tf.nn.softmax(score, axis=1) # 上下文向量 (context_vector) 求和之后的形状 == (批大小,隐藏层大小) context_vector = attention_weights * values context_vector = tf.reduce_sum(context_vector, axis=1) return context_vector, attention_weights attention_layer = BahdanauAttention(10) attention_result, attention_weights = attention_layer(sample_hidden, sample_output) print("Attention result shape: (batch size, units) {}".format(attention_result.shape)) print("Attention weights shape: (batch_size, sequence_length, 1) {}".format(attention_weights.shape)) class Decoder(tf.keras.Model): def __init__(self, vocab_size, embedding_dim, dec_units, batch_sz): super(Decoder, self).__init__() self.batch_sz = batch_sz self.dec_units = dec_units self.embedding = tf.keras.layers.Embedding(vocab_size, embedding_dim) self.gru = tf.keras.layers.GRU(self.dec_units, return_sequences=True, return_state=True, recurrent_initializer='glorot_uniform') self.fc = tf.keras.layers.Dense(vocab_size) # 用于注意力 self.attention = BahdanauAttention(self.dec_units) def call(self, x, hidden, enc_output): # 编码器输出 (enc_output) 的形状 == (批大小,最大长度,隐藏层大小) context_vector, attention_weights = self.attention(hidden, enc_output) # x 在通过嵌入层后的形状 == (批大小,1,嵌入维度) x = self.embedding(x) # x 在拼接 (concatenation) 后的形状 == (批大小,1,嵌入维度 + 隐藏层大小) x = tf.concat([tf.expand_dims(context_vector, 1), x], axis=-1) # 将合并后的向量传送到 GRU output, state = self.gru(x) # 输出的形状 == (批大小 * 1,隐藏层大小) output = tf.reshape(output, (-1, output.shape[2])) # 输出的形状 == (批大小,vocab) x = self.fc(output) return x, state, attention_weights decoder = Decoder(vocab_tar_size, embedding_dim, units, BATCH_SIZE) sample_decoder_output, _, _ = decoder(tf.random.uniform((64, 1)), sample_hidden, sample_output) print ('Decoder output shape: (batch_size, vocab size) {}'.format(sample_decoder_output.shape)) ###Output _____no_output_____ ###Markdown 定义优化器和损失函数 ###Code optimizer = tf.keras.optimizers.Adam() loss_object = tf.keras.losses.SparseCategoricalCrossentropy( from_logits=True, reduction='none') def loss_function(real, pred): mask = tf.math.logical_not(tf.math.equal(real, 0)) loss_ = loss_object(real, pred) mask = tf.cast(mask, dtype=loss_.dtype) loss_ *= mask return tf.reduce_mean(loss_) ###Output _____no_output_____ ###Markdown 检查点(基于对象保存) ###Code checkpoint_dir = './training_checkpoints' checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt") checkpoint = tf.train.Checkpoint(optimizer=optimizer, encoder=encoder, decoder=decoder) ###Output _____no_output_____ ###Markdown 训练1. 将 *输入* 传送至 *编码器*,编码器返回 *编码器输出* 和 *编码器隐藏层状态*。2. 将编码器输出、编码器隐藏层状态和解码器输入(即 *开始标记*)传送至解码器。3. 解码器返回 *预测* 和 *解码器隐藏层状态*。4. 解码器隐藏层状态被传送回模型,预测被用于计算损失。5. 使用 *教师强制 (teacher forcing)* 决定解码器的下一个输入。6. *教师强制* 是将 *目标词* 作为 *下一个输入* 传送至解码器的技术。7. 最后一步是计算梯度,并将其应用于优化器和反向传播。 ###Code @tf.function def train_step(inp, targ, enc_hidden): loss = 0 with tf.GradientTape() as tape: enc_output, enc_hidden = encoder(inp, enc_hidden) dec_hidden = enc_hidden dec_input = tf.expand_dims([targ_lang.word_index['<start>']] * BATCH_SIZE, 1) # 教师强制 - 将目标词作为下一个输入 for t in range(1, targ.shape[1]): # 将编码器输出 (enc_output) 传送至解码器 predictions, dec_hidden, _ = decoder(dec_input, dec_hidden, enc_output) loss += loss_function(targ[:, t], predictions) # 使用教师强制 dec_input = tf.expand_dims(targ[:, t], 1) batch_loss = (loss / int(targ.shape[1])) variables = encoder.trainable_variables + decoder.trainable_variables gradients = tape.gradient(loss, variables) optimizer.apply_gradients(zip(gradients, variables)) return batch_loss EPOCHS = 10 for epoch in range(EPOCHS): start = time.time() enc_hidden = encoder.initialize_hidden_state() total_loss = 0 for (batch, (inp, targ)) in enumerate(dataset.take(steps_per_epoch)): batch_loss = train_step(inp, targ, enc_hidden) total_loss += batch_loss if batch % 100 == 0: print('Epoch {} Batch {} Loss {:.4f}'.format(epoch + 1, batch, batch_loss.numpy())) # 每 2 个周期(epoch),保存(检查点)一次模型 if (epoch + 1) % 2 == 0: checkpoint.save(file_prefix = checkpoint_prefix) print('Epoch {} Loss {:.4f}'.format(epoch + 1, total_loss / steps_per_epoch)) print('Time taken for 1 epoch {} sec\n'.format(time.time() - start)) ###Output _____no_output_____ ###Markdown 翻译* 评估函数类似于训练循环,不同之处在于在这里我们不使用 *教师强制*。每个时间步的解码器输入是其先前的预测、隐藏层状态和编码器输出。* 当模型预测 *结束标记* 时停止预测。* 存储 *每个时间步的注意力权重*。请注意:对于一个输入,编码器输出仅计算一次。 ###Code def evaluate(sentence): attention_plot = np.zeros((max_length_targ, max_length_inp)) sentence = preprocess_sentence(sentence) inputs = [inp_lang.word_index[i] for i in sentence.split(' ')] inputs = tf.keras.preprocessing.sequence.pad_sequences([inputs], maxlen=max_length_inp, padding='post') inputs = tf.convert_to_tensor(inputs) result = '' hidden = [tf.zeros((1, units))] enc_out, enc_hidden = encoder(inputs, hidden) dec_hidden = enc_hidden dec_input = tf.expand_dims([targ_lang.word_index['<start>']], 0) for t in range(max_length_targ): predictions, dec_hidden, attention_weights = decoder(dec_input, dec_hidden, enc_out) # 存储注意力权重以便后面制图 attention_weights = tf.reshape(attention_weights, (-1, )) attention_plot[t] = attention_weights.numpy() predicted_id = tf.argmax(predictions[0]).numpy() result += targ_lang.index_word[predicted_id] + ' ' if targ_lang.index_word[predicted_id] == '<end>': return result, sentence, attention_plot # 预测的 ID 被输送回模型 dec_input = tf.expand_dims([predicted_id], 0) return result, sentence, attention_plot # 注意力权重制图函数 def plot_attention(attention, sentence, predicted_sentence): fig = plt.figure(figsize=(10,10)) ax = fig.add_subplot(1, 1, 1) ax.matshow(attention, cmap='viridis') fontdict = {'fontsize': 14} ax.set_xticklabels([''] + sentence, fontdict=fontdict, rotation=90) ax.set_yticklabels([''] + predicted_sentence, fontdict=fontdict) ax.xaxis.set_major_locator(ticker.MultipleLocator(1)) ax.yaxis.set_major_locator(ticker.MultipleLocator(1)) plt.show() def translate(sentence): result, sentence, attention_plot = evaluate(sentence) print('Input: %s' % (sentence)) print('Predicted translation: {}'.format(result)) attention_plot = attention_plot[:len(result.split(' ')), :len(sentence.split(' '))] plot_attention(attention_plot, sentence.split(' '), result.split(' ')) ###Output _____no_output_____ ###Markdown 恢复最新的检查点并验证 ###Code # 恢复检查点目录 (checkpoint_dir) 中最新的检查点 checkpoint.restore(tf.train.latest_checkpoint(checkpoint_dir)) translate(u'hace mucho frio aqui.') translate(u'esta es mi vida.') translate(u'¿todavia estan en casa?') # 错误的翻译 translate(u'trata de averiguarlo.') ###Output _____no_output_____
S06Pandas/L07GroupBy.ipynb
###Markdown PANDAS - GROUPBY ###Code import numpy as np import pandas as pd data = {'Company':['GOOGL','GOOGL','MSFT','MSFT','FB','FB'], 'Person':['Sam','Charlie','Amy','Vanessa','Carl','Sarah'], 'Sales':[200,120,340,124,243,350]} df=pd.DataFrame(data) df bycomp = df.groupby('Company') # pass the column name and it returns groupby object bycomp # call aggregate functions on groupby object bycomp.mean() # it will automatically ignore non-numeric columns lik 'Person' bycomp.sum() bycomp.std() bycomp.min() bycomp.max() bycomp.std() bycomp.sum().loc['FB'] df.groupby('Company').count() # common way to use grouby df.groupby('Company').describe() # describe returns a dataframe. descriptive statistics df.groupby('Company').describe().transpose() # switch rows and columns df.groupby('Company').describe().transpose()['FB'] # select whichever company you are interested in ###Output _____no_output_____
tests/notebooks/lazy_pipeline.ipynb
###Markdown Test notebook lazy pipeline ###Code # Installed packages import pandas as pd # Testing from IPython.utils.capture import capture_output # Our package from pandas_profiling import ProfileReport from pandas_profiling.utils.cache import cache_file # Read the Titanic Dataset file_name = cache_file( "titanic.csv", "https://raw.githubusercontent.com/datasciencedojo/datasets/master/titanic.csv", ) df = pd.read_csv(file_name) # Generate the Profiling Report (with progress bar) with capture_output() as out: profile = ProfileReport(df, title="Titanic Dataset", progress_bar=True, lazy=False) assert all( any(v in s.data["text/plain"] for v in ["%|", "FloatProgress"]) for s in out.outputs ) assert len(out.outputs) == 2 # Generate the Profiling Report (without progress bar) with capture_output() as out: profile = df.profile_report( title="Titanic Dataset", html={"style": {"full_width": True}}, progress_bar=True, lazy=True, ) assert len(out.outputs) == 0 with capture_output() as out: _ = profile.to_html() assert all( any(v in s.data["text/plain"] for v in ["%|", "FloatProgress"]) for s in out.outputs ) assert len(out.outputs) == 3 with capture_output() as out: _ = profile.to_file("/tmp/tmpfile.html") assert "Export report to file" in out.outputs[0].data["text/plain"] assert len(out.outputs) == 1 # Test caching of the iterative building process with capture_output() as out: profile = ProfileReport(df, title="Titanic Dataset", progress_bar=True, lazy=True) assert len(out.outputs) == 0 with capture_output() as out: profile.description_set assert len(out.outputs) == 1 with capture_output() as out: profile.report assert len(out.outputs) == 1 with capture_output() as out: profile.html assert len(out.outputs) == 1 with capture_output() as out: profile.config.html.style.theme = "united" profile.invalidate_cache("rendering") profile.to_file("/tmp/cache1.html") assert len(out.outputs) == 2 with capture_output() as out: profile.config.pool_size = 1 profile.html assert len(out.outputs) == 0 with capture_output() as out: profile.config.pool_size = 0 profile.config.samples.head = 5 profile.config.samples.tail = 15 profile.invalidate_cache() profile.to_file("/tmp/cache2.html") assert len(out.outputs) == 4 ###Output _____no_output_____
mazes/maze-traversal.ipynb
###Markdown Maze Traversal- - -This notebook generates a maze and then populates it with three autonomous agents.Each agent leverages a unique strategy for trying to escape the maze. The Agents The Clueless WalkerThe clueless walker simply walks in a straight line until encountering a wall.Once a wall is hit, the agent tries to turn right. If it can't, then it tries to turn left. If it cannot, then it turns around. The Wall FollowerThe wall follower leverages the [wall following](https://en.wikipedia.org/wiki/Maze-solving_algorithmWall_follower) algorithm. The Path FinderThe path finder uses the [A* algorithm](https://en.wikipedia.org/wiki/A*_search_algorithm) to chart a path to the exit. It does not consider the entrance. **Resources**- [Smoothstep](https://smoothstep.io/)- Breadth First Search- Dijkstra's algorithm/Fast Marching Method for solving the Eikonal equation?- BFS is generalized as Disjkstra, which is generalized as Fast Marching, then as Ordered Upwind method, then as Anisotropic Fast Marching- Bellman-Ford Algorithm- Fast Marching Algorithm (FMM), Eikonal equation- [Maze Art](https://troika.uk.com/work/troika-labyrinth/)- [Lee Algorithm](https://en.wikipedia.org/wiki/Lee_algorithm)- [Procedural Content Generation: Mazes](http://pcg.wikidot.com/pcg-algorithm:maze)- [Wikipedia Maze Generation Algorithms](https://en.wikipedia.org/wiki/Maze_generation_algorithm)- [Smart Move: Intelligent Path Finding](https://www.gamedeveloper.com/programming/smart-move-intelligent-path-finding)- [Toward more Realistic Path Finding](https://www.gamedeveloper.com/programming/toward-more-realistic-pathfinding)- [AI Wisdom A* Articles](http://www.aiwisdom.com/ai_astar.html)- [Maze Solving Algorithm](https://en.wikipedia.org/wiki/Maze-solving_algorithm)- [Maze Routing Algorithm](https://en.wikipedia.org/wiki/Maze-solving_algorithmMaze-routing_algorithm)- [Shortest Path Algorithms](https://en.wikipedia.org/wiki/Maze-solving_algorithmShortest_path_algorithm)Game Programming Gems 1 (PDF)- Simple Implementation: Page 248- Optimized Implementation: Page 279 - Fuzzy Logic for Video Games: Page 313- A Neural Net Primer: Page 324 ###Code # Load python code. %load_ext autoreload %autoreload 2 # All imports from __future__ import annotations from IPython.display import display from ipycanvas import Canvas, hold_canvas, MultiCanvas import time from typing import List from generation.structures import Corner, Point from generation.maze import Maze from generation.generators.random_backtracer import generate_maze_walls from generation.npc import Agent from generation.renderers.units import AGENT_SIZE, ROOM_SIZE_WIDTH, ROOM_SIZE_HEIGHT from generation.renderers.wall_drawer import draw_maze from generation.renderers.agents_renderer import draw_agents from generation.direction import Direction from generation.walkers.clueless import clueless_walk from generation.walkers.wall_follower import wall_follower_walk from generation.walkers.a_star import find_path as find_path_with_a_star, build_path_walker # Rendering Functions def draw_path(path: List[Point], canvas: Canvas, line_color: str) -> None: """Renders a list of points as a solid line""" canvas_points = [] # Create an array of tuples for canvas to render in a single draw call. print(f'Path has {len(path)} steps') for location in path: # Find the upper left corner for the room. upper_left_corner = Corner(location.x * ROOM_SIZE_WIDTH, location.y * ROOM_SIZE_HEIGHT) # Find the midpoint of the room. horizontal_offset = ROOM_SIZE_WIDTH/2.0 vertical_offset = ROOM_SIZE_HEIGHT/2.0 midpoint = Point(upper_left_corner.x + horizontal_offset, upper_left_corner.y + vertical_offset) canvas_points.append((midpoint.x, midpoint.y)) canvas.stroke_style = line_color canvas.stroke_lines(canvas_points) def draw_legend(canvas: Canvas, frame) -> None: """Renders a legend for the maze.""" LINE_HEIGHT = 14 FIRST_LINE = 20 HORIZONTAL_OFFSET = 450 with hold_canvas(canvas): canvas.text_baseline = "top" canvas.clear() # Draw the frame count canvas.fill_style = 'black' canvas.fill_text(f'Frame: {frame}', HORIZONTAL_OFFSET, FIRST_LINE) # Draw Wall Walker Legend canvas.fill_style = 'blue' canvas.fill_rect(HORIZONTAL_OFFSET,FIRST_LINE+LINE_HEIGHT, AGENT_SIZE) canvas.fill_style = 'black' canvas.fill_text(f'- Wall Follower', HORIZONTAL_OFFSET + AGENT_SIZE + 5, FIRST_LINE+LINE_HEIGHT) # Draw Clueless Walker Legend canvas.fill_style = 'green' canvas.fill_rect(HORIZONTAL_OFFSET,FIRST_LINE+LINE_HEIGHT*2, AGENT_SIZE) canvas.fill_style = 'black' canvas.fill_text(f'- Clueless Walker', HORIZONTAL_OFFSET + AGENT_SIZE + 5, FIRST_LINE+LINE_HEIGHT*2) # Draw A* Walker Legend canvas.fill_style = 'yellow' canvas.fill_rect(HORIZONTAL_OFFSET,FIRST_LINE+LINE_HEIGHT*3, AGENT_SIZE) canvas.fill_style = 'black' canvas.fill_text(f'- Path Finder', HORIZONTAL_OFFSET + AGENT_SIZE + 5, FIRST_LINE+LINE_HEIGHT*3) # The Main Cell # 1000/125 = 8 FPS SLEEP_TIME_SEC:float = 0.125 # Generate a maze. maze: Maze = Maze(20, 20) generate_maze_walls(maze) # Create 4 layers of canvases. 0: Maze, 1: A* Path, 2: Agents, 3: HUD mc = MultiCanvas(n_canvases=4, width=800, height=400) display(mc) # Create NPCs with different strategies common_starting_point = Point(int(maze.width/2), int(maze.height/2)) wall_follower = Agent('blue') wall_follower.maze_strategy(wall_follower_walk) wall_follower.move_to(common_starting_point) wall_follower.face(Direction.SOUTH) random_walker = Agent('green') random_walker.maze_strategy(clueless_walk) random_walker.move_to(common_starting_point) random_walker.face(Direction.SOUTH) # Calculate a path using A* path_finder = Agent('yellow') path_finder.move_to(common_starting_point) path_finder.face(Direction.SOUTH) found_path, escape_path = find_path_with_a_star(path_finder, maze, maze.exit_cell.location) if not found_path: raise Exception('Failed to find a path.') path_walker = build_path_walker(escape_path) path_finder.maze_strategy(path_walker) agents = [wall_follower, random_walker, path_finder] # Initial Render time.sleep(2) draw_maze(maze, mc[0]) draw_path(escape_path, mc[1], 'red') draw_agents(agents, mc[2]) time.sleep(2) for frame in range(200): for agent in agents: agent.explore(maze) draw_agents(agents, mc[2]) draw_legend(mc[3],frame) time.sleep(SLEEP_TIME_SEC) ###Output Exit cell: Point(x=17, y=19)
aata/domains-sage.ipynb
###Markdown **Important:** to view this notebook properly you will need to execute the cell above, which assumes you have an Internet connection. It should already be selected, or place your cursor anywhere above to select. Then press the "Run" button in the menu bar above (the right-pointing arrowhead), or press Shift-Enter on your keyboard. $\newcommand{\identity}{\mathrm{id}}\newcommand{\notdivide}{\nmid}\newcommand{\notsubset}{\not\subset}\newcommand{\lcm}{\operatorname{lcm}}\newcommand{\gf}{\operatorname{GF}}\newcommand{\inn}{\operatorname{Inn}}\newcommand{\aut}{\operatorname{Aut}}\newcommand{\Hom}{\operatorname{Hom}}\newcommand{\cis}{\operatorname{cis}}\newcommand{\chr}{\operatorname{char}}\newcommand{\Null}{\operatorname{Null}}\newcommand{\lt}{<}\newcommand{\gt}{>}\newcommand{\amp}{&}$ Section18.5Sage¶ We have already seen some integral domains and unique factorizations in the previous two chapters. In addition to what we have already seen, Sage has support for some of the topics from this section, but the coverage is limited. Some functions will work for some rings and not others, while some functions are not yet part of Sage. So we will give some examples, but this is far from comprehensive. SubsectionField of Fractions Sage is frequently able to construct a field of fractions, or identify a certain field as the field of fractions. For example, the ring of integers and the field of rational numbers are both implemented in Sage, and the integers “know” that the rationals is it's field of fractions. ###Code Q = ZZ.fraction_field(); Q Q == QQ ###Output _____no_output_____ ###Markdown In other cases Sage will construct a fraction field, in the spirit of Lemma 18.3. So it is then possible to do basic calculations in the constructed field. ###Code R.<x> = ZZ[] P = R.fraction_field();P f = P((x^2+3)/(7*x+4)) g = P((4*x^2)/(3*x^2-5*x+4)) h = P((-2*x^3+4*x^2+3)/(x^2+1)) ((f+g)/h).numerator() ((f+g)/h).denominator() ###Output _____no_output_____ ###Markdown SubsectionPrime Subfields Corollary 18.7 says every field of characteristic $p$ has a subfield isomorphic to ${\mathbb Z}_p\text{.}$ For a finite field, the exact nature of this subfield is not a surprise, but Sage will allow us to extract it easily. ###Code F.<c> = FiniteField(3^5) F.characteristic() G = F.prime_subfield(); G G.list() ###Output _____no_output_____ ###Markdown More generally, the fields mentioned in the conclusions of Corollary 18.6 and Corollary 18.7 are known as the “prime subfield” of the ring containing them. Here is an example of the characteristic zero case. ###Code K.<y>=QuadraticField(-7); K K.prime_subfield() ###Output _____no_output_____ ###Markdown In a rough sense, every characteristic zero field contains a copy of the rational numbers (the fraction field of the integers), which can explain Sage's extensive support for rings and fields that extend the integers and the rationals. SubsectionIntegral Domains Sage can determine if some rings are integral domains and we can test products in them. However, notions of units, irreducibles or prime elements are not generally supported (outside of what we have seen for polynomials in the previous chapter). Worse, the construction below creates a ring within a larger field and so some functions (such as .is_unit()) pass through and give misleading results. This is because the construction below creates a ring known as an “order in a number field.” ###Code K.<x> = ZZ[sqrt(-3)]; K K.is_integral_domain() K.basis() x (1+x)*(1-x) == 2*2 ###Output _____no_output_____ ###Markdown The following is a bit misleading, since $4\text{,}$ as an element of ${\mathbb Z}[\sqrt{3}i]$ does not have a multiplicative inverse, though seemingly we can compute one. ###Code four = K(4) four.is_unit() four^-1 ###Output _____no_output_____ ###Markdown SubsectionPrincipal Ideals When a ring is a principal ideal domain, such as the integers, or polynomials over a field, Sage works well. Beyond that, support begins to weaken. ###Code T.<x>=ZZ[] T.is_integral_domain() J = T.ideal(5, x); J Q = T.quotient(J); Q J.is_principal() Q.is_field() ###Output _____no_output_____
notebooks/timing.ipynb
###Markdown Testing Order of Growth [Click here to run this chapter on Colab](https://colab.research.google.com/github/AllenDowney/DSIRP/blob/main/notebooks/timing.ipynb) Analysis of algorithms makes it possible to predict how run time will grow as the size of a problem increases.But this kind of analysis ignores leading coefficients and non-leading terms.So the behavior for small and medium problems might not be what the analysis predicts.To see how run time really behaves for a range of problem sizes, we can run the algorithm and measure.To do the measurement, we'll use the [times](https://docs.python.org/3/library/os.htmlos.times) function from the `os` module. ###Code import os def etime(): """Measures user and system time this process has used. Returns the sum of user and system time.""" user, sys, chuser, chsys, real = os.times() return user+sys start = etime() t = [x**2 for x in range(10000)] end = etime() end - start ###Output _____no_output_____ ###Markdown Exercise: Use `etime` to measure the computation time used by `sleep`. ###Code from time import sleep sleep(1) def time_func(func, n): """Run a function and return the elapsed time. func: function n: problem size, passed as an argument to func returns: user+sys time in seconds """ start = etime() func(n) end = etime() elapsed = end - start return elapsed ###Output _____no_output_____ ###Markdown One of the things that makes timing tricky is that many operations are too fast to measure accurately.`%timeit` handles this by running enough times get a precise estimate, even for things that run very fast.We'll handle it by running over a wide range of problem sizes, hoping to find sizes that run long enough to measure, but not more than a few seconds. The following function takes a size, `n`, creates an empty list, and calls `list.append` `n` times. ###Code def list_append(n): t = [] [t.append(x) for x in range(n)] ###Output _____no_output_____ ###Markdown `timeit` can time this function accurately. ###Code %timeit list_append(10000) ###Output _____no_output_____ ###Markdown But our `time_func` is not that smart. ###Code time_func(list_append, 10000) ###Output _____no_output_____ ###Markdown Exercise: Increase the number of iterations until the run time is measureable. List appendThe following function gradually increases `n` and records the total time. ###Code def run_timing_test(func, max_time=1): """Tests the given function with a range of values for n. func: function object returns: list of ns and a list of run times. """ ns = [] ts = [] for i in range(10, 28): n = 2**i t = time_func(func, n) print(n, t) if t > 0: ns.append(n) ts.append(t) if t > max_time: break return ns, ts ns, ts = run_timing_test(list_append) import matplotlib.pyplot as plt plt.plot(ns, ts, 'o-') plt.xlabel('Problem size (n)') plt.ylabel('Runtime (seconds)'); ###Output _____no_output_____ ###Markdown This one looks pretty linear, but it won't always be so clear.It will help to plot a straight line that goes through the last data point. ###Code def fit(ns, ts, exp=1.0, index=-1): """Fits a curve with the given exponent. ns: sequence of problem sizes ts: sequence of times exp: exponent of the fitted curve index: index of the element the fitted line should go through returns: sequence of fitted times """ # Use the element with the given index as a reference point, # and scale all other points accordingly. nref = ns[index] tref = ts[index] tfit = [] for n in ns: ratio = n / nref t = ratio**exp * tref tfit.append(t) return tfit ts_fit = fit(ns, ts) ts_fit ###Output _____no_output_____ ###Markdown The following function plots the actual results and the fitted line. ###Code def plot_timing_test(ns, ts, label='', color='C0', exp=1.0, scale='log'): """Plots data and a fitted curve. ns: sequence of n (problem size) ts: sequence of t (run time) label: string label for the data curve color: string color for the data curve exp: exponent (slope) for the fitted curve scale: string passed to xscale and yscale """ ts_fit = fit(ns, ts, exp) fit_label = 'exp = %d' % exp plt.plot(ns, ts_fit, label=fit_label, color='0.7', linestyle='dashed') plt.plot(ns, ts, 'o-', label=label, color=color, alpha=0.7) plt.xlabel('Problem size (n)') plt.ylabel('Runtime (seconds)') plt.xscale(scale) plt.yscale(scale) plt.legend() plot_timing_test(ns, ts, scale='linear') plt.title('list append'); ###Output _____no_output_____ ###Markdown From these results, what can we conclude about the order of growth of `list.append`? Before we go on, let's also look at the results on a log-log scale. ###Code plot_timing_test(ns, ts, scale='log') plt.title('list append'); ###Output _____no_output_____ ###Markdown Why might we prefer this scale? List popNow let's do the same for `list.pop` (which pops from the end of the list by default).Notice that we have to make the list before we pop things from it, so we will have to think about how to interpret the results. ###Code def list_pop(n): t = [] [t.append(x) for x in range(n)] [t.pop() for _ in range(n)] ns, ts = run_timing_test(list_pop) plot_timing_test(ns, ts, scale='log') plt.title('list pop'); ###Output _____no_output_____ ###Markdown What can we conclude?What about `pop(0)`, which pops from the beginning of the list?Note: You might have to adjust `exp` to make the fitted line fit. ###Code def list_pop0(n): t = [] [t.append(x) for x in range(n)] [t.pop(0) for _ in range(n)] ns, ts = run_timing_test(list_pop0) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('list pop(0)'); ###Output _____no_output_____ ###Markdown Searching a list`list.index` searches a list and returns the index of the first element that matches the target.What do we expect if we always search for the first element? ###Code def list_index0(n): t = [] [t.append(x) for x in range(n)] [t.index(0) for _ in range(n)] ns, ts = run_timing_test(list_index0) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('list index(0)'); ###Output _____no_output_____ ###Markdown What if we always search for the last element? ###Code def list_index_n(n): t = [] [t.append(x) for x in range(n)] [t.index(n-1) for _ in range(n)] ns, ts = run_timing_test(list_index_n) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('list index(n-1)'); ###Output _____no_output_____ ###Markdown Dictionary add ###Code def dict_add(n): d = {} [d.setdefault(x, x) for x in range(n)] ns, ts = run_timing_test(dict_add) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('dict add'); ###Output _____no_output_____ ###Markdown Dictionary lookup ###Code def dict_lookup(n): d = {} [d.setdefault(x, x) for x in range(n)] [d[x] for x in range(n)] ns, ts = run_timing_test(dict_lookup) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('dict lookup'); ###Output _____no_output_____ ###Markdown Testing Order of Growth *Data Structures and Information Retrieval in Python*Copyright 2021 Allen DowneyLicense: [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International](https://creativecommons.org/licenses/by-nc-sa/4.0/) [Click here to run this chapter on Colab](https://colab.research.google.com/github/AllenDowney/DSIRP/blob/main/chapters/timing.ipynb) Read the [documentation of os.times](https://docs.python.org/3/library/os.htmlos.times) ###Code import os def etime(): """Measures user and system time this process has used. Returns the sum of user and system time.""" user, sys, chuser, chsys, real = os.times() return user+sys start = etime() t = [x**2 for x in range(10000)] end = etime() end - start ###Output _____no_output_____ ###Markdown Exercise: Use `etime` to measure the computation time used by `sleep`. ###Code from time import sleep sleep(1) # Solution goes here def time_func(func, n): """Run a function and return the elapsed time. func: function n: problem size, passed as an argument to func returns: user+sys time in seconds """ start = etime() func(n) end = etime() elapsed = end - start return elapsed ###Output _____no_output_____ ###Markdown One of the things that makes timing tricky is that many operations are too fast to measure accurately.`%timeit` handles this by running enough times get a precise estimate, even for things that run very fast.We'll handle it by running over a wide range of problem sizes, hoping to sizes that run long enough to measure, but not more than a few seconds. The following function takes a size, `n`, creates an empty list, and calls `list.append` `n` times. ###Code def list_append(n): t = [] [t.append(x) for x in range(n)] ###Output _____no_output_____ ###Markdown `timeit` can time this function accurately. ###Code %timeit list_append(10000) ###Output _____no_output_____ ###Markdown But our `time_func` is not that smart. ###Code time_func(list_append, 10000) ###Output _____no_output_____ ###Markdown Exercise: Increase the number of iterations until the run time is measureable. ###Code # Solution goes here ###Output _____no_output_____ ###Markdown List appendThe following function gradually increases `n` and records the total time. ###Code def run_timing_test(func, max_time=1): """Tests the given function with a range of values for n. func: function object returns: list of ns and a list of run times. """ ns = [] ts = [] for i in range(10, 28): n = 2**i t = time_func(func, n) print(n, t) if t > 0: ns.append(n) ts.append(t) if t > max_time: break return ns, ts ns, ts = run_timing_test(list_append) import matplotlib.pyplot as plt plt.plot(ns, ts, 'o-') plt.xlabel('Problem size (n)') plt.ylabel('Runtime (seconds)'); ###Output _____no_output_____ ###Markdown This one looks pretty linear, but it won't always be so clear.It will help to plot a straight line that goes through the last data point. ###Code def fit(ns, ts, exp=1.0, index=-1): """Fits a curve with the given exponent. ns: sequence of problem sizes ts: sequence of times exp: exponent of the fitted curve index: index of the element the fitted line should go through returns: sequence of fitted times """ # Use the element with the given index as a reference point, # and scale all other points accordingly. nref = ns[index] tref = ts[index] tfit = [] for n in ns: ratio = n / nref t = ratio**exp * tref tfit.append(t) return tfit ts_fit = fit(ns, ts) ts_fit ###Output _____no_output_____ ###Markdown The following function plots the actual results and the fitted line. ###Code def plot_timing_test(ns, ts, label='', color='C0', exp=1.0, scale='log'): """Plots data and a fitted curve. ns: sequence of n (problem size) ts: sequence of t (run time) label: string label for the data curve color: string color for the data curve exp: exponent (slope) for the fitted curve scale: string passed to xscale and yscale """ ts_fit = fit(ns, ts, exp) fit_label = 'exp = %d' % exp plt.plot(ns, ts_fit, label=fit_label, color='0.7', linestyle='dashed') plt.plot(ns, ts, 'o-', label=label, color=color, alpha=0.7) plt.xlabel('Problem size (n)') plt.ylabel('Runtime (seconds)') plt.xscale(scale) plt.yscale(scale) plt.legend() plot_timing_test(ns, ts, scale='linear') plt.title('list append'); ###Output _____no_output_____ ###Markdown From these results, what can we conclude about the order of growth of `list.append`? Before we go on, let's also look at the results on a log-log scale. ###Code plot_timing_test(ns, ts, scale='log') plt.title('list append'); ###Output _____no_output_____ ###Markdown Why might we prefer this scale? List popNow let's do the same for `list.pop` (which pops from the end of the list by default).Notice that we have to make the list before we pop things from it, so we will have to think about how to interpret the results. ###Code def list_pop(n): t = [] [t.append(x) for x in range(n)] [t.pop() for _ in range(n)] ns, ts = run_timing_test(list_pop) plot_timing_test(ns, ts, scale='log') plt.title('list pop'); ###Output _____no_output_____ ###Markdown What can we conclude?What about `pop(0)`, which pops from the beginning of the list?Note: You might have to adjust `exp` to make the fitted line fit. ###Code def list_pop0(n): t = [] [t.append(x) for x in range(n)] [t.pop(0) for _ in range(n)] ns, ts = run_timing_test(list_pop0) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('list pop(0)'); ###Output _____no_output_____ ###Markdown Searching a list`list.index` searches a list and returns the index of the first element that matches the target.What do we expect if we always search for the first element? ###Code def list_index0(n): t = [] [t.append(x) for x in range(n)] [t.index(0) for _ in range(n)] ns, ts = run_timing_test(list_index0) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('list index(0)'); ###Output _____no_output_____ ###Markdown What if we always search for the last element? ###Code def list_index_n(n): t = [] [t.append(x) for x in range(n)] [t.index(n-1) for _ in range(n)] ns, ts = run_timing_test(list_index_n) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('list index(n-1)'); ###Output _____no_output_____ ###Markdown Dictionary add ###Code def dict_add(n): d = {} [d.setdefault(x, x) for x in range(n)] ns, ts = run_timing_test(dict_add) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('dict add'); ###Output _____no_output_____ ###Markdown Dictionary lookup ###Code def dict_lookup(n): d = {} [d.setdefault(x, x) for x in range(n)] [d[x] for x in range(n)] ns, ts = run_timing_test(dict_lookup) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('dict lookup'); ###Output _____no_output_____ ###Markdown Testing Order of Growth [Click here to run this chapter on Colab](https://colab.research.google.com/github/AllenDowney/DSIRP/blob/main/notebooks/timing.ipynb) Analysis of algorithms makes it possible to predict how run time will grow as the size of a problem increases.But this kind of analysis ignores leading coefficients and non-leading terms.So the behavior for small and medium problems might not be what the analysis predicts.To see how run time really behaves for a range of problem sizes, we can run the algorithm and measure.To do the measurement, we'll use the [times](https://docs.python.org/3/library/os.htmlos.times) function from the `os` module. ###Code import os def etime(): """Measures user and system time this process has used. Returns the sum of user and system time.""" user, sys, chuser, chsys, real = os.times() return user+sys start = etime() t = [x**2 for x in range(10000)] end = etime() end - start ###Output _____no_output_____ ###Markdown Exercise: Use `etime` to measure the computation time used by `sleep`. ###Code from time import sleep sleep(1) def time_func(func, n): """Run a function and return the elapsed time. func: function n: problem size, passed as an argument to func returns: user+sys time in seconds """ start = etime() func(n) end = etime() elapsed = end - start return elapsed ###Output _____no_output_____ ###Markdown One of the things that makes timing tricky is that many operations are too fast to measure accurately.`%timeit` handles this by running enough times get a precise estimate, even for things that run very fast.We'll handle it by running over a wide range of problem sizes, hoping to sizes that run long enough to measure, but not more than a few seconds. The following function takes a size, `n`, creates an empty list, and calls `list.append` `n` times. ###Code def list_append(n): t = [] [t.append(x) for x in range(n)] ###Output _____no_output_____ ###Markdown `timeit` can time this function accurately. ###Code %timeit list_append(10000) ###Output _____no_output_____ ###Markdown But our `time_func` is not that smart. ###Code time_func(list_append, 10000) ###Output _____no_output_____ ###Markdown Exercise: Increase the number of iterations until the run time is measureable. List appendThe following function gradually increases `n` and records the total time. ###Code def run_timing_test(func, max_time=1): """Tests the given function with a range of values for n. func: function object returns: list of ns and a list of run times. """ ns = [] ts = [] for i in range(10, 28): n = 2**i t = time_func(func, n) print(n, t) if t > 0: ns.append(n) ts.append(t) if t > max_time: break return ns, ts ns, ts = run_timing_test(list_append) import matplotlib.pyplot as plt plt.plot(ns, ts, 'o-') plt.xlabel('Problem size (n)') plt.ylabel('Runtime (seconds)'); ###Output _____no_output_____ ###Markdown This one looks pretty linear, but it won't always be so clear.It will help to plot a straight line that goes through the last data point. ###Code def fit(ns, ts, exp=1.0, index=-1): """Fits a curve with the given exponent. ns: sequence of problem sizes ts: sequence of times exp: exponent of the fitted curve index: index of the element the fitted line should go through returns: sequence of fitted times """ # Use the element with the given index as a reference point, # and scale all other points accordingly. nref = ns[index] tref = ts[index] tfit = [] for n in ns: ratio = n / nref t = ratio**exp * tref tfit.append(t) return tfit ts_fit = fit(ns, ts) ts_fit ###Output _____no_output_____ ###Markdown The following function plots the actual results and the fitted line. ###Code def plot_timing_test(ns, ts, label='', color='C0', exp=1.0, scale='log'): """Plots data and a fitted curve. ns: sequence of n (problem size) ts: sequence of t (run time) label: string label for the data curve color: string color for the data curve exp: exponent (slope) for the fitted curve scale: string passed to xscale and yscale """ ts_fit = fit(ns, ts, exp) fit_label = 'exp = %d' % exp plt.plot(ns, ts_fit, label=fit_label, color='0.7', linestyle='dashed') plt.plot(ns, ts, 'o-', label=label, color=color, alpha=0.7) plt.xlabel('Problem size (n)') plt.ylabel('Runtime (seconds)') plt.xscale(scale) plt.yscale(scale) plt.legend() plot_timing_test(ns, ts, scale='linear') plt.title('list append'); ###Output _____no_output_____ ###Markdown From these results, what can we conclude about the order of growth of `list.append`? Before we go on, let's also look at the results on a log-log scale. ###Code plot_timing_test(ns, ts, scale='log') plt.title('list append'); ###Output _____no_output_____ ###Markdown Why might we prefer this scale? List popNow let's do the same for `list.pop` (which pops from the end of the list by default).Notice that we have to make the list before we pop things from it, so we will have to think about how to interpret the results. ###Code def list_pop(n): t = [] [t.append(x) for x in range(n)] [t.pop() for _ in range(n)] ns, ts = run_timing_test(list_pop) plot_timing_test(ns, ts, scale='log') plt.title('list pop'); ###Output _____no_output_____ ###Markdown What can we conclude?What about `pop(0)`, which pops from the beginning of the list?Note: You might have to adjust `exp` to make the fitted line fit. ###Code def list_pop0(n): t = [] [t.append(x) for x in range(n)] [t.pop(0) for _ in range(n)] ns, ts = run_timing_test(list_pop0) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('list pop(0)'); ###Output _____no_output_____ ###Markdown Searching a list`list.index` searches a list and returns the index of the first element that matches the target.What do we expect if we always search for the first element? ###Code def list_index0(n): t = [] [t.append(x) for x in range(n)] [t.index(0) for _ in range(n)] ns, ts = run_timing_test(list_index0) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('list index(0)'); ###Output _____no_output_____ ###Markdown What if we always search for the last element? ###Code def list_index_n(n): t = [] [t.append(x) for x in range(n)] [t.index(n-1) for _ in range(n)] ns, ts = run_timing_test(list_index_n) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('list index(n-1)'); ###Output _____no_output_____ ###Markdown Dictionary add ###Code def dict_add(n): d = {} [d.setdefault(x, x) for x in range(n)] ns, ts = run_timing_test(dict_add) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('dict add'); ###Output _____no_output_____ ###Markdown Dictionary lookup ###Code def dict_lookup(n): d = {} [d.setdefault(x, x) for x in range(n)] [d[x] for x in range(n)] ns, ts = run_timing_test(dict_lookup) plot_timing_test(ns, ts, scale='log', exp=1) plt.title('dict lookup'); ###Output _____no_output_____
nCoV-counts-data.ipynb
###Markdown Data sources of 2019 nCoVAcknowledgement: - API from https://lab.isaaclin.cn/nCoV/- https://github.com/jianxu305/nCov2019_analysis/blob/master/src/demo.ipynb 实时获取需要的数据 ###Code def parse_time(data): df = pd.DataFrame(data) try: if np.any(['Time' in name for name in df.columns]): for time_name in df.columns[['Time' in name for name in df.columns]]: df[time_name] = pd.to_datetime(df.loc[:, time_name], unit='ms') except Exception as e: print(e) # add new column of the updating "Date" instead of concrete time if 'updateTime' in df.columns: df['updateDate'] = pd.Series([pd.to_datetime(item).date() for item in df['updateTime']]) elif 'pubDate' in df.columns: df['pubTime'] = pd.to_datetime(df.loc[:, 'pubDate'], unit='ms') df['pubDate'] = pd.Series([pd.to_datetime(item).date() for item in df['pubTime']]) return df def query_counts_data(category='area', archival=False, province='all'): ''' API for retrieving data from https://lab.isaaclin.cn/nCoV/. Parameters: Category (str): available options are 'overall', 'area'. Check the above website for more. archival (bool): whether retrieve archival time-series data. Default is False, only retrieve today's data. province (str): name of specific province. Use 'all' to get data from all provinces and countries. Notice: full name is required ("湖北省", instead of "湖北"). Returns: df (pandas.DataFrame): dataframe object. ''' import requests import pandas as pd assert isinstance(category, str), 'Input "catecory" must be a string!' url = 'https://lab.isaaclin.cn/nCoV/api/' + category url += '?latest={}'.format(int(not archival)) if province is not 'all': url += '&province=' + province req = requests.get(url) if req.status_code != 200 or req.json()['success'] is False: raise ValueError('The connection fails! Please check input arguments.') return False else: results = req.json()['results'] df = parse_time(results) return df def aggregate_Daily(df): frm_list = [] for key, frm in df.sort_values(['updateDate']).groupby(['provinceName', 'updateDate']): frm_list.append(frm.sort_values(['updateTime'])[-1:]) return pd.concat(frm_list).sort_values(['updateTime', 'provinceName']).loc[::-1] def parse_city(city_row): return pd.DataFrame(city_row.values[0]) area = query_counts_data(category='area', archival=True, province='湖北省') # not very slow area[:3] df = aggregate_Daily(area)[::-1] df[::-1] mask = df['updateDate'] == datetime.date(year=2020, month=2, day=5) city = parse_city(df[mask]['cities']) city plt.rcParams['font.size'] = 12.0 fig, [ax1, ax2] = plt.subplots(2, 1, figsize=(13, 10)) df.plot(y=['confirmedCount'], x='updateDate', style='-*', ax=ax1, grid=True, logy=False, color='black', marker='o') ax1.set_ylabel("Confirmed") df.plot(y=['deadCount', 'curedCount'], x='updateDate', style='-*', grid=True, ax=ax2, sharex=True) ax2.set_ylabel("Counts") plt.subplots_adjust(hspace=0.0) ax1.tick_params(direction='in') ax2.tick_params(direction='in') ###Output _____no_output_____ ###Markdown 利用已有的数据 ###Code df = pd.read_csv('https://github.com/BlankerL/DXY-2019-nCoV-Data/raw/master/csv/DXYArea.csv') df = parse_time(df) def aggregate_daily_csv(df): frm_list = [] for key, frm in df.sort_values(['updateDate']).groupby(['provinceName', 'cityName', 'updateDate']): frm_list.append(frm.sort_values(['updateTime'])[-1:]) return pd.concat(frm_list).sort_values(['updateTime', 'provinceName', 'cityName']).loc[::-1] jingmen = df[df['cityName'] == '荆门'] jingmen_daily = aggregate_daily_csv(jingmen) jingmen_daily.plot(y='city_confirmedCount', x='updateDate', style='-*', figsize=(10, 6), title='Jingmen') df = pd.read_csv('./data/DXYOverall.csv') df.columns df = df[['currentConfirmedCount', 'confirmedCount', 'updateTime']] df.iloc[719] ###Output _____no_output_____ ###Markdown 新闻报道?? ###Code def query_news_data(category='news', num='all', province='all'): ''' API for retrieving news data from https://lab.isaaclin.cn/nCoV/. Parameters: Category (str): available options are 'news'. Check the above website for more. archival (bool): whether retrieve archival time-series data. Default is False, only retrieve today's data. province (str): name of specific province. Use 'all' to get data from all provinces and countries. Notice: full name is required ("湖北省", instead of "湖北"). Returns: df (pandas.DataFrame): dataframe object. ''' import requests import pandas as pd assert isinstance(category, str), 'Input "catecory" must be a string!' url = 'https://lab.isaaclin.cn/nCoV/api/' + category url += '?num={}'.format(num) if province is not 'all': url += '&province=' + province req = requests.get(url) if req.status_code != 200 or req.json()['success'] is False: raise ValueError('The connection fails! Please check input arguments.') return False else: results = req.json()['results'] df = parse_time(results) return df news = query_news_data(category='news', num='all') news.groupby('provinceName').count().sort_values('title')[::-1] # today's news mask = (news['pubDate'] == datetime.date(year=2020, month=2, day=8)) news[mask] ###Output _____no_output_____
notebooks/Coverage Analysis.ipynb
###Markdown Coverage Analysis ###Code import sys import os sys.path.insert(0, os.path.abspath('../')) import numpy as np import matplotlib.pyplot as plt import astropy.units as u import poliastro import CtllDes from CtllDes.core import ctll, satellite ###Output WARNING: AstropyDeprecationWarning: The private astropy._erfa module has been made into its own package, pyerfa, which is a dependency of astropy and can be imported directly using "import erfa" [astropy._erfa] ###Markdown Building test satellite ###Code from poliastro.bodies import Earth sat = satellite.Sat.from_vectors([8000,0,0]*u.km, [0,5,2.5]*u.km/u.s, attractor=Earth) ###Output _____no_output_____ ###Markdown Add Coverage Instrument Camera symmetric FOV ###Code from CtllDes.core.instrument import Instrument, Camera cam = Camera(10,3) sat.update_instruments(cam,f=True) #check if Camera is a Coverage instrument, more on this later sat.cov_instruments ###Output _____no_output_____ ###Markdown Push Broom Instrument ###Code from CtllDes.core.instrument import PushBroom pixel_width = 7*1E-6*u.m n_pixels = 12288 sensor_width = n_pixels*pixel_width f_length = 0.42*u.m broom = PushBroom(f_length, sensor_width) sat.update_instruments(broom,f=True) #check if PushBroom is a Coverage instrument, more on this later sat.cov_instruments ###Output _____no_output_____ ###Markdown Defining targets ###Code #In order to do a coverage analysis you must have targets. The module targets is the one in charge of that. from CtllDes.targets.targets import Targets, Target from shapely.geometry import Point #simple target tgt = Target(0,0) #multiple targets tgts = Targets([Target(i,i) for i in range(0,180,10)],tag='linear targets') #define targets from country, administration level 0. tgts = Targets.from_country('Argentina') figc = tgts.plot() plt.title("Argentina, N=50") plt.grid() plt.xlabel("longitude [°]") plt.ylabel("latitude [°]") plt.show() #define targets from state name, administration level 1 tgts = Targets.from_state('Río Negro', N=100) figs = tgts.plot() plt.title("Río Negro, N=100") plt.xlabel("longitude [°]") plt.ylabel("latitude [°]") plt.grid() plt.show() #define single Target from city name bs_as = Target.from_city('Buenos Aires',country='AR') # less points for country targets tgts = Targets.from_country('Peru', N=6) ###Output _____no_output_____ ###Markdown Building CoveragesCoverages is the main container for Coverage analysis, it consist on Coverage (singular) objects. This objects are defined by covs, an array with length = T*3600*24/dt containing ones or zeroes depending if the target is on sight or not. Targets described earlier in this notebook T == Time of propagation analysis dt == Time interval of integration Merit figuresIf you want more information on the merit figures calculated for each target, I recommend reading the chapter 9 of O.C.D.M. from James R. Wertz. ###Code from CtllDes.requests.coverage import Coverages #Build Coverages from satellite and single target covs = Coverages.from_sat(sat, tgt, 10, dt=10, J2=True, drag=False) #transform coverages into dataframe covs.to_df() from CtllDes.requests.coverage import Coverages import time covs = Coverages.from_sat(sat, tgts, 10, dt=100, J2=True, drag=False) dfcov = covs.to_df() dfcov lons,lats = sat.ssps(10, dt=5, J2=True, drag=False) lons = lons*180/np.pi lats = lats*180/np.pi %matplotlib qt5 target_lons = [tgts.targets[i].lon + 180 for i in range(len(tgts.targets))] target_lats = [tgts.targets[i].lat for i in range(len(tgts.targets))] plt.figure(figsize=(10,10)) plt.scatter(lons,lats,c='red',s=1) plt.ylim(-90,90) plt.scatter(target_lons,target_lats,c='k', s=5) accum = dfcov['accumulated'].values accum = np.array([float(i) for i in accum]) fig = plt.figure(figsize=(10,10)) ax = fig.add_subplot(projection='3d') ax.plot_trisurf(target_lons, target_lats, accum, antialiased=False) ax.set_xlim(min(target_lons),max(target_lons)) ax.set_ylim(min(target_lats),max(target_lats)) #ax.scatter(lons,lats, np.zeros(len(lats)),s=1) ax.scatter(target_lons,target_lats, np.zeros(len(target_lons)),s=100,c='k') ###Output _____no_output_____ ###Markdown What is a Coverage Instrument?In order to be a Coverage Instrument first of all the object must be an Instrument. The coverage ability is defined by the interface of the library, i.e. a coverage method must be overwritten. See the example below. ###Code #first lets check out the coverage method requirements to be correcly overwritten. help(Instrument.coverage) #So if you want to build a taylor made instrument, first you must specify #the correct arguments to the coverage method. And most importantly, return #an Iterable containing ones or zeroes depending on the target being seen or not #at that r,v. class GodInstrument(Instrument): def __init__(self): super().__init__() def coverage(self, lons, lats, r, v, target, R): return [1 for _ in range(len(r))] #as you can see this is a silly example, God sees it all. from CtllDes.requests.coverage import symmetric_disk #What does exactly symmetric_disk do? help(symmetric_disk) #A more realistic Instrument that uses one of the few coverage methods already written. class DiskInstrument(Instrument): def __init__(self): super().__init__() self.FOV_min = 0.1*u.rad self.FOV_max = 0.2*u.rad def coverage(self, lons, lats, r, v, target, R): return coverage.symmetric_disk(self.FOV_min, self.FOV_max, lons, lats, r, v, target, R) ###Output Help on function symmetric_disk in module CtllDes.requests.coverage: symmetric_disk(FOV_min, FOV_max, lons, lats, r, target, R) coverage method. Disk of coverage centered on subsatellite point. Parameters ---------- FOV_min : ~astropy.units.quantity.Quantity minimum field of view in radians FOV_max : ~astropy.units.quantity.Quantity maximum field of view in radians * : default coverage parameters help(CtllDes.request.coverage.Instrument.coverage) for more info. ###Markdown So the intuition here you must get is that the interface is the coverage method, with the default parameters needed to compute coverage figures.If you have extra parameters that define the coverage, for example, a roll angle allowed, this must be included as a parameter of the specific Instrument child class. ###Code #Define your own parameters. from CtllDes.utils import trigsf class OnOffCamera(Instrument): def __init__(self, thresh): super().__init__() self.threshold = thresh self._FOV = np.pi*u.rad/8 @property def threshold(self): return self._threshold @threshold.setter def threshold(self,thresh): if not isinstance(thresh,u.Quantity): thresh = thresh * u.km elif thresh.unit.physical_type != 'length': raise ValueError("threshold must be length quantity") self._threshold = thresh.to(u.km) @property def FOV(self): return self._FOV def coverage(self,lons,lats,r,v,target,R): lams = trigsf.get_lam(r,self.FOV,R) angles = trigsf.get_angles(lons,lats,(target.x*u.deg).to(u.rad), (target.y*u.deg).to(u.rad)) radiis = np.sqrt(np.sum(r**2,axis=1)) cov = [] for lam,angle,radii in zip(lams,angles,radiis): if angle < lam: if self.threshold < radii < 2*R : cov.append(1) else: cov.append(0) else: cov.append(0) return cov onoffcam = OnOffCamera(300) sat.update_instruments(onoffcam,f=True) sat.instruments newcovs = Coverages.from_sat(sat, tgts, 5, dt=5, J2=True) ###Output /home/vancii/Documents/Instituto Balseiro/6to semestre/pi/CtllDes/venv/lib/python3.8/site-packages/astropy/units/quantity.py:477: RuntimeWarning: invalid value encountered in arcsin result = super().__array_ufunc__(function, method, *arrays, **kwargs) ###Markdown More spherical trigonometry calculations will be added (in development right now) to create coverage methods easier and faster. ###Code newcovs.to_df() constellation = ctll.Ctll.from_sats(sat) help(Camera) from CtllDes.requests.coverage import symmetric_with_roll #What does exactly symmetric_disk do? help(symmetric_with_roll) #A more realistic Instrument that uses one of the few coverage methods already written. class RollCamera(Instrument): def __init__(self,FOV,roll_angle): """Constructor for RollCamera. Parameters ---------- FOV : ~astropy.units.quantity.Quantity field of view, angle quantity roll_angle : ~astropy.units.quantity.Quantity maximum rolling angle """ super().__init__() self.FOV = FOV.to(u.rad) self.roll = roll_angle.to(u.rad) def coverage(self, lons, lats, r, v, target, R): return symmetric_with_roll(self.FOV, lons, lats, r, v, target, R, roll_angle = self.roll) roll_cam = RollCamera(0.15*u.rad,15*u.deg) sat.update_instruments(roll_cam, f=True) sat.instruments[0] roll_cov = Coverages.from_sat(sat,tgts,20, dt=5, drag=False, J2=True) rollcovdf = roll_cov.to_df() rollcovdf roll_accum = rollcovdf['accumulated'].to_numpy(dtype=float) roll_accum /= max(roll_accum) response_time = rollcovdf['response time'].to_numpy() response_time = 1/response_time response_time -= min(response_time) response_time /= max(response_time) roll_avg = rollcovdf['average time gap'].to_numpy(dtype=float) roll_avg = 1/roll_avg roll_avg -= min(roll_avg) roll_avg /= max(roll_avg) fig1 = plt.figure(figsize=(10,10)) ax1 = fig1.add_subplot(projection='3d') ax1.plot_trisurf(target_lons, target_lats, roll_accum, antialiased=False,cmap='viridis') ax1.set_title("Coverage over Perú") ax1.set_xlabel("longitude [°]") ax1.set_ylabel("latitude [°]") ax1.set_zlabel("Accumulated time of coverage, normalized") ax1.set_xlim(min(target_lons),max(target_lons)) ax1.set_ylim(min(target_lats),max(target_lats)) ax1.set_zlim(min(roll_accum),max(roll_accum)) ax1.scatter(target_lons,target_lats, np.zeros(len(target_lons)), s=100,c='k') fig2 = plt.figure(figsize=(10,10)) ax2 = fig2.add_subplot(projection='3d') ax2.plot_trisurf(target_lons, target_lats, response_time, antialiased=False,cmap='viridis') ax2.set_title("Response time over Perú") ax2.set_xlabel("longitude [°]") ax2.set_ylabel("latitude [°]") ax2.set_zlabel("1/tᵣ normalized") ax2.set_xlim(min(target_lons),max(target_lons)) ax2.set_ylim(min(target_lats),max(target_lats)) ax2.scatter(target_lons,target_lats, np.zeros(len(target_lons)), s=100,c='k') fig3 = plt.figure(figsize=(10,10)) ax3 = fig3.add_subplot(projection='3d') ax3.plot_trisurf(target_lons, target_lats, roll_avg, antialiased=False,cmap='viridis') ax3.set_title("Average time gap over Perú") ax3.set_xlabel("longitude [°]") ax3.set_ylabel("latitude [°]") ax3.set_zlabel("Averaget time gap normalized") ax3.set_xlim(min(target_lons),max(target_lons)) ax3.set_ylim(min(target_lats),max(target_lats)) ax3.scatter(target_lons,target_lats, np.zeros(len(target_lons)), s=100,c='k') ###Output _____no_output_____
data-science-tutorial-for-beginners.ipynb
###Markdown \*Contents in this Jupyter Notebook are from and can be found in [DATAI's kaggle kernel](https://www.kaggle.com/kanncaa1). The order of the content were changed for this workshop session.Data scientist need to have these skills:1. Basic Tools: Like python, R or SQL. You do not need to know everything. What you only need is to learn how to use **python**1. Basic Statistics: Like mean, median or standart deviation. If you know basic statistics, you can use **python** easily. 1. Data Munging: Working with messy and difficult data. Like a inconsistent date and string formatting. As you guess, **python** helps us.1. Data Visualization: Title is actually explanatory. We will visualize the data with **python** like matplot and seaborn libraries.1. Machine Learning: You do not need to understand math behind the machine learning technique. You only need is understanding basics of machine learning and learning how to implement it while using **python**.**Content:**1. [Introduction to Python:](1) 1. [Dictionaries ](3) 1. [Loop data structures](6) 1. [User defined function](8) 1. [Scope](9) 1. [Nested function](10) 1. [Default and flexible arguments](11) 1. [Lambda function](12) 1. [Anonymous function](13) 1. [Iterators](14) 1. [List comprehension](15)1. [Python Data Science Toolbox:](7) 1. [Pandas](4) 1. [Data types](23) 1. [Logic, control flow and filtering](5) 1. [Matplotlib](2)1. [Cleaning Data](16) 1. [Exploratory data analysis](18) 1. [Visual exploratory data analysis](19) 1. [Diagnose data for cleaning](17) 1. [Tidy data](20) 1. [Pivoting data](21) 1. [Concatenating data](22) 1. [Missing data and testing with assert](24)1. [Pandas Foundation](25) 1. [Review of pandas](26) 1. [Building data frames from scratch](27) 1. [Visual exploratory data analysis](28) 1. [Statistical explatory data analysis](29) 1. [Indexing pandas time series](30) 1. [Resampling pandas time series](31)1. [Manipulating Data Frames with Pandas](32) 1. [Indexing data frames](33) 1. [Slicing data frames](34) 1. [Filtering data frames](35) 1. [Transforming data frames](36) 1. [Index objects and labeled data](37) 1. [Hierarchical indexing](38) 1. [Pivoting data frames](39) 1. [Stacking and unstacking data frames](40) 1. [Melting data frames](41) 1. [Categoricals and groupby](42)1. Data Visualization 1. Seaborn: https://www.kaggle.com/kanncaa1/seaborn-for-beginners 1. Bokeh 1: https://www.kaggle.com/kanncaa1/interactive-bokeh-tutorial-part-1 1. Rare Visualization: https://www.kaggle.com/kanncaa1/rare-visualization-tools 1. Plotly: https://www.kaggle.com/kanncaa1/plotly-tutorial-for-beginners1. Machine Learning 1. https://www.kaggle.com/kanncaa1/machine-learning-tutorial-for-beginners/1. Deep Learning 1. https://www.kaggle.com/kanncaa1/deep-learning-tutorial-for-beginners1. Time Series Prediction 1. https://www.kaggle.com/kanncaa1/time-series-prediction-tutorial-with-eda1. Statistic 1. https://www.kaggle.com/kanncaa1/basic-statistic-tutorial-for-beginners1. Deep Learning with Pytorch 1. Artificial Neural Network: https://www.kaggle.com/kanncaa1/pytorch-tutorial-for-deep-learning-lovers 1. Convolutional Neural Network: https://www.kaggle.com/kanncaa1/pytorch-tutorial-for-deep-learning-lovers 1. Recurrent Neural Network: https://www.kaggle.com/kanncaa1/recurrent-neural-network-with-pytorch 1. INTRODUCTION TO PYTHON * Basic dictionary features* While and for loops* User defined function * Scope* Nested function* Default and flexible arguments* Lambda function* Anonymous function* Iterators* List comprehension DICTIONARYWhy we need dictionary?* It has 'key' and 'value'* Faster than listsWhat is key and value. Example:* dictionary = {'spain' : 'madrid'}* Key is spain.* Values is madrid.**It's that easy.**Lets practice some other properties like keys(), values(), update, add, check, remove key, remove all entries and remove dicrionary. ###Code #create dictionary and look its keys and values dictionary = {'spain' : 'madrid','usa' : 'vegas'} print(dictionary.keys()) print(dictionary.values()) # Keys have to be immutable objects like string, boolean, float, integer or tubles # List is not immutable # Keys are unique dictionary['spain'] = "barcelona" # update existing entry print(dictionary) dictionary['france'] = "paris" # Add new entry print(dictionary) del dictionary['spain'] # remove entry with key 'spain' print(dictionary) print('france' in dictionary) # check include or not dictionary.clear() # remove all entries in dict print(dictionary) # In order to run all code you need to take comment this line # del dictionary # delete entire dictionary print(dictionary) # it gives error because dictionary is deleted ###Output _____no_output_____ ###Markdown WHILE and FOR LOOPSWe will learn most basic while and for loops ###Code # Stay in loop if condition( i is not equal 5) is true i = 0 while i != 5 : print('i is: ',i) i +=1 print(i,' is equal to 5') # Stay in loop if condition( i is not equal 5) is true lis = [1,2,3,4,5] for i in lis: print('i is: ',i) print('') # Enumerate index and value of list # index : value = 0:1, 1:2, 2:3, 3:4, 4:5 for index, value in enumerate(lis): print(index," : ",value) print('') # For dictionaries # We can use for loop to achive key and value of dictionary. We learnt key and value at dictionary part. dictionary = {'spain':'madrid','france':'paris'} for key,value in dictionary.items(): print(key," : ",value) print('') ###Output _____no_output_____ ###Markdown USER DEFINED FUNCTIONWhat we need to know about functions:* docstrings: documentation for functions. Example:for f(): """This is docstring for documentation of function f"""* tuble: sequence of immutable python objects. cant modify valuestuble uses paranthesis like tuble = (1,2,3)unpack tuble into several variables like a,b,c = tuble ###Code # example of what we learn above def tuble_ex(): """ return defined t tuble""" t = (1,2,3) return t a,b,c = tuble_ex() print(a,b,c) ###Output _____no_output_____ ###Markdown SCOPEWhat we need to know about scope:* global: defined main body in script* local: defined in a function* built in scope: names in predefined built in scope module such as print, lenLets make some basic examples ###Code # guess print what x = 2 def f(): x = 3 return x print(x) # x = 2 global scope print(f()) # x = 3 local scope # What if there is no local scope x = 5 def f(): y = 2*x # there is no local scope x return y print(f()) # it uses global scope x # First local scopesearched, then global scope searched, if two of them cannot be found lastly built in scope searched. # How can we learn what is built in scope import builtins dir(builtins) ###Output _____no_output_____ ###Markdown NESTED FUNCTION* function inside function.* There is a LEGB rule that is search local scope, enclosing function, global and built in scopes, respectively. ###Code #nested function def square(): """ return square of value """ def add(): """ add two local variable """ x = 2 y = 3 z = x + y return z return add()**2 print(square()) ###Output _____no_output_____ ###Markdown DEFAULT and FLEXIBLE ARGUMENTS* Default argument example: def f(a, b=1): """ b = 1 is default argument"""* Flexible argument example: def f(*args): """ *args can be one or more"""def f(** kwargs) """ **kwargs is a dictionary""" lets write some code to practice ###Code # default arguments def f(a, b = 1, c = 2): y = a + b + c return y print(f(5)) # what if we want to change default arguments print(f(5,4,3)) # flexible arguments *args def f(*args): for i in args: print(i) f(1) print("") f(1,2,3,4) # flexible arguments **kwargs that is dictionary def f(**kwargs): """ print key and value of dictionary""" for key, value in kwargs.items(): # If you do not understand this part turn for loop part and look at dictionary in for loop print(key, " ", value) f(country = 'spain', capital = 'madrid', population = 123456) ###Output _____no_output_____ ###Markdown LAMBDA FUNCTIONFaster way of writing function ###Code # lambda function square = lambda x: x**2 # where x is name of argument print(square(4)) tot = lambda x,y,z: x+y+z # where x,y,z are names of arguments print(tot(1,2,3)) ###Output _____no_output_____ ###Markdown ANONYMOUS FUNCTİONLike lambda function but it can take more than one arguments.* map(func,seq) : applies a function to all the items in a list ###Code number_list = [1,2,3] y = map(lambda x:x**2,number_list) print(list(y)) ###Output _____no_output_____ ###Markdown ITERATORS* iterable is an object that can return an iterator* iterable: an object with an associated iter() method example: list, strings and dictionaries* iterator: produces next value with next() method ###Code # iteration example name = "ronaldo" it = iter(name) print(next(it)) # print next iteration print(*it) # print remaining iteration ###Output _____no_output_____ ###Markdown zip(): zip lists ###Code # zip example list1 = [1,2,3,4] list2 = [5,6,7,8] z = zip(list1,list2) print(z) z_list = list(z) print(z_list) un_zip = zip(*z_list) un_list1,un_list2 = list(un_zip) # unzip returns tuble print(un_list1) print(un_list2) print(type(un_list2)) ###Output _____no_output_____ ###Markdown LIST COMPREHENSİON**One of the most important topic of this kernel**We use list comprehension for data analysis often. list comprehension: collapse for loops for building lists into a single lineEx: num1 = [1,2,3] and we want to make it num2 = [2,3,4]. This can be done with for loop. However it is unnecessarily long. We can make it one line code that is list comprehension. ###Code # Example of list comprehension num1 = [1,2,3] num2 = [i + 1 for i in num1 ] print(num2) ###Output _____no_output_____ ###Markdown [i + 1 for i in num1 ]: list of comprehension i +1: list comprehension syntax for i in num1: for loop syntax i: iterator num1: iterable object ###Code # Conditionals on iterable num1 = [5,10,15] num2 = [i**2 if i == 10 else i-5 if i < 7 else i+5 for i in num1] print(num2) ###Output _____no_output_____ ###Markdown 2. PYTHON DATA SCIENCE TOOLBOX In this part, you learn:* how to import csv file* data types* basic pandas features like filtering that is actually something always used and main for being data scientist* plotting line,scatter and histogram In programming, a [module](https://www.learnpython.org/en/Modules_and_Packages) is a piece of software that has a specific functionality. For example, when building a ping pong game, one module would be responsible for the game logic, andanother module would be responsible for drawing the game on the screen. Each module is a different file, which can be edited separately.Packages are namespaces which contain multiple packages and modules themselves. ###Code import numpy as np # linear algebra import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv) import matplotlib.pyplot as plt import seaborn as sns # visualization tool # Input data files are available in the "../input/" directory. # For example, running this (by clicking run or pressing Shift+Enter) will list the files in the input directory from subprocess import check_output print(check_output(["ls", "../input"]).decode("utf8")) # Any results you write to the current directory are saved as output. ###Output _____no_output_____ ###Markdown PANDAS* CSV: comma - separated values We can use head, tail, columns, shape and info methods to overview the data ###Code data = pd.read_csv('../input/pokemon.csv') series = data['Defense'] # data['Defense'] = series print(type(series)) data_frame = data[['Defense']] # data[['Defense']] = data frame print(type(data_frame)) # head shows first 5 rows data.head() # tail shows last 5 rows data.tail() # columns gives column names of features data.columns # shape gives number of rows and columns in a tuble data.shape # info gives data type like dataframe, number of sample or row, number of feature or column, feature types and memory usage data.info() # We can also use describe to see the basic statistics about the data data.describe() ###Output _____no_output_____ ###Markdown DATA TYPESThere are 5 basic data types: object(string),booleab, integer, float and categorical. We can make conversion data types like from str to categorical or from int to float Why is category important: * make dataframe smaller in memory * can be utilized for anlaysis especially for sklearn(we will learn later) ###Code data.dtypes # lets convert object(str) to categorical and int to float. data['Type 1'] = data['Type 1'].astype('category') data['Speed'] = data['Speed'].astype('float') # As you can see Type 1 is converted from object to categorical # And Speed ,s converted from int to float data.dtypes ###Output _____no_output_____ ###Markdown Before continue with pandas, we need to learn **logic, control flow** and **filtering.**Comparison operator: ==, , <=Boolean operators: and, or ,not Filtering pandas ###Code # Comparison operator print(3 > 2) print(3!=2) # Boolean operators print(True and False) print(True or False) # 1 - Filtering Pandas data frame x = data['Defense']>200 # There are only 3 pokemons who have higher defense value than 200 data[x] # 2 - Filtering pandas with logical_and # There are only 2 pokemons who have higher defence value than 2oo and higher attack value than 100 data[np.logical_and(data['Defense']>200, data['Attack']>100 )] # This is also same with previous code line. Therefore we can also use '&' for filtering. data[(data['Defense']>200) & (data['Attack']>100)] ###Output _____no_output_____ ###Markdown We can use what we learned so far to do some calculatation and manipulation on the data. ###Code # lets return pokemon csv and make one more list comprehension example # lets classify pokemons whether they have high or low speed. Our threshold is average speed. threshold = sum(data.Speed)/len(data.Speed) data["speed_level"] = ["high" if i > threshold else "low" for i in data.Speed] data.loc[:10,["speed_level","Speed"]] # we will learn loc more detailed later ###Output _____no_output_____ ###Markdown MATPLOTLIBMatplot is a python library that help us to plot data. The easiest and basic plots are line, scatter and histogram plots.* Line plot is better when x axis is time.* Scatter is better when there is correlation between two variables* Histogram is better when we need to see distribution of numerical data.* Customization: Colors,labels,thickness of line, title, opacity, grid, figsize, ticks of axis and linestyle ###Code data.corr() ###Output _____no_output_____ ###Markdown See `matplotlib` documentation for using [`.subplots()`](https://matplotlib.org/api/_as_gen/matplotlib.pyplot.subplots.html) ###Code #correlation map f,ax = plt.subplots(figsize=(18, 18)) # create a canvas sns.heatmap(data.corr(), annot=True, linewidths=.5, fmt= '.1f',ax=ax) # plt.show() # Line Plot # color = color, label = label, linewidth = width of line, alpha = opacity, grid = grid, linestyle = sytle of line data['Speed'].plot(kind = 'line', color = 'g',label = 'Speed',linewidth=1,alpha = 0.5,grid = True,linestyle = ':') data['Defense'].plot(color = 'r',label = 'Defense',linewidth=1, alpha = 0.5,grid = True,linestyle = '-.') plt.legend(loc='upper right') # legend = puts label into plot plt.xlabel('x axis') # label = name of label plt.ylabel('y axis') plt.title('Line Plot') # title = title of plot plt.show() # Scatter Plot # x = attack, y = defense data.plot(kind='scatter', x='Attack', y='Defense',alpha = 0.5,color = 'red') plt.xlabel('Attack') # label = name of label plt.ylabel('Defence') plt.title('Attack Defense Scatter Plot') # title = title of plot # Histogram # bins = number of bar in figure data['Speed'].plot(kind = 'hist',bins = 50,figsize = (12,12)) plt.show() # clf() = cleans it up again you can start a fresh data['Speed'].plot(kind = 'hist',bins = 50) plt.clf() # We cannot see plot due to clf() ###Output _____no_output_____ ###Markdown 3. EXPLORING AND CLEANING DATA In this part, you will learn:* Exploratory data analysis* Visual exploratory data analysis* Diagnose data for cleaning* Tidy data* Pivoting data* Concatenating data* Missing data and testing with assert EXPLORATORY DATA ANALYSISvalue_counts(): Frequency countsoutliers: the value that is considerably higher or lower from rest of the data* Lets say value at 75% is Q3 and value at 25% is Q1. * Outlier are smaller than Q1 - 1.5(Q3-Q1) and bigger than Q3 + 1.5(Q3-Q1). (Q3-Q1) = IQRWe will use describe() method. Describe method includes:* count: number of entries* mean: average of entries* std: standart deviation* min: minimum entry* 25%: first quantile* 50%: median or second quantile* 75%: third quantile* max: maximum entry What is quantile?* 1,4,5,6,8,9,11,12,13,14,15,16,17* The median is the number that is in **middle** of the sequence. In this case it would be 11.* The lower quartile is the median in between the smallest number and the median i.e. in between 1 and 11, which is 6.* The upper quartile, you find the median between the median and the largest number i.e. between 11 and 17, which will be 14 according to the question above. ###Code # For example lets look frequency of pokemom types print(data['Type 1'].value_counts(dropna =False)) # if there are nan values that also be counted # As it can be seen below there are 112 water pokemon or 70 grass pokemon # For example max HP is 255 or min defense is 5 data.describe() #ignore null entries ###Output _____no_output_____ ###Markdown VISUAL EXPLORATORY DATA ANALYSIS* Box plots: visualize basic statistics like outliers, min/max or quantiles ###Code # For example: compare attack of pokemons that are legendary or not # Black line at top is max # Blue line at top is 75% # Red line is median (50%) # Blue line at bottom is 25% # Black line at bottom is min # There are no outliers data.boxplot(column='Attack',by = 'Legendary') ###Output _____no_output_____ ###Markdown DIAGNOSE DATA for CLEANINGUnclean data:* Column name inconsistency like upper-lower case letter or space between words* missing data* different language* extream values* ...Another short [tutorial](https://realpython.com/python-data-cleaning-numpy-pandas/) you can read about cleaning data in python. TIDY DATAWe tidy data with melt().Describing melt is confusing. Therefore lets make example to understand it. ###Code # Firstly I create new data from pokemons data to explain melt nore easily. data_new = data.head() # I only take 5 rows into new data data_new # lets melt # id_vars = what we do not wish to melt # value_vars = what we want to melt melted = pd.melt(frame=data_new,id_vars = 'Name', value_vars= ['Attack','Defense']) melted ###Output _____no_output_____ ###Markdown PIVOTING DATAReverse of melting. ###Code # Index is name # I want to make that columns are variable # Finally values in columns are value melted.pivot(index = 'Name', columns = 'variable',values='value') ###Output _____no_output_____ ###Markdown CONCATENATING DATAWe can concatenate two dataframe ###Code # Firstly lets create 2 data frame data1 = data.head() data2= data.tail() conc_data_row = pd.concat([data1,data2],axis =0,ignore_index =True) # axis = 0 : adds dataframes in row conc_data_row data1 = data['Attack'].head() data2= data['Defense'].head() conc_data_col = pd.concat([data1,data2],axis =1) # axis = 0 : adds dataframes in row conc_data_col ###Output _____no_output_____ ###Markdown MISSING DATA and TESTING WITH ASSERTIf we encounter with missing data, what we can do:* leave as is* drop them with dropna()* fill missing value with fillna()* fill missing values with test statistics like meanAssert statement: check that you can turn on or turn off when you are done with your testing of the program ###Code # Lets look at does pokemon data have nan value # As you can see there are 800 entries. However Type 2 has 414 non-null object so it has 386 null object. data.info() # Lets chech Type 2 data["Type 2"].value_counts(dropna =False) # As you can see, there are 386 NAN value # Lets drop nan values data1=data # also we will use data to fill missing value so I assign it to data1 variable data1["Type 2"].dropna(inplace = True) # inplace = True means we do not assign it to new variable. Changes automatically assigned to data # So does it work ? # Lets check with assert statement # Assert statement: assert 1==1 # return nothing because it is true # In order to run all code, we need to make this line comment # assert 1==2 # return error because it is false assert data['Type 2'].notnull().all() # returns nothing because we drop nan values data["Type 2"].fillna('empty',inplace = True) assert data['Type 2'].notnull().all() # returns nothing because we do not have nan values # # With assert statement we can check a lot of thing. For example # assert data.columns[1] == 'Name' # assert data.Speed.dtypes == np.int ###Output _____no_output_____ ###Markdown 4. PANDAS FOUNDATION REVİEW of PANDASAs you notice, I do not give all idea in a same time. Although, we learn some basics of pandas, we will go deeper in pandas.* single column = series* NaN = not a number* dataframe.values = numpy BUILDING DATA FRAMES FROM SCRATCH* We can build data frames from csv as we did earlier.* Also we can build dataframe from dictionaries * zip() method: This function returns a list of tuples, where the i-th tuple contains the i-th element from each of the argument sequences or iterables.* Adding new column* Broadcasting: Create new column and assign a value to entire column ###Code # data frames from dictionary country = ["Spain","France"] population = ["11","12"] list_label = ["country","population"] list_col = [country,population] zipped = list(zip(list_label,list_col)) data_dict = dict(zipped) df = pd.DataFrame(data_dict) df # Add new columns df["capital"] = ["madrid","paris"] df # Broadcasting df["income"] = 0 #Broadcasting entire column df ###Output _____no_output_____ ###Markdown VISUAL EXPLORATORY DATA ANALYSIS* Plot* Subplot* Histogram: * bins: number of bins * range(tuble): min and max values of bins * normed(boolean): normalize or not * cumulative(boolean): compute cumulative distribution ###Code # Plotting all data data1 = data.loc[:,["Attack","Defense","Speed"]] data1.plot() # it is confusing # subplots data1.plot(subplots = True) plt.show() # scatter plot data1.plot(kind = "scatter",x="Attack",y = "Defense") plt.show() # hist plot data1.plot(kind = "hist",y = "Defense",bins = 50,range= (0,250),normed = True) # histogram subplot with non cumulative and cumulative fig, axes = plt.subplots(nrows=2,ncols=1) data1.plot(kind = "hist",y = "Defense",bins = 50,range= (0,250),normed = True,ax = axes[0]) data1.plot(kind = "hist",y = "Defense",bins = 50,range= (0,250),normed = True,ax = axes[1],cumulative = True) plt.savefig('graph.png') plt ###Output _____no_output_____ ###Markdown STATISTICAL EXPLORATORY DATA ANALYSISI already explained it at previous parts. However lets look at one more time.* count: number of entries* mean: average of entries* std: standart deviation* min: minimum entry* 25%: first quantile* 50%: median or second quantile* 75%: third quantile* max: maximum entry ###Code data.describe() ###Output _____no_output_____ ###Markdown INDEXING PANDAS TIME SERIES* datetime = object* parse_dates(boolean): Transform date to ISO 8601 (yyyy-mm-dd hh:mm:ss ) format ###Code time_list = ["1992-03-08","1992-04-12"] print(type(time_list[1])) # As you can see date is string # however we want it to be datetime object datetime_object = pd.to_datetime(time_list) print(type(datetime_object)) # close warning import warnings warnings.filterwarnings("ignore") # In order to practice lets take head of pokemon data and add it a time list data2 = data.head() date_list = ["1992-01-10","1992-02-10","1992-03-10","1993-03-15","1993-03-16"] datetime_object = pd.to_datetime(date_list) data2["date"] = datetime_object # lets make date as index data2= data2.set_index("date") data2 # Now we can select according to our date index print(data2.loc["1993-03-16"]) print(data2.loc["1992-03-10":"1993-03-16"]) ###Output _____no_output_____ ###Markdown RESAMPLING PANDAS TIME SERIES* Resampling: statistical method over different time intervals * Needs string to specify frequency like "M" = month or "A" = year* Downsampling: reduce date time rows to slower frequency like from daily to weekly* Upsampling: increase date time rows to faster frequency like from daily to hourly* Interpolate: Interpolate values according to different methods like ‘linear’, ‘time’ or index’ * https://pandas.pydata.org/pandas-docs/stable/generated/pandas.Series.interpolate.html ###Code data2.resample("A") # We will use data2 that we create at previous part data2.resample("A").mean() # Lets resample with month data2.resample("M").mean() # As you can see there are a lot of nan because data2 does not include all months # In real life (data is real. Not created from us like data2) we can solve this problem with interpolate # We can interpolete from first value resampled = data2.resample("M").first() resampled.interpolate("ffill") # Or we can interpolate with mean() data2.resample("M").mean().interpolate("linear") ###Output _____no_output_____ ###Markdown MANIPULATING DATA FRAMES WITH PANDAS INDEXING DATA FRAMES* Indexing using square brackets* Using column attribute and row label* Using loc accessor* Selecting only some columns ###Code # read data data = pd.read_csv('../input/pokemon.csv') data= data.set_index("#") data.head() # indexing using square brackets data["HP"][1] # using column attribute and row label data.HP[1] # using loc accessor data.loc[1,["HP"]] # Selecting only some columns data[["HP","Attack"]] ###Output _____no_output_____ ###Markdown SLICING DATA FRAME* Difference between selecting columns * Series and data frames* Slicing and indexing series* Reverse slicing * From something to end ###Code # Difference between selecting columns: series and dataframes print(type(data["HP"])) # series print(type(data[["HP"]])) # data frames # Slicing and indexing series data.loc[1:10,"HP":"Defense"] # 10 and "Defense" are inclusive # Reverse slicing data.loc[10:1:-1,"HP":"Defense"] # From something to end data.loc[1:10,"Speed":] ###Output _____no_output_____ ###Markdown FILTERING DATA FRAMESCreating boolean seriesCombining filtersFiltering column based others ###Code # Creating boolean series boolean = data.HP > 200 data[boolean] # Combining filters first_filter = data.HP > 150 second_filter = data.Speed > 35 data[first_filter & second_filter] # Filtering column based others data.HP[data.Speed<15] ###Output _____no_output_____ ###Markdown TRANSFORMING DATA* Plain python functions* Lambda function: to apply arbitrary python function to every element* Defining column using other columns ###Code # Plain python functions def div(n): return n/2 data.HP.apply(div) # Or we can use lambda function data.HP.apply(lambda n : n/2) # Defining column using other columns data["total_power"] = data.Attack + data.Defense data.head() ###Output _____no_output_____ ###Markdown INDEX OBJECTS AND LABELED DATAindex: sequence of label ###Code # our index name is this: print(data.index.name) # lets change it data.index.name = "index_name" data.head() # Overwrite index # if we want to modify index we need to change all of them. data.head() # first copy of our data to data3 then change index data3 = data.copy() # lets make index start from 100. It is not remarkable change but it is just example data3.index = range(100,900,1) data3.head() # We can make one of the column as index. I actually did it at the beginning of manipulating data frames with pandas section # It was like this # data= data.set_index("#") # also you can use # data.index = data["#"] ###Output _____no_output_____ ###Markdown HIERARCHICAL INDEXING* Setting indexing ###Code # lets read data frame one more time to start from beginning data = pd.read_csv('../input/pokemon.csv') data.head() # As you can see there is index. However we want to set one or more column to be index # Setting index : type 1 is outer type 2 is inner index data1 = data.set_index(["Type 1","Type 2"]) data1.head(100) # data1.loc["Fire","Flying"] # howw to use indexes ###Output _____no_output_____ ###Markdown PIVOTING DATA FRAMES* pivoting: reshape tool ###Code dic = {"treatment":["A","A","B","B"],"gender":["F","M","F","M"],"response":[10,45,5,9],"age":[15,4,72,65]} df = pd.DataFrame(dic) df # pivoting df.pivot(index="treatment",columns = "gender",values="response") ###Output _____no_output_____ ###Markdown STACKING and UNSTACKING DATAFRAME* deal with multi label indexes* level: position of unstacked index* swaplevel: change inner and outer level index positionRead more about unstacking [here](https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.unstack.html)> Pivot a level of the (necessarily hierarchical) index labels, returning a DataFrame having a new level of column labels whose inner-most level consists of the pivoted index labels. ###Code df1 = df.set_index(["treatment","gender"]) df1 # lets unstack it # level determines indexes df1.unstack(level=0) df1.unstack(level=1) # change inner and outer level index position df2 = df1.swaplevel(0,1) df2 ###Output _____no_output_____ ###Markdown MELTING DATA FRAMES* Reverse of pivoting ###Code df # df.pivot(index="treatment",columns = "gender",values="response") pd.melt(df,id_vars="treatment",value_vars=["age","response"]) ###Output _____no_output_____ ###Markdown CATEGORICALS AND GROUPBY ###Code # We will use df df # according to treatment take means of other features df.groupby("treatment").mean() # mean is aggregation / reduction method # there are other methods like sum, std,max or min # we can only choose one of the feature df.groupby("treatment").age.max() # Or we can choose multiple features df.groupby("treatment")[["age","response"]].min() df.info() # as you can see gender is object # However if we use groupby, we can convert it categorical data. # Because categorical data uses less memory, speed up operations like groupby #df["gender"] = df["gender"].astype("category") #df["treatment"] = df["treatment"].astype("category") #df.info() ###Output _____no_output_____
RNN_Lab.ipynb
###Markdown ###Code from google.colab import drive drive.mount('/content/drive') ###Output Drive already mounted at /content/drive; to attempt to forcibly remount, call drive.mount("/content/drive", force_remount=True). ###Markdown Part-of-Speech Tagging with Recurrent Neural Networks Your task in this assignment is to implement a simple part-of-speech tagger based on recurrent neural networks. Problem specification Your task in this assignment is1. to build a part-of-speech tagger based on a recurrent neural network architecture2. to train this tagger on the provided training data and identify a good model2. to evaluate the performance of this model on the provided test dataTo identify a good model, you can use the provided development (validation) data. Part-of-speech tagging Part-of-speech (POS) tagging is the task of labelling words (tokens) with [parts of speech](https://en.wikipedia.org/wiki/Part_of_speech). To give an example, consider the sentence *Parker hates parsnips*. In this sentence, the word *Parker* should be labelled as a proper noun (a noun that is the name of a person), *hates* should be labelled as a verb, and *parsnips* should be labelled as a (common) noun. Part-of-speech tagging is an essential ingredient of many state-of-the-art natural language understanding systems.Part-of-speech tagging can be cast as a supervised machine learning problem where the gold-standard data consists of sentences whose words have been manually annotated with parts of speech. For the present assignment you will be using a corpus built over the source material of the [English Web Treebank](https://catalog.ldc.upenn.edu/ldc2012t13), consisting of approximately 16,000&nbsp;sentences with 254,000&nbsp;tokens. The corpus has been released by the [Universal Dependencies Project](http://universaldependencies.org).To make it easier to compare systems, the gold-standard data has been split into three parts: training, development (validation), and test. The following cell provides a function that can be used to load the data. ###Code def read_data(path): with open(path, encoding='utf-8') as fp: result = [] for line in fp: line = line.rstrip() if len(line) == 0: yield result result = [] elif not line.startswith('#'): columns = line.split() if columns[0].isdigit(): result.append((columns[1], columns[3])) ###Output _____no_output_____ ###Markdown The next cell loads the data: ###Code train_data = list(read_data('/content/drive/My Drive/Colab Notebooks/RNN/en_ewt-ud-train.conllu')) print('Number of sentences in the training data: {}'.format(len(train_data))) dev_data = list(read_data('/content/drive/My Drive/Colab Notebooks/RNN/en_ewt-ud-dev.conllu')) print('Number of sentences in the development data: {}'.format(len(dev_data))) test_data = list(read_data('/content/drive/My Drive/Colab Notebooks/RNN/en_ewt-ud-test.conllu')) print('Number of sentences in the test data: {}'.format(len(test_data))) ###Output Number of sentences in the training data: 12543 Number of sentences in the development data: 2002 Number of sentences in the test data: 2077 ###Markdown From a Python perspective, each of the data sets is a list of what we shall refer to as *tagged sentences*. A tagged sentence, in turn, is a list of pairs $(w,t)$, where $w$ is a word token and $t$ is the word&rsquo;s POS tag. Here is an example from the training data to show you how this looks like: ###Code train_data[42] ###Output _____no_output_____ ###Markdown You will see part-of-speech tags such as `VERB` for verb, `NOUN` for noun, and `ADV` for adverb. If you are interested in learning more about the tag set used in the gold-standard data, you can have a look at the documentation of the [Universal POS tags](http://universaldependencies.org/u/pos/all.html). However, you do not need to understand the meaning of the POS tags to solve this assignment; you can simply treat them as labels drawn from a finite set of alternatives. Network architecture The proposed network architecture for your tagger is a sequential model with three layers, illustrated below: an embedding, a bidirectional LSTM, and a softmax layer. The embedding turns word indexes (integers representing words) into fixed-size dense vectors which are then fed into the bidirectional LSTM. The output of the LSTM at each position of the sentence is passed to a softmax layer which predicts the POS tag for the word at that position.![architecture.png](attachment:architecture.png)To implement the network architecture, you will use [Keras](https://keras.io/). Keras comes with an extensive online documentation, and reading the relevant parts of this documentation will be essential when working on this assignment. We suggest to start with the tutorial [Getting started with the Keras Sequential model](https://keras.io/getting-started/sequential-model-guide/). After that, you should have a look at some of the examples mentioned in that tutorial, and in particular the [Bidirectional LSTM](https://keras.io/examples/imdb_bidirectional_lstm/) example. Evaluation The most widely-used evaluation measure for part-of-speech tagging is per-word accuracy, which is the percentage of words to which the tagger assigns the correct tag (according to the gold standard). This is one of the default metrics in Keras.One problem that you will encounter during evaluation is that the evaluation data contains words that you did not see (and did not add to your index) during training. The simplest solution to this problem is to introduce a special &lsquo;word&rsquo; `` and replace each unknown word with this pseudoword. Part 1: Pre-process the data Before you can start to implement the network architecture as such, you will have to bring the tagged sentences from the gold-standard data into a form that can be used with the network. One important step in this is to map the words and tags (strings) to integers. Here is code that illustrates the idea: ###Code word_to_index = {} for tagged_sentence in train_data: for word, tag in tagged_sentence: if word not in word_to_index: word_to_index[word] = len(word_to_index) print('Number of unique words in the training data: {}'.format(len(word_to_index))) print('Index of the word "hates": {}'.format(word_to_index['hates'])) ###Output Number of unique words in the training data: 19672 Index of the word "hates": 4579 ###Markdown Once you have indexes for the words and the tags, you can construct the input and the gold-standard output tensor required to train the network. Constructing the input tensorThe input tensor should be of shape $(N, n)$ where $N$ is the total number of sentences in the training data and $n$ is the length of the longest sentence. Note that Keras requires all sequences in an input tensor to have the same length, which means that you will have to pad all sequences to that length. You can use the helper function [`pad_sequences`](https://www.tensorflow.org/api_docs/python/tf/keras/preprocessing/sequence/pad_sequences) for this, which by default will front-pad sequences with the value&nbsp;0. It is essential then that you do not use this special padding value as the index of actual words. Constructing the target output tensorThe target output tensor should be of shape $(N, n, T)$ where $T$ is the number of unique tags in the training data, plus one to cater for the special padding value. The additional dimension corresponds to the fact that the softmax layer of the network will output one $T$-dimensional vector for each position of an input sentence. To construct this vector, you can use the helper function [`to_categorical`](https://www.tensorflow.org/api_docs/python/tf/keras/utils/to_categorical). ###Code # Define a help function to build index from a list of words or tags, each word / tag will have a unique number def build_index(strings, init=[]): string_to_index = {s: i for i, s in enumerate(init)} # Loop over strings in 'strings' for string in strings: # Check if string exists in variable 'string_to_index', # if string does not exist, add a new element to 'string_to_index': the current length of 'string_to_index' if string not in string_to_index: string_to_index[string]=len(string_to_index) return string_to_index # Convert all words and tags in train_data to lists, start with empty lists and use '.append()' # to add one word / tag at a time, similar to the cell below 'pre-process the data' words, tags = [], [] for tagged_sentence in train_data: for word,tag in tagged_sentence: words.append(word) tags.append(tag) # Call the help function you made, to build an index for words (word_to_index), and one index for tags (tag_to_index) word_to_index=build_index(words,['<pad>','<unk>']) tag_to_index=build_index(tags,['<pad>']) # Check number of words and tags num_words = len(word_to_index) num_tags = len(tag_to_index) print(f'Number of unique words in the training data: {num_words}') print(f'Number of unique tags in the training_data: {num_tags}') from tensorflow.keras.preprocessing.sequence import pad_sequences from tensorflow.keras.utils import to_categorical # Make a function that converts the tagged sentences, word indices and tag indices to # X and Y, that can be used when training the RNN def encode(tagged_sentences, word_to_index, tag_to_index): # Start with empty lists that will contain all training examples and corresponding output X, Y = [], [] # Loop over tagged sentences for current_tagged_sentence in tagged_sentences: Xcurrent, Ycurrent = [], [] for word,tag in current_tagged_sentence:# Loop over words and tags in current sentence if word not in word_to_index: Xcurrent.append(word_to_index.get('<unk>'))#adding an unkown word index else: Xcurrent.append(word_to_index.get(word))#adding the index of the word if tag not in tag_to_index: Ycurrent.append(tag_to_index.get('<unk>'))#adding an unkown tag index else: Ycurrent.append(tag_to_index.get(tag))#adding the index of an exitsing tag # Append X with Xcurrent, and Y with Ycurrent X.append(Xcurrent) Y.append(Ycurrent) # Pad the sequences, so that all have the same length X=pad_sequences(sequences=X,padding='post') Y=pad_sequences(sequences=Y,padding='post') # Convert labels to categorical, as you did in the CNN lab Y=to_categorical(Y,num_classes=num_tags,dtype= 'float32') return X, Y # Use your 'encode' function to create X and Y from train_data, word_to_index, tag_to_index X,Y=encode(train_data,word_to_index,tag_to_index) # Print the shape of X and Y print('Shape of X:',X.shape) print('Shape of Y:',Y.shape) ###Output Shape of X: (12543, 159) Shape of Y: (12543, 159, 18) ###Markdown Part 2: Construct the model To implement the network architecture, you need to find and instantiate the relevant building blocks from the Keras library. Note that Keras layers support a large number of optional parameters; use the default values unless you have a good reason not to. Two mandatory parameters that you will have to specify are the dimensionality of the embedding and the dimensionality of the output of the LSTM layer. The following values are reasonable starting points, but do try a number of different settings.* dimensionality of the embedding: 100* dimensionality of the output of the bidirectional LSTM layer: 100You will also have to choose an appropriate loss function. For training we recommend the Adam optimiser. ###Code from tensorflow.keras import Sequential # Import necessary layers from tensorflow.keras.layers import Dense,Embedding, LSTM, Bidirectional from keras.losses import categorical_crossentropy embedding_dim = 100 hidden_dim = 100 model = Sequential() model.add(Embedding(input_dim=num_words,output_dim=embedding_dim)) model.add(Bidirectional(LSTM(units=hidden_dim))) model.add(Dense(num_tags, activation='softmax')) # Compile model model.compile(loss='categorical_crossentropy', optimizer='Adam', metrics=['accuracy']) # Print a summary of the model model.summary() type(dev_data) type(train_data) ###Output _____no_output_____ ###Markdown Part 3: Train the network The next step is to train the network. Use the following parameters:* number of epochs: 10* batch size: 32Training will print the average running loss on the training data after each minibatch. In addition to that, we ask you to also print the loss and accuracy on the development data after each epoch. You can do so by providing the `validation_data` argument to the `fit` method.Note that the `fit` method returns a [`History`](https://keras.io/callbacks/history) object that contains useful information about the training. We will use that information in the next step. ###Code # Encode the development (validation data) using the 'encode' function you created before batch_size=32 epochs=10 #splitting the dev data into Xval and Yval to train the network Xval,Yval=encode(dev_data,word_to_index,tag_to_index) # Train the model and save the history, as you did in the DNN and CNN labs, provide validation data history=model.fit(X,Y,validation_data = (Xval, Yval),batch_size = batch_size, epochs = epochs) ###Output _____no_output_____ ###Markdown Part 4: Identify a good model The following code will plot the loss on the training data and the loss on the validation data after each epoch: ###Code # Lets define a help function for plotting the training results import matplotlib.pyplot as plt def plot_results(history): val_loss = history.history['val_loss'] acc = history.history['accuracy'] loss = history.history['loss'] val_acc = history.history['val_accuracy'] plt.figure(figsize=(10,4)) plt.xlabel('Epochs') plt.ylabel('Loss') plt.plot(loss) plt.plot(val_loss) plt.legend(['Training','Validation']) plt.figure(figsize=(10,4)) plt.xlabel('Epochs') plt.ylabel('Accuracy') plt.plot(acc) plt.plot(val_acc) plt.legend(['Training','Validation']) plt.show() plot_results(history) ###Output _____no_output_____ ###Markdown Look at the plot and determine the epoch after which the model starts to overfit. Then, re-train your model using that many epochs and compute the accuracy of the tagger on the test data. ###Code # Encode the test_data using the 'encode' function you created before # Evaluate the model on test data, as you did in the DNN and CNN lab ###Output _____no_output_____
recursion_dynamic/n_pairs_parentheses/n_pairs_parentheses_solution.ipynb
###Markdown This notebook was prepared by [Rishi Rajasekaran](https://github.com/rishihot55). Source and license info is on [GitHub](https://github.com/donnemartin/interactive-coding-challenges). Solution Notebook Problem: Print all valid combinations of n-pairs of parentheses.* [Constraints](Constraints)* [Test Cases](Test-Cases)* [Algorithm](Algorithm)* [Code](Code)* [Unit Test](Unit-Test) Constraints* None Test Cases* 0 -> ' '* 1 -> ()* 2 -> (()), ()()* 3 -> ((())), (()()), (())(), ()(()), ()()() AlgorithmLet `l` and `r` denote the number of left and right parentheses remaining at any given point. The algorithm makes use of the following conditions applied recursively:* Left braces can be inserted any time, as long as we do not exhaust them i.e. `l > 0`.* Right braces can be inserted, as long as the number of right braces remaining is greater than the left braces remaining i.e. `r > l`. Violation of the aforementioned condition produces an unbalanced string of parentheses.* If both left and right braces have been exhausted i.e. `l = 0 and r = 0`, then the resultant string produced is balanced.The algorithm can be rephrased as:* Base case: `l = 0 and r = 0` - Add the string generated to the result set* Case 1: `l > 0` - Add a left parenthesis to the parentheses string. - Call parentheses_util(l - 1, r, new_string, result_set)* Case 2: `r > l` - Add a right parenthesis to the parentheses string. - Call parentheses_util(l, r - 1, new_string, result_set)Complexity:* Time: `O(4^n/n^(3/2))`. See [Catalan numbers](https://en.wikipedia.org/wiki/Catalan_numberApplications_in_combinatorics)* Space complexity: `O(n)` (Due to the implicit call stack storing a maximum of 2n function calls) Code ###Code def parentheses_util(no_left, no_right, pair_string, result): if no_left == 0 and no_right == 0: result.add(pair_string) else: if no_left > 0: parentheses_util(no_left - 1, no_right, pair_string + '(', result) if no_right > no_left: parentheses_util(no_left, no_right - 1, pair_string + ')', result) def pair_parentheses(n): result_set = set() if n == 0: return result_set parentheses_util(n, n, '', result_set) return result_set ###Output _____no_output_____ ###Markdown Unit Test ###Code %%writefile test_n_pairs_parentheses.py from nose.tools import assert_equal class TestPairParentheses(object): def test_pair_parentheses(self, solution): assert_equal(solution(0), set([])) assert_equal(solution(1), set(['()'])) assert_equal(solution(2), set(['(())', '()()'])) assert_equal(solution(3), set(['((()))', '(()())', '(())()', '()(())', '()()()'])) print('Success: test_pair_parentheses') def main(): test = TestPairParentheses() test.test_pair_parentheses(pair_parentheses) if __name__ == '__main__': main() %run -i test_n_pairs_parentheses.py ###Output Success: test_pair_parentheses ###Markdown Unit Test ###Code %%writefile test_n_pairs_parentheses.py from nose.tools import assert_equal class TestPairParentheses(object): def test_pair_parentheses(self, solution): assert_equal(solution(0), set([])) assert_equal(solution(1), set(['()'])) assert_equal(solution(2), set(['(())', '()()'])) assert_equal(solution(3), set(['((()))', '(()())', '(())()', '()(())', '()()()'])) print('Success: test_pair_parentheses') def main(): test = TestPairParentheses() test.test_pair_parentheses(pair_parentheses) if __name__ == '__main__': main() %run -i test_n_pairs_parentheses.py ###Output Success: test_pair_parentheses ###Markdown This notebook was prepared by [Rishi Rajasekaran](https://github.com/rishihot55). Source and license info is on [GitHub](https://github.com/donnemartin/interactive-coding-challenges). Solution Notebook Problem: Find all valid combinations of n-pairs of parentheses.* [Constraints](Constraints)* [Test Cases](Test-Cases)* [Algorithm](Algorithm)* [Code](Code)* [Unit Test](Unit-Test) Constraints* Is the input an integer representing the number of pairs? * Yes* Can we assume the inputs are valid? * No* Is the output a list of valid combinations? * Yes* Should the output have duplicates? * No* Can we assume this fits memory? * Yes Test Cases* None -> Exception* Negative -> Exception* 0 -> []* 1 -> ['()']* 2 -> ['(())', '()()']* 3 -> ['((()))', '(()())', '(())()', '()(())', '()()()'] AlgorithmLet `l` and `r` denote the number of left and right parentheses remaining at any given point. The algorithm makes use of the following conditions applied recursively:* Left braces can be inserted any time, as long as we do not exhaust them i.e. `l > 0`.* Right braces can be inserted, as long as the number of right braces remaining is greater than the left braces remaining i.e. `r > l`. Violation of the aforementioned condition produces an unbalanced string of parentheses.* If both left and right braces have been exhausted i.e. `l = 0 and r = 0`, then the resultant string produced is balanced.The algorithm can be rephrased as:* Base case: `l = 0 and r = 0` - Add the string generated to the result set* Case 1: `l > 0` - Add a left parenthesis to the parentheses string. - Recurse (l - 1, r, new_string, result_set)* Case 2: `r > l` - Add a right parenthesis to the parentheses string. - Recurse (l, r - 1, new_string, result_set)Complexity:* Time: `O(4^n/n^(3/2))`, see [Catalan numbers](https://en.wikipedia.org/wiki/Catalan_numberApplications_in_combinatorics) - 1, 1, 2, 5, 14, 42, 132...* Space complexity: `O(n)`, due to the implicit call stack storing a maximum of 2n function calls) Code ###Code class Parentheses(object): def find_pair(self, num_pairs): if num_pairs is None: raise TypeError('num_pairs cannot be None') if num_pairs < 0: raise ValueError('num_pairs cannot be < 0') if not num_pairs: return [] results = [] curr_results = [] self._find_pair(num_pairs, num_pairs, curr_results, results) return results def _find_pair(self, nleft, nright, curr_results, results): if nleft == 0 and nright == 0: results.append(''.join(curr_results)) else: if nleft >= 0: self._find_pair(nleft-1, nright, curr_results+['('], results) if nright > nleft: self._find_pair(nleft, nright-1, curr_results+[')'], results) ###Output _____no_output_____ ###Markdown Unit Test ###Code %%writefile test_n_pairs_parentheses.py import unittest class TestPairParentheses(unittest.TestCase): def test_pair_parentheses(self): parentheses = Parentheses() self.assertRaises(TypeError, parentheses.find_pair, None) self.assertRaises(ValueError, parentheses.find_pair, -1) self.assertEqual(parentheses.find_pair(0), []) self.assertEqual(parentheses.find_pair(1), ['()']) self.assertEqual(parentheses.find_pair(2), ['(())', '()()']) self.assertEqual(parentheses.find_pair(3), ['((()))', '(()())', '(())()', '()(())', '()()()']) print('Success: test_pair_parentheses') def main(): test = TestPairParentheses() test.test_pair_parentheses() if __name__ == '__main__': main() %run -i test_n_pairs_parentheses.py ###Output Success: test_pair_parentheses ###Markdown This notebook was prepared by [Rishi Rajasekaran](https://github.com/rishihot55). Source and license info is on [GitHub](https://github.com/donnemartin/interactive-coding-challenges). Solution Notebook Problem: Print all valid combinations of n-pairs of parentheses.* [Constraints](Constraints)* [Test Cases](Test-Cases)* [Algorithm](Algorithm)* [Code](Code)* [Unit Test](Unit-Test) Constraints* None Test Cases* 0 -> ' '* 1 -> ()* 2 -> (()), ()()* 3 -> ((())), (()()), (())(), ()(()), ()()() AlgorithmLet `l` and `r` denote the number of left and right parentheses remaining at any given point. The algorithm makes use of the following conditions applied recursively:* Left braces can be inserted any time, as long as we do not exhaust them i.e. `l > 0`.* Right braces can be inserted, as long as the number of right braces remaining is greater than the left braces remaining i.e. `r > l`. Violation of the aforementioned condition produces an unbalanced string of parentheses.* If both left and right braces have been exhausted i.e. `l = 0 and r = 0`, then the resultant string produced is balanced.The algorithm can be rephrased as:* Base case: `l = 0 and r = 0` - Add the string generated to the result set* Case 1: `l > 0` - Add a left parenthesis to the parentheses string. - Call parentheses_util(l - 1, r, new_string, result_set)* Case 2: `r > l` - Add a right parenthesis to the parentheses string. - Call parentheses_util(l, r - 1, new_string, result_set)Complexity:* Time: `O(4^n/n^(3/2))`. See [Catalan numbers](https://en.wikipedia.org/wiki/Catalan_numberApplications_in_combinatorics)* Space complexity: `O(n)` (Due to the implicit call stack storing a maximum of 2n function calls) Code ###Code def parentheses_util(no_left, no_right, pair_string, result): if no_left == 0 and no_right == 0: result.add(pair_string) else: if no_left > 0: parentheses_util(no_left - 1, no_right, pair_string + '(', result) if no_right > no_left: parentheses_util(no_left, no_right - 1, pair_string + ')', result) def pair_parentheses(n): result_set = set() if n == 0: return result_set parentheses_util(n, n, '', result_set) return result_set ###Output _____no_output_____ ###Markdown This notebook was prepared by [Rishi Rajasekaran](https://github.com/rishihot55). Source and license info is on [GitHub](https://github.com/donnemartin/interactive-coding-challenges). Solution Notebook Problem: Find all valid combinations of n-pairs of parentheses.* [Constraints](Constraints)* [Test Cases](Test-Cases)* [Algorithm](Algorithm)* [Code](Code)* [Unit Test](Unit-Test) Constraints* Is the input an integer representing the number of pairs? * Yes* Can we assume the inputs are valid? * No* Is the output a list of valid combinations? * Yes* Should the output have duplicates? * No* Can we assume this fits memory? * Yes Test Cases* None -> Exception* Negative -> Exception* 0 -> []* 1 -> ['()']* 2 -> ['(())', '()()']* 3 -> ['((()))', '(()())', '(())()', '()(())', '()()()'] AlgorithmLet `l` and `r` denote the number of left and right parentheses remaining at any given point. The algorithm makes use of the following conditions applied recursively:* Left braces can be inserted any time, as long as we do not exhaust them i.e. `l > 0`.* Right braces can be inserted, as long as the number of right braces remaining is greater than the left braces remaining i.e. `r > l`. Violation of the aforementioned condition produces an unbalanced string of parentheses.* If both left and right braces have been exhausted i.e. `l = 0 and r = 0`, then the resultant string produced is balanced.The algorithm can be rephrased as:* Base case: `l = 0 and r = 0` - Add the string generated to the result set* Case 1: `l > 0` - Add a left parenthesis to the parentheses string. - Recurse (l - 1, r, new_string, result_set)* Case 2: `r > l` - Add a right parenthesis to the parentheses string. - Recurse (l, r - 1, new_string, result_set)Complexity:* Time: `O(4^n/n^(3/2))`, see [Catalan numbers](https://en.wikipedia.org/wiki/Catalan_numberApplications_in_combinatorics) - 1, 1, 2, 5, 14, 42, 132...* Space complexity: `O(n)`, due to the implicit call stack storing a maximum of 2n function calls) Code ###Code class Parentheses(object): def find_pair(self, num_pairs): if num_pairs is None: raise TypeError('num_pairs cannot be None') if num_pairs < 0: raise ValueError('num_pairs cannot be < 0') if not num_pairs: return [] results = [] curr_results = [] self._find_pair(num_pairs, num_pairs, curr_results, results) return results def _find_pair(self, nleft, nright, curr_results, results): if nleft == 0 and nright == 0: results.append(''.join(curr_results)) else: if nleft >= 0: self._find_pair(nleft-1, nright, curr_results+['('], results) if nright > nleft: self._find_pair(nleft, nright-1, curr_results+[')'], results) ###Output _____no_output_____ ###Markdown Unit Test ###Code %%writefile test_n_pairs_parentheses.py from nose.tools import assert_equal, assert_raises class TestPairParentheses(object): def test_pair_parentheses(self): parentheses = Parentheses() assert_raises(TypeError, parentheses.find_pair, None) assert_raises(ValueError, parentheses.find_pair, -1) assert_equal(parentheses.find_pair(0), []) assert_equal(parentheses.find_pair(1), ['()']) assert_equal(parentheses.find_pair(2), ['(())', '()()']) assert_equal(parentheses.find_pair(3), ['((()))', '(()())', '(())()', '()(())', '()()()']) print('Success: test_pair_parentheses') def main(): test = TestPairParentheses() test.test_pair_parentheses() if __name__ == '__main__': main() %run -i test_n_pairs_parentheses.py ###Output Success: test_pair_parentheses
Sistema Completo.ipynb
###Markdown __Treinamento do sistema ajustado por usuário.__O sistema deve treinar e ajustar um modelo para cada usuário. O modelo treinado será salvo em uma pasta. ###Code models = [] for USER in manager.users: print(f"Training models for user {USER}") # 1. Treina os modelos de classificação e regressão. # Aqui os modelos serão especificados manualmente, mas eles poderiam ser escolhidos utilizado os scripts # do sistema de avaliação de modelos. print('Training classification and regression models.') classifier, regressor = train_evaluation_system(manager, USER) evaluator_model = Valuer(classification_model = classifier, regression_model = regressor) # 2. Treina o modelo generativo. print('Training generative model.') skorch_model, scaler = train_generative_system(manager, USER, verbose = False) models.append(skorch_model) # para gerar as curvas de aprendizado generative_model = skorch_model.module_ # 3. Criação do sistema de recomendação adaptado ao usuário print("Creating recommender model.") recommender = Recommender(generativeModel = generative_model, evaluationModel = evaluator_model, scaler = scaler, user = USER) print() # 4. Salva o modelo path = Path(OUTPUT_PATH) / USER path.mkdir(parents = True, exist_ok = True) print(f"Saving recommender model to: {path}") filehandler = open(path / 'recommender.pickle', "wb") pickle.dump(recommender, filehandler) print(f"Finishing model adjusment for user {USER}") print() plt.style.use('ggplot') fig, axs = plt.subplots(ncols = 2, figsize = (15, 5)) train_loss = models[0].history[:, 'train_loss'] valid_loss = models[0].history[:, 'valid_loss'] X = range(len(train_loss)) axs[0].plot(X, train_loss, label = 'Erro de Treinamento', linewidth = 3.5) axs[0].plot(X, valid_loss, label = 'Erro de Validação', linewidth = 3.5) axs[0].set_ylabel("Custo", fontsize = 22) axs[0].set_xlabel("Iteração", fontsize = 22) axs[0].set_title(f"Curvas de aprendizado do modelo generativo. {manager.users[0]}", fontsize = 15) axs[0].tick_params(axis='both', which='major', labelsize=20) axs[0].legend(fontsize = 15) train_loss = models[1].history[:, 'train_loss'] valid_loss = models[1].history[:, 'valid_loss'] X = range(len(train_loss)) axs[1].plot(X, train_loss, label = 'Erro de Treinamento', linewidth = 3.5) axs[1].plot(X, valid_loss, label = 'Erro de Validação', linewidth = 3.5) axs[1].set_ylabel("Custo", fontsize = 22) axs[1].set_xlabel("Iteração", fontsize = 22) axs[1].set_title(f"Curvas de aprendizado do modelo generativo. {manager.users[1]}", fontsize = 15) axs[1].tick_params(axis='both', which='major', labelsize=20) axs[1].legend(fontsize = 15) fig.savefig("curvas_aprendizado_autoencoder.png", bbox_inches = 'tight') ###Output _____no_output_____ ###Markdown __Teste do sistema generativo carregado da memória após o ajuste.__ ###Code USER = manager.users[0] model_path = Path(OUTPUT_PATH) / manager.users[0] / 'recommender.pickle' file = open(model_path, "rb") recommender = pickle.load(file) recommendation_list = recommender.getMusicList(20, manager.data.drop(columns = ['id_cliente', 'data_curtida', 'n_reproducao', 'gostou'])) recommendation_list.drop_duplicates() fig, ax = plt.subplots(ncols = 2, figsize = (15, 5)) user_data_liked = manager.user_data(USER) user_data_liked = user_data_liked[user_data_liked['gostou'] == 1] recommendation_list['VolMedio'].plot.density(ax = ax[0], label = 'Lista de Recomendação') user_data_liked['VolMedio'].plot.density(ax = ax[0], label = 'Dados do Usuário') recommendation_list['PctCantada'].plot.density(ax = ax[1], label = 'Lista de Recomendação') user_data_liked['PctCantada'].plot.density(ax = ax[1], label = 'Dados do Usuário') ax[0].legend(fontsize = 12) ax[1].legend(fontsize = 12) ax[0].set_xlabel("Volume Médio", fontsize = 22) ax[1].set_xlabel("Porcentagem com Vocal", fontsize = 22) ax[0].set_xlim([0, 30.0]) ax[1].set_xlim([0, 1.0]) ax[0].tick_params(axis='both', which='major', labelsize=20) ax[1].tick_params(axis='both', which='major', labelsize=20) ax[0].set_title(f"Curvas de densidade do Volume Médio. {manager.users[0]}", fontsize = 15) ax[1].set_title(f"Curvas de densidade da Porcentagem de Vocal. {manager.users[0]}", fontsize = 15) fig.savefig("densidades_lista_recomendacao.png", bbox_inches = 'tight') ###Output _____no_output_____ ###Markdown __Avaliação do desempenho do sistema.__ ###Code USER = manager.users[0] model_path = Path(OUTPUT_PATH) / manager.users[0] / 'recommender.pickle' file = open(model_path, "rb") recommender = pickle.load(file) classification_errors = [] regression_errors = [] for iteration in range(500): a, b = recommender.test_model(manager.user_data(USER)) classification_errors.append(a) regression_errors.append(b) np.array(classification_errors).mean(), np.array(classification_errors).std() np.array(regression_errors).mean(), np.array(regression_errors).std() USER = manager.users[1] model_path = Path(OUTPUT_PATH) / manager.users[1] / 'recommender.pickle' file = open(model_path, "rb") recommender = pickle.load(file) classification_errors = [] regression_errors = [] for iteration in range(500): a, b = recommender.test_model(manager.user_data(USER)) classification_errors.append(a) regression_errors.append(b) np.array(classification_errors).mean(), np.array(classification_errors).std() np.array(regression_errors).mean(), np.array(regression_errors).std() ###Output _____no_output_____
Lec9.ipynb
###Markdown Installed Python Libraries now import them ###Code import pymongo from pymongo import MongoClient import json import tweepy import twitter from pprint import pprint import configparser import pandas as pd ###Output _____no_output_____ ###Markdown Load Authorization Information ###Code config = configparser.ConfigParser() config.read('config.ini') CONSUMER_KEY = config['mytwitter']['api_key'] CONSUMER_SECRET = config['mytwitter']['api_secrete'] OAUTH_TOKEN = config['mytwitter']['access_token'] OATH_TOKEN_SECRET = config['mytwitter']['access_secrete'] mongod_connect = config['mymongo']['connection'] ###Output _____no_output_____ ###Markdown Connect to MongoDB Cluster ###Code client = MongoClient(mongod_connect) db = client.demo # use or create a database named demo tweet_collection = db.tweet_collection #use or create a collection named tweet_collection tweet_collection.create_index([("id", pymongo.ASCENDING)],unique = True) # make sure the collected tweets are unique ###Output _____no_output_____ ###Markdown Use the Rest API to Collect Tweets Authorization ###Code rest_auth = twitter.oauth.OAuth(OAUTH_TOKEN,OATH_TOKEN_SECRET,CONSUMER_KEY,CONSUMER_SECRET) rest_api = twitter.Twitter(auth=rest_auth) ###Output _____no_output_____ ###Markdown Define query for rest API ###Code count = 100 #number of returned tweets, default and max is 100 geocode = "38.4392897,-78.9412224,50mi" # defin the location, in Harrisonburg, VA q = "election" #define the keywords, tweets contain election ###Output _____no_output_____ ###Markdown The retained tweets will contain election and be located in harrisonburg ###Code search_results = rest_api.search.tweets( count=count,q=q, geocode=geocode) #you can use both q and geocode statuses = search_results["statuses"] since_id_new = statuses[-1]['id'] for statuse in statuses: try: tweet_collection.insert_one(statuse) pprint(statuse['created_at'])# print the date of the collected tweets except: pass ###Output 'Tue Nov 02 19:42:13 +0000 2021' 'Tue Nov 02 19:37:15 +0000 2021' 'Tue Nov 02 19:35:47 +0000 2021' 'Tue Nov 02 19:32:12 +0000 2021' 'Tue Nov 02 19:32:10 +0000 2021' 'Tue Nov 02 19:30:32 +0000 2021' 'Tue Nov 02 19:30:00 +0000 2021' 'Tue Nov 02 19:27:38 +0000 2021' 'Tue Nov 02 19:17:19 +0000 2021' 'Tue Nov 02 19:12:16 +0000 2021' 'Tue Nov 02 19:03:15 +0000 2021' 'Tue Nov 02 18:49:27 +0000 2021' 'Tue Nov 02 18:49:27 +0000 2021' 'Tue Nov 02 18:44:51 +0000 2021' 'Tue Nov 02 18:42:59 +0000 2021' 'Tue Nov 02 18:38:57 +0000 2021' 'Tue Nov 02 18:38:01 +0000 2021' 'Tue Nov 02 18:35:28 +0000 2021' 'Tue Nov 02 18:33:48 +0000 2021' 'Tue Nov 02 18:31:06 +0000 2021' 'Tue Nov 02 18:30:28 +0000 2021' 'Tue Nov 02 18:26:18 +0000 2021' 'Tue Nov 02 18:24:35 +0000 2021' 'Tue Nov 02 18:12:19 +0000 2021' 'Tue Nov 02 18:07:27 +0000 2021' 'Tue Nov 02 18:05:11 +0000 2021' 'Tue Nov 02 18:02:30 +0000 2021' 'Tue Nov 02 18:00:54 +0000 2021' 'Tue Nov 02 17:53:10 +0000 2021' 'Tue Nov 02 17:50:08 +0000 2021' 'Tue Nov 02 17:47:18 +0000 2021' 'Tue Nov 02 17:44:44 +0000 2021' 'Tue Nov 02 17:41:43 +0000 2021' 'Tue Nov 02 17:35:11 +0000 2021' 'Tue Nov 02 17:25:32 +0000 2021' 'Tue Nov 02 17:25:25 +0000 2021' 'Tue Nov 02 17:23:34 +0000 2021' 'Tue Nov 02 17:18:47 +0000 2021' 'Tue Nov 02 17:15:23 +0000 2021' 'Tue Nov 02 17:14:14 +0000 2021' 'Tue Nov 02 17:12:22 +0000 2021' 'Tue Nov 02 17:08:41 +0000 2021' 'Tue Nov 02 17:00:02 +0000 2021' 'Tue Nov 02 16:59:07 +0000 2021' 'Tue Nov 02 16:57:43 +0000 2021' 'Tue Nov 02 16:55:00 +0000 2021' 'Tue Nov 02 16:54:39 +0000 2021' 'Tue Nov 02 16:53:24 +0000 2021' 'Tue Nov 02 16:53:22 +0000 2021' 'Tue Nov 02 16:52:52 +0000 2021' 'Tue Nov 02 16:46:40 +0000 2021' 'Tue Nov 02 16:44:01 +0000 2021' 'Tue Nov 02 16:43:00 +0000 2021' 'Tue Nov 02 16:41:30 +0000 2021' 'Tue Nov 02 16:37:30 +0000 2021' 'Tue Nov 02 16:35:28 +0000 2021' 'Tue Nov 02 16:13:30 +0000 2021' 'Tue Nov 02 16:12:32 +0000 2021' 'Tue Nov 02 16:11:17 +0000 2021' 'Tue Nov 02 16:05:59 +0000 2021' 'Tue Nov 02 16:02:08 +0000 2021' 'Tue Nov 02 16:01:17 +0000 2021' 'Tue Nov 02 16:00:22 +0000 2021' 'Tue Nov 02 15:58:41 +0000 2021' 'Tue Nov 02 15:56:47 +0000 2021' 'Tue Nov 02 15:54:42 +0000 2021' 'Tue Nov 02 15:53:50 +0000 2021' 'Tue Nov 02 15:53:08 +0000 2021' 'Tue Nov 02 15:52:09 +0000 2021' 'Tue Nov 02 15:48:39 +0000 2021' 'Tue Nov 02 15:44:57 +0000 2021' 'Tue Nov 02 15:40:04 +0000 2021' 'Tue Nov 02 15:34:46 +0000 2021' 'Tue Nov 02 15:30:35 +0000 2021' 'Tue Nov 02 15:28:56 +0000 2021' 'Tue Nov 02 15:22:59 +0000 2021' 'Tue Nov 02 15:19:52 +0000 2021' 'Tue Nov 02 15:19:44 +0000 2021' 'Tue Nov 02 15:10:47 +0000 2021' 'Tue Nov 02 15:07:50 +0000 2021' 'Tue Nov 02 15:07:12 +0000 2021' 'Tue Nov 02 15:02:17 +0000 2021' 'Tue Nov 02 15:01:29 +0000 2021' 'Tue Nov 02 15:00:52 +0000 2021' 'Tue Nov 02 14:54:28 +0000 2021' 'Tue Nov 02 14:51:39 +0000 2021' 'Tue Nov 02 14:49:11 +0000 2021' 'Tue Nov 02 14:48:25 +0000 2021' 'Tue Nov 02 14:46:30 +0000 2021' 'Tue Nov 02 14:44:41 +0000 2021' 'Tue Nov 02 14:44:29 +0000 2021' 'Tue Nov 02 14:44:25 +0000 2021' 'Tue Nov 02 14:43:03 +0000 2021' 'Tue Nov 02 14:41:22 +0000 2021' 'Tue Nov 02 14:41:18 +0000 2021' 'Tue Nov 02 14:28:51 +0000 2021' 'Tue Nov 02 14:26:02 +0000 2021' 'Tue Nov 02 14:24:42 +0000 2021' 'Tue Nov 02 14:23:49 +0000 2021' 'Tue Nov 02 14:20:57 +0000 2021' ###Markdown Use Rest API to Collect Tweets ###Code rest_auth = twitter.oauth.OAuth(OAUTH_TOKEN,OATH_TOKEN_SECRET,CONSUMER_KEY,CONSUMER_SECRET) rest_api = twitter.Twitter(auth=rest_auth) count = 100 #number of returned tweets, default and max is 100 geocode = "38.4392897,-78.9412224,50mi" # defin the location, in Harrisonburg, VA q = "election" #define the keywords, tweets contain election search_results = rest_api.search.tweets( count=count,q=q, geocode=geocode) #you can use both q and geocode statuses = search_results["statuses"] since_id_new = statuses[-1]['id'] for statuse in statuses: try: tweet_collection.insert_one(statuse) pprint(statuse['created_at'])# print the date of the collected tweets except: pass ###Output 'Mon Nov 01 21:00:59 +0000 2021' 'Mon Nov 01 20:53:30 +0000 2021' 'Mon Nov 01 20:53:02 +0000 2021' 'Mon Nov 01 20:52:26 +0000 2021' 'Mon Nov 01 20:37:07 +0000 2021' 'Mon Nov 01 20:22:42 +0000 2021' 'Mon Nov 01 20:05:11 +0000 2021' 'Mon Nov 01 19:25:39 +0000 2021' 'Mon Nov 01 19:13:57 +0000 2021' 'Mon Nov 01 19:08:46 +0000 2021' 'Mon Nov 01 19:07:19 +0000 2021' 'Mon Nov 01 19:00:47 +0000 2021' 'Mon Nov 01 18:49:22 +0000 2021' 'Mon Nov 01 18:33:51 +0000 2021' 'Mon Nov 01 18:30:09 +0000 2021' 'Mon Nov 01 18:17:19 +0000 2021' 'Mon Nov 01 18:08:49 +0000 2021' 'Mon Nov 01 18:02:33 +0000 2021' 'Mon Nov 01 18:00:10 +0000 2021' 'Mon Nov 01 17:59:53 +0000 2021' 'Mon Nov 01 17:57:53 +0000 2021' 'Mon Nov 01 17:27:06 +0000 2021' 'Mon Nov 01 17:18:56 +0000 2021' 'Mon Nov 01 17:12:46 +0000 2021' 'Mon Nov 01 17:07:01 +0000 2021' 'Mon Nov 01 16:57:09 +0000 2021' 'Mon Nov 01 16:48:57 +0000 2021' 'Mon Nov 01 16:45:07 +0000 2021' 'Mon Nov 01 16:42:37 +0000 2021' 'Mon Nov 01 16:40:18 +0000 2021' 'Mon Nov 01 16:29:51 +0000 2021' 'Mon Nov 01 16:29:08 +0000 2021' 'Mon Nov 01 16:28:58 +0000 2021' 'Mon Nov 01 16:05:00 +0000 2021' 'Mon Nov 01 16:02:05 +0000 2021' 'Mon Nov 01 15:38:09 +0000 2021' 'Mon Nov 01 15:35:33 +0000 2021' 'Mon Nov 01 15:12:24 +0000 2021' 'Mon Nov 01 15:04:06 +0000 2021' 'Mon Nov 01 14:59:37 +0000 2021' 'Mon Nov 01 14:32:26 +0000 2021' 'Mon Nov 01 14:30:26 +0000 2021' 'Mon Nov 01 13:51:08 +0000 2021' 'Mon Nov 01 13:47:27 +0000 2021' 'Mon Nov 01 13:38:55 +0000 2021' 'Mon Nov 01 13:36:56 +0000 2021' 'Mon Nov 01 13:13:38 +0000 2021' 'Mon Nov 01 13:09:25 +0000 2021' 'Mon Nov 01 13:03:43 +0000 2021' 'Mon Nov 01 12:56:32 +0000 2021' 'Mon Nov 01 12:55:41 +0000 2021' 'Mon Nov 01 12:47:27 +0000 2021' 'Mon Nov 01 12:47:25 +0000 2021' 'Mon Nov 01 12:43:24 +0000 2021' 'Mon Nov 01 12:42:01 +0000 2021' 'Mon Nov 01 12:41:47 +0000 2021' 'Mon Nov 01 12:29:40 +0000 2021' 'Mon Nov 01 12:28:45 +0000 2021' 'Mon Nov 01 12:18:11 +0000 2021' 'Mon Nov 01 12:12:53 +0000 2021' 'Mon Nov 01 12:04:15 +0000 2021' 'Mon Nov 01 12:02:05 +0000 2021' 'Mon Nov 01 10:00:31 +0000 2021' 'Mon Nov 01 06:55:33 +0000 2021' 'Mon Nov 01 02:52:58 +0000 2021' 'Mon Nov 01 02:52:55 +0000 2021' 'Mon Nov 01 02:52:41 +0000 2021' 'Mon Nov 01 02:52:40 +0000 2021' 'Mon Nov 01 02:40:00 +0000 2021' 'Mon Nov 01 01:19:06 +0000 2021' 'Mon Nov 01 00:48:40 +0000 2021' 'Mon Nov 01 00:31:28 +0000 2021' 'Sun Oct 31 23:42:36 +0000 2021' 'Sun Oct 31 23:41:28 +0000 2021' 'Sun Oct 31 23:36:00 +0000 2021' 'Sun Oct 31 23:34:29 +0000 2021' 'Sun Oct 31 23:24:49 +0000 2021' 'Sun Oct 31 23:23:36 +0000 2021' 'Sun Oct 31 23:22:06 +0000 2021' 'Sun Oct 31 22:47:38 +0000 2021' 'Sun Oct 31 21:56:21 +0000 2021' 'Sun Oct 31 21:45:29 +0000 2021' 'Sun Oct 31 21:36:54 +0000 2021' 'Sun Oct 31 21:33:15 +0000 2021' 'Sun Oct 31 20:21:24 +0000 2021' 'Sun Oct 31 20:06:35 +0000 2021' 'Sun Oct 31 19:13:01 +0000 2021' 'Sun Oct 31 19:05:07 +0000 2021' 'Sun Oct 31 18:53:15 +0000 2021' 'Sun Oct 31 18:43:43 +0000 2021' 'Sun Oct 31 18:41:39 +0000 2021' 'Sun Oct 31 18:16:00 +0000 2021' 'Sun Oct 31 17:57:46 +0000 2021' 'Sun Oct 31 17:24:00 +0000 2021' 'Sun Oct 31 17:11:33 +0000 2021' 'Sun Oct 31 17:08:45 +0000 2021' 'Sun Oct 31 16:36:12 +0000 2021' 'Sun Oct 31 16:21:43 +0000 2021' 'Sun Oct 31 16:20:05 +0000 2021' 'Sun Oct 31 16:07:03 +0000 2021' ###Markdown Collect Tweets into MongoDB Install Python librariesYou may need to restart your Jupyter Notebook instance after installed those libraries. ###Code !pip install pymongo !pip install pymongo[srv] !pip install dnspython !pip install tweepy !pip install twitter ###Output Collecting twitter Downloading twitter-1.18.0-py2.py3-none-any.whl (54 kB)  |████████████████████████████████| 54 kB 491 kB/s eta 0:00:01 [?25hInstalling collected packages: twitter Successfully installed twitter-1.18.0 WARNING: You are using pip version 20.0.2; however, version 20.2.4 is available. You should consider upgrading via the '/home/ec2-user/anaconda3/envs/python3/bin/python -m pip install --upgrade pip' command. ###Markdown Import Python libraries ###Code import pymongo from pymongo import MongoClient import json import tweepy import twitter from pprint import pprint import configparser import pandas as pd ###Output _____no_output_____ ###Markdown Load the Authorization Info Save database connection info and API Keys in a config.ini file and use the configparse to load the authorization info. ###Code config = configparser.ConfigParser() config.read('config.ini') CONSUMER_KEY = config['mytwitter']['api_key'] CONSUMER_SECRET = config['mytwitter']['api_secrete'] OAUTH_TOKEN = config['mytwitter']['access_token'] OATH_TOKEN_SECRET = config['mytwitter']['access_secrete'] mongod_connect = config['mymongo']['connection'] ###Output _____no_output_____ ###Markdown Connect to the MongoDB Cluster ###Code client = MongoClient(mongod_connect) db = client.gp25 # use or create a database named demo tweet_collection = db.tweet_collection #use or create a collection named tweet_collection tweet_collection.create_index([("id", pymongo.ASCENDING)],unique = True) # make sure the collected tweets are unique ###Output _____no_output_____ ###Markdown Use the Streaming API to Collect Tweets Authorize the Stream API ###Code stream_auth = tweepy.OAuthHandler(CONSUMER_KEY, CONSUMER_SECRET) stream_auth.set_access_token(OAUTH_TOKEN, OATH_TOKEN_SECRET) strem_api = tweepy.API(stream_auth) ###Output _____no_output_____ ###Markdown Define the query for the Stream API ###Code track = ['election'] # define the keywords, tweets contain election locations = [-78.9326449,38.4150904,-78.8816972,38.4450731] #defin the location, in Harrisonburg, VA ###Output _____no_output_____ ###Markdown The collected tweets will contain 'election' OR are located in Harrisonburg, VA ###Code class MyStreamListener(tweepy.StreamListener): def on_status(self, status): print (status.id_str) try: tweet_collection.insert_one(status._json) except: pass def on_error(self, status_code): if status_code == 420: #returning False in on_data disconnects the stream return False myStreamListener = MyStreamListener() myStream = tweepy.Stream(auth = strem_api.auth, listener=myStreamListener) myStream.filter(track=track)# (locations = locations) #Use either track or locations ###Output 1326627846752768000 1326627846756818944 1326627846807146496 1326627847168020482 1326627847046303746 1326627847209951239 1326627847172202501 1326627847134302208 1326627847482593280 1326627847482433536 1326627847390322689 1326627847386099721 1326627847570681858 1326627847654563851 1326627847897800708 1326627847847342080 1326627848002523136 1326627847960662018 1326627848170295296 1326627848258543617 1326627848329814017 1326627848329850881 1326627848187224065 1326627848459853825 1326627848493309952 1326627848564641793 1326627848539549697 1326627848833159172 1326627848912859136 1326627848983994369 1326627848954634242 1326627849034403841 1326627849105629185 1326627849185488897 1326627849252593666 1326627849051246593 1326627849332092928 1326627849386647552 1326627849365823488 1326627849416175616 1326627849416085506 1326627849642663936 1326627849659314176 1326627849709740033 1326627849852235776 1326627849868988417 1326627849848016896 1326627849760083970 1326627849923538944 1326627848002613261 1326627850124980227 1326627850187927552 1326627850150154240 1326627850204700672 1326627850284363776 1326627850246479872 1326627850393415680 1326627850405998595 1326627850569588736 1326627850707984387 1326627850766733317 1326627850724597760 1326627850770808834 1326627850900926464 1326627850829492224 1326627850854797316 1326627851127414784 1326627851139952641 1326627851324502016 1326627851387461637 1326627851471302659 1326627851567845378 1326627851362213889 1326627851391684614 1326627851639119872 1326627851580436481 1326627851626569728 1326627851731423239 1326627851827773443 1326627851907571714 1326627851840364545 1326627851810955264 1326627851882405890 1326627852066967553 1326627852041777153 1326627852129771520 1326627851651739648 1326627852117274625 1326627852259889154 1326627852410761216 1326627852578660352 1326627852008136704 1326627852750626816 1326627852855382017 1326627852884828160 1326627852909858816 1326627852779982852 1326627852998078466 1326627852926607361 1326627853035859970 1326627854357041159 1326627854411583498 1326627854449209344 1326627854487080961 1326627854638059521 1326627854612881410 1326627854650650627 1326627854717612032 1326627854357057537 1326627854671601666 1326627854281551873 1326627854877032448 1326627854923259906 1326627854801571844 1326627855191695362 1326627855288164353 1326627855225253890 1326627855330136067 1326627855384645632 1326627855418060801 1326627855346905091 1326627855514562560 1326627855573397507 1326627855527256071 1326627855573381121 1326627855669719042 1326627855581802496 1326627855544053766 1326627855866925065 1326627855912980480 1326627855833440257 1326627855988613120 1326627856047337474 1326627856101695488 1326627856005308416 1326627856164663296 1326627856198361094 1326627856273846272 1326627856244383744 1326627856273842180 1326627856164728833 1326627856403722241 1326627856391303169 1326627856567382019 1326627856592609282 1326627856592457729 1326627856684888064 1326627856898777095 1326627857007845378 1326627857083330565 1326627858421313536 1326627858446491649 1326627858391863296 1326627858467450881 1326627858416988160 1326627858626867201 1326627858794438656 1326627857779585025 1326627859197276163 1326627859037827074 1326627859302141952 1326627859411177473 1326627859478302722 1326627859432148992 1326627859503476739 1326627859650080768 1326627859717386242 1326627859868217344 1326627860056952832 1326627860149395464 1326627860069617665 1326627860094701568 1326627860216504320 1326627860480741376 1326627860371689472 1326627860518461440 1326627860598190083 1326627860724011011 1326627860778344450 1326627860795318273 1326627860862349312 1326627860908548096 1326627860937928707 1326627861013389312 1326627861013426187 1326627861084704769 1326627861097140224 1326627861189423105 1326627861235720192 1326627861248303105 1326627861176872961 1326627860849676288 1326627861437042688 1326627861369741312 1326627861445320704 1326627861575458817 1326627861445406720 1326627861470601228 1326627861613211648 1326627861696970752 1326627862732918784 1326627862632402946 1326627862745583618 1326627862762442752 1326627862854688770 1326627862582079488 1326627862934253568 1326627863106359300 1326627863047647234 1326627863181844480 1326627863584518145 1326627863588724736 1326627863508873216 1326627863592841222 1326627863655690240 1326627863697747968 1326627863651553289 1326627863760674821 1326627863747964929 1326627863869665281 1326627863978708999 1326627863836110853 1326627864087687168 1326627864192671744 1326627864196882433 1326627864226099200 1326627864347750400 1326627864419201029 1326627864477773824 1326627864519831553 1326627864553299969 1326627864700219392 1326627864670838786 1326627864670834688 1326627864712794112 1326627864817627137 1326627864981204994 1326627865060765696 1326627865304051713 1326627865232896001 1326627865312415744 1326627865337737216 1326627865341857793 1326627865358643200 1326627865484546055 1326627865555767296 1326627865480355840 1326627865681670144 1326627865681649664 1326627865530675201 1326627866826600449 1326627866935701504 1326627866935635973 1326627866889641990 1326627867061592067 1326627867007053831 1326627867149656065 1326627867220992001 1326627867308994560 1326627867367624705 1326627867577495554 1326627867619450880 1326627867564838914 1326627867669696518 1326627867585875969 1326627867669786631 1326627867770433537 1326627867673944064 1326627867883663363 1326627867950804995 1326627867980148738 1326627867858448385 1326627867954962432 1326627868030365696 1326627868206641153 1326627868160483328 1326627868382670848 1326627868361678848 1326627868416208896 1326627868416356352 1326627868554694662 1326627868382797826 1326627868588306440 1326627868613472256 1326627868990955520 1326627869066485761 1326627868965744646 1326627869188034561 1326627869284511745 1326627869242617857 1326627869158727686 1326627869351579649 1326627869343309826 1326627869422985217 1326627869490094081 1326627869691432962 1326627869506801664 1326627869771128834 1326627869792067587 1326627869968248833 1326627871029391362 1326627871071367172 1326627871151038465 1326627871104802816 1326627871222161408 1326627871159422976 1326627871373332486 1326627871432073216 1326627871557890055 1326627871683715072 1326627871671029760 1326627871859695616 1326627871838892034 1326627871977320451 1326627871880867843 1326627872015065096 1326627872061202434 1326627871855685636 1326627872161853448 1326627872073674752 1326627872199507968 1326627872262426624 1326627872157528064 1326627872157675529 1326627872304472064 1326627872295903232 1326627872363130884 1326627872132493317 1326627872489021441 1326627872463855617 1326627872325373952 1326627872665112579 1326627872673468416 1326627872661000199 1326627872598073346 1326627872799404032 1326627872786812930 1326627872811835392 1326627872899956737 1326627872887496706 1326627872153481216 1326627872929320961 1326627872845537283 1326627872895868929 1326627873059450883 1326627872954601477 1326627873071960067 1326627872996548612 1326627873004818432 1326627873025912833 1326627875328385024 1326627875169198083 1326627875311792129 1326627875441815557 1326627875487961088 1326627875420827648 1326627875567513601 1326627875513114624 1326627875588501504 1326627875684904960 1326627875668205568 1326627875571773440 1326627875886428162 1326627876125487104 1326627876129697793 1326627875890618368 1326627876347678720 1326627876507152387 1326627876557479938 1326627876599369740 1326627875886346245 1326627876687339520 1326627876670640128 1326627876775616514 1326627876595175426 1326627876763033601 1326627876557500416 1326627876821737473 1326627876997885956 1326627876846891015 1326627877203271680 1326627877127925761 1326627877291495425 1326627877203349505 1326627877283098630 1326627877379575814 1326627877480263687 1326627877505232897 1326627877496893441 1326627877652205572 1326627877727531008 1326627877593473034 1326627877878640641 1326627878037934081 1326627877962600450 1326627878017126401 1326627878142844928 1326627878025351168 1326627878184775680 1326627878428160003 1326627879401254915 1326627879879278593 1326627879757688833 1326627879942230017 1326627879954735104 1326627879988330496 1326627879908593665 1326627880005144578 1326627879829073921 1326627880294522880 ###Markdown Use the REST API to Collect Tweets Authorize the REST API ###Code rest_auth = twitter.oauth.OAuth(OAUTH_TOKEN,OATH_TOKEN_SECRET,CONSUMER_KEY,CONSUMER_SECRET) rest_api = twitter.Twitter(auth=rest_auth) ###Output _____no_output_____ ###Markdown Define the query for the REST API ###Code count = 100 #number of returned tweets, default and max is 100 geocode = "38.4392897,-78.9412224,50mi" # defin the location, in Harrisonburg, VA q = "election" #define the keywords, tweets contain election ###Output _____no_output_____ ###Markdown The collected tweets will contain 'election' AND are located in Harrisonburg, VA ###Code search_results = rest_api.search.tweets( count=count,q=q, geocode=geocode) #you can use both q and geocode statuses = search_results["statuses"] since_id_new = statuses[-1]['id'] for statuse in statuses: try: tweet_collection.insert_one(statuse) pprint(statuse['created_at'])# print the date of the collected tweets except: pass ###Output 'Wed Nov 11 20:52:30 +0000 2020' 'Wed Nov 11 20:51:50 +0000 2020' 'Wed Nov 11 20:51:04 +0000 2020' 'Wed Nov 11 20:49:04 +0000 2020' 'Wed Nov 11 20:46:29 +0000 2020' 'Wed Nov 11 20:44:41 +0000 2020' 'Wed Nov 11 20:42:18 +0000 2020' 'Wed Nov 11 20:38:24 +0000 2020' 'Wed Nov 11 20:37:59 +0000 2020' 'Wed Nov 11 20:36:48 +0000 2020' 'Wed Nov 11 20:35:38 +0000 2020' 'Wed Nov 11 20:34:41 +0000 2020' 'Wed Nov 11 20:34:35 +0000 2020' 'Wed Nov 11 20:33:18 +0000 2020' 'Wed Nov 11 20:33:18 +0000 2020' 'Wed Nov 11 20:32:59 +0000 2020' 'Wed Nov 11 20:32:47 +0000 2020' 'Wed Nov 11 20:32:26 +0000 2020' 'Wed Nov 11 20:29:56 +0000 2020' 'Wed Nov 11 20:28:32 +0000 2020' 'Wed Nov 11 20:27:26 +0000 2020' 'Wed Nov 11 20:25:01 +0000 2020' 'Wed Nov 11 20:24:05 +0000 2020' 'Wed Nov 11 20:21:47 +0000 2020' 'Wed Nov 11 20:20:17 +0000 2020' 'Wed Nov 11 20:16:41 +0000 2020' 'Wed Nov 11 20:16:05 +0000 2020' 'Wed Nov 11 20:16:05 +0000 2020' 'Wed Nov 11 20:15:31 +0000 2020' 'Wed Nov 11 20:14:42 +0000 2020' 'Wed Nov 11 20:13:35 +0000 2020' 'Wed Nov 11 20:12:58 +0000 2020' 'Wed Nov 11 20:12:51 +0000 2020' 'Wed Nov 11 20:10:04 +0000 2020' 'Wed Nov 11 20:06:46 +0000 2020' 'Wed Nov 11 20:06:29 +0000 2020' 'Wed Nov 11 20:05:36 +0000 2020' 'Wed Nov 11 20:04:29 +0000 2020' ###Markdown Continue fetching early tweets with the same query. YOU WILL REACH YOUR RATE LIMIT VERY FAST ###Code since_id_old = 0 while(since_id_new != since_id_old): since_id_old = since_id_new search_results = rest_api.search.tweets( count=count,q=q, geocode=geocode, max_id= since_id_new) statuses = search_results["statuses"] since_id_new = statuses[-1]['id'] for statuse in statuses: try: tweet_collection.insert_one(statuse) pprint(statuse['created_at']) # print the date of the collected tweets except: pass ###Output 'Wed Nov 11 20:03:40 +0000 2020' 'Wed Nov 11 20:03:28 +0000 2020' 'Wed Nov 11 20:02:35 +0000 2020' 'Wed Nov 11 20:01:59 +0000 2020' 'Wed Nov 11 20:01:12 +0000 2020' 'Wed Nov 11 20:01:11 +0000 2020' 'Wed Nov 11 20:00:57 +0000 2020' 'Wed Nov 11 20:00:28 +0000 2020' 'Wed Nov 11 19:54:33 +0000 2020' 'Wed Nov 11 19:52:54 +0000 2020' 'Wed Nov 11 19:51:20 +0000 2020' 'Wed Nov 11 19:48:32 +0000 2020' 'Wed Nov 11 19:48:13 +0000 2020' 'Wed Nov 11 19:48:01 +0000 2020' 'Wed Nov 11 19:47:31 +0000 2020' 'Wed Nov 11 19:47:06 +0000 2020' 'Wed Nov 11 19:45:40 +0000 2020' 'Wed Nov 11 19:45:10 +0000 2020' 'Wed Nov 11 19:44:43 +0000 2020' 'Wed Nov 11 19:44:29 +0000 2020' 'Wed Nov 11 19:41:12 +0000 2020' 'Wed Nov 11 19:41:12 +0000 2020' 'Wed Nov 11 19:41:05 +0000 2020' 'Wed Nov 11 19:38:37 +0000 2020' 'Wed Nov 11 19:36:27 +0000 2020' 'Wed Nov 11 19:33:42 +0000 2020' 'Wed Nov 11 19:33:05 +0000 2020' 'Wed Nov 11 19:32:20 +0000 2020' 'Wed Nov 11 19:30:45 +0000 2020' 'Wed Nov 11 19:30:07 +0000 2020' 'Wed Nov 11 19:29:47 +0000 2020' 'Wed Nov 11 19:28:00 +0000 2020' 'Wed Nov 11 19:26:42 +0000 2020' 'Wed Nov 11 19:25:33 +0000 2020' 'Wed Nov 11 19:25:11 +0000 2020' 'Wed Nov 11 19:25:00 +0000 2020' 'Wed Nov 11 19:24:20 +0000 2020' 'Wed Nov 11 19:21:37 +0000 2020' 'Wed Nov 11 19:21:19 +0000 2020' 'Wed Nov 11 19:20:49 +0000 2020' 'Wed Nov 11 19:20:10 +0000 2020' 'Wed Nov 11 19:18:11 +0000 2020' 'Wed Nov 11 19:18:09 +0000 2020' 'Wed Nov 11 19:17:35 +0000 2020' 'Wed Nov 11 19:15:34 +0000 2020' 'Wed Nov 11 19:15:11 +0000 2020' 'Wed Nov 11 19:14:58 +0000 2020' 'Wed Nov 11 19:14:46 +0000 2020' 'Wed Nov 11 19:14:22 +0000 2020' 'Wed Nov 11 19:13:53 +0000 2020' 'Wed Nov 11 19:13:46 +0000 2020' 'Wed Nov 11 19:13:30 +0000 2020' 'Wed Nov 11 19:13:25 +0000 2020' 'Wed Nov 11 19:13:03 +0000 2020' 'Wed Nov 11 19:11:54 +0000 2020' 'Wed Nov 11 19:11:20 +0000 2020' 'Wed Nov 11 19:11:17 +0000 2020' 'Wed Nov 11 19:11:15 +0000 2020' 'Wed Nov 11 19:10:49 +0000 2020' 'Wed Nov 11 19:10:06 +0000 2020' 'Wed Nov 11 19:09:51 +0000 2020' 'Wed Nov 11 19:09:43 +0000 2020' 'Wed Nov 11 19:09:09 +0000 2020' 'Wed Nov 11 19:08:14 +0000 2020' 'Wed Nov 11 19:08:00 +0000 2020' 'Wed Nov 11 19:06:12 +0000 2020' 'Wed Nov 11 19:03:30 +0000 2020' 'Wed Nov 11 19:03:29 +0000 2020' 'Wed Nov 11 19:00:47 +0000 2020' 'Wed Nov 11 19:00:18 +0000 2020' 'Wed Nov 11 18:55:44 +0000 2020' 'Wed Nov 11 18:55:16 +0000 2020' 'Wed Nov 11 18:54:21 +0000 2020' 'Wed Nov 11 18:53:17 +0000 2020' 'Wed Nov 11 18:53:16 +0000 2020' 'Wed Nov 11 18:53:09 +0000 2020' 'Wed Nov 11 18:52:46 +0000 2020' 'Wed Nov 11 18:52:37 +0000 2020' 'Wed Nov 11 18:51:49 +0000 2020' 'Wed Nov 11 18:51:41 +0000 2020' 'Wed Nov 11 18:48:52 +0000 2020' 'Wed Nov 11 18:48:13 +0000 2020' 'Wed Nov 11 18:47:39 +0000 2020' 'Wed Nov 11 18:46:42 +0000 2020' 'Wed Nov 11 18:46:40 +0000 2020' 'Wed Nov 11 18:46:23 +0000 2020' 'Wed Nov 11 18:45:55 +0000 2020' 'Wed Nov 11 18:45:16 +0000 2020' 'Wed Nov 11 18:45:05 +0000 2020' 'Wed Nov 11 18:44:39 +0000 2020' 'Wed Nov 11 18:44:34 +0000 2020' 'Wed Nov 11 18:43:30 +0000 2020' 'Wed Nov 11 18:43:13 +0000 2020' 'Wed Nov 11 18:43:03 +0000 2020' 'Wed Nov 11 18:41:34 +0000 2020' 'Wed Nov 11 18:40:17 +0000 2020' 'Wed Nov 11 18:40:06 +0000 2020' 'Wed Nov 11 18:39:55 +0000 2020' 'Wed Nov 11 18:39:53 +0000 2020' 'Wed Nov 11 18:39:45 +0000 2020' 'Wed Nov 11 18:39:25 +0000 2020' 'Wed Nov 11 18:39:11 +0000 2020' 'Wed Nov 11 18:39:06 +0000 2020' 'Wed Nov 11 18:38:55 +0000 2020' 'Wed Nov 11 18:38:46 +0000 2020' 'Wed Nov 11 18:38:39 +0000 2020' 'Wed Nov 11 18:38:32 +0000 2020' 'Wed Nov 11 18:38:26 +0000 2020' 'Wed Nov 11 18:38:03 +0000 2020' 'Wed Nov 11 18:37:33 +0000 2020' 'Wed Nov 11 18:37:30 +0000 2020' 'Wed Nov 11 18:36:59 +0000 2020' 'Wed Nov 11 18:36:52 +0000 2020' 'Wed Nov 11 18:36:48 +0000 2020' 'Wed Nov 11 18:36:42 +0000 2020' 'Wed Nov 11 18:36:41 +0000 2020' 'Wed Nov 11 18:36:32 +0000 2020' 'Wed Nov 11 18:35:17 +0000 2020' 'Wed Nov 11 18:34:26 +0000 2020' 'Wed Nov 11 18:34:21 +0000 2020' 'Wed Nov 11 18:34:09 +0000 2020' 'Wed Nov 11 18:33:59 +0000 2020' 'Wed Nov 11 18:33:47 +0000 2020' 'Wed Nov 11 18:32:40 +0000 2020' 'Wed Nov 11 18:32:34 +0000 2020' 'Wed Nov 11 18:32:02 +0000 2020' 'Wed Nov 11 18:31:58 +0000 2020' 'Wed Nov 11 18:31:46 +0000 2020' 'Wed Nov 11 18:31:43 +0000 2020' 'Wed Nov 11 18:31:00 +0000 2020' 'Wed Nov 11 18:30:32 +0000 2020' 'Wed Nov 11 18:30:18 +0000 2020' 'Wed Nov 11 18:29:30 +0000 2020' 'Wed Nov 11 18:28:11 +0000 2020' 'Wed Nov 11 18:28:06 +0000 2020' 'Wed Nov 11 18:27:59 +0000 2020' 'Wed Nov 11 18:27:45 +0000 2020' 'Wed Nov 11 18:27:42 +0000 2020' 'Wed Nov 11 18:27:37 +0000 2020' 'Wed Nov 11 18:27:32 +0000 2020' 'Wed Nov 11 18:27:06 +0000 2020' 'Wed Nov 11 18:27:05 +0000 2020' 'Wed Nov 11 18:26:45 +0000 2020' 'Wed Nov 11 18:26:27 +0000 2020' 'Wed Nov 11 18:26:13 +0000 2020' 'Wed Nov 11 18:26:11 +0000 2020' 'Wed Nov 11 18:25:33 +0000 2020' 'Wed Nov 11 18:25:30 +0000 2020' 'Wed Nov 11 18:25:29 +0000 2020' 'Wed Nov 11 18:25:24 +0000 2020' 'Wed Nov 11 18:25:07 +0000 2020' 'Wed Nov 11 18:24:59 +0000 2020' 'Wed Nov 11 18:24:55 +0000 2020' 'Wed Nov 11 18:24:44 +0000 2020' 'Wed Nov 11 18:24:44 +0000 2020' 'Wed Nov 11 18:24:37 +0000 2020' 'Wed Nov 11 18:24:21 +0000 2020' 'Wed Nov 11 18:24:16 +0000 2020' 'Wed Nov 11 18:24:16 +0000 2020' 'Wed Nov 11 18:24:14 +0000 2020' 'Wed Nov 11 18:24:12 +0000 2020' 'Wed Nov 11 18:23:52 +0000 2020' 'Wed Nov 11 18:23:50 +0000 2020' 'Wed Nov 11 18:23:35 +0000 2020' 'Wed Nov 11 18:23:34 +0000 2020' 'Wed Nov 11 18:23:29 +0000 2020' 'Wed Nov 11 18:23:28 +0000 2020' 'Wed Nov 11 18:23:24 +0000 2020' 'Wed Nov 11 18:23:16 +0000 2020' 'Wed Nov 11 18:23:07 +0000 2020' 'Wed Nov 11 18:23:03 +0000 2020' 'Wed Nov 11 18:23:00 +0000 2020' 'Wed Nov 11 18:22:50 +0000 2020' 'Wed Nov 11 18:22:48 +0000 2020' 'Wed Nov 11 18:22:44 +0000 2020' 'Wed Nov 11 18:22:41 +0000 2020' 'Wed Nov 11 18:22:37 +0000 2020' 'Wed Nov 11 18:22:31 +0000 2020' 'Wed Nov 11 18:22:26 +0000 2020' 'Wed Nov 11 18:22:01 +0000 2020' 'Wed Nov 11 18:21:59 +0000 2020' 'Wed Nov 11 18:21:58 +0000 2020' 'Wed Nov 11 18:21:57 +0000 2020' 'Wed Nov 11 18:21:51 +0000 2020' 'Wed Nov 11 18:21:43 +0000 2020' 'Wed Nov 11 18:21:41 +0000 2020' 'Wed Nov 11 18:21:31 +0000 2020' 'Wed Nov 11 18:21:20 +0000 2020' 'Wed Nov 11 18:21:14 +0000 2020' 'Wed Nov 11 18:21:13 +0000 2020' 'Wed Nov 11 18:21:12 +0000 2020' 'Wed Nov 11 18:21:06 +0000 2020' 'Wed Nov 11 18:21:03 +0000 2020' 'Wed Nov 11 18:21:03 +0000 2020' 'Wed Nov 11 18:21:02 +0000 2020' 'Wed Nov 11 18:20:53 +0000 2020' 'Wed Nov 11 18:20:50 +0000 2020' 'Wed Nov 11 18:20:45 +0000 2020' 'Wed Nov 11 18:20:42 +0000 2020' 'Wed Nov 11 18:20:35 +0000 2020' 'Wed Nov 11 18:20:29 +0000 2020' 'Wed Nov 11 18:20:25 +0000 2020' 'Wed Nov 11 18:20:06 +0000 2020' 'Wed Nov 11 18:19:59 +0000 2020' 'Wed Nov 11 18:19:45 +0000 2020' 'Wed Nov 11 18:19:41 +0000 2020' 'Wed Nov 11 18:19:38 +0000 2020' 'Wed Nov 11 18:19:38 +0000 2020' 'Wed Nov 11 18:19:28 +0000 2020' 'Wed Nov 11 18:19:25 +0000 2020' 'Wed Nov 11 18:19:22 +0000 2020' 'Wed Nov 11 18:19:19 +0000 2020' 'Wed Nov 11 18:19:18 +0000 2020' 'Wed Nov 11 18:19:16 +0000 2020' 'Wed Nov 11 18:19:12 +0000 2020' 'Wed Nov 11 18:19:12 +0000 2020' 'Wed Nov 11 18:18:54 +0000 2020' 'Wed Nov 11 18:18:53 +0000 2020' 'Wed Nov 11 18:18:51 +0000 2020' 'Wed Nov 11 18:18:49 +0000 2020' 'Wed Nov 11 18:18:48 +0000 2020' 'Wed Nov 11 18:18:42 +0000 2020' 'Wed Nov 11 18:18:41 +0000 2020' 'Wed Nov 11 18:18:40 +0000 2020' 'Wed Nov 11 18:18:18 +0000 2020' 'Wed Nov 11 18:17:58 +0000 2020' 'Wed Nov 11 18:17:23 +0000 2020' 'Wed Nov 11 18:16:17 +0000 2020' 'Wed Nov 11 18:15:14 +0000 2020' 'Wed Nov 11 18:15:12 +0000 2020' 'Wed Nov 11 18:15:03 +0000 2020' 'Wed Nov 11 18:13:05 +0000 2020' 'Wed Nov 11 18:12:31 +0000 2020' 'Wed Nov 11 18:11:40 +0000 2020' 'Wed Nov 11 18:11:08 +0000 2020' 'Wed Nov 11 18:10:55 +0000 2020' 'Wed Nov 11 18:10:43 +0000 2020' 'Wed Nov 11 18:10:08 +0000 2020' 'Wed Nov 11 18:09:20 +0000 2020' 'Wed Nov 11 18:09:02 +0000 2020' 'Wed Nov 11 18:08:57 +0000 2020' 'Wed Nov 11 18:08:04 +0000 2020' 'Wed Nov 11 18:06:20 +0000 2020' 'Wed Nov 11 18:06:02 +0000 2020' 'Wed Nov 11 18:05:20 +0000 2020' 'Wed Nov 11 18:04:21 +0000 2020' 'Wed Nov 11 18:04:11 +0000 2020' 'Wed Nov 11 18:04:04 +0000 2020' 'Wed Nov 11 18:03:41 +0000 2020' 'Wed Nov 11 18:02:58 +0000 2020' 'Wed Nov 11 18:01:43 +0000 2020' 'Wed Nov 11 18:01:17 +0000 2020' 'Wed Nov 11 18:01:02 +0000 2020' 'Wed Nov 11 18:00:19 +0000 2020' 'Wed Nov 11 17:59:59 +0000 2020' 'Wed Nov 11 17:59:15 +0000 2020' 'Wed Nov 11 17:59:08 +0000 2020' ###Markdown View the Collected Tweets Print the number of tweets and unique twitter users ###Code print(tweet_collection.estimated_document_count())# number of tweets collected user_cursor = tweet_collection.distinct("user.id") print (len(user_cursor)) # number of unique Twitter users ###Output 1086 1038 ###Markdown Create a text index and print the Tweets containing specific keywords. ###Code tweet_collection.create_index([("text", pymongo.TEXT)], name='text_index', default_language='english') # create a text index ###Output _____no_output_____ ###Markdown Create a cursor to query tweets with the created index ###Code tweet_cursor = tweet_collection.find({"$text": {"$search": "vote"}}) # return tweets contain vote ###Output _____no_output_____ ###Markdown Use pprint to display tweets ###Code for document in tweet_cursor[0:10]: # display the first 10 tweets from the query try: print ('----') # pprint (document) # use pprint to print the entire tweet document print ('name:', document["user"]["name"]) # user name print ('text:', document["text"]) # tweets except: print ("***error in encoding") pass tweet_cursor = tweet_collection.find({"$text": {"$search": "vote"}}) # return tweets contain vote ###Output _____no_output_____ ###Markdown Use pandas to display tweets ###Code tweet_df = pd.DataFrame(list(tweet_cursor )) tweet_df[:10] #display the first 10 tweets tweet_df["favorite_count"].hist() # create a histogram show the favorite count ###Output _____no_output_____ ###Markdown Know second and third lines for quiz ###Code client = MongoClient(mongod_connect) db = client.demo # use or create a database named demo tweet_collection = db.tweet_collection #use or create a collection named tweet_collection tweet_collection.create_index([("id", pymongo.ASCENDING)],unique = True) # make sure the collected tweets are unique rest_auth = twitter.oauth.OAuth(OAUTH_TOKEN,OATH_TOKEN_SECRET,CONSUMER_KEY,CONSUMER_SECRET) rest_api = twitter.Twitter(auth=rest_auth) count = 100 #number of returned tweets, default and max is 100 geocode = "38.4392897,-78.9412224,50mi" # defin the location, in Harrisonburg, VA q = "election" #define the keywords, tweets contain election search_results = rest_api.search.tweets( count=count,q=q, geocode=geocode) #you can use both q and geocode statuses = search_results["statuses"] since_id_new = statuses[-1]['id'] for statuse in statuses: try: tweet_collection.insert_one(statuse) pprint(statuse['created_at'])# print the date of the collected tweets except: pass ###Output 'Mon Nov 01 21:00:59 +0000 2021' 'Mon Nov 01 20:53:30 +0000 2021' 'Mon Nov 01 20:53:02 +0000 2021' 'Mon Nov 01 20:52:26 +0000 2021' 'Mon Nov 01 20:37:07 +0000 2021' 'Mon Nov 01 20:22:42 +0000 2021' 'Mon Nov 01 20:05:11 +0000 2021' 'Mon Nov 01 19:25:39 +0000 2021' 'Mon Nov 01 19:13:57 +0000 2021' 'Mon Nov 01 19:08:46 +0000 2021' 'Mon Nov 01 19:07:19 +0000 2021' 'Mon Nov 01 19:00:47 +0000 2021' 'Mon Nov 01 18:49:22 +0000 2021' 'Mon Nov 01 18:33:51 +0000 2021' 'Mon Nov 01 18:30:09 +0000 2021' 'Mon Nov 01 18:17:19 +0000 2021' 'Mon Nov 01 18:08:49 +0000 2021' 'Mon Nov 01 18:02:33 +0000 2021' 'Mon Nov 01 18:00:10 +0000 2021' 'Mon Nov 01 17:59:53 +0000 2021' 'Mon Nov 01 17:57:53 +0000 2021' 'Mon Nov 01 17:27:06 +0000 2021' 'Mon Nov 01 17:18:56 +0000 2021' 'Mon Nov 01 17:12:46 +0000 2021' 'Mon Nov 01 17:07:01 +0000 2021' 'Mon Nov 01 16:57:09 +0000 2021' 'Mon Nov 01 16:48:57 +0000 2021' 'Mon Nov 01 16:45:07 +0000 2021' 'Mon Nov 01 16:42:37 +0000 2021' 'Mon Nov 01 16:40:18 +0000 2021' 'Mon Nov 01 16:29:51 +0000 2021' 'Mon Nov 01 16:29:08 +0000 2021' 'Mon Nov 01 16:28:58 +0000 2021' 'Mon Nov 01 16:05:00 +0000 2021' 'Mon Nov 01 16:02:05 +0000 2021' 'Mon Nov 01 15:38:09 +0000 2021' 'Mon Nov 01 15:35:33 +0000 2021' 'Mon Nov 01 15:12:24 +0000 2021' 'Mon Nov 01 15:04:06 +0000 2021' 'Mon Nov 01 14:59:37 +0000 2021' 'Mon Nov 01 14:32:26 +0000 2021' 'Mon Nov 01 14:30:26 +0000 2021' 'Mon Nov 01 13:51:08 +0000 2021' 'Mon Nov 01 13:47:27 +0000 2021' 'Mon Nov 01 13:38:55 +0000 2021' 'Mon Nov 01 13:36:56 +0000 2021' 'Mon Nov 01 13:13:38 +0000 2021' 'Mon Nov 01 13:09:25 +0000 2021' 'Mon Nov 01 13:03:43 +0000 2021' 'Mon Nov 01 12:56:32 +0000 2021' 'Mon Nov 01 12:55:41 +0000 2021' 'Mon Nov 01 12:47:27 +0000 2021' 'Mon Nov 01 12:47:25 +0000 2021' 'Mon Nov 01 12:43:24 +0000 2021' 'Mon Nov 01 12:42:01 +0000 2021' 'Mon Nov 01 12:41:47 +0000 2021' 'Mon Nov 01 12:29:40 +0000 2021' 'Mon Nov 01 12:28:45 +0000 2021' 'Mon Nov 01 12:18:11 +0000 2021' 'Mon Nov 01 12:12:53 +0000 2021' 'Mon Nov 01 12:04:15 +0000 2021' 'Mon Nov 01 12:02:05 +0000 2021' 'Mon Nov 01 10:00:31 +0000 2021' 'Mon Nov 01 06:55:33 +0000 2021' 'Mon Nov 01 02:52:58 +0000 2021' 'Mon Nov 01 02:52:55 +0000 2021' 'Mon Nov 01 02:52:41 +0000 2021' 'Mon Nov 01 02:52:40 +0000 2021' 'Mon Nov 01 02:40:00 +0000 2021' 'Mon Nov 01 01:19:06 +0000 2021' 'Mon Nov 01 00:48:40 +0000 2021' 'Mon Nov 01 00:31:28 +0000 2021' 'Sun Oct 31 23:42:36 +0000 2021' 'Sun Oct 31 23:41:28 +0000 2021' 'Sun Oct 31 23:36:00 +0000 2021' 'Sun Oct 31 23:34:29 +0000 2021' 'Sun Oct 31 23:24:49 +0000 2021' 'Sun Oct 31 23:23:36 +0000 2021' 'Sun Oct 31 23:22:06 +0000 2021' 'Sun Oct 31 22:47:38 +0000 2021' 'Sun Oct 31 21:56:21 +0000 2021' 'Sun Oct 31 21:45:29 +0000 2021' 'Sun Oct 31 21:36:54 +0000 2021' 'Sun Oct 31 21:33:15 +0000 2021' 'Sun Oct 31 20:21:24 +0000 2021' 'Sun Oct 31 20:06:35 +0000 2021' 'Sun Oct 31 19:13:01 +0000 2021' 'Sun Oct 31 19:05:07 +0000 2021' 'Sun Oct 31 18:53:15 +0000 2021' 'Sun Oct 31 18:43:43 +0000 2021' 'Sun Oct 31 18:41:39 +0000 2021' 'Sun Oct 31 18:16:00 +0000 2021' 'Sun Oct 31 17:57:46 +0000 2021' 'Sun Oct 31 17:24:00 +0000 2021' 'Sun Oct 31 17:11:33 +0000 2021' 'Sun Oct 31 17:08:45 +0000 2021' 'Sun Oct 31 16:36:12 +0000 2021' 'Sun Oct 31 16:21:43 +0000 2021' 'Sun Oct 31 16:20:05 +0000 2021' 'Sun Oct 31 16:07:03 +0000 2021' ###Markdown Lec 9 Install Python Libraries ###Code !pip install pymongo !pip install pymongo[srv] !pip install dnspython !pip install tweepy !pip install twitter import pymongo from pymongo import MongoClient import json import tweepy import twitter from pprint import pprint import configparser import pandas as pd ###Output _____no_output_____ ###Markdown Load the Authorization Info ###Code config = configparser.ConfigParser() config.read('config.ini') CONSUMER_KEY = config['mytwitter']['api_key'] CONSUMER_SECRET = config['mytwitter']['api_secrete'] OAUTH_TOKEN = config['mytwitter']['access_token'] OATH_TOKEN_SECRET = config['mytwitter']['access_secrete'] mongod_connect = config['mymongo']['connection'] ###Output _____no_output_____ ###Markdown Connect to the MongoDB Cluster ###Code client = MongoClient(mongod_connect) db = client.demo # use or create a database named demo tweet_collection = db.tweet_collection #use or create a collection named tweet_collection tweet_collection.create_index([("id", pymongo.ASCENDING)],unique = True) # make sure the collected tweets are unique ###Output _____no_output_____ ###Markdown Use the REST API to Collect Tweets ###Code rest_auth = twitter.oauth.OAuth(OAUTH_TOKEN,OATH_TOKEN_SECRET,CONSUMER_KEY,CONSUMER_SECRET) rest_api = twitter.Twitter(auth=rest_auth) count = 100 #number of returned tweets, default and max is 100 geocode = "38.4392897,-78.9412224,50mi" # defin the location, in Harrisonburg, VA q = "election" #define the keywords, tweets contain election search_results = rest_api.search.tweets( count=count,q=q, geocode=geocode) #you can use both q and geocode statuses = search_results["statuses"] since_id_new = statuses[-1]['id'] for statuse in statuses: try: tweet_collection.insert_one(statuse) pprint(statuse['created_at'])# print the date of the collected tweets except: pass ###Output 'Mon Nov 01 21:00:59 +0000 2021' 'Mon Nov 01 20:53:30 +0000 2021' 'Mon Nov 01 20:53:02 +0000 2021' 'Mon Nov 01 20:52:26 +0000 2021' 'Mon Nov 01 20:37:07 +0000 2021' 'Mon Nov 01 20:22:42 +0000 2021' 'Mon Nov 01 20:05:11 +0000 2021' 'Mon Nov 01 19:25:39 +0000 2021' 'Mon Nov 01 19:13:57 +0000 2021' 'Mon Nov 01 19:08:46 +0000 2021' 'Mon Nov 01 19:07:19 +0000 2021' 'Mon Nov 01 19:00:47 +0000 2021' 'Mon Nov 01 18:49:22 +0000 2021' 'Mon Nov 01 18:33:51 +0000 2021' 'Mon Nov 01 18:30:09 +0000 2021' 'Mon Nov 01 18:17:19 +0000 2021' 'Mon Nov 01 18:08:49 +0000 2021' 'Mon Nov 01 18:02:33 +0000 2021' 'Mon Nov 01 18:00:10 +0000 2021' 'Mon Nov 01 17:59:53 +0000 2021' 'Mon Nov 01 17:57:53 +0000 2021' 'Mon Nov 01 17:27:06 +0000 2021' 'Mon Nov 01 17:18:56 +0000 2021' 'Mon Nov 01 17:12:46 +0000 2021' 'Mon Nov 01 17:07:01 +0000 2021' 'Mon Nov 01 16:57:09 +0000 2021' 'Mon Nov 01 16:48:57 +0000 2021' 'Mon Nov 01 16:45:07 +0000 2021' 'Mon Nov 01 16:42:37 +0000 2021' 'Mon Nov 01 16:40:18 +0000 2021' 'Mon Nov 01 16:29:51 +0000 2021' 'Mon Nov 01 16:29:08 +0000 2021' 'Mon Nov 01 16:28:58 +0000 2021' 'Mon Nov 01 16:05:00 +0000 2021' 'Mon Nov 01 16:02:05 +0000 2021' 'Mon Nov 01 15:38:09 +0000 2021' 'Mon Nov 01 15:35:33 +0000 2021' 'Mon Nov 01 15:12:24 +0000 2021' 'Mon Nov 01 15:04:06 +0000 2021' 'Mon Nov 01 14:59:37 +0000 2021' 'Mon Nov 01 14:32:26 +0000 2021' 'Mon Nov 01 14:30:26 +0000 2021' 'Mon Nov 01 13:51:08 +0000 2021' 'Mon Nov 01 13:47:27 +0000 2021' 'Mon Nov 01 13:38:55 +0000 2021' 'Mon Nov 01 13:36:56 +0000 2021' 'Mon Nov 01 13:13:38 +0000 2021' 'Mon Nov 01 13:09:25 +0000 2021' 'Mon Nov 01 13:03:43 +0000 2021' 'Mon Nov 01 12:56:32 +0000 2021' 'Mon Nov 01 12:55:41 +0000 2021' 'Mon Nov 01 12:47:27 +0000 2021' 'Mon Nov 01 12:47:25 +0000 2021' 'Mon Nov 01 12:43:24 +0000 2021' 'Mon Nov 01 12:42:01 +0000 2021' 'Mon Nov 01 12:41:47 +0000 2021' 'Mon Nov 01 12:29:40 +0000 2021' 'Mon Nov 01 12:28:45 +0000 2021' 'Mon Nov 01 12:18:11 +0000 2021' 'Mon Nov 01 12:12:53 +0000 2021' 'Mon Nov 01 12:04:15 +0000 2021' 'Mon Nov 01 12:02:05 +0000 2021' 'Mon Nov 01 10:00:31 +0000 2021' 'Mon Nov 01 06:55:33 +0000 2021' 'Mon Nov 01 02:52:58 +0000 2021' 'Mon Nov 01 02:52:55 +0000 2021' 'Mon Nov 01 02:52:41 +0000 2021' 'Mon Nov 01 02:52:40 +0000 2021' 'Mon Nov 01 02:40:00 +0000 2021' 'Mon Nov 01 01:19:06 +0000 2021' 'Mon Nov 01 00:48:40 +0000 2021' 'Mon Nov 01 00:31:28 +0000 2021' 'Sun Oct 31 23:42:36 +0000 2021' 'Sun Oct 31 23:41:28 +0000 2021' 'Sun Oct 31 23:36:00 +0000 2021' 'Sun Oct 31 23:34:29 +0000 2021' 'Sun Oct 31 23:24:49 +0000 2021' 'Sun Oct 31 23:23:36 +0000 2021' 'Sun Oct 31 23:22:06 +0000 2021' 'Sun Oct 31 22:47:38 +0000 2021' 'Sun Oct 31 21:56:21 +0000 2021' 'Sun Oct 31 21:45:29 +0000 2021' 'Sun Oct 31 21:36:54 +0000 2021' 'Sun Oct 31 21:33:15 +0000 2021' 'Sun Oct 31 20:21:24 +0000 2021' 'Sun Oct 31 20:06:35 +0000 2021' 'Sun Oct 31 19:13:01 +0000 2021' 'Sun Oct 31 19:05:07 +0000 2021' 'Sun Oct 31 18:53:15 +0000 2021' 'Sun Oct 31 18:43:43 +0000 2021' 'Sun Oct 31 18:41:39 +0000 2021' 'Sun Oct 31 18:16:00 +0000 2021' 'Sun Oct 31 17:57:46 +0000 2021' 'Sun Oct 31 17:24:00 +0000 2021' 'Sun Oct 31 17:11:33 +0000 2021' 'Sun Oct 31 17:08:45 +0000 2021' 'Sun Oct 31 16:36:12 +0000 2021' 'Sun Oct 31 16:21:43 +0000 2021' 'Sun Oct 31 16:20:05 +0000 2021' 'Sun Oct 31 16:07:03 +0000 2021' ###Markdown Collect Tweets into MongoDB Install Python librariesYou may need to restart your Jupyter Notebook instance after installed those libraries. ###Code !pip install pymongo !pip install pymongo[srv] !pip install dnspython !pip install tweepy !pip install twitter ###Output Requirement already satisfied: twitter in /home/ec2-user/anaconda3/envs/python3/lib/python3.6/site-packages (1.18.0) WARNING: You are using pip version 20.0.2; however, version 20.2.4 is available. You should consider upgrading via the '/home/ec2-user/anaconda3/envs/python3/bin/python -m pip install --upgrade pip' command. ###Markdown Import Python libraries ###Code import pymongo from pymongo import MongoClient import json import tweepy import twitter from pprint import pprint import configparser import pandas as pd ###Output _____no_output_____ ###Markdown Load the Authorization Info Save database connection info and API Keys in a config.ini file and use the configparse to load the authorization info. ###Code config = configparser.ConfigParser() config.read('config.ini') CONSUMER_KEY = config['mytwitter']['api_key'] CONSUMER_SECRET = config['mytwitter']['api_secrete'] OAUTH_TOKEN = config['mytwitter']['access_token'] OATH_TOKEN_SECRET = config['mytwitter']['access_secrete'] mongod_connect = config['mymongo']['connection'] ###Output _____no_output_____ ###Markdown Connect to the MongoDB Cluster ###Code client = MongoClient(mongod_connect) db = client.gp5 # use or create a database named demo tweet_collection = db.tweet_collection #use or create a collection named tweet_collection tweet_collection.create_index([("id", pymongo.ASCENDING)],unique = True) # make sure the collected tweets are unique ###Output _____no_output_____ ###Markdown Use the Streaming API to Collect Tweets Authorize the Stream API ###Code stream_auth = tweepy.OAuthHandler(CONSUMER_KEY, CONSUMER_SECRET) stream_auth.set_access_token(OAUTH_TOKEN, OATH_TOKEN_SECRET) strem_api = tweepy.API(stream_auth) ###Output _____no_output_____ ###Markdown Define the query for the Stream API ###Code track = ['election'] # define the keywords, tweets contain election locations = [-78.9326449,38.4150904,-78.8816972,38.4450731] #defin the location, in Harrisonburg, VA ###Output _____no_output_____ ###Markdown The collected tweets will contain 'election' OR are located in Harrisonburg, VA ###Code class MyStreamListener(tweepy.StreamListener): def on_status(self, status): print (status.id_str) try: tweet_collection.insert_one(status._json) except: pass def on_error(self, status_code): if status_code == 420: #returning False in on_data disconnects the stream return False myStreamListener = MyStreamListener() myStream = tweepy.Stream(auth = strem_api.auth, listener=myStreamListener) myStream.filter(track=track)# (locations = locations) #Use either track or locations ###Output 1323304991533240327 1323304991453491202 1323304991415742466 1323304991424188422 1323304991675863040 1323304991604486145 1323304991659036675 1323304991470325760 1323304991596032000 1323304991696846851 1323304991793303555 1323304991948333056 1323304991843639296 1323304991982051329 1323304991998824451 1323304992137240576 1323304992221134849 1323304992145625088 1323304992300769280 1323304992225153024 1323304992296431617 1323304992447614978 1323304992501977089 1323304992690888705 1323304992623779840 1323304992426598400 1323304992770502662 1323304992909021185 1323304992904806401 1323304992992841735 1323304993068376066 1323304992648953858 1323304993001263104 1323304993059794945 1323304993110319104 1323304993231790080 1323304993236070402 1323304993181630464 1323304993156464643 1323304993235951617 1323304993064054784 1323304993152225280 1323304993353584645 1323304993353461761 1323304993299062784 1323304993349345281 1323304993399709698 1323304993567338498 1323304993588420608 1323304993487790080 1323304993290653697 1323304993697505289 1323304993693200385 1323304993806520320 1323304993928155136 1323304993907220485 1323304993882034181 1323304994079043585 1323304994037239814 1323304994053910529 1323304993974267905 1323304994142113794 1323304994079010817 1323304994221682688 1323304994284597249 1323304994276282374 1323304994314067968 1323304994204954624 1323304994267930625 1323304994263638017 1323304994507051009 1323304996671291392 1323304996750921729 1323304996675485717 1323304996809678848 1323304996738400259 1323304996830588928 1323304996801286147 1323304996906164225 1323304997044604928 1323304997107331073 1323304997233328130 1323304997124276230 1323304997384265729 1323304997329817601 1323304997333929991 1323304997384314880 1323304997442985984 1323304997476573194 1323304997426237441 1323304997392687104 1323304997635936257 1323304997602365441 1323304997614952453 1323304997736665090 1323304997837180930 1323304997866606592 1323304997853880320 1323304997979901952 1323304997912674305 1323304997925257216 1323304997967196160 1323304998210478080 1323304998357372931 1323304997862346752 1323304998349012992 1323304998357401600 1323304998432919557 1323304998491619329 1323304998474813441 1323304998495768578 1323304998474829824 1323304998558748674 1323304998684545025 1323304998709522433 1323304998827012097 1323304998856515585 1323304998739058689 1323304998835523586 1323304998902603776 1323304998885859329 1323305000894803969 1323305000789925888 1323305000945225733 1323305001087864832 1323305000890626050 1323305000936722433 1323305000962019330 1323305000878198785 1323305001108819969 1323305001180016641 1323305001276575744 1323305001209516039 1323305001419157505 1323305001360465922 1323305001557643265 1323305001582764038 1323305001607929857 1323305001536655360 1323305001687678976 1323305001721057280 1323305001033281537 1323305001922502661 1323305001981235200 1323305001964441603 1323305001998032896 1323305001918365698 1323305002149040128 1323305001909932032 1323305002216075264 1323305002522288128 1323305002497118213 1323305002455150592 1323305002631270401 1323305002753052677 1323305002769674240 1323305002924998667 1323305002979393536 1323305003025649672 1323305003159834625 1323305002861879296 1323305003151339520 1323305003193454594 1323305003100983296 1323305003172417537 1323305003201650688 1323305003226976256 1323305003239485441 1323305003256143873 1323305003176599552 1323305002744471553 1323305004967493632 1323305005135273984 1323305005261058048 1323305004950695938 1323305005303140355 1323305005361876999 1323305005579935744 1323305005512859648 1323305005588369411 1323305005634461697 1323305005894565889 1323305005898715138 1323305005861011458 1323305006129434624 1323305006137827329 1323305006234136577 1323305006401990659 1323305006313934850 1323305006422847488 1323305006389481474 1323305006204932097 1323305006511083520 1323305006636945408 1323305006653603841 1323305006733402113 1323305006603317248 1323305006792122368 1323305006796328962 1323305006766956545 1323305006901141506 1323305006867587073 1323305006959919106 1323305007039586309 1323305006741757954 1323305007337279488 1323305007307984898 1323305007253463040 1323305007429672961 1323305007404453894 1323305007253475330 1323305007156928512 1323305007651848192 1323305007618363396 1323305007890997249 1323305007857414144 1323305007979106309 1323305008176181249 1323305008251633665 1323305008184627202 1323305008373387264 1323305009233231874 1323305008448696320 1323305009245728768 1323305009346392065 1323305009375830019 1323305009283551234 1323305009388441600 1323305009606397952 1323305009526804482 1323305009619107841 1323305009706983424 1323305009744936960 1323305009778319360 1323305009828777985 1323305009635840001 1323305009958834186 1323305009996582913 1323305010101317632 1323305009996537863 1323305009988071424 1323305010197835776 1323305010210480130 1323305010248241152 1323305010092978177 1323305010273267713 1323305010399125504 1323305010600566784 1323305010613030912 1323305010713649152 1323305010638295040 1323305010755739655 1323305010843639812 1323305010868948993 1323305010822864896 1323305010835394560 1323305010814484481 1323305011003228161 1323305010957033475 1323305010923524106 1323305011166633986 1323305011141582849 1323305011267268608 1323305011460231168 1323305011431002119 1323305011007393792 1323305011502342146 1323305011506565120 1323305011808489472 1323305011867242496 1323305011980365824 1323305013406519299 1323305013431705603 1323305013502971905 1323305012655783937 1323305013603635206 1323305013683212288 1323305013691768834 1323305013473472512 1323305013632999426 1323305013859536899 1323305013863538688 1323305013758885888 1323305013960126464 1323305014014709770 1323305014077591553 1323305013997830144 1323305014123728898 1323305014291501059 1323305014278918146 1323305014195064832 1323305014257946627 1323305014207676417 1323305013502976001 1323305014308171776 1323305014476099589 1323305014413111297 1323305014648053761 1323305014736097288 1323305014857662470 1323305014908145664 1323305014861901834 1323305014769586176 1323305014731964416 1323305014060670976 1323305015008702464 1323305015159738371 1323305015214223361 1323305015298109440 1323305015340109824 1323305015327531008 1323305015755288577 1323305015591796738 1323305015700836361 1323305015814070273 1323305015918927874 1323305015969259520 1323305015977672704 1323305015965065217 1323305016136904704 1323305016111865856 ###Markdown Use the REST API to Collect Tweets Authorize the REST API ###Code rest_auth = twitter.oauth.OAuth(OAUTH_TOKEN,OATH_TOKEN_SECRET,CONSUMER_KEY,CONSUMER_SECRET) rest_api = twitter.Twitter(auth=rest_auth) ###Output _____no_output_____ ###Markdown Define the query for the REST API ###Code count = 100 #number of returned tweets, default and max is 100 geocode = "38.4392897,-78.9412224,50mi" # defin the location, in Harrisonburg, VA q = "election" #define the keywords, tweets contain election ###Output _____no_output_____ ###Markdown The collected tweets will contain 'election' AND are located in Harrisonburg, VA ###Code search_results = rest_api.search.tweets( count=count,q=q, geocode=geocode) #you can use both q and geocode statuses = search_results["statuses"] since_id_new = statuses[-1]['id'] for statuse in statuses: try: tweet_collection.insert_one(statuse) pprint(statuse['created_at'])# print the date of the collected tweets except: pass ###Output 'Mon Nov 02 16:44:48 +0000 2020' 'Mon Nov 02 16:44:40 +0000 2020' 'Mon Nov 02 16:44:23 +0000 2020' 'Mon Nov 02 16:42:23 +0000 2020' 'Mon Nov 02 16:41:27 +0000 2020' 'Mon Nov 02 16:41:03 +0000 2020' 'Mon Nov 02 16:39:15 +0000 2020' 'Mon Nov 02 16:37:36 +0000 2020' 'Mon Nov 02 16:36:23 +0000 2020' 'Mon Nov 02 16:35:59 +0000 2020' 'Mon Nov 02 16:35:51 +0000 2020' 'Mon Nov 02 16:35:42 +0000 2020' 'Mon Nov 02 16:35:39 +0000 2020' 'Mon Nov 02 16:34:16 +0000 2020' 'Mon Nov 02 16:33:57 +0000 2020' 'Mon Nov 02 16:32:29 +0000 2020' 'Mon Nov 02 16:32:14 +0000 2020' 'Mon Nov 02 16:32:07 +0000 2020' 'Mon Nov 02 16:31:01 +0000 2020' 'Mon Nov 02 16:30:45 +0000 2020' 'Mon Nov 02 16:30:23 +0000 2020' 'Mon Nov 02 16:29:40 +0000 2020' 'Mon Nov 02 16:29:33 +0000 2020' 'Mon Nov 02 16:29:21 +0000 2020' 'Mon Nov 02 16:29:08 +0000 2020' 'Mon Nov 02 16:29:05 +0000 2020' 'Mon Nov 02 16:27:55 +0000 2020' 'Mon Nov 02 16:27:55 +0000 2020' 'Mon Nov 02 16:27:16 +0000 2020' 'Mon Nov 02 16:27:03 +0000 2020' 'Mon Nov 02 16:26:33 +0000 2020' 'Mon Nov 02 16:26:31 +0000 2020' 'Mon Nov 02 16:25:07 +0000 2020' 'Mon Nov 02 16:25:06 +0000 2020' 'Mon Nov 02 16:24:45 +0000 2020' 'Mon Nov 02 16:24:44 +0000 2020' 'Mon Nov 02 16:24:34 +0000 2020' 'Mon Nov 02 16:24:29 +0000 2020' 'Mon Nov 02 16:23:49 +0000 2020' 'Mon Nov 02 16:23:49 +0000 2020' 'Mon Nov 02 16:23:32 +0000 2020' 'Mon Nov 02 16:23:31 +0000 2020' 'Mon Nov 02 16:23:19 +0000 2020' 'Mon Nov 02 16:22:32 +0000 2020' 'Mon Nov 02 16:22:12 +0000 2020' 'Mon Nov 02 16:22:07 +0000 2020' 'Mon Nov 02 16:21:55 +0000 2020' 'Mon Nov 02 16:21:33 +0000 2020' 'Mon Nov 02 16:20:25 +0000 2020' 'Mon Nov 02 16:20:05 +0000 2020' 'Mon Nov 02 16:19:54 +0000 2020' 'Mon Nov 02 16:19:36 +0000 2020' 'Mon Nov 02 16:18:40 +0000 2020' 'Mon Nov 02 16:16:33 +0000 2020' 'Mon Nov 02 16:15:26 +0000 2020' 'Mon Nov 02 16:13:50 +0000 2020' 'Mon Nov 02 16:13:24 +0000 2020' 'Mon Nov 02 16:13:05 +0000 2020' 'Mon Nov 02 16:11:48 +0000 2020' 'Mon Nov 02 16:10:50 +0000 2020' 'Mon Nov 02 16:09:48 +0000 2020' 'Mon Nov 02 16:09:42 +0000 2020' 'Mon Nov 02 16:08:32 +0000 2020' 'Mon Nov 02 16:07:02 +0000 2020' 'Mon Nov 02 16:06:55 +0000 2020' 'Mon Nov 02 16:06:43 +0000 2020' 'Mon Nov 02 16:05:33 +0000 2020' 'Mon Nov 02 16:05:19 +0000 2020' 'Mon Nov 02 16:04:53 +0000 2020' 'Mon Nov 02 16:04:41 +0000 2020' 'Mon Nov 02 16:04:13 +0000 2020' 'Mon Nov 02 16:03:24 +0000 2020' 'Mon Nov 02 16:03:04 +0000 2020' 'Mon Nov 02 16:02:25 +0000 2020' 'Mon Nov 02 16:02:21 +0000 2020' 'Mon Nov 02 16:01:04 +0000 2020' ###Markdown Continue fetching early tweets with the same query. YOU WILL REACH YOUR RATE LIMIT VERY FAST ###Code since_id_old = 0 while(since_id_new != since_id_old): since_id_old = since_id_new search_results = rest_api.search.tweets( count=count,q=q, geocode=geocode, max_id= since_id_new) statuses = search_results["statuses"] since_id_new = statuses[-1]['id'] for statuse in statuses: try: tweet_collection.insert_one(statuse) pprint(statuse['created_at']) # print the date of the collected tweets except: pass ###Output 'Mon Nov 02 00:52:26 +0000 2020' 'Mon Nov 02 00:51:16 +0000 2020' 'Mon Nov 02 00:50:44 +0000 2020' 'Mon Nov 02 00:49:20 +0000 2020' 'Mon Nov 02 00:46:02 +0000 2020' 'Mon Nov 02 00:43:02 +0000 2020' 'Mon Nov 02 00:41:22 +0000 2020' 'Mon Nov 02 00:41:07 +0000 2020' 'Mon Nov 02 00:40:21 +0000 2020' 'Mon Nov 02 00:36:32 +0000 2020' 'Mon Nov 02 00:35:59 +0000 2020' 'Mon Nov 02 00:34:24 +0000 2020' 'Mon Nov 02 00:33:37 +0000 2020' 'Mon Nov 02 00:31:33 +0000 2020' 'Mon Nov 02 00:28:27 +0000 2020' 'Mon Nov 02 00:28:15 +0000 2020' 'Mon Nov 02 00:28:13 +0000 2020' 'Mon Nov 02 00:27:17 +0000 2020' 'Mon Nov 02 00:26:04 +0000 2020' 'Mon Nov 02 00:23:07 +0000 2020' 'Mon Nov 02 00:22:05 +0000 2020' 'Mon Nov 02 00:21:26 +0000 2020' 'Mon Nov 02 00:20:43 +0000 2020' 'Mon Nov 02 00:19:05 +0000 2020' 'Mon Nov 02 00:18:37 +0000 2020' 'Mon Nov 02 00:16:46 +0000 2020' 'Mon Nov 02 00:13:57 +0000 2020' 'Mon Nov 02 00:13:24 +0000 2020' 'Mon Nov 02 00:11:26 +0000 2020' 'Mon Nov 02 00:09:37 +0000 2020' 'Mon Nov 02 00:07:41 +0000 2020' 'Mon Nov 02 00:07:30 +0000 2020' 'Mon Nov 02 00:05:45 +0000 2020' 'Mon Nov 02 00:04:19 +0000 2020' 'Mon Nov 02 00:03:43 +0000 2020' 'Mon Nov 02 00:02:55 +0000 2020' 'Mon Nov 02 00:00:14 +0000 2020' 'Sun Nov 01 23:54:07 +0000 2020' 'Sun Nov 01 23:53:49 +0000 2020' 'Sun Nov 01 23:52:43 +0000 2020' 'Sun Nov 01 23:49:21 +0000 2020' 'Sun Nov 01 23:48:45 +0000 2020' 'Sun Nov 01 23:46:48 +0000 2020' 'Sun Nov 01 23:46:37 +0000 2020' 'Sun Nov 01 23:46:05 +0000 2020' 'Sun Nov 01 23:44:56 +0000 2020' 'Sun Nov 01 23:44:20 +0000 2020' 'Sun Nov 01 23:42:40 +0000 2020' 'Sun Nov 01 23:42:21 +0000 2020' 'Sun Nov 01 23:40:35 +0000 2020' 'Sun Nov 01 23:39:17 +0000 2020' 'Sun Nov 01 23:34:40 +0000 2020' 'Sun Nov 01 23:33:08 +0000 2020' 'Sun Nov 01 23:29:44 +0000 2020' 'Sun Nov 01 23:28:10 +0000 2020' 'Sun Nov 01 23:28:02 +0000 2020' 'Sun Nov 01 23:27:21 +0000 2020' 'Sun Nov 01 23:25:59 +0000 2020' 'Sun Nov 01 23:21:22 +0000 2020' 'Sun Nov 01 23:21:01 +0000 2020' 'Sun Nov 01 23:15:20 +0000 2020' 'Sun Nov 01 23:08:14 +0000 2020' 'Sun Nov 01 23:07:47 +0000 2020' 'Sun Nov 01 23:04:53 +0000 2020' 'Sun Nov 01 23:00:47 +0000 2020' 'Sun Nov 01 23:00:39 +0000 2020' 'Sun Nov 01 22:56:52 +0000 2020' 'Sun Nov 01 22:51:29 +0000 2020' 'Sun Nov 01 22:50:04 +0000 2020' 'Sun Nov 01 22:50:01 +0000 2020' 'Sun Nov 01 22:48:34 +0000 2020' 'Sun Nov 01 22:46:53 +0000 2020' 'Sun Nov 01 22:43:52 +0000 2020' 'Sun Nov 01 22:38:00 +0000 2020' 'Sun Nov 01 22:37:24 +0000 2020' 'Sun Nov 01 22:36:20 +0000 2020' 'Sun Nov 01 22:36:14 +0000 2020' 'Sun Nov 01 22:35:33 +0000 2020' 'Sun Nov 01 22:34:34 +0000 2020' 'Sun Nov 01 22:34:18 +0000 2020' 'Sun Nov 01 22:32:35 +0000 2020' 'Sun Nov 01 22:32:14 +0000 2020' 'Sun Nov 01 22:32:02 +0000 2020' 'Sun Nov 01 22:30:59 +0000 2020' 'Sun Nov 01 22:28:19 +0000 2020' 'Sun Nov 01 22:27:33 +0000 2020' 'Sun Nov 01 22:26:58 +0000 2020' 'Sun Nov 01 22:26:51 +0000 2020' 'Sun Nov 01 22:26:19 +0000 2020' 'Sun Nov 01 22:26:00 +0000 2020' 'Sun Nov 01 22:25:20 +0000 2020' 'Sun Nov 01 22:24:46 +0000 2020' 'Sun Nov 01 22:24:09 +0000 2020' 'Sun Nov 01 22:24:00 +0000 2020' 'Sun Nov 01 22:23:59 +0000 2020' 'Sun Nov 01 22:21:12 +0000 2020' 'Sun Nov 01 22:20:37 +0000 2020' 'Sun Nov 01 22:20:03 +0000 2020' 'Sun Nov 01 22:18:28 +0000 2020' 'Sun Nov 01 22:18:15 +0000 2020' 'Sun Nov 01 22:17:10 +0000 2020' 'Sun Nov 01 22:16:29 +0000 2020' 'Sun Nov 01 22:16:25 +0000 2020' 'Sun Nov 01 22:15:46 +0000 2020' 'Sun Nov 01 22:14:04 +0000 2020' 'Sun Nov 01 22:10:42 +0000 2020' 'Sun Nov 01 22:04:01 +0000 2020' 'Sun Nov 01 22:03:56 +0000 2020' 'Sun Nov 01 22:03:06 +0000 2020' 'Sun Nov 01 22:00:08 +0000 2020' 'Sun Nov 01 21:58:00 +0000 2020' 'Sun Nov 01 21:55:15 +0000 2020' 'Sun Nov 01 21:52:31 +0000 2020' 'Sun Nov 01 21:51:47 +0000 2020' 'Sun Nov 01 21:50:26 +0000 2020' 'Sun Nov 01 21:46:03 +0000 2020' 'Sun Nov 01 21:41:54 +0000 2020' 'Sun Nov 01 21:40:25 +0000 2020' 'Sun Nov 01 21:40:13 +0000 2020' 'Sun Nov 01 21:39:35 +0000 2020' 'Sun Nov 01 21:39:16 +0000 2020' 'Sun Nov 01 21:38:46 +0000 2020' 'Sun Nov 01 21:36:51 +0000 2020' 'Sun Nov 01 21:36:28 +0000 2020' 'Sun Nov 01 21:35:53 +0000 2020' 'Sun Nov 01 21:34:59 +0000 2020' 'Sun Nov 01 21:34:03 +0000 2020' 'Sun Nov 01 21:33:47 +0000 2020' 'Sun Nov 01 21:32:58 +0000 2020' 'Sun Nov 01 21:32:47 +0000 2020' 'Sun Nov 01 21:31:35 +0000 2020' 'Sun Nov 01 21:30:39 +0000 2020' 'Sun Nov 01 21:29:37 +0000 2020' 'Sun Nov 01 21:29:19 +0000 2020' 'Sun Nov 01 21:28:56 +0000 2020' 'Sun Nov 01 21:28:22 +0000 2020' 'Sun Nov 01 21:27:34 +0000 2020' 'Sun Nov 01 21:25:46 +0000 2020' 'Sun Nov 01 21:25:25 +0000 2020' 'Sun Nov 01 21:25:20 +0000 2020' 'Sun Nov 01 21:23:31 +0000 2020' 'Sun Nov 01 21:21:59 +0000 2020' 'Sun Nov 01 21:20:27 +0000 2020' 'Sun Nov 01 21:18:29 +0000 2020' 'Sun Nov 01 21:17:56 +0000 2020' 'Sun Nov 01 21:15:23 +0000 2020' 'Sun Nov 01 21:13:30 +0000 2020' 'Sun Nov 01 21:11:20 +0000 2020' 'Sun Nov 01 21:09:51 +0000 2020' 'Sun Nov 01 21:09:43 +0000 2020' 'Sun Nov 01 21:08:11 +0000 2020' 'Sun Nov 01 21:07:20 +0000 2020' 'Sun Nov 01 21:06:45 +0000 2020' 'Sun Nov 01 21:06:15 +0000 2020' 'Sun Nov 01 21:01:16 +0000 2020' 'Sun Nov 01 20:59:48 +0000 2020' 'Sun Nov 01 20:58:20 +0000 2020' 'Sun Nov 01 20:57:30 +0000 2020' 'Sun Nov 01 20:57:05 +0000 2020' 'Sun Nov 01 20:54:38 +0000 2020' 'Sun Nov 01 20:53:55 +0000 2020' 'Sun Nov 01 20:53:54 +0000 2020' 'Sun Nov 01 20:51:48 +0000 2020' 'Sun Nov 01 20:51:28 +0000 2020' 'Sun Nov 01 20:51:23 +0000 2020' 'Sun Nov 01 20:50:47 +0000 2020' 'Sun Nov 01 20:50:08 +0000 2020' 'Sun Nov 01 20:48:25 +0000 2020' 'Sun Nov 01 20:48:04 +0000 2020' 'Sun Nov 01 20:47:33 +0000 2020' 'Sun Nov 01 20:44:43 +0000 2020' 'Sun Nov 01 20:44:34 +0000 2020' 'Sun Nov 01 20:44:33 +0000 2020' 'Sun Nov 01 20:43:37 +0000 2020' 'Sun Nov 01 20:42:32 +0000 2020' 'Sun Nov 01 20:41:51 +0000 2020' 'Sun Nov 01 20:41:08 +0000 2020' 'Sun Nov 01 20:40:12 +0000 2020' 'Sun Nov 01 20:37:23 +0000 2020' 'Sun Nov 01 20:36:44 +0000 2020' 'Sun Nov 01 20:32:07 +0000 2020' 'Sun Nov 01 20:31:39 +0000 2020' 'Sun Nov 01 20:31:21 +0000 2020' 'Sun Nov 01 20:30:11 +0000 2020' 'Sun Nov 01 20:30:00 +0000 2020' 'Sun Nov 01 20:29:50 +0000 2020' 'Sun Nov 01 20:29:10 +0000 2020' 'Sun Nov 01 20:27:51 +0000 2020' 'Sun Nov 01 20:26:01 +0000 2020' 'Sun Nov 01 20:21:29 +0000 2020' 'Sun Nov 01 20:19:41 +0000 2020' 'Sun Nov 01 20:18:46 +0000 2020' 'Sun Nov 01 20:17:45 +0000 2020' 'Sun Nov 01 20:15:06 +0000 2020' 'Sun Nov 01 20:14:09 +0000 2020' 'Sun Nov 01 20:13:22 +0000 2020' 'Sun Nov 01 20:12:15 +0000 2020' 'Sun Nov 01 20:11:18 +0000 2020' 'Sun Nov 01 20:10:09 +0000 2020' 'Sun Nov 01 20:08:39 +0000 2020' 'Sun Nov 01 20:08:16 +0000 2020' 'Sun Nov 01 20:06:30 +0000 2020' 'Sun Nov 01 20:06:00 +0000 2020' 'Sun Nov 01 20:04:35 +0000 2020' 'Sun Nov 01 20:03:42 +0000 2020' 'Sun Nov 01 20:02:37 +0000 2020' 'Sun Nov 01 19:57:53 +0000 2020' 'Sun Nov 01 19:56:36 +0000 2020' 'Sun Nov 01 19:54:37 +0000 2020' 'Sun Nov 01 19:54:22 +0000 2020' 'Sun Nov 01 19:53:59 +0000 2020' 'Sun Nov 01 19:53:27 +0000 2020' 'Sun Nov 01 19:52:38 +0000 2020' 'Sun Nov 01 19:51:04 +0000 2020' 'Sun Nov 01 19:50:51 +0000 2020' 'Sun Nov 01 19:50:45 +0000 2020' 'Sun Nov 01 19:49:58 +0000 2020' 'Sun Nov 01 19:49:03 +0000 2020' 'Sun Nov 01 19:48:32 +0000 2020' 'Sun Nov 01 19:48:08 +0000 2020' 'Sun Nov 01 19:47:43 +0000 2020' 'Sun Nov 01 19:46:43 +0000 2020' 'Sun Nov 01 19:45:06 +0000 2020' 'Sun Nov 01 19:44:55 +0000 2020' 'Sun Nov 01 19:44:39 +0000 2020' 'Sun Nov 01 19:43:43 +0000 2020' 'Sun Nov 01 19:43:05 +0000 2020' 'Sun Nov 01 19:41:40 +0000 2020' 'Sun Nov 01 19:40:36 +0000 2020' 'Sun Nov 01 19:40:27 +0000 2020' 'Sun Nov 01 19:40:19 +0000 2020' 'Sun Nov 01 19:37:46 +0000 2020' 'Sun Nov 01 19:36:39 +0000 2020' 'Sun Nov 01 19:36:33 +0000 2020' 'Sun Nov 01 19:35:04 +0000 2020' 'Sun Nov 01 19:34:50 +0000 2020' 'Sun Nov 01 19:34:11 +0000 2020' 'Sun Nov 01 19:33:51 +0000 2020' 'Sun Nov 01 19:30:48 +0000 2020' 'Sun Nov 01 19:30:24 +0000 2020' 'Sun Nov 01 19:30:18 +0000 2020' 'Sun Nov 01 19:26:49 +0000 2020' 'Sun Nov 01 19:25:43 +0000 2020' 'Sun Nov 01 19:25:39 +0000 2020' 'Sun Nov 01 19:25:22 +0000 2020' 'Sun Nov 01 19:24:26 +0000 2020' 'Sun Nov 01 19:21:34 +0000 2020' 'Sun Nov 01 19:18:00 +0000 2020' 'Sun Nov 01 19:17:54 +0000 2020' 'Sun Nov 01 19:15:40 +0000 2020' 'Sun Nov 01 19:13:18 +0000 2020' 'Sun Nov 01 19:11:10 +0000 2020' 'Sun Nov 01 19:08:05 +0000 2020' 'Sun Nov 01 19:07:47 +0000 2020' 'Sun Nov 01 19:06:03 +0000 2020' 'Sun Nov 01 19:05:03 +0000 2020' 'Sun Nov 01 19:05:02 +0000 2020' 'Sun Nov 01 19:03:34 +0000 2020' 'Sun Nov 01 19:01:24 +0000 2020' 'Sun Nov 01 19:00:07 +0000 2020' 'Sun Nov 01 19:00:01 +0000 2020' 'Sun Nov 01 18:59:12 +0000 2020' 'Sun Nov 01 18:58:10 +0000 2020' 'Sun Nov 01 18:57:35 +0000 2020' 'Sun Nov 01 18:57:13 +0000 2020' 'Sun Nov 01 18:56:02 +0000 2020' 'Sun Nov 01 18:55:58 +0000 2020' 'Sun Nov 01 18:53:38 +0000 2020' 'Sun Nov 01 18:53:01 +0000 2020' 'Sun Nov 01 18:51:16 +0000 2020' 'Sun Nov 01 18:51:04 +0000 2020' 'Sun Nov 01 18:50:32 +0000 2020' 'Sun Nov 01 18:50:21 +0000 2020' 'Sun Nov 01 18:48:10 +0000 2020' 'Sun Nov 01 18:47:21 +0000 2020' 'Sun Nov 01 18:46:52 +0000 2020' 'Sun Nov 01 18:46:05 +0000 2020' 'Sun Nov 01 18:45:00 +0000 2020' 'Sun Nov 01 18:43:19 +0000 2020' 'Sun Nov 01 18:40:57 +0000 2020' ###Markdown View the Collected Tweets Print the number of tweets and unique twitter users ###Code print(tweet_collection.estimated_document_count())# number of tweets collected user_cursor = tweet_collection.distinct("user.id") print (len(user_cursor)) # number of unique Twitter users ###Output 3529 3275 ###Markdown Create a text index and print the Tweets containing specific keywords. ###Code tweet_collection.create_index([("text", pymongo.TEXT)], name='text_index', default_language='english') # create a text index ###Output _____no_output_____ ###Markdown Create a cursor to query tweets with the created index ###Code tweet_cursor = tweet_collection.find({"$text": {"$search": "vote"}}) # return tweets contain vote ###Output _____no_output_____ ###Markdown Use pprint to display tweets ###Code for document in tweet_cursor[0:10]: # display the first 10 tweets from the query try: print ('----') #pprint (document) # use pprint to print the entire tweet document print ('name:', document["user"]["name"]) # user name print ('text:', document["text"]) # tweets except: print ("***error in encoding") pass tweet_cursor = tweet_collection.find({"$text": {"$search": "vote"}}) # return tweets contain vote ###Output _____no_output_____ ###Markdown Use pandas to display tweets ###Code tweet_df = pd.DataFrame(list(tweet_cursor )) tweet_df[:10] #display the first 10 tweets tweet_df["favorite_count"].hist() # create a histogram show the favorite count ###Output _____no_output_____
module4/ssignment_applied_modeling_4.ipynb
###Markdown Setting up for model ###Code #Setting up matrices for models: target = ['critic_score'] leaky = ['critic_count','user_score','user_count','name'] features = train.columns.drop(target+leaky) features X_train = train[features] y_train = np.ravel(train[target]) X_val = val[features] y_val = np.ravel(val[target]) X_test = test[features] y_test = np.ravel(test[target]) #Some verification: train.shape,X_train.shape ###Output _____no_output_____ ###Markdown Baseline ###Code from sklearn.metrics import mean_absolute_error as mae from sklearn.metrics import r2_score X_train.head() y_baseline = [y_train.mean()]*len(y_val) print(f'Mean absolute error: {mae(y_val,y_baseline)}') print(f'R2 score: {r2_score(y_val,y_baseline)}') ###Output Mean absolute error: 11.458943440022388 R2 score: -9.199689323002858e-06 ###Markdown Model ###Code np.ravel(y_val) import category_encoders as ce from sklearn.linear_model import LinearRegression from sklearn.ensemble import RandomForestRegressor from sklearn.pipeline import make_pipeline from sklearn.impute import SimpleImputer from sklearn.model_selection import RandomizedSearchCV process = make_pipeline( ce.OrdinalEncoder(), SimpleImputer(), ) X_train_processed = process.fit_transform(X_train) X_val_processed = process.transform(X_val) model = RandomForestRegressor( n_estimators=100 ) model.fit(X_train_processed,y_train) model.score(X_val_processed,y_val) #Get one observation to put through the predictor: print(y_train[0]) X_train.iloc[[0]] # process.transform(X_val) train v ###Output _____no_output_____
L3E3.ipynb
###Markdown Assignment 3 - Exercise 3Bruno Kiyoshi Ynumaru - 201805995 ###Code prob_statement = """Uma empresa est´a considerando 5 oportunidades de investimento distintas. A sa´ıda de caixa e o valor presente l´ıquido (VPL) destes investimentos s˜ao dados na tabela abaixo (em milh˜oes de d´olares). A empresa possui $40 milh˜oes para investimento no instante atual (instante 0); ela estima que em um ano a partir de agora (instante 1) $20 milh˜oes estar˜ao dispon´ıveis para investimento. Os investimentos podem ser comprados em qualquer fra¸c˜ao. Neste caso, a sa´ıda de caixa e o VPL s˜ao ajustados na mesma propor¸c˜ao. Formule e resolva um PL que ajude a empresa a maximizar o VPL que pode ser gerado investindo nos investimentos 1-5. """ def fix_statement(str_statement): list_replacements = [("¸c", "ç"), ("´a", "á"), ("´e", "é"), ("´ı", "í"), ("´o", "ó"), ("´u", "ú"), ("˜a", "ã"), ("˜o", "õ"), ("$", "\\\$")] for replacement in list_replacements: str_statement = str_statement.replace(replacement[0], replacement[1]) return str_statement prob_statement = fix_statement(prob_statement) print(prob_statement) ###Output Uma empresa está considerando 5 oportunidades de investimento distintas. A saída de caixa e o valor presente líquido (VPL) destes investimentos são dados na tabela abaixo (em milhões de dólares). A empresa possui \\$40 milhões para investimento no instante atual (instante 0); ela estima que em um ano a partir de agora (instante 1) \\$20 milhões estarão disponíveis para investimento. Os investimentos podem ser comprados em qualquer fração. Neste caso, a saída de caixa e o VPL são ajustados na mesma proporção. Formule e resolva um PL que ajude a empresa a maximizar o VPL que pode ser gerado investindo nos investimentos 1-5. ###Markdown Uma empresa está considerando 5 oportunidades de investimento distintas. A saída de caixa e o valorpresente líquido (VPL) destes investimentos são dados na tabela abaixo (em milhões de dólares). Aempresa possui \\$40 milhões para investimento no instante atual (instante 0); ela estima que em umano a partir de agora (instante 1) \\$20 milhões estarão disponíveis para investimento. Os investimentospodem ser comprados em qualquer fração. Neste caso, a saída de caixa e o VPL são ajustados namesma proporção. Formule e resolva um PL que ajude a empresa a maximizar o VPL que pode sergerado investindo nos investimentos 1-5.| Investimentos|1| 2| 3| 4| 5|| --- | --- | --- | --- | --- | --- ||Saída de caixa (instante 0)| 11| 53| 5 |5 |29||Saída de caixa (instante 1) |3 |6| 5| 1 |34||VPL| 13| 16| 10| 14| 39| ###Code import gurobipy as gp from gurobipy import GRB, Model # Create a new model m = Model("Wyndor_Glass") # Create variables x1 = m.addVar(lb=0, vtype=GRB.CONTINUOUS, name="investment 1 (share)") x2 = m.addVar(lb=0, vtype=GRB.CONTINUOUS, name="investment 2 (share)") x3 = m.addVar(lb=0, vtype=GRB.CONTINUOUS, name="investment 3 (share)") x4 = m.addVar(lb=0, vtype=GRB.CONTINUOUS, name="investment 4 (share)") x5 = m.addVar(lb=0, vtype=GRB.CONTINUOUS, name="investment 5 (share)") m.setObjective(x1 * 13 + x2 * 16 + x3 * 10 + x4 * 14 + x5 * 39, GRB.MAXIMIZE) # Add constraints m.addConstr(x1 * 11 + x2 * 53 + x3 * 5 + x4 * 5 + x5 * 29 <= 40, 'Funds on time 0') m.addConstr(x1 * 3 + x2 * 6 + x3 * 5 + x4 * 1 + x5 * 34 <= 20, 'Funds on time 1') m.addConstr(x1 <= 1, 'max share of investment 1') m.addConstr(x2 <= 1, 'max share of investment 2') m.addConstr(x3 <= 1, 'max share of investment 3') m.addConstr(x4 <= 1, 'max share of investment 4') m.addConstr(x5 <= 1, 'max share of investment 5') m.optimize() for v in m.getVars(): print(f'{v.varName}, {v.x}') print(f'Obj: {m.objVal}') ###Output investment 1 (share), 1.0 investment 2 (share), 0.20085995085995084 investment 3 (share), 1.0 investment 4 (share), 1.0 investment 5 (share), 0.2880835380835381 Obj: 51.449017199017206
_site/notes/notes_ipynb/docs/lecture3-linear-regression.ipynb
###Markdown Lecture 3: Optimization and Linear Regression Applied Machine Learning__Volodymyr Kuleshov__Cornell Tech Part 1: Optimization and Calculus BackgroundIn the previous lecture, we learned what is a supervised machine learning problem.Before we turn our attention to Linear Regression, we will first dive deeper into the question of optimization. Review: Components of A Supervised Machine Learning ProblemAt a high level, a supervised machine learning problem has the following structure:$$ \text{Dataset} + \underbrace{\text{Learning Algorithm}}_\text{Model Class + Objective + Optimizer } \to \text{Predictive Model} $$The predictive model is chosen to model the relationship between inputs and targets. For instance, it can predict future targets. Optimizer: NotationAt a high-level an optimizer takes * an objective $J$ (also called a loss function) and * a model class $\mathcal{M}$ and finds a model $f \in \mathcal{M}$ with the smallest value of the objective $J$.\begin{align*}\min_{f \in \mathcal{M}} J(f)\end{align*}Intuitively, this is the function that bests "fits" the data on the training dataset $\mathcal{D} = \{(x^{(i)}, y^{(i)}) \mid i = 1,2,...,n\}$. We will use the a quadratic function as our running example for an objective $J$. ###Code import numpy as np import matplotlib.pyplot as plt plt.rcParams['figure.figsize'] = [8, 4] def quadratic_function(theta): """The cost function, J(theta).""" return 0.5*(2*theta-1)**2 ###Output _____no_output_____ ###Markdown We can visualize it. ###Code # First construct a grid of theta1 parameter pairs and their corresponding # cost function values. thetas = np.linspace(-0.2,1,10) f_vals = quadratic_function(thetas[:,np.newaxis]) plt.plot(thetas, f_vals) plt.xlabel('Theta') plt.ylabel('Objective value') plt.title('Simple quadratic function') ###Output _____no_output_____ ###Markdown Calculus Review: DerivativesRecall that the derivative $$\frac{d f(\theta_0)}{d \theta}$$ of a univariate function $f : \mathbb{R} \to \mathbb{R}$ is the instantaneous rate of change of the function $f(\theta)$ with respect to its parameter $\theta$ at the point $\theta_0$. ###Code def quadratic_derivative(theta): return (2*theta-1)*2 df0 = quadratic_derivative(np.array([[0]])) # derivative at zero f0 = quadratic_function(np.array([[0]])) line_length = 0.2 plt.plot(thetas, f_vals) plt.annotate('', xytext=(0-line_length, f0-line_length*df0), xy=(0+line_length, f0+line_length*df0), arrowprops={'arrowstyle': '-', 'lw': 1.5}, va='center', ha='center') plt.xlabel('Theta') plt.ylabel('Objective value') plt.title('Simple quadratic function') pts = np.array([[0, 0.5, 0.8]]).reshape((3,1)) df0s = quadratic_derivative(pts) f0s = quadratic_function(pts) plt.plot(thetas, f_vals) for pt, f0, df0 in zip(pts.flatten(), f0s.flatten(), df0s.flatten()): plt.annotate('', xytext=(pt-line_length, f0-line_length*df0), xy=(pt+line_length, f0+line_length*df0), arrowprops={'arrowstyle': '-', 'lw': 1}, va='center', ha='center') plt.xlabel('Theta') plt.ylabel('Objective value') plt.title('Simple quadratic function') ###Output _____no_output_____ ###Markdown Calculus Review: Partial DerivativesThe partial derivative $$\frac{\partial f(\theta_0)}{\partial \theta_j}$$ of a multivariate function $f : \mathbb{R}^d \to \mathbb{R}$ is the derivative of $f$ with respect to $\theta_j$ while all othe other inputs $\theta_k$ for $k\neq j$ are fixed. Calculus Review: The GradientThe gradient $\nabla_\theta f$ further extends the derivative to multivariate functions $f : \mathbb{R}^d \to \mathbb{R}$, and is defined at a point $\theta_0$ as$$ \nabla_\theta f (\theta_0) = \begin{bmatrix}\frac{\partial f(\theta_0)}{\partial \theta_1} \\\frac{\partial f(\theta_0)}{\partial \theta_2} \\\vdots \\\frac{\partial f(\theta_0)}{\partial \theta_d}\end{bmatrix}.$$The $j$-th entry of the vector $\nabla_\theta f (\theta_0)$ is the partial derivative $\frac{\partial f(\theta_0)}{\partial \theta_j}$ of $f$ with respect to the $j$-th component of $\theta$. We will use a quadratic function as a running example. ###Code def quadratic_function2d(theta0, theta1): """Quadratic objective function, J(theta0, theta1). The inputs theta0, theta1 are 2d arrays and we evaluate the objective at each value theta0[i,j], theta1[i,j]. We implement it this way so it's easier to plot the level curves of the function in 2d. Parameters: theta0 (np.array): 2d array of first parameter theta0 theta1 (np.array): 2d array of second parameter theta1 Returns: fvals (np.array): 2d array of objective function values fvals is the same dimension as theta0 and theta1. fvals[i,j] is the value at theta0[i,j] and theta1[i,j]. """ theta0 = np.atleast_2d(np.asarray(theta0)) theta1 = np.atleast_2d(np.asarray(theta1)) return 0.5*((2*theta1-2)**2 + (theta0-3)**2) ###Output _____no_output_____ ###Markdown Let's visualize this function. ###Code theta0_grid = np.linspace(-4,7,101) theta1_grid = np.linspace(-1,4,101) theta_grid = theta0_grid[np.newaxis,:], theta1_grid[:,np.newaxis] J_grid = quadratic_function2d(theta0_grid[np.newaxis,:], theta1_grid[:,np.newaxis]) X, Y = np.meshgrid(theta0_grid, theta1_grid) contours = plt.contour(X, Y, J_grid, 10) plt.clabel(contours) plt.axis('equal') ###Output _____no_output_____ ###Markdown Let's write down the derivative of the quadratic function. ###Code def quadratic_derivative2d(theta0, theta1): """Derivative of quadratic objective function. The inputs theta0, theta1 are 1d arrays and we evaluate the derivative at each value theta0[i], theta1[i]. Parameters: theta0 (np.array): 1d array of first parameter theta0 theta1 (np.array): 1d array of second parameter theta1 Returns: grads (np.array): 2d array of partial derivatives grads is of the same size as theta0 and theta1 along first dimension and of size two along the second dimension. grads[i,j] is the j-th partial derivative at input theta0[i], theta1[i]. """ # this is the gradient of 0.5*((2*theta1-2)**2 + (theta0-3)**2) grads = np.stack([theta0-3, (2*theta1-2)*2], axis=1) grads = grads.reshape([len(theta0), 2]) return grads ###Output _____no_output_____ ###Markdown We can visualize the derivative. ###Code theta0_pts, theta1_pts = np.array([2.3, -1.35, -2.3]), np.array([2.4, -0.15, 2.75]) dfs = quadratic_derivative2d(theta0_pts, theta1_pts) line_length = 0.2 contours = plt.contour(X, Y, J_grid, 10) for theta0_pt, theta1_pt, df0 in zip(theta0_pts, theta1_pts, dfs): plt.annotate('', xytext=(theta0_pt, theta1_pt), xy=(theta0_pt-line_length*df0[0], theta1_pt-line_length*df0[1]), arrowprops={'arrowstyle': '->', 'lw': 2}, va='center', ha='center') plt.scatter(theta0_pts, theta1_pts) plt.clabel(contours) plt.xlabel('Theta0') plt.ylabel('Theta1') plt.title('Gradients of the quadratic function') plt.axis('equal') ###Output _____no_output_____ ###Markdown Part 1b: Gradient DescentNext, we will use gradients to define an important algorithm called *gradient descent*. Calculus Review: The GradientThe gradient $\nabla_\theta f$ further extends the derivative to multivariate functions $f : \mathbb{R}^d \to \mathbb{R}$, and is defined at a point $\theta_0$ as$$ \nabla_\theta f (\theta_0) = \begin{bmatrix}\frac{\partial f(\theta_0)}{\partial \theta_1} \\\frac{\partial f(\theta_0)}{\partial \theta_2} \\\vdots \\\frac{\partial f(\theta_0)}{\partial \theta_d}\end{bmatrix}.$$The $j$-th entry of the vector $\nabla_\theta f (\theta_0)$ is the partial derivative $\frac{\partial f(\theta_0)}{\partial \theta_j}$ of $f$ with respect to the $j$-th component of $\theta$. ###Code theta0_pts, theta1_pts = np.array([2.3, -1.35, -2.3]), np.array([2.4, -0.15, 2.75]) dfs = quadratic_derivative2d(theta0_pts, theta1_pts) line_length = 0.2 contours = plt.contour(X, Y, J_grid, 10) for theta0_pt, theta1_pt, df0 in zip(theta0_pts, theta1_pts, dfs): plt.annotate('', xytext=(theta0_pt, theta1_pt), xy=(theta0_pt-line_length*df0[0], theta1_pt-line_length*df0[1]), arrowprops={'arrowstyle': '->', 'lw': 2}, va='center', ha='center') plt.scatter(theta0_pts, theta1_pts) plt.clabel(contours) plt.xlabel('Theta0') plt.ylabel('Theta1') plt.title('Gradients of the quadratic function') plt.axis('equal') ###Output _____no_output_____ ###Markdown Gradient Descent: IntuitionGradient descent is a very common optimization algorithm used in machine learning.The intuition behind gradient descent is to repeatedly obtain the gradient to determine the direction in which the function decreases most steeply and take a step in that direction. Gradient Descent: NotationMore formally, if we want to optimize $J(\theta)$, we start with an initial guess $\theta_0$ for the parameters and repeat the following update until $\theta$ is no longer changing:$$ \theta_i := \theta_{i-1} - \alpha \cdot \nabla_\theta J(\theta_{i-1}). $$As code, this method may look as follows:```pythontheta, theta_prev = random_initialization()while norm(theta - theta_prev) > convergence_threshold: theta_prev = theta theta = theta_prev - step_size * gradient(theta_prev)```In the above algorithm, we stop when $||\theta_i - \theta_{i-1}||$ is small. It's easy to implement this function in numpy. ###Code convergence_threshold = 2e-1 step_size = 2e-1 theta, theta_prev = np.array([[-2], [3]]), np.array([[0], [0]]) opt_pts = [theta.flatten()] opt_grads = [] while np.linalg.norm(theta - theta_prev) > convergence_threshold: # we repeat this while the value of the function is decreasing theta_prev = theta gradient = quadratic_derivative2d(*theta).reshape([2,1]) theta = theta_prev - step_size * gradient opt_pts += [theta.flatten()] opt_grads += [gradient.flatten()] ###Output _____no_output_____ ###Markdown We can now visualize gradient descent. ###Code opt_pts = np.array(opt_pts) opt_grads = np.array(opt_grads) contours = plt.contour(X, Y, J_grid, 10) plt.clabel(contours) plt.scatter(opt_pts[:,0], opt_pts[:,1]) for opt_pt, opt_grad in zip(opt_pts, opt_grads): plt.annotate('', xytext=(opt_pt[0], opt_pt[1]), xy=(opt_pt[0]-0.8*step_size*opt_grad[0], opt_pt[1]-0.8*step_size*opt_grad[1]), arrowprops={'arrowstyle': '->', 'lw': 2}, va='center', ha='center') plt.axis('equal') ###Output _____no_output_____ ###Markdown Part 2: Gradient Descent in Linear ModelsLet's now use gradient descent to derive a supervised learning algorithm for linear models. Review: Gradient DescentIf we want to optimize $J(\theta)$, we start with an initial guess $\theta_0$ for the parameters and repeat the following update:$$ \theta_i := \theta_{i-1} - \alpha \cdot \nabla_\theta J(\theta_{i-1}). $$As code, this method may look as follows:```pythontheta, theta_prev = random_initialization()while norm(theta - theta_prev) > convergence_threshold: theta_prev = theta theta = theta_prev - step_size * gradient(theta_prev)``` Review: Linear Model FamilyRecall that a linear model has the form\begin{align*}y & = \theta_0 + \theta_1 \cdot x_1 + \theta_2 \cdot x_2 + ... + \theta_d \cdot x_d\end{align*}where $x \in \mathbb{R}^d$ is a vector of features and $y$ is the target. The $\theta_j$ are the *parameters* of the model.By using the notation $x_0 = 1$, we can represent the model in a vectorized form$$ f_\theta(x) = \sum_{j=0}^d \theta_j \cdot x_j = \theta^\top x. $$ Let's define our model in Python. ###Code def f(X, theta): """The linear model we are trying to fit. Parameters: theta (np.array): d-dimensional vector of parameters X (np.array): (n,d)-dimensional data matrix Returns: y_pred (np.array): n-dimensional vector of predicted targets """ return X.dot(theta) ###Output _____no_output_____ ###Markdown An Objective: Mean Squared ErrorWe pick $\theta$ to minimize the mean squared error (MSE). Slight variants of this objective are also known as the residual sum of squares (RSS) or the sum of squared residuals (SSR).$$J(\theta)= \frac{1}{2n} \sum_{i=1}^n(y^{(i)}-\theta^\top x^{(i)})^2$$In other words, we are looking for the best compromise in $\theta$ over all the data points. Let's implement mean squared error. ###Code def mean_squared_error(theta, X, y): """The cost function, J, describing the goodness of fit. Parameters: theta (np.array): d-dimensional vector of parameters X (np.array): (n,d)-dimensional design matrix y (np.array): n-dimensional vector of targets """ return 0.5*np.mean((y-f(X, theta))**2) ###Output _____no_output_____ ###Markdown Mean Squared Error: Partial DerivativesLet's work out what a partial derivative is for the MSE error loss for a linear model.\begin{align*}\frac{\partial J(\theta)}{\partial \theta_j} & = \frac{\partial}{\partial \theta_j} \frac{1}{2} \left( f_\theta(x) - y \right)^2 \\& = \left( f_\theta(x) - y \right) \cdot \frac{\partial}{\partial \theta_j} \left( f_\theta(x) - y \right) \\& = \left( f_\theta(x) - y \right) \cdot \frac{\partial}{\partial \theta_j} \left( \sum_{k=0}^d \theta_k \cdot x_k - y \right) \\& = \left( f_\theta(x) - y \right) \cdot x_j\end{align*} Mean Squared Error: The GradientWe can use this derivation to obtain an expression for the gradient of the MSE for a linear model\begin{align*}\nabla_\theta J (\theta) = \begin{bmatrix}\frac{\partial f(\theta)}{\partial \theta_1} \\\frac{\partial f(\theta)}{\partial \theta_2} \\\vdots \\\frac{\partial f(\theta)}{\partial \theta_d}\end{bmatrix}=\begin{bmatrix}\left( f_\theta(x) - y \right) \cdot x_1 \\\left( f_\theta(x) - y \right) \cdot x_2 \\\vdots \\\left( f_\theta(x) - y \right) \cdot x_d\end{bmatrix}=\left( f_\theta(x) - y \right) \cdot \bf{x}.\end{align*} Let's implement the gradient. ###Code def mse_gradient(theta, X, y): """The gradient of the cost function. Parameters: theta (np.array): d-dimensional vector of parameters X (np.array): (n,d)-dimensional design matrix y (np.array): n-dimensional vector of targets Returns: grad (np.array): d-dimensional gradient of the MSE """ return np.mean((f(X, theta) - y) * X.T, axis=1) ###Output _____no_output_____ ###Markdown The UCI Diabetes DatasetIn this section, we are going to again use the UCI Diabetes Dataset.* For each patient we have a access to a measurement of their body mass index (BMI) and a quantiative diabetes risk score (from 0-300). * We are interested in understanding how BMI affects an individual's diabetes risk. ###Code %matplotlib inline import matplotlib.pyplot as plt plt.rcParams['figure.figsize'] = [8, 4] import numpy as np import pandas as pd from sklearn import datasets # Load the diabetes dataset X, y = datasets.load_diabetes(return_X_y=True, as_frame=True) # add an extra column of onens X['one'] = 1 # Collect 20 data points and only use bmi dimension X_train = X.iloc[-20:].loc[:, ['bmi', 'one']] y_train = y.iloc[-20:] / 300 plt.scatter(X_train.loc[:,['bmi']], y_train, color='black') plt.xlabel('Body Mass Index (BMI)') plt.ylabel('Diabetes Risk') ###Output _____no_output_____ ###Markdown Gradient Descent for Linear RegressionPutting this together with the gradient descent algorithm, we obtain a learning method for training linear models.```pythontheta, theta_prev = random_initialization()while abs(J(theta) - J(theta_prev)) > conv_threshold: theta_prev = theta theta = theta_prev - step_size * (f(x, theta)-y) * x```This update rule is also known as the Least Mean Squares (LMS) or Widrow-Hoff learning rule. ###Code threshold = 1e-3 step_size = 4e-1 theta, theta_prev = np.array([2,1]), np.ones(2,) opt_pts = [theta] opt_grads = [] iter = 0 while np.linalg.norm(theta - theta_prev) > threshold: if iter % 100 == 0: print('Iteration %d. MSE: %.6f' % (iter, mean_squared_error(theta, X_train, y_train))) theta_prev = theta gradient = mse_gradient(theta, X_train, y_train) theta = theta_prev - step_size * gradient opt_pts += [theta] opt_grads += [gradient] iter += 1 x_line = np.stack([np.linspace(-0.1, 0.1, 10), np.ones(10,)]) y_line = opt_pts[-1].dot(x_line) plt.scatter(X_train.loc[:,['bmi']], y_train, color='black') plt.plot(x_line[0], y_line) plt.xlabel('Body Mass Index (BMI)') plt.ylabel('Diabetes Risk') ###Output _____no_output_____ ###Markdown Part 3: Ordinary Least SquaresIn practice, there is a more effective way than gradient descent to find linear model parameters.We will see this method here, which will lead to our first non-toy algorithm: Ordinary Least Squares. Review: The GradientThe gradient $\nabla_\theta f$ further extends the derivative to multivariate functions $f : \mathbb{R}^d \to \mathbb{R}$, and is defined at a point $\theta_0$ as$$ \nabla_\theta f (\theta_0) = \begin{bmatrix}\frac{\partial f(\theta_0)}{\partial \theta_1} \\\frac{\partial f(\theta_0)}{\partial \theta_2} \\\vdots \\\frac{\partial f(\theta_0)}{\partial \theta_d}\end{bmatrix}.$$In other words, the $j$-th entry of the vector $\nabla_\theta f (\theta_0)$ is the partial derivative $\frac{\partial f(\theta_0)}{\partial \theta_j}$ of $f$ with respect to the $j$-th component of $\theta$. The UCI Diabetes DatasetIn this section, we are going to again use the UCI Diabetes Dataset.* For each patient we have a access to a measurement of their body mass index (BMI) and a quantiative diabetes risk score (from 0-300). * We are interested in understanding how BMI affects an individual's diabetes risk. ###Code %matplotlib inline import matplotlib.pyplot as plt plt.rcParams['figure.figsize'] = [8, 4] import numpy as np import pandas as pd from sklearn import datasets # Load the diabetes dataset X, y = datasets.load_diabetes(return_X_y=True, as_frame=True) # add an extra column of onens X['one'] = 1 # Collect 20 data points X_train = X.iloc[-20:] y_train = y.iloc[-20:] plt.scatter(X_train.loc[:,['bmi']], y_train, color='black') plt.xlabel('Body Mass Index (BMI)') plt.ylabel('Diabetes Risk') ###Output _____no_output_____ ###Markdown Notation: Design MatrixMachine learning algorithms are most easily defined in the language of linear algebra. Therefore, it will be useful to represent the entire dataset as one matrix $X \in \mathbb{R}^{n \times d}$, of the form:$$ X = \begin{bmatrix}x^{(1)}_1 & x^{(1)}_2 & \ldots & x^{(1)}_d \\x^{(2)}_1 & x^{(2)}_2 & \ldots & x^{(2)}_d \\\vdots \\x^{(n)}_1 & x^{(n)}_2 & \ldots & x^{(n)}_d\end{bmatrix}=\begin{bmatrix}- & (x^{(1)})^\top & - \\- & (x^{(2)})^\top & - \\& \vdots & \\- & (x^{(n)})^\top & - \\\end{bmatrix}.$$ We can view the design matrix for the diabetes dataset. ###Code X_train.head() ###Output _____no_output_____ ###Markdown Notation: Design MatrixSimilarly, we can vectorize the target variables into a vector $y \in \mathbb{R}^n$ of the form$$ y = \begin{bmatrix}y^{(1)} \\y^{(2)} \\\vdots \\y^{(n)}\end{bmatrix}.$$ Squared Error in Matrix FormRecall that we may fit a linear model by choosing $\theta$ that minimizes the squared error:$$J(\theta)=\frac{1}{2}\sum_{i=1}^n(y^{(i)}-\theta^\top x^{(i)})^2$$In other words, we are looking for the best compromise in $\beta$ over all the data points. We can write this sum in matrix-vector form as:$$J(\theta) = \frac{1}{2} (y-X\theta)^\top(y-X\theta) = \frac{1}{2} \|y-X\theta\|^2,$$where $X$ is the design matrix and $\|\cdot\|$ denotes the Euclidean norm. The Gradient of the Squared ErrorWe can a gradient for the mean squared error as follows.\begin{align*}\nabla_\theta J(\theta) & = \nabla_\theta \frac{1}{2} (X \theta - y)^\top (X \theta - y) \\& = \frac{1}{2} \nabla_\theta \left( (X \theta)^\top (X \theta) - (X \theta)^\top y - y^\top (X \theta) + y^\top y \right) \\& = \frac{1}{2} \nabla_\theta \left( \theta^\top (X^\top X) \theta - 2(X \theta)^\top y \right) \\& = \frac{1}{2} \left( 2(X^\top X) \theta - 2X^\top y \right) \\& = (X^\top X) \theta - X^\top y\end{align*}We used the facts that $a^\top b = b^\top a$ (line 3), that $\nabla_x b^\top x = b$ (line 4), and that $\nabla_x x^\top A x = 2 A x$ for a symmetric matrix $A$ (line 4). Normal Equations<!-- We know from calculus that a function is minimized when its derivative is set to zero. In our case, our objective function is a (multivariate) quadratic; hence it only has one minimum, which is the global minimum. -->Setting the above derivative to zero, we obtain the *normal equations*:$$ (X^\top X) \theta = X^\top y.$$Hence, the value $\theta^*$ that minimizes this objective is given by:$$ \theta^* = (X^\top X)^{-1} X^\top y.$$ Note that we assumed that the matrix $(X^\top X)$ is invertible; if this is not the case, there are easy ways of addressing this issue. Let's apply the normal equations. ###Code import numpy as np theta_best = np.linalg.inv(X_train.T.dot(X_train)).dot(X_train.T).dot(y_train) theta_best_df = pd.DataFrame(data=theta_best[np.newaxis, :], columns=X.columns) theta_best_df ###Output _____no_output_____ ###Markdown We can now use our estimate of theta to compute predictions for 3 new data points. ###Code # Collect 3 data points for testing X_test = X.iloc[:3] y_test = y.iloc[:3] # generate predictions on the new patients y_test_pred = X_test.dot(theta_best) ###Output _____no_output_____ ###Markdown Let's visualize these predictions. ###Code # visualize the results plt.xlabel('Body Mass Index (BMI)') plt.ylabel('Diabetes Risk') plt.scatter(X_train.loc[:, ['bmi']], y_train) plt.scatter(X_test.loc[:, ['bmi']], y_test, color='red', marker='o') plt.plot(X_test.loc[:, ['bmi']], y_test_pred, 'x', color='red', mew=3, markersize=8) plt.legend(['Model', 'Prediction', 'Initial patients', 'New patients']) ###Output _____no_output_____ ###Markdown Algorithm: Ordinary Least Squares* __Type__: Supervised learning (regression)* __Model family__: Linear models* __Objective function__: Mean squared error* __Optimizer__: Normal equations Part 4: Non-Linear Least SquaresSo far, we have learned about a very simple linear model. These can capture only simple linear relationships in the data. How can we use what we learned so far to model more complex relationships?We will now see a simple approach to model complex non-linear relationships called *least squares*. Review: Polynomial FunctionsRecall that a polynomial of degree $p$ is a function of the form$$a_p x^p + a_{p-1} x^{p-1} + ... + a_{1} x + a_0.$$Below are some examples of polynomial functions. ###Code import warnings warnings.filterwarnings("ignore") plt.figure(figsize=(16,4)) x_vars = np.linspace(-2, 2) plt.subplot('131') plt.title('Quadratic Function') plt.plot(x_vars, x_vars**2) plt.legend(["$x^2$"]) plt.subplot('132') plt.title('Cubic Function') plt.plot(x_vars, x_vars**3) plt.legend(["$x^3$"]) plt.subplot('133') plt.title('Third Degree Polynomial') plt.plot(x_vars, x_vars**3 + 2*x_vars**2 + x_vars + 1) plt.legend(["$x^3 + 2 x^2 + x + 1$"]) ###Output _____no_output_____ ###Markdown Modeling Non-Linear Relationships With Polynomial Regression<!-- Note that the set of $p$-th degree polynomials forms a linear model with parameters $a_p, a_{p-1}, ..., a_0$.This means we can use our algorithms for linear models to learn non-linear features! -->Specifically, given a one-dimensional continuous variable $x$, we can defining a feature function $\phi : \mathbb{R} \to \mathbb{R}^{p+1}$ as$$ \phi(x) = \begin{bmatrix}1 \\x \\x^2 \\\vdots \\x^p\end{bmatrix}.$$ The class of models of the form$$ f_\theta(x) := \sum_{j=0}^p \theta_p x^p = \theta^\top \phi(x) $$with parameters $\theta$ and polynomial features $\phi$ is the set of $p$-degree polynomials. * This model is non-linear in the input variable $x$, meaning that we can model complex data relationships. * It is a linear model as a function of the parameters $\theta$, meaning that we can use our familiar ordinary least squares algorithm to learn these features. The UCI Diabetes DatasetIn this section, we are going to again use the UCI Diabetes Dataset.* For each patient we have a access to a measurement of their body mass index (BMI) and a quantiative diabetes risk score (from 0-300). * We are interested in understanding how BMI affects an individual's diabetes risk. ###Code %matplotlib inline import matplotlib.pyplot as plt plt.rcParams['figure.figsize'] = [8, 4] import numpy as np import pandas as pd from sklearn import datasets # Load the diabetes dataset X, y = datasets.load_diabetes(return_X_y=True, as_frame=True) # add an extra column of onens X['one'] = 1 # Collect 20 data points X_train = X.iloc[-20:] y_train = y.iloc[-20:] plt.scatter(X_train.loc[:,['bmi']], y_train, color='black') plt.xlabel('Body Mass Index (BMI)') plt.ylabel('Diabetes Risk') ###Output _____no_output_____ ###Markdown Diabetes Dataset: A Non-Linear FeaturizationLet's now obtain linear features for this dataset. ###Code X_bmi = X_train.loc[:, ['bmi']] X_bmi_p3 = pd.concat([X_bmi, X_bmi**2, X_bmi**3], axis=1) X_bmi_p3.columns = ['bmi', 'bmi2', 'bmi3'] X_bmi_p3['one'] = 1 X_bmi_p3.head() ###Output _____no_output_____ ###Markdown Diabetes Dataset: A Polynomial ModelBy training a linear model on this featurization of the diabetes set, we can obtain a polynomial model of diabetest risk as a function of BMI. ###Code # Fit a linear regression theta = np.linalg.inv(X_bmi_p3.T.dot(X_bmi_p3)).dot(X_bmi_p3.T).dot(y_train) # Show the learned polynomial curve x_line = np.linspace(-0.1, 0.1, 10) x_line_p3 = np.stack([x_line, x_line**2, x_line**3, np.ones(10,)], axis=1) y_train_pred = x_line_p3.dot(theta) plt.xlabel('Body Mass Index (BMI)') plt.ylabel('Diabetes Risk') plt.scatter(X_bmi, y_train) plt.plot(x_line, y_train_pred) ###Output _____no_output_____ ###Markdown Multivariate Polynomial RegressionWe can also take this approach to construct non-linear function of multiples variable by using multivariate polynomials.For example, a polynomial of degree $2$ over two variables $x_1, x_2$ is a function of the form<!-- $$a_{20} x_1^2 + a_{10} x_1 + a_{02} x_2^2 + a_{01} x_2 + a_{22} x_1^2 x_2^2 + a_{21} x_1^2 x_2 + a_{12} x_1 x_2^2 + a_11 x_1 x_2 + a_{00}.$$ -->$$a_{20} x_1^2 + a_{10} x_1 + a_{02} x_2^2 + a_{01} x_2 + a_{11} x_1 x_2 + a_{00}.$$ In general, a polynomial of degree $p$ over two variables $x_1, x_2$ is a function of the form$$f(x_1, x_2) = \sum_{i,j \geq 0 : i+j \leq p} a_{ij} x_1^i x_2^j.$$ In our two-dimensional example, this corresponds to a feature function $\phi : \mathbb{R}^2 \to \mathbb{R}^6$ of the form$$ \phi(x) = \begin{bmatrix}1 \\x_1 \\x_1^2 \\x_2 \\x_2^2 \\x_1 x_2\end{bmatrix}.$$The same approach holds for polynomials of an degree and any number of variables. Towards General Non-Linear FeaturesAny non-linear feature map $\phi(x) : \mathbb{R}^d \to \mathbb{R}^p$ can be used in this way to obtain general models of the form$$ f_\theta(x) := \theta^\top \phi(x) $$that are highly non-linear in $x$ but linear in $\theta$. For example, here is a way of modeling complex periodic functions via a sum of sines and cosines. ###Code import warnings warnings.filterwarnings("ignore") plt.figure(figsize=(16,4)) x_vars = np.linspace(-5, 5) plt.subplot('131') plt.title('Cosine Function') plt.plot(x_vars, np.cos(x_vars)) plt.legend(["$cos(x)$"]) plt.subplot('132') plt.title('Sine Function') plt.plot(x_vars, np.sin(2*x_vars)) plt.legend(["$x^3$"]) plt.subplot('133') plt.title('Combination of Sines and Cosines') plt.plot(x_vars, np.cos(x_vars) + np.sin(2*x_vars) + np.cos(4*x_vars)) plt.legend(["$cos(x) + sin(2x) + cos(4x)$"]) ###Output _____no_output_____
churn_project.ipynb
###Markdown Data Manipulation: a. Extract the 5th column & store it in ‘customer_5’b. Extract the 15th column & store it in ‘customer_15’c. Extract all the male senior citizens whose Payment Method is Electronic check & store the result in ‘senior_male_electronic’d. Extract all those customers whose tenure is greater than 70 months or their Monthly charges is more than 100$ & store the result in ‘customer_total_tenure’e. Extract all the customers whose Contract is of two years, payment method is Mailed check & the value of Churn is ‘Yes’ & store the result in ‘two_mail_yes’f. Extract 333 random records from the customer_churndataframe& store the result in ‘customer_333’g. Get the count of different levels from the ‘Churn’ column ###Code customer_5 = df.iloc[:,4] print(customer_5) customer_15 = df.iloc[:,14] customer_15 senior_male_electronic = (df['gender']=='male') & (df['PaymentMethod']=="Electronic check") & (df['SeniorCitizen']==1) senior_male_electronic customer_total_tenure = df[(df['tenure']>70) | (df['MonthlyCharges']>100)] customer_total_tenure two_mail_yes = df[(df['Contract']=='Two year') & (df['PaymentMethod']=='Mailed check') & (df['Churn']=='Yes')] two_mail_yes customer_333 = df.sample(n=333) customer_333 df['Churn'].value_counts() ###Output _____no_output_____ ###Markdown Data Visualization: a. Build a bar-plot for the ’InternetService’ column:i. Set x-axis label to ‘Categories of Internet Service’ii. Set y-axis label to ‘Count of Categories’iii. Set the title of plot to be ‘Distribution of Internet Service’iv. Set the color of the bars to be ‘orange’b. Build a histogram for the ‘tenure’ column:i. Set the number of bins to be 30 ii. Set the color of the bins to be ‘green’iii. Assign the title ‘Distribution of tenure’c. Build a scatter-plot between ‘MonthlyCharges’ & ‘tenure’. Map ‘MonthlyCharges’ to the y-axis & ‘tenure’ to the ‘x-axis’:i. Assign the points a color of ‘brown’ii. Set the x-axis label to ‘Tenure of customer’iii. Set the y-axis label to ‘Monthly Charges of customer’iv. Set the title to ‘Tenure vs Monthly Charges’d. Build a box-plot between ‘tenure’ & ‘Contract’. Map ‘tenure’ on the y-axis & ‘Contract’ on the x-axis ###Code df['InternetService'].value_counts().tolist() # for understanding df['InternetService'].value_counts().keys().tolist() #for understanding plt.figure(figsize=(10,7)) plt.bar(df['InternetService'].value_counts().keys().tolist(),df['InternetService'].value_counts().tolist(),color='orange') plt.xlabel("Categories of Internet Service") plt.ylabel("Count of Categories") plt.title("Distribution of Internet Service") plt.show() plt.figure(figsize=(10,7)) plt.hist(df['tenure'],bins=30,color='g') plt.title('‘Distribution of tenure’') plt.show() plt.figure(figsize=(10,7)) plt.scatter(x=df['tenure'],y=df['MonthlyCharges'],color='brown') plt.xlabel("Tenure of customer") plt.ylabel("Monthly Charges of customer") plt.title("Tenure vs Monthly Charges") plt.show() df.boxplot(column=['tenure'],by=['Contract']) plt.show() ###Output _____no_output_____ ###Markdown Linear Regression: a. Build a simple linear model where dependent variable is ‘MonthlyCharges’ and independent variable is ‘tenure’i. Divide the dataset into train and test sets in 70:30 ratio. ii. Build the model on train set and predict the values on test set iii. After predicting the values, find the root mean square error iv. Find out the error in prediction & store the result in ‘error’v. Find the root mean square error ###Code from sklearn.linear_model import LinearRegression from sklearn.model_selection import train_test_split x=df[['tenure']] y=df['MonthlyCharges'] x.head() y.head() x_train,x_test,y_train,y_test = train_test_split(x,y,test_size=0.30,random_state=0) print("Independent Variable for Training the model", x_train.shape, '\n', x_train.head(2)) print("Target Variable for Training the model", y_train.shape, '\n', y_train.head(2)) print("Independent Variable for Testing the model", x_test.shape, '\n', x_test.head(2)) print("Target Variable for Comparing the Prediction the model's prediction", y_test.shape, '\n', y_test.head(2)) model=LinearRegression() model.fit(x_train,y_train) # Intercept inter = model.intercept_ inter # Slope slope = model.coef_ slope y_pred = model.predict(x_test) y_pred[:5] aftermodel = pd.DataFrame({"Actual Sales": y_test, "Predicted Sales" : np.around(y_pred, 2)}) aftermodel[:10] #Error in predicted value aftermodel['Error']= aftermodel['Actual Sales'] - aftermodel['Predicted Sales'] aftermodel from sklearn.metrics import mean_squared_error # Calculating Mean Squared Error and Root Mean Squared Error mse = mean_squared_error(y_test, y_pred) rmse = round(np.sqrt(mse), 2) print("Root mean square error : ",rmse,"\n") print(y_pred.mean()) ###Output Root mean square error : 29.39 64.85362533013019 ###Markdown Logistic Regression: a. Build a simple logistic regression modelwhere dependent variable is ‘Churn’ & independent variable is ‘MonthlyCharges’i. Divide the dataset in 65:35 ratio ii. Build the model on train set and predict the values on test set iii. Build the confusion matrix and get the accuracy score ###Code from sklearn.linear_model import LogisticRegression x = df[['MonthlyCharges']] y = df['Churn'] x_train,x_test,y_train,y_test = train_test_split(x,y,test_size=0.35,random_state=150) log_model = LogisticRegression() log_model.fit(x_train,y_train) predict = log_model.predict(x_test) predict[:5] # Comparing Actual and Predicted Data comp = pd.DataFrame({"Actual" : y_test, "Predicted": predict}) comp.head() from sklearn.metrics import confusion_matrix, accuracy_score print("confusion_matrix\n",confusion_matrix(y_test,predict)) print("accuracy_score\n",accuracy_score(y_test,predict)) ###Output accuracy_score 0.7493917274939172 ###Markdown Multiple Logistic Regression: b. Build a multiple logistic regression model where dependent variable is ‘Churn’ & independent variables are ‘tenure’ & ‘MonthlyCharges’i. Divide the dataset in 80:20 ratio ii. Build the model on train set and predict the values on test set iii. Build the confusion matrix and get the accuracy score ###Code x = df[['tenure','MonthlyCharges']] y = df['Churn'] x_train,x_test,y_train,y_test = train_test_split(x,y,test_size=0.20,random_state=150) multi_log_model = LogisticRegression() multi_log_model.fit(x_train,y_train) pred = multi_log_model.predict(x_test) pred[:10] # Comparing Actual and Predicted Data comp = pd.DataFrame({"Actual" : y_test, "Predicted": pred}) comp.head() print("confusion_matrix\n",confusion_matrix(y_test,pred)) print("accuracy_score\n",accuracy_score(y_test,pred)) ###Output accuracy_score 0.8076650106458482 ###Markdown Decision Tree: a. Build a decision tree model where dependent variable is ‘Churn’ & independent variable is ‘tenure’i. Divide the dataset in 80:20 ratio ii. Build the model on train set and predict the values on test setiii. Build the confusion matrix and calculate the accuracy ###Code x = df[['tenure']] y = df['Churn'] from sklearn.tree import DecisionTreeClassifier x_train,x_test,y_train,y_test = train_test_split(x,y,test_size=0.20,random_state=150) tree = DecisionTreeClassifier() tree.fit(x_train,y_train) pred = tree.predict(x_test) print("confusion_matrix \n",confusion_matrix(y_test,pred)) print("accuracy_score",accuracy_score(y_test,pred)) ###Output accuracy_score 0.7629524485450674 ###Markdown Random Forest: a. Build a Random Forest model where dependent variable is ‘Churn’ & independent variables are ‘tenure’ and ‘MonthlyCharges’i. Divide the dataset in 70:30 ratio ii. Build the model on train set and predict the values on test set iii. Build the confusion matrix and calculate the accuracy ###Code x = df[['tenure','MonthlyCharges']] y = df['Churn'] from sklearn.ensemble import RandomForestClassifier x_train,x_test,y_train,y_test = train_test_split(x,y,test_size=0.30,random_state=0) forest = RandomForestClassifier() forest.fit(x_train,y_train) pred = forest.predict(x_test) pred[:5] print("confusion_matrix \n",confusion_matrix(y_test,pred)) print("accuracy_score",accuracy_score(y_test,pred)) from sklearn import metrics print('Error Metrics') em = metrics.classification_report(y_test, pred) print(em) ###Output Error Metrics precision recall f1-score support No 0.80 0.86 0.83 1560 Yes 0.51 0.41 0.45 553 accuracy 0.74 2113 macro avg 0.66 0.63 0.64 2113 weighted avg 0.73 0.74 0.73 2113 ###Markdown Customer bank churn project The dataset consists of a randomly sampled population of a banking customers detailing demographics (independent variables) and whether a customer left (or stayed) the bank within the last 6 months (dependent variable).Project goal is to predict whether will customer leave or not. Data exploring ###Code import pandas as pd import numpy as np import matplotlib.pyplot as plt %matplotlib inline import seaborn as sns from matplotlib import cm df=pd.read_csv("Churn-Modelling.csv") df.head() ###Output _____no_output_____ ###Markdown Column information:- RowNumber - number of rows- CustomerId - customer id number- Surname - customer's surname- CreditScore - number between 300–850 that depicts a consumer's creditworthiness. The higher the score, the better a borrower looks to potential lenders- Geography - customer's state- Gender - female/male- Age - customer's age- Tenure - period or duration for which the loan amount is sanctioned- Balance - the amount of money held in a bank account at a given moment- NumOfProducts - number any facilities or services related to cash management, including treasury, depository, overdraft, credit or debit card, purchase card, electronic funds transfer and other cash management arrangements- HasCrCard - 1-has credit card ,0 - hasn't credit card- IsActiveMember - 1 - is active member , 0 - is not active member- EstimatedSalary - estimated salary- Exited - 1 - customer left , 0 - customer stayed ###Code df.info() ###Output <class 'pandas.core.frame.DataFrame'> RangeIndex: 10000 entries, 0 to 9999 Data columns (total 14 columns): # Column Non-Null Count Dtype --- ------ -------------- ----- 0 RowNumber 10000 non-null int64 1 CustomerId 10000 non-null int64 2 Surname 10000 non-null object 3 CreditScore 10000 non-null int64 4 Geography 10000 non-null object 5 Gender 10000 non-null object 6 Age 10000 non-null int64 7 Tenure 10000 non-null int64 8 Balance 10000 non-null float64 9 NumOfProducts 10000 non-null int64 10 HasCrCard 10000 non-null int64 11 IsActiveMember 10000 non-null int64 12 EstimatedSalary 10000 non-null float64 13 Exited 10000 non-null int64 dtypes: float64(2), int64(9), object(3) memory usage: 976.6+ KB ###Markdown We dont have any missing values. But we do have columns that we don't need like "RowNumber", "CustomerId","Surname" so we will just drop them and also we will rename some of columns name for better understanding. ###Code df2=df.drop(["RowNumber","CustomerId","Surname"],axis=1) #Renaming columns df2.columns=['credit_score', 'state', 'gender', 'age', 'tenure', 'balance', 'number_of_products', 'credit_card', 'active_member', 'estimated_salary', 'churn'] df2.head() ###Output _____no_output_____ ###Markdown Data visualization Balance column We want to see how many customers leave depend on their amount of money in bank account. ###Code #balance of customers which are not leaving balance_churn_no=df2[df2.churn==0].balance #balance of customers which are leaving balance_churn_yes=df2[df2.churn==1].balance #Plot histogram plt.hist([balance_churn_yes, balance_churn_no], color=['red', 'blue'], label=['Churn=Yes', 'Churn=No']) plt.legend() plt.xlabel('balance') plt.ylabel('Number of customers') ###Output _____no_output_____ ###Markdown As we can see that over 3000 customers hasn't left even though they don't have no money in bank account. We also see that he most customers left with no money and around 100 000$. Tenure column We want to see if customers with longer tenure would stay and vice versa. ###Code #tenure of customers which are not leaving tenure_churn_no=df2[df2.churn==0].tenure #tenure of customers whicht are not leaving tenure_churn_yes=df2[df2.churn==1].tenure #Plot histogram plt.hist([tenure_churn_yes, tenure_churn_no], color=['red', 'blue'], label=['Churn=Yes', 'Churn=No']) plt.legend() plt.xlabel('tenure') plt.ylabel('Number of customers') ###Output _____no_output_____ ###Markdown As we could assume the customers with longer tenure have stayed. Gender column Let's see how many females and males are leaving and how many are staying. ###Code # Count how many females and males stayed count_churn=df2.groupby('gender')['churn'].apply(lambda x: (x==0).sum()).reset_index(name='Number of customers') color = cm.viridis(np.linspace(.4, .8, 30)) count_churn= count_churn.sort_values("Number of customers" , ascending=[False]) count_churn.plot.bar(x='gender', y='Number of customers', color=color , figsize=(12,7)) ###Output _____no_output_____ ###Markdown We can see that more males have stayed then females. Age column Let's see how ages affects on staying or leaving. ###Code #age of customers which are not leaving age_churn_no=df2[df2.churn==0].age #age of customers which are not leaving age_churn_yes=df2[df2.churn==1].age #plot histogram plt.hist([age_churn_yes, age_churn_no], color=['red', 'blue'], label=['Churn=Yes', 'Churn=No']) plt.legend() plt.xlabel('age') plt.ylabel('Number of customers') ###Output _____no_output_____ ###Markdown We can see that the customer between 30 and 40 years old are staying. ###Code # Count percentage of churns by state count_churn=df2.groupby('state')['churn'].apply(lambda x: (x==1).sum()).reset_index(name='Percentage of churn') count_churn= count_churn.sort_values("Percentage of churn" , ascending=[False]) count_churn.plot.pie(x="state",y='Percentage of churn', autopct='%1.1f%%',labels=df2["state"].unique(),figsize=(10,7)) ###Output _____no_output_____ ###Markdown We can see that France and Spain have the most customers that left. Preparing dataset for model ###Code df2.head() # define a function to discover columns with categorical varibales def discover_categorical_columns(df): """ This function takes dataframe as an input, goes through columns in a dataframe to check if column is of an object type. If the colunm is of an object type it means column contains categorical variables. Function than prints all unique categorical values of a column Args: df (pd.DataFrame) - only requried argument for the function Returns: Prints unique categorical values for object type columns. """ for column in df: if df[column].dtype=='object': print('{} : {}'.format(column, df[column].unique())) discover_categorical_columns(df2) ###Output state : ['France' 'Spain' 'Germany'] gender : ['Female' 'Male'] ###Markdown We only have 2 object type columns with maximum 3 unique values. We can replace 'Female' and 'Male' categorical values of 'gender' column to 1's and 0's. And for "state" column we will use pd.get_dummies() ###Code df2.info() # replace 'Female' with 1; replace 'Male' with 0 df2.replace({'Female':1, 'Male':0}, value=None, inplace=True) #function for creating dummy column def create_dummies(df,column_name): dummies=pd.get_dummies(df[column_name],prefix=column_name) df=pd.concat([df,dummies],axis=1) return df #dummy columns df3=create_dummies(df2,"state") df4=df3.drop(["state"],axis=1) df4.head() # visualizing correlations plt.figure(figsize=(10,10)) sns.heatmap(df4.corr(), annot=True, cmap='coolwarm') ###Output _____no_output_____ ###Markdown Everything look fine and we can now scale data for better result. Scaling data ###Code # min-max scaling from sklearn.preprocessing import MinMaxScaler, RobustScaler scaler=MinMaxScaler() data_scaled_array=scaler.fit_transform(df4) df5=pd.DataFrame(data_scaled_array, columns=df4.columns) df5.head() ###Output _____no_output_____ ###Markdown Building Models ###Code from sklearn.model_selection import train_test_split # Defining features and target column X = df5.drop(columns='churn', axis ='columns') y = df5.churn #Splitting our data on train and test train_X,test_X,train_y,test_y=train_test_split(X,y,train_size=0.8,random_state=1) ###Output _____no_output_____ ###Markdown Balance data ###Code train_y.value_counts() ###Output _____no_output_____ ###Markdown As we can see we have a lot more customers that are staying then that are leaving. Because of this we need to equalize those parameters to get better result. We will use SMOTE technique.MOTE (Synthetic Minority Oversampling TEchnique) consists of synthesizing elements for the minority class, based on those that already exist. It works randomly picingk a point from the minority class and computing the k-nearest neighbors for this point. The synthetic points are added between the chosen point and its neighbors. ###Code #Importing SMOTE from imblearn.over_sampling import SMOTE smote = SMOTE(sampling_strategy='minority') X_s, y_s = smote.fit_sample(X, y) # check value counts y_s.value_counts() ###Output _____no_output_____ ###Markdown As we can see our target column is now balanced. So we can now build our models. ###Code #Splitting our data on train and test train_X, test_X, train_y,test_y = train_test_split(X_s, y_s, test_size=0.2, random_state=15, stratify=y_s) train_y.value_counts() ###Output _____no_output_____ ###Markdown Random Forest classifier model ###Code # Importing RandomForestClassifier from sklearn.ensemble import RandomForestClassifier #Defining model rf=RandomForestClassifier(n_estimators=200, random_state=1, min_samples_leaf=2) # Fitting the model rf.fit(train_X,train_y) # predicting values on test set predictions_rf=rf.predict(test_X) ###Output _____no_output_____ ###Markdown Random Forest classifier accuracy ###Code #Importing cross_val_score and accuracy_score from sklearn.model_selection import cross_val_score from sklearn.metrics import accuracy_score # calculating accuracy with accuracy_score() accuracy_rf=accuracy_score(test_y, predictions_rf) accuracy_rf # calculating accuracy result with cross_val_score() accuracy_cross_val_rf=cross_val_score(rf, X_s, y_s, cv=10) accuracy_cross_val_rf #calculating cross_val_score mean accuracy_cross_val_rf=np.mean(accuracy_cross_val_rf) accuracy_cross_val_rf ###Output _____no_output_____ ###Markdown Calculating f-1 score in RandomForest model ###Code #Importing classification_report from sklearn.metrics import classification_report print(classification_report(test_y,predictions_rf)) ###Output precision recall f1-score support 0.0 0.90 0.89 0.89 1593 1.0 0.89 0.90 0.90 1593 accuracy 0.89 3186 macro avg 0.89 0.89 0.89 3186 weighted avg 0.89 0.89 0.89 3186 ###Markdown We can see that our model can accurately predict 90% of churns. Artificial neural networks model ###Code # import MLPClassifier and make an instance from sklearn.neural_network import MLPClassifier # Defining our model mlp=MLPClassifier(hidden_layer_sizes=(10,10), activation="relu",max_iter=1000) # fitting the model mlp.fit(train_X, train_y) # predicting values on test set predictions_mlp=mlp.predict(test_X) ###Output _____no_output_____ ###Markdown Artificial neural networks accuracy ###Code # calculating accuracy with accuracy_score() accuracy_mlp=accuracy_score(test_y, predictions_mlp) accuracy_mlp # calculating accuracy result with cross_val_score() accuracy_cross_val_mlp=cross_val_score(mlp, X_s, y_s, cv=10) accuracy_cross_val_mlp #calculating cross_val_score mean accuracy_cross_val_mlp=np.mean(accuracy_cross_val_mlp) accuracy_cross_val_mlp ###Output _____no_output_____ ###Markdown Calculating f-1 score in Artifical Neural networks ###Code print(classification_report(test_y,predictions_mlp)) ###Output precision recall f1-score support 0.0 0.78 0.83 0.80 1593 1.0 0.82 0.77 0.79 1593 accuracy 0.80 3186 macro avg 0.80 0.80 0.80 3186 weighted avg 0.80 0.80 0.80 3186
gray2color.ipynb
###Markdown Mounting google drive to save trained model ###Code from google.colab import drive drive.mount('/gdrive') ###Output Go to this URL in a browser: https://accounts.google.com/o/oauth2/auth?client_id=947318989803-6bn6qk8qdgf4n4g3pfee6491hc0brc4i.apps.googleusercontent.com&redirect_uri=urn%3aietf%3awg%3aoauth%3a2.0%3aoob&response_type=code&scope=email%20https%3a%2f%2fwww.googleapis.com%2fauth%2fdocs.test%20https%3a%2f%2fwww.googleapis.com%2fauth%2fdrive%20https%3a%2f%2fwww.googleapis.com%2fauth%2fdrive.photos.readonly%20https%3a%2f%2fwww.googleapis.com%2fauth%2fpeopleapi.readonly Enter your authorization code: ·········· Mounted at /gdrive ###Markdown Importing all Dependencies ###Code import tensorflow as tf import tensorflow.keras as keras import matplotlib.pyplot as plt import numpy as np import os from PIL import Image from skimage.color import rgb2lab, lab2rgb, rgb2gray ###Output _____no_output_____ ###Markdown Downloading training dataset using curlI found a alternative [kaggle dataset](https://www.kaggle.com/greatgamedota/ffhq-face-data-set) paste download link of dataset here. These link get obselete very soon so paste new download link.Link can be get when you download it into local and in download section of browser you can get link for download. Size: Approx 2GB ###Code !curl -o "/content/faces.zip" "https://storage.googleapis.com/kaggle-data-sets/379454/735991/bundle/archive.zip?GoogleAccessId=web-data@kaggle-161607.iam.gserviceaccount.com&Expires=1590056491&Signature=iZbEXd9tLS6urEio8l4sGXEPbUdBOvwoyvqweIq%2BSEprSrQYGQ6AwS3Us93g%2FsV7OHMXI2dtXl0ZILOnA95nAjZv2u9DbCEjjsmqdZU3zuGXQpthruhAJ2ybVyFIBeTCuQdx1%2FoPHp3K%2FUSz03SODoeJG6zTg1QOEcp2vfIytplNIYIEVbHYuzokWN2ahDfW4JOQsytO%2F8TTJxB7fn2Yu16STvEf%2FKSpM7zHqvWqNyhh7d3hPgBIc1M6hZUHUXXWXT5PkJWeSa3gXSysvCOhemVWJyfhqX0tDs%2FM4MUtg1ZMe0V80KdzXrueBnIr5aVzuRn1PDFQvzANmGAkj1hrrQ%3D%3D&response-content-disposition=attachment%3B+filename%3Dffhq-face-data-set.zip" ###Output % Total % Received % Xferd Average Speed Time Time Time Current Dload Upload Total Spent Left Speed 100 2015M 100 2015M 0 0 68.3M 0 0:00:29 0:00:29 --:--:-- 65.5M ###Markdown Unzipping the dataset ###Code !mkdir "/content/faces" !unzip "/content/faces.zip" -d "/content/faces" !rm faces.zip ###Output Streaming output truncated to the last 5000 lines. inflating: /content/faces/thumbnails128x128/65001.png inflating: /content/faces/thumbnails128x128/65002.png inflating: /content/faces/thumbnails128x128/65003.png inflating: /content/faces/thumbnails128x128/65004.png inflating: /content/faces/thumbnails128x128/65005.png inflating: /content/faces/thumbnails128x128/65006.png inflating: /content/faces/thumbnails128x128/65007.png inflating: /content/faces/thumbnails128x128/65008.png inflating: /content/faces/thumbnails128x128/65009.png inflating: /content/faces/thumbnails128x128/65010.png inflating: /content/faces/thumbnails128x128/65011.png inflating: /content/faces/thumbnails128x128/65012.png inflating: /content/faces/thumbnails128x128/65013.png inflating: /content/faces/thumbnails128x128/65014.png inflating: /content/faces/thumbnails128x128/65015.png inflating: /content/faces/thumbnails128x128/65016.png inflating: /content/faces/thumbnails128x128/65017.png inflating: /content/faces/thumbnails128x128/65018.png inflating: /content/faces/thumbnails128x128/65019.png inflating: /content/faces/thumbnails128x128/65020.png inflating: /content/faces/thumbnails128x128/65021.png inflating: /content/faces/thumbnails128x128/65022.png inflating: /content/faces/thumbnails128x128/65023.png inflating: /content/faces/thumbnails128x128/65024.png inflating: /content/faces/thumbnails128x128/65025.png inflating: /content/faces/thumbnails128x128/65026.png inflating: /content/faces/thumbnails128x128/65027.png inflating: /content/faces/thumbnails128x128/65028.png inflating: /content/faces/thumbnails128x128/65029.png inflating: /content/faces/thumbnails128x128/65030.png inflating: /content/faces/thumbnails128x128/65031.png inflating: /content/faces/thumbnails128x128/65032.png inflating: /content/faces/thumbnails128x128/65033.png inflating: /content/faces/thumbnails128x128/65034.png inflating: /content/faces/thumbnails128x128/65035.png inflating: /content/faces/thumbnails128x128/65036.png inflating: /content/faces/thumbnails128x128/65037.png inflating: /content/faces/thumbnails128x128/65038.png inflating: /content/faces/thumbnails128x128/65039.png inflating: /content/faces/thumbnails128x128/65040.png inflating: /content/faces/thumbnails128x128/65041.png inflating: /content/faces/thumbnails128x128/65042.png inflating: /content/faces/thumbnails128x128/65043.png inflating: /content/faces/thumbnails128x128/65044.png inflating: /content/faces/thumbnails128x128/65045.png inflating: /content/faces/thumbnails128x128/65046.png inflating: /content/faces/thumbnails128x128/65047.png inflating: /content/faces/thumbnails128x128/65048.png inflating: /content/faces/thumbnails128x128/65049.png inflating: /content/faces/thumbnails128x128/65050.png inflating: /content/faces/thumbnails128x128/65051.png inflating: /content/faces/thumbnails128x128/65052.png inflating: /content/faces/thumbnails128x128/65053.png inflating: /content/faces/thumbnails128x128/65054.png inflating: /content/faces/thumbnails128x128/65055.png inflating: /content/faces/thumbnails128x128/65056.png inflating: /content/faces/thumbnails128x128/65057.png inflating: /content/faces/thumbnails128x128/65058.png inflating: /content/faces/thumbnails128x128/65059.png inflating: /content/faces/thumbnails128x128/65060.png inflating: /content/faces/thumbnails128x128/65061.png inflating: /content/faces/thumbnails128x128/65062.png inflating: /content/faces/thumbnails128x128/65063.png inflating: /content/faces/thumbnails128x128/65064.png inflating: /content/faces/thumbnails128x128/65065.png inflating: /content/faces/thumbnails128x128/65066.png inflating: /content/faces/thumbnails128x128/65067.png inflating: /content/faces/thumbnails128x128/65068.png inflating: /content/faces/thumbnails128x128/65069.png inflating: /content/faces/thumbnails128x128/65070.png inflating: /content/faces/thumbnails128x128/65071.png inflating: /content/faces/thumbnails128x128/65072.png inflating: /content/faces/thumbnails128x128/65073.png inflating: /content/faces/thumbnails128x128/65074.png inflating: /content/faces/thumbnails128x128/65075.png inflating: /content/faces/thumbnails128x128/65076.png inflating: /content/faces/thumbnails128x128/65077.png inflating: /content/faces/thumbnails128x128/65078.png inflating: /content/faces/thumbnails128x128/65079.png inflating: /content/faces/thumbnails128x128/65080.png inflating: /content/faces/thumbnails128x128/65081.png inflating: /content/faces/thumbnails128x128/65082.png inflating: /content/faces/thumbnails128x128/65083.png inflating: /content/faces/thumbnails128x128/65084.png inflating: /content/faces/thumbnails128x128/65085.png inflating: /content/faces/thumbnails128x128/65086.png inflating: /content/faces/thumbnails128x128/65087.png inflating: /content/faces/thumbnails128x128/65088.png inflating: /content/faces/thumbnails128x128/65089.png inflating: /content/faces/thumbnails128x128/65090.png inflating: /content/faces/thumbnails128x128/65091.png inflating: /content/faces/thumbnails128x128/65092.png inflating: /content/faces/thumbnails128x128/65093.png inflating: /content/faces/thumbnails128x128/65094.png inflating: /content/faces/thumbnails128x128/65095.png inflating: /content/faces/thumbnails128x128/65096.png inflating: /content/faces/thumbnails128x128/65097.png inflating: /content/faces/thumbnails128x128/65098.png inflating: /content/faces/thumbnails128x128/65099.png inflating: /content/faces/thumbnails128x128/65100.png inflating: /content/faces/thumbnails128x128/65101.png inflating: /content/faces/thumbnails128x128/65102.png inflating: /content/faces/thumbnails128x128/65103.png inflating: /content/faces/thumbnails128x128/65104.png inflating: /content/faces/thumbnails128x128/65105.png inflating: /content/faces/thumbnails128x128/65106.png inflating: /content/faces/thumbnails128x128/65107.png inflating: /content/faces/thumbnails128x128/65108.png inflating: /content/faces/thumbnails128x128/65109.png inflating: /content/faces/thumbnails128x128/65110.png inflating: /content/faces/thumbnails128x128/65111.png inflating: /content/faces/thumbnails128x128/65112.png inflating: /content/faces/thumbnails128x128/65113.png inflating: /content/faces/thumbnails128x128/65114.png inflating: /content/faces/thumbnails128x128/65115.png inflating: /content/faces/thumbnails128x128/65116.png inflating: /content/faces/thumbnails128x128/65117.png inflating: /content/faces/thumbnails128x128/65118.png inflating: /content/faces/thumbnails128x128/65119.png inflating: /content/faces/thumbnails128x128/65120.png inflating: /content/faces/thumbnails128x128/65121.png inflating: /content/faces/thumbnails128x128/65122.png inflating: /content/faces/thumbnails128x128/65123.png inflating: /content/faces/thumbnails128x128/65124.png inflating: /content/faces/thumbnails128x128/65125.png inflating: /content/faces/thumbnails128x128/65126.png inflating: /content/faces/thumbnails128x128/65127.png inflating: /content/faces/thumbnails128x128/65128.png inflating: /content/faces/thumbnails128x128/65129.png inflating: /content/faces/thumbnails128x128/65130.png inflating: /content/faces/thumbnails128x128/65131.png inflating: /content/faces/thumbnails128x128/65132.png inflating: /content/faces/thumbnails128x128/65133.png inflating: /content/faces/thumbnails128x128/65134.png inflating: /content/faces/thumbnails128x128/65135.png inflating: /content/faces/thumbnails128x128/65136.png inflating: /content/faces/thumbnails128x128/65137.png inflating: /content/faces/thumbnails128x128/65138.png inflating: /content/faces/thumbnails128x128/65139.png inflating: /content/faces/thumbnails128x128/65140.png inflating: /content/faces/thumbnails128x128/65141.png inflating: /content/faces/thumbnails128x128/65142.png inflating: /content/faces/thumbnails128x128/65143.png inflating: /content/faces/thumbnails128x128/65144.png inflating: /content/faces/thumbnails128x128/65145.png inflating: /content/faces/thumbnails128x128/65146.png inflating: /content/faces/thumbnails128x128/65147.png inflating: /content/faces/thumbnails128x128/65148.png inflating: /content/faces/thumbnails128x128/65149.png inflating: /content/faces/thumbnails128x128/65150.png inflating: /content/faces/thumbnails128x128/65151.png inflating: /content/faces/thumbnails128x128/65152.png inflating: /content/faces/thumbnails128x128/65153.png inflating: /content/faces/thumbnails128x128/65154.png inflating: /content/faces/thumbnails128x128/65155.png inflating: /content/faces/thumbnails128x128/65156.png inflating: /content/faces/thumbnails128x128/65157.png inflating: /content/faces/thumbnails128x128/65158.png inflating: /content/faces/thumbnails128x128/65159.png inflating: /content/faces/thumbnails128x128/65160.png inflating: /content/faces/thumbnails128x128/65161.png inflating: /content/faces/thumbnails128x128/65162.png inflating: /content/faces/thumbnails128x128/65163.png inflating: /content/faces/thumbnails128x128/65164.png inflating: /content/faces/thumbnails128x128/65165.png inflating: /content/faces/thumbnails128x128/65166.png inflating: /content/faces/thumbnails128x128/65167.png inflating: /content/faces/thumbnails128x128/65168.png inflating: /content/faces/thumbnails128x128/65169.png inflating: /content/faces/thumbnails128x128/65170.png inflating: /content/faces/thumbnails128x128/65171.png inflating: /content/faces/thumbnails128x128/65172.png inflating: /content/faces/thumbnails128x128/65173.png inflating: /content/faces/thumbnails128x128/65174.png inflating: /content/faces/thumbnails128x128/65175.png inflating: /content/faces/thumbnails128x128/65176.png inflating: /content/faces/thumbnails128x128/65177.png inflating: /content/faces/thumbnails128x128/65178.png inflating: /content/faces/thumbnails128x128/65179.png inflating: /content/faces/thumbnails128x128/65180.png inflating: /content/faces/thumbnails128x128/65181.png inflating: /content/faces/thumbnails128x128/65182.png inflating: /content/faces/thumbnails128x128/65183.png inflating: /content/faces/thumbnails128x128/65184.png inflating: /content/faces/thumbnails128x128/65185.png inflating: /content/faces/thumbnails128x128/65186.png inflating: /content/faces/thumbnails128x128/65187.png inflating: /content/faces/thumbnails128x128/65188.png inflating: /content/faces/thumbnails128x128/65189.png inflating: /content/faces/thumbnails128x128/65190.png inflating: /content/faces/thumbnails128x128/65191.png inflating: /content/faces/thumbnails128x128/65192.png inflating: /content/faces/thumbnails128x128/65193.png inflating: /content/faces/thumbnails128x128/65194.png inflating: /content/faces/thumbnails128x128/65195.png inflating: /content/faces/thumbnails128x128/65196.png inflating: /content/faces/thumbnails128x128/65197.png inflating: /content/faces/thumbnails128x128/65198.png inflating: /content/faces/thumbnails128x128/65199.png inflating: /content/faces/thumbnails128x128/65200.png inflating: /content/faces/thumbnails128x128/65201.png inflating: /content/faces/thumbnails128x128/65202.png inflating: /content/faces/thumbnails128x128/65203.png inflating: /content/faces/thumbnails128x128/65204.png inflating: /content/faces/thumbnails128x128/65205.png inflating: /content/faces/thumbnails128x128/65206.png inflating: /content/faces/thumbnails128x128/65207.png inflating: /content/faces/thumbnails128x128/65208.png inflating: /content/faces/thumbnails128x128/65209.png inflating: /content/faces/thumbnails128x128/65210.png inflating: /content/faces/thumbnails128x128/65211.png inflating: /content/faces/thumbnails128x128/65212.png inflating: /content/faces/thumbnails128x128/65213.png inflating: /content/faces/thumbnails128x128/65214.png inflating: /content/faces/thumbnails128x128/65215.png inflating: /content/faces/thumbnails128x128/65216.png inflating: /content/faces/thumbnails128x128/65217.png inflating: /content/faces/thumbnails128x128/65218.png inflating: /content/faces/thumbnails128x128/65219.png inflating: /content/faces/thumbnails128x128/65220.png inflating: /content/faces/thumbnails128x128/65221.png inflating: /content/faces/thumbnails128x128/65222.png inflating: /content/faces/thumbnails128x128/65223.png inflating: /content/faces/thumbnails128x128/65224.png inflating: /content/faces/thumbnails128x128/65225.png inflating: /content/faces/thumbnails128x128/65226.png inflating: /content/faces/thumbnails128x128/65227.png inflating: /content/faces/thumbnails128x128/65228.png inflating: /content/faces/thumbnails128x128/65229.png inflating: /content/faces/thumbnails128x128/65230.png inflating: /content/faces/thumbnails128x128/65231.png inflating: /content/faces/thumbnails128x128/65232.png inflating: /content/faces/thumbnails128x128/65233.png inflating: /content/faces/thumbnails128x128/65234.png inflating: /content/faces/thumbnails128x128/65235.png inflating: /content/faces/thumbnails128x128/65236.png inflating: /content/faces/thumbnails128x128/65237.png inflating: /content/faces/thumbnails128x128/65238.png inflating: /content/faces/thumbnails128x128/65239.png inflating: /content/faces/thumbnails128x128/65240.png inflating: /content/faces/thumbnails128x128/65241.png inflating: /content/faces/thumbnails128x128/65242.png inflating: /content/faces/thumbnails128x128/65243.png inflating: /content/faces/thumbnails128x128/65244.png inflating: /content/faces/thumbnails128x128/65245.png inflating: /content/faces/thumbnails128x128/65246.png inflating: /content/faces/thumbnails128x128/65247.png inflating: /content/faces/thumbnails128x128/65248.png inflating: /content/faces/thumbnails128x128/65249.png inflating: /content/faces/thumbnails128x128/65250.png inflating: /content/faces/thumbnails128x128/65251.png inflating: /content/faces/thumbnails128x128/65252.png inflating: /content/faces/thumbnails128x128/65253.png inflating: /content/faces/thumbnails128x128/65254.png inflating: /content/faces/thumbnails128x128/65255.png inflating: /content/faces/thumbnails128x128/65256.png inflating: /content/faces/thumbnails128x128/65257.png inflating: /content/faces/thumbnails128x128/65258.png inflating: /content/faces/thumbnails128x128/65259.png inflating: /content/faces/thumbnails128x128/65260.png inflating: /content/faces/thumbnails128x128/65261.png inflating: /content/faces/thumbnails128x128/65262.png inflating: /content/faces/thumbnails128x128/65263.png inflating: /content/faces/thumbnails128x128/65264.png inflating: /content/faces/thumbnails128x128/65265.png inflating: /content/faces/thumbnails128x128/65266.png inflating: /content/faces/thumbnails128x128/65267.png inflating: /content/faces/thumbnails128x128/65268.png inflating: /content/faces/thumbnails128x128/65269.png inflating: /content/faces/thumbnails128x128/65270.png inflating: /content/faces/thumbnails128x128/65271.png inflating: /content/faces/thumbnails128x128/65272.png inflating: /content/faces/thumbnails128x128/65273.png inflating: /content/faces/thumbnails128x128/65274.png inflating: /content/faces/thumbnails128x128/65275.png inflating: /content/faces/thumbnails128x128/65276.png inflating: /content/faces/thumbnails128x128/65277.png inflating: /content/faces/thumbnails128x128/65278.png inflating: /content/faces/thumbnails128x128/65279.png inflating: /content/faces/thumbnails128x128/65280.png inflating: /content/faces/thumbnails128x128/65281.png inflating: /content/faces/thumbnails128x128/65282.png inflating: /content/faces/thumbnails128x128/65283.png inflating: /content/faces/thumbnails128x128/65284.png inflating: /content/faces/thumbnails128x128/65285.png inflating: /content/faces/thumbnails128x128/65286.png inflating: /content/faces/thumbnails128x128/65287.png inflating: /content/faces/thumbnails128x128/65288.png inflating: /content/faces/thumbnails128x128/65289.png inflating: /content/faces/thumbnails128x128/65290.png inflating: /content/faces/thumbnails128x128/65291.png inflating: /content/faces/thumbnails128x128/65292.png inflating: /content/faces/thumbnails128x128/65293.png inflating: /content/faces/thumbnails128x128/65294.png inflating: /content/faces/thumbnails128x128/65295.png inflating: /content/faces/thumbnails128x128/65296.png inflating: /content/faces/thumbnails128x128/65297.png inflating: /content/faces/thumbnails128x128/65298.png inflating: /content/faces/thumbnails128x128/65299.png inflating: /content/faces/thumbnails128x128/65300.png inflating: /content/faces/thumbnails128x128/65301.png inflating: /content/faces/thumbnails128x128/65302.png inflating: /content/faces/thumbnails128x128/65303.png inflating: /content/faces/thumbnails128x128/65304.png inflating: /content/faces/thumbnails128x128/65305.png inflating: /content/faces/thumbnails128x128/65306.png inflating: /content/faces/thumbnails128x128/65307.png inflating: /content/faces/thumbnails128x128/65308.png inflating: /content/faces/thumbnails128x128/65309.png inflating: /content/faces/thumbnails128x128/65310.png inflating: /content/faces/thumbnails128x128/65311.png inflating: /content/faces/thumbnails128x128/65312.png inflating: /content/faces/thumbnails128x128/65313.png inflating: /content/faces/thumbnails128x128/65314.png inflating: /content/faces/thumbnails128x128/65315.png inflating: /content/faces/thumbnails128x128/65316.png inflating: /content/faces/thumbnails128x128/65317.png inflating: /content/faces/thumbnails128x128/65318.png inflating: /content/faces/thumbnails128x128/65319.png inflating: /content/faces/thumbnails128x128/65320.png inflating: /content/faces/thumbnails128x128/65321.png inflating: /content/faces/thumbnails128x128/65322.png inflating: /content/faces/thumbnails128x128/65323.png inflating: /content/faces/thumbnails128x128/65324.png inflating: /content/faces/thumbnails128x128/65325.png inflating: /content/faces/thumbnails128x128/65326.png inflating: /content/faces/thumbnails128x128/65327.png inflating: /content/faces/thumbnails128x128/65328.png inflating: /content/faces/thumbnails128x128/65329.png inflating: /content/faces/thumbnails128x128/65330.png inflating: /content/faces/thumbnails128x128/65331.png inflating: /content/faces/thumbnails128x128/65332.png inflating: /content/faces/thumbnails128x128/65333.png inflating: /content/faces/thumbnails128x128/65334.png inflating: /content/faces/thumbnails128x128/65335.png inflating: /content/faces/thumbnails128x128/65336.png inflating: /content/faces/thumbnails128x128/65337.png inflating: /content/faces/thumbnails128x128/65338.png inflating: /content/faces/thumbnails128x128/65339.png inflating: /content/faces/thumbnails128x128/65340.png inflating: /content/faces/thumbnails128x128/65341.png inflating: /content/faces/thumbnails128x128/65342.png inflating: /content/faces/thumbnails128x128/65343.png inflating: /content/faces/thumbnails128x128/65344.png inflating: /content/faces/thumbnails128x128/65345.png inflating: /content/faces/thumbnails128x128/65346.png inflating: /content/faces/thumbnails128x128/65347.png inflating: /content/faces/thumbnails128x128/65348.png inflating: /content/faces/thumbnails128x128/65349.png inflating: /content/faces/thumbnails128x128/65350.png inflating: /content/faces/thumbnails128x128/65351.png inflating: /content/faces/thumbnails128x128/65352.png inflating: /content/faces/thumbnails128x128/65353.png inflating: /content/faces/thumbnails128x128/65354.png inflating: /content/faces/thumbnails128x128/65355.png inflating: /content/faces/thumbnails128x128/65356.png inflating: /content/faces/thumbnails128x128/65357.png inflating: /content/faces/thumbnails128x128/65358.png inflating: /content/faces/thumbnails128x128/65359.png inflating: /content/faces/thumbnails128x128/65360.png inflating: /content/faces/thumbnails128x128/65361.png inflating: /content/faces/thumbnails128x128/65362.png inflating: /content/faces/thumbnails128x128/65363.png inflating: /content/faces/thumbnails128x128/65364.png inflating: /content/faces/thumbnails128x128/65365.png inflating: /content/faces/thumbnails128x128/65366.png inflating: /content/faces/thumbnails128x128/65367.png inflating: /content/faces/thumbnails128x128/65368.png inflating: /content/faces/thumbnails128x128/65369.png inflating: /content/faces/thumbnails128x128/65370.png inflating: /content/faces/thumbnails128x128/65371.png inflating: /content/faces/thumbnails128x128/65372.png inflating: /content/faces/thumbnails128x128/65373.png inflating: /content/faces/thumbnails128x128/65374.png inflating: /content/faces/thumbnails128x128/65375.png inflating: /content/faces/thumbnails128x128/65376.png inflating: /content/faces/thumbnails128x128/65377.png inflating: /content/faces/thumbnails128x128/65378.png inflating: /content/faces/thumbnails128x128/65379.png inflating: /content/faces/thumbnails128x128/65380.png inflating: /content/faces/thumbnails128x128/65381.png inflating: /content/faces/thumbnails128x128/65382.png inflating: /content/faces/thumbnails128x128/65383.png inflating: /content/faces/thumbnails128x128/65384.png inflating: /content/faces/thumbnails128x128/65385.png inflating: /content/faces/thumbnails128x128/65386.png inflating: /content/faces/thumbnails128x128/65387.png inflating: /content/faces/thumbnails128x128/65388.png inflating: /content/faces/thumbnails128x128/65389.png inflating: /content/faces/thumbnails128x128/65390.png inflating: /content/faces/thumbnails128x128/65391.png inflating: /content/faces/thumbnails128x128/65392.png inflating: /content/faces/thumbnails128x128/65393.png inflating: /content/faces/thumbnails128x128/65394.png inflating: /content/faces/thumbnails128x128/65395.png inflating: /content/faces/thumbnails128x128/65396.png inflating: /content/faces/thumbnails128x128/65397.png inflating: /content/faces/thumbnails128x128/65398.png inflating: /content/faces/thumbnails128x128/65399.png inflating: /content/faces/thumbnails128x128/65400.png inflating: /content/faces/thumbnails128x128/65401.png inflating: /content/faces/thumbnails128x128/65402.png inflating: /content/faces/thumbnails128x128/65403.png inflating: /content/faces/thumbnails128x128/65404.png inflating: /content/faces/thumbnails128x128/65405.png inflating: /content/faces/thumbnails128x128/65406.png inflating: /content/faces/thumbnails128x128/65407.png inflating: /content/faces/thumbnails128x128/65408.png inflating: /content/faces/thumbnails128x128/65409.png inflating: /content/faces/thumbnails128x128/65410.png inflating: /content/faces/thumbnails128x128/65411.png inflating: /content/faces/thumbnails128x128/65412.png inflating: /content/faces/thumbnails128x128/65413.png inflating: /content/faces/thumbnails128x128/65414.png inflating: /content/faces/thumbnails128x128/65415.png inflating: /content/faces/thumbnails128x128/65416.png inflating: /content/faces/thumbnails128x128/65417.png inflating: /content/faces/thumbnails128x128/65418.png inflating: /content/faces/thumbnails128x128/65419.png inflating: /content/faces/thumbnails128x128/65420.png inflating: /content/faces/thumbnails128x128/65421.png inflating: /content/faces/thumbnails128x128/65422.png inflating: /content/faces/thumbnails128x128/65423.png inflating: /content/faces/thumbnails128x128/65424.png inflating: /content/faces/thumbnails128x128/65425.png inflating: /content/faces/thumbnails128x128/65426.png inflating: /content/faces/thumbnails128x128/65427.png inflating: /content/faces/thumbnails128x128/65428.png inflating: /content/faces/thumbnails128x128/65429.png inflating: /content/faces/thumbnails128x128/65430.png inflating: /content/faces/thumbnails128x128/65431.png inflating: /content/faces/thumbnails128x128/65432.png inflating: /content/faces/thumbnails128x128/65433.png inflating: /content/faces/thumbnails128x128/65434.png inflating: /content/faces/thumbnails128x128/65435.png inflating: /content/faces/thumbnails128x128/65436.png inflating: /content/faces/thumbnails128x128/65437.png inflating: /content/faces/thumbnails128x128/65438.png inflating: /content/faces/thumbnails128x128/65439.png inflating: /content/faces/thumbnails128x128/65440.png inflating: /content/faces/thumbnails128x128/65441.png inflating: /content/faces/thumbnails128x128/65442.png inflating: /content/faces/thumbnails128x128/65443.png inflating: /content/faces/thumbnails128x128/65444.png inflating: /content/faces/thumbnails128x128/65445.png inflating: /content/faces/thumbnails128x128/65446.png inflating: /content/faces/thumbnails128x128/65447.png inflating: /content/faces/thumbnails128x128/65448.png inflating: /content/faces/thumbnails128x128/65449.png inflating: /content/faces/thumbnails128x128/65450.png inflating: /content/faces/thumbnails128x128/65451.png inflating: /content/faces/thumbnails128x128/65452.png inflating: /content/faces/thumbnails128x128/65453.png inflating: /content/faces/thumbnails128x128/65454.png inflating: /content/faces/thumbnails128x128/65455.png inflating: /content/faces/thumbnails128x128/65456.png inflating: /content/faces/thumbnails128x128/65457.png inflating: /content/faces/thumbnails128x128/65458.png inflating: /content/faces/thumbnails128x128/65459.png inflating: /content/faces/thumbnails128x128/65460.png inflating: /content/faces/thumbnails128x128/65461.png inflating: /content/faces/thumbnails128x128/65462.png inflating: /content/faces/thumbnails128x128/65463.png inflating: /content/faces/thumbnails128x128/65464.png inflating: /content/faces/thumbnails128x128/65465.png inflating: /content/faces/thumbnails128x128/65466.png inflating: /content/faces/thumbnails128x128/65467.png inflating: /content/faces/thumbnails128x128/65468.png inflating: /content/faces/thumbnails128x128/65469.png inflating: /content/faces/thumbnails128x128/65470.png inflating: /content/faces/thumbnails128x128/65471.png inflating: /content/faces/thumbnails128x128/65472.png inflating: /content/faces/thumbnails128x128/65473.png inflating: /content/faces/thumbnails128x128/65474.png inflating: /content/faces/thumbnails128x128/65475.png inflating: /content/faces/thumbnails128x128/65476.png inflating: /content/faces/thumbnails128x128/65477.png inflating: /content/faces/thumbnails128x128/65478.png inflating: /content/faces/thumbnails128x128/65479.png inflating: /content/faces/thumbnails128x128/65480.png inflating: /content/faces/thumbnails128x128/65481.png inflating: /content/faces/thumbnails128x128/65482.png inflating: /content/faces/thumbnails128x128/65483.png inflating: /content/faces/thumbnails128x128/65484.png inflating: /content/faces/thumbnails128x128/65485.png inflating: /content/faces/thumbnails128x128/65486.png inflating: /content/faces/thumbnails128x128/65487.png inflating: /content/faces/thumbnails128x128/65488.png inflating: /content/faces/thumbnails128x128/65489.png inflating: /content/faces/thumbnails128x128/65490.png inflating: /content/faces/thumbnails128x128/65491.png inflating: /content/faces/thumbnails128x128/65492.png inflating: /content/faces/thumbnails128x128/65493.png inflating: /content/faces/thumbnails128x128/65494.png inflating: /content/faces/thumbnails128x128/65495.png inflating: /content/faces/thumbnails128x128/65496.png inflating: /content/faces/thumbnails128x128/65497.png inflating: /content/faces/thumbnails128x128/65498.png inflating: /content/faces/thumbnails128x128/65499.png inflating: /content/faces/thumbnails128x128/65500.png inflating: /content/faces/thumbnails128x128/65501.png inflating: /content/faces/thumbnails128x128/65502.png inflating: /content/faces/thumbnails128x128/65503.png inflating: /content/faces/thumbnails128x128/65504.png inflating: /content/faces/thumbnails128x128/65505.png inflating: /content/faces/thumbnails128x128/65506.png inflating: /content/faces/thumbnails128x128/65507.png inflating: /content/faces/thumbnails128x128/65508.png inflating: /content/faces/thumbnails128x128/65509.png inflating: /content/faces/thumbnails128x128/65510.png inflating: /content/faces/thumbnails128x128/65511.png inflating: /content/faces/thumbnails128x128/65512.png inflating: /content/faces/thumbnails128x128/65513.png inflating: /content/faces/thumbnails128x128/65514.png inflating: /content/faces/thumbnails128x128/65515.png inflating: /content/faces/thumbnails128x128/65516.png inflating: /content/faces/thumbnails128x128/65517.png inflating: /content/faces/thumbnails128x128/65518.png inflating: /content/faces/thumbnails128x128/65519.png inflating: /content/faces/thumbnails128x128/65520.png inflating: /content/faces/thumbnails128x128/65521.png inflating: /content/faces/thumbnails128x128/65522.png inflating: /content/faces/thumbnails128x128/65523.png inflating: /content/faces/thumbnails128x128/65524.png inflating: /content/faces/thumbnails128x128/65525.png inflating: /content/faces/thumbnails128x128/65526.png inflating: /content/faces/thumbnails128x128/65527.png inflating: /content/faces/thumbnails128x128/65528.png inflating: /content/faces/thumbnails128x128/65529.png inflating: /content/faces/thumbnails128x128/65530.png inflating: /content/faces/thumbnails128x128/65531.png inflating: /content/faces/thumbnails128x128/65532.png inflating: /content/faces/thumbnails128x128/65533.png inflating: /content/faces/thumbnails128x128/65534.png inflating: /content/faces/thumbnails128x128/65535.png inflating: /content/faces/thumbnails128x128/65536.png inflating: /content/faces/thumbnails128x128/65537.png inflating: /content/faces/thumbnails128x128/65538.png inflating: /content/faces/thumbnails128x128/65539.png inflating: /content/faces/thumbnails128x128/65540.png inflating: /content/faces/thumbnails128x128/65541.png inflating: /content/faces/thumbnails128x128/65542.png inflating: /content/faces/thumbnails128x128/65543.png inflating: /content/faces/thumbnails128x128/65544.png inflating: /content/faces/thumbnails128x128/65545.png inflating: /content/faces/thumbnails128x128/65546.png inflating: /content/faces/thumbnails128x128/65547.png inflating: /content/faces/thumbnails128x128/65548.png inflating: /content/faces/thumbnails128x128/65549.png inflating: /content/faces/thumbnails128x128/65550.png inflating: /content/faces/thumbnails128x128/65551.png inflating: /content/faces/thumbnails128x128/65552.png inflating: /content/faces/thumbnails128x128/65553.png inflating: /content/faces/thumbnails128x128/65554.png inflating: /content/faces/thumbnails128x128/65555.png inflating: /content/faces/thumbnails128x128/65556.png inflating: /content/faces/thumbnails128x128/65557.png inflating: /content/faces/thumbnails128x128/65558.png inflating: /content/faces/thumbnails128x128/65559.png inflating: /content/faces/thumbnails128x128/65560.png inflating: /content/faces/thumbnails128x128/65561.png inflating: /content/faces/thumbnails128x128/65562.png inflating: /content/faces/thumbnails128x128/65563.png inflating: /content/faces/thumbnails128x128/65564.png inflating: /content/faces/thumbnails128x128/65565.png inflating: /content/faces/thumbnails128x128/65566.png inflating: /content/faces/thumbnails128x128/65567.png inflating: /content/faces/thumbnails128x128/65568.png inflating: /content/faces/thumbnails128x128/65569.png inflating: /content/faces/thumbnails128x128/65570.png inflating: /content/faces/thumbnails128x128/65571.png inflating: /content/faces/thumbnails128x128/65572.png inflating: /content/faces/thumbnails128x128/65573.png inflating: /content/faces/thumbnails128x128/65574.png inflating: /content/faces/thumbnails128x128/65575.png inflating: /content/faces/thumbnails128x128/65576.png inflating: /content/faces/thumbnails128x128/65577.png inflating: /content/faces/thumbnails128x128/65578.png inflating: /content/faces/thumbnails128x128/65579.png inflating: /content/faces/thumbnails128x128/65580.png inflating: /content/faces/thumbnails128x128/65581.png inflating: /content/faces/thumbnails128x128/65582.png inflating: /content/faces/thumbnails128x128/65583.png inflating: /content/faces/thumbnails128x128/65584.png inflating: /content/faces/thumbnails128x128/65585.png inflating: /content/faces/thumbnails128x128/65586.png inflating: /content/faces/thumbnails128x128/65587.png inflating: /content/faces/thumbnails128x128/65588.png inflating: /content/faces/thumbnails128x128/65589.png inflating: /content/faces/thumbnails128x128/65590.png inflating: /content/faces/thumbnails128x128/65591.png inflating: /content/faces/thumbnails128x128/65592.png inflating: /content/faces/thumbnails128x128/65593.png inflating: /content/faces/thumbnails128x128/65594.png inflating: /content/faces/thumbnails128x128/65595.png inflating: /content/faces/thumbnails128x128/65596.png inflating: /content/faces/thumbnails128x128/65597.png inflating: /content/faces/thumbnails128x128/65598.png inflating: /content/faces/thumbnails128x128/65599.png inflating: /content/faces/thumbnails128x128/65600.png inflating: /content/faces/thumbnails128x128/65601.png inflating: /content/faces/thumbnails128x128/65602.png inflating: /content/faces/thumbnails128x128/65603.png inflating: /content/faces/thumbnails128x128/65604.png inflating: /content/faces/thumbnails128x128/65605.png inflating: /content/faces/thumbnails128x128/65606.png inflating: /content/faces/thumbnails128x128/65607.png inflating: /content/faces/thumbnails128x128/65608.png inflating: /content/faces/thumbnails128x128/65609.png inflating: /content/faces/thumbnails128x128/65610.png inflating: /content/faces/thumbnails128x128/65611.png inflating: /content/faces/thumbnails128x128/65612.png inflating: /content/faces/thumbnails128x128/65613.png inflating: /content/faces/thumbnails128x128/65614.png inflating: /content/faces/thumbnails128x128/65615.png inflating: /content/faces/thumbnails128x128/65616.png inflating: /content/faces/thumbnails128x128/65617.png inflating: /content/faces/thumbnails128x128/65618.png inflating: /content/faces/thumbnails128x128/65619.png inflating: /content/faces/thumbnails128x128/65620.png inflating: /content/faces/thumbnails128x128/65621.png inflating: /content/faces/thumbnails128x128/65622.png inflating: /content/faces/thumbnails128x128/65623.png inflating: /content/faces/thumbnails128x128/65624.png inflating: /content/faces/thumbnails128x128/65625.png inflating: /content/faces/thumbnails128x128/65626.png inflating: /content/faces/thumbnails128x128/65627.png inflating: /content/faces/thumbnails128x128/65628.png inflating: /content/faces/thumbnails128x128/65629.png inflating: /content/faces/thumbnails128x128/65630.png inflating: /content/faces/thumbnails128x128/65631.png inflating: /content/faces/thumbnails128x128/65632.png inflating: /content/faces/thumbnails128x128/65633.png inflating: /content/faces/thumbnails128x128/65634.png inflating: /content/faces/thumbnails128x128/65635.png inflating: /content/faces/thumbnails128x128/65636.png inflating: /content/faces/thumbnails128x128/65637.png inflating: /content/faces/thumbnails128x128/65638.png inflating: /content/faces/thumbnails128x128/65639.png inflating: /content/faces/thumbnails128x128/65640.png inflating: /content/faces/thumbnails128x128/65641.png inflating: /content/faces/thumbnails128x128/65642.png inflating: /content/faces/thumbnails128x128/65643.png inflating: /content/faces/thumbnails128x128/65644.png inflating: /content/faces/thumbnails128x128/65645.png inflating: /content/faces/thumbnails128x128/65646.png inflating: /content/faces/thumbnails128x128/65647.png inflating: /content/faces/thumbnails128x128/65648.png inflating: /content/faces/thumbnails128x128/65649.png inflating: /content/faces/thumbnails128x128/65650.png inflating: /content/faces/thumbnails128x128/65651.png inflating: /content/faces/thumbnails128x128/65652.png inflating: /content/faces/thumbnails128x128/65653.png inflating: /content/faces/thumbnails128x128/65654.png inflating: /content/faces/thumbnails128x128/65655.png inflating: /content/faces/thumbnails128x128/65656.png inflating: /content/faces/thumbnails128x128/65657.png inflating: /content/faces/thumbnails128x128/65658.png inflating: /content/faces/thumbnails128x128/65659.png inflating: /content/faces/thumbnails128x128/65660.png inflating: /content/faces/thumbnails128x128/65661.png inflating: /content/faces/thumbnails128x128/65662.png inflating: /content/faces/thumbnails128x128/65663.png inflating: /content/faces/thumbnails128x128/65664.png inflating: /content/faces/thumbnails128x128/65665.png inflating: /content/faces/thumbnails128x128/65666.png inflating: /content/faces/thumbnails128x128/65667.png inflating: /content/faces/thumbnails128x128/65668.png inflating: /content/faces/thumbnails128x128/65669.png inflating: /content/faces/thumbnails128x128/65670.png inflating: /content/faces/thumbnails128x128/65671.png inflating: /content/faces/thumbnails128x128/65672.png inflating: /content/faces/thumbnails128x128/65673.png inflating: /content/faces/thumbnails128x128/65674.png inflating: /content/faces/thumbnails128x128/65675.png inflating: /content/faces/thumbnails128x128/65676.png inflating: /content/faces/thumbnails128x128/65677.png inflating: /content/faces/thumbnails128x128/65678.png inflating: /content/faces/thumbnails128x128/65679.png inflating: /content/faces/thumbnails128x128/65680.png inflating: /content/faces/thumbnails128x128/65681.png inflating: /content/faces/thumbnails128x128/65682.png inflating: /content/faces/thumbnails128x128/65683.png inflating: /content/faces/thumbnails128x128/65684.png inflating: /content/faces/thumbnails128x128/65685.png inflating: /content/faces/thumbnails128x128/65686.png inflating: /content/faces/thumbnails128x128/65687.png inflating: /content/faces/thumbnails128x128/65688.png inflating: /content/faces/thumbnails128x128/65689.png inflating: /content/faces/thumbnails128x128/65690.png inflating: /content/faces/thumbnails128x128/65691.png inflating: /content/faces/thumbnails128x128/65692.png inflating: /content/faces/thumbnails128x128/65693.png inflating: /content/faces/thumbnails128x128/65694.png inflating: /content/faces/thumbnails128x128/65695.png inflating: /content/faces/thumbnails128x128/65696.png inflating: /content/faces/thumbnails128x128/65697.png inflating: /content/faces/thumbnails128x128/65698.png inflating: /content/faces/thumbnails128x128/65699.png inflating: /content/faces/thumbnails128x128/65700.png inflating: /content/faces/thumbnails128x128/65701.png inflating: /content/faces/thumbnails128x128/65702.png inflating: /content/faces/thumbnails128x128/65703.png inflating: /content/faces/thumbnails128x128/65704.png inflating: /content/faces/thumbnails128x128/65705.png inflating: /content/faces/thumbnails128x128/65706.png inflating: /content/faces/thumbnails128x128/65707.png inflating: /content/faces/thumbnails128x128/65708.png inflating: /content/faces/thumbnails128x128/65709.png inflating: /content/faces/thumbnails128x128/65710.png inflating: /content/faces/thumbnails128x128/65711.png inflating: /content/faces/thumbnails128x128/65712.png inflating: /content/faces/thumbnails128x128/65713.png inflating: /content/faces/thumbnails128x128/65714.png inflating: /content/faces/thumbnails128x128/65715.png inflating: /content/faces/thumbnails128x128/65716.png inflating: /content/faces/thumbnails128x128/65717.png inflating: /content/faces/thumbnails128x128/65718.png inflating: /content/faces/thumbnails128x128/65719.png inflating: /content/faces/thumbnails128x128/65720.png inflating: /content/faces/thumbnails128x128/65721.png inflating: /content/faces/thumbnails128x128/65722.png inflating: /content/faces/thumbnails128x128/65723.png inflating: /content/faces/thumbnails128x128/65724.png inflating: /content/faces/thumbnails128x128/65725.png inflating: /content/faces/thumbnails128x128/65726.png inflating: /content/faces/thumbnails128x128/65727.png inflating: /content/faces/thumbnails128x128/65728.png inflating: /content/faces/thumbnails128x128/65729.png inflating: /content/faces/thumbnails128x128/65730.png inflating: /content/faces/thumbnails128x128/65731.png inflating: /content/faces/thumbnails128x128/65732.png inflating: /content/faces/thumbnails128x128/65733.png inflating: /content/faces/thumbnails128x128/65734.png inflating: /content/faces/thumbnails128x128/65735.png inflating: /content/faces/thumbnails128x128/65736.png inflating: /content/faces/thumbnails128x128/65737.png inflating: /content/faces/thumbnails128x128/65738.png inflating: /content/faces/thumbnails128x128/65739.png inflating: /content/faces/thumbnails128x128/65740.png inflating: /content/faces/thumbnails128x128/65741.png inflating: /content/faces/thumbnails128x128/65742.png inflating: /content/faces/thumbnails128x128/65743.png inflating: /content/faces/thumbnails128x128/65744.png inflating: /content/faces/thumbnails128x128/65745.png inflating: /content/faces/thumbnails128x128/65746.png inflating: /content/faces/thumbnails128x128/65747.png inflating: /content/faces/thumbnails128x128/65748.png inflating: /content/faces/thumbnails128x128/65749.png inflating: /content/faces/thumbnails128x128/65750.png inflating: /content/faces/thumbnails128x128/65751.png inflating: /content/faces/thumbnails128x128/65752.png inflating: /content/faces/thumbnails128x128/65753.png inflating: /content/faces/thumbnails128x128/65754.png inflating: /content/faces/thumbnails128x128/65755.png inflating: /content/faces/thumbnails128x128/65756.png inflating: /content/faces/thumbnails128x128/65757.png inflating: /content/faces/thumbnails128x128/65758.png inflating: /content/faces/thumbnails128x128/65759.png inflating: /content/faces/thumbnails128x128/65760.png inflating: /content/faces/thumbnails128x128/65761.png inflating: /content/faces/thumbnails128x128/65762.png inflating: /content/faces/thumbnails128x128/65763.png inflating: /content/faces/thumbnails128x128/65764.png inflating: /content/faces/thumbnails128x128/65765.png inflating: /content/faces/thumbnails128x128/65766.png inflating: /content/faces/thumbnails128x128/65767.png inflating: /content/faces/thumbnails128x128/65768.png inflating: /content/faces/thumbnails128x128/65769.png inflating: /content/faces/thumbnails128x128/65770.png inflating: /content/faces/thumbnails128x128/65771.png inflating: /content/faces/thumbnails128x128/65772.png inflating: /content/faces/thumbnails128x128/65773.png inflating: /content/faces/thumbnails128x128/65774.png inflating: /content/faces/thumbnails128x128/65775.png inflating: /content/faces/thumbnails128x128/65776.png inflating: /content/faces/thumbnails128x128/65777.png inflating: /content/faces/thumbnails128x128/65778.png inflating: /content/faces/thumbnails128x128/65779.png inflating: /content/faces/thumbnails128x128/65780.png inflating: /content/faces/thumbnails128x128/65781.png inflating: /content/faces/thumbnails128x128/65782.png inflating: /content/faces/thumbnails128x128/65783.png inflating: /content/faces/thumbnails128x128/65784.png inflating: /content/faces/thumbnails128x128/65785.png inflating: /content/faces/thumbnails128x128/65786.png inflating: /content/faces/thumbnails128x128/65787.png inflating: /content/faces/thumbnails128x128/65788.png inflating: /content/faces/thumbnails128x128/65789.png inflating: /content/faces/thumbnails128x128/65790.png inflating: /content/faces/thumbnails128x128/65791.png inflating: /content/faces/thumbnails128x128/65792.png inflating: /content/faces/thumbnails128x128/65793.png inflating: /content/faces/thumbnails128x128/65794.png inflating: /content/faces/thumbnails128x128/65795.png inflating: /content/faces/thumbnails128x128/65796.png inflating: /content/faces/thumbnails128x128/65797.png inflating: /content/faces/thumbnails128x128/65798.png inflating: /content/faces/thumbnails128x128/65799.png inflating: /content/faces/thumbnails128x128/65800.png inflating: /content/faces/thumbnails128x128/65801.png inflating: /content/faces/thumbnails128x128/65802.png inflating: /content/faces/thumbnails128x128/65803.png inflating: /content/faces/thumbnails128x128/65804.png inflating: /content/faces/thumbnails128x128/65805.png inflating: /content/faces/thumbnails128x128/65806.png inflating: /content/faces/thumbnails128x128/65807.png inflating: /content/faces/thumbnails128x128/65808.png inflating: /content/faces/thumbnails128x128/65809.png inflating: /content/faces/thumbnails128x128/65810.png inflating: /content/faces/thumbnails128x128/65811.png inflating: /content/faces/thumbnails128x128/65812.png inflating: /content/faces/thumbnails128x128/65813.png inflating: /content/faces/thumbnails128x128/65814.png inflating: /content/faces/thumbnails128x128/65815.png inflating: /content/faces/thumbnails128x128/65816.png inflating: /content/faces/thumbnails128x128/65817.png inflating: /content/faces/thumbnails128x128/65818.png inflating: /content/faces/thumbnails128x128/65819.png inflating: /content/faces/thumbnails128x128/65820.png inflating: /content/faces/thumbnails128x128/65821.png inflating: /content/faces/thumbnails128x128/65822.png inflating: /content/faces/thumbnails128x128/65823.png inflating: /content/faces/thumbnails128x128/65824.png inflating: /content/faces/thumbnails128x128/65825.png inflating: /content/faces/thumbnails128x128/65826.png inflating: /content/faces/thumbnails128x128/65827.png inflating: /content/faces/thumbnails128x128/65828.png inflating: /content/faces/thumbnails128x128/65829.png inflating: /content/faces/thumbnails128x128/65830.png inflating: /content/faces/thumbnails128x128/65831.png inflating: /content/faces/thumbnails128x128/65832.png inflating: /content/faces/thumbnails128x128/65833.png inflating: /content/faces/thumbnails128x128/65834.png inflating: /content/faces/thumbnails128x128/65835.png inflating: /content/faces/thumbnails128x128/65836.png inflating: /content/faces/thumbnails128x128/65837.png inflating: /content/faces/thumbnails128x128/65838.png inflating: /content/faces/thumbnails128x128/65839.png inflating: /content/faces/thumbnails128x128/65840.png inflating: /content/faces/thumbnails128x128/65841.png inflating: /content/faces/thumbnails128x128/65842.png inflating: /content/faces/thumbnails128x128/65843.png inflating: /content/faces/thumbnails128x128/65844.png inflating: /content/faces/thumbnails128x128/65845.png inflating: /content/faces/thumbnails128x128/65846.png inflating: /content/faces/thumbnails128x128/65847.png inflating: /content/faces/thumbnails128x128/65848.png inflating: /content/faces/thumbnails128x128/65849.png inflating: /content/faces/thumbnails128x128/65850.png inflating: /content/faces/thumbnails128x128/65851.png inflating: /content/faces/thumbnails128x128/65852.png inflating: /content/faces/thumbnails128x128/65853.png inflating: /content/faces/thumbnails128x128/65854.png inflating: /content/faces/thumbnails128x128/65855.png inflating: /content/faces/thumbnails128x128/65856.png inflating: /content/faces/thumbnails128x128/65857.png inflating: /content/faces/thumbnails128x128/65858.png inflating: /content/faces/thumbnails128x128/65859.png inflating: /content/faces/thumbnails128x128/65860.png inflating: /content/faces/thumbnails128x128/65861.png inflating: /content/faces/thumbnails128x128/65862.png inflating: /content/faces/thumbnails128x128/65863.png inflating: /content/faces/thumbnails128x128/65864.png inflating: /content/faces/thumbnails128x128/65865.png inflating: /content/faces/thumbnails128x128/65866.png inflating: /content/faces/thumbnails128x128/65867.png inflating: /content/faces/thumbnails128x128/65868.png inflating: /content/faces/thumbnails128x128/65869.png inflating: /content/faces/thumbnails128x128/65870.png inflating: /content/faces/thumbnails128x128/65871.png inflating: /content/faces/thumbnails128x128/65872.png inflating: /content/faces/thumbnails128x128/65873.png inflating: /content/faces/thumbnails128x128/65874.png inflating: /content/faces/thumbnails128x128/65875.png inflating: /content/faces/thumbnails128x128/65876.png inflating: /content/faces/thumbnails128x128/65877.png inflating: /content/faces/thumbnails128x128/65878.png inflating: /content/faces/thumbnails128x128/65879.png inflating: /content/faces/thumbnails128x128/65880.png inflating: /content/faces/thumbnails128x128/65881.png inflating: /content/faces/thumbnails128x128/65882.png inflating: /content/faces/thumbnails128x128/65883.png inflating: /content/faces/thumbnails128x128/65884.png inflating: /content/faces/thumbnails128x128/65885.png inflating: /content/faces/thumbnails128x128/65886.png inflating: /content/faces/thumbnails128x128/65887.png inflating: /content/faces/thumbnails128x128/65888.png inflating: /content/faces/thumbnails128x128/65889.png inflating: /content/faces/thumbnails128x128/65890.png inflating: /content/faces/thumbnails128x128/65891.png inflating: /content/faces/thumbnails128x128/65892.png inflating: /content/faces/thumbnails128x128/65893.png inflating: /content/faces/thumbnails128x128/65894.png inflating: /content/faces/thumbnails128x128/65895.png inflating: /content/faces/thumbnails128x128/65896.png inflating: /content/faces/thumbnails128x128/65897.png inflating: /content/faces/thumbnails128x128/65898.png inflating: /content/faces/thumbnails128x128/65899.png inflating: /content/faces/thumbnails128x128/65900.png inflating: /content/faces/thumbnails128x128/65901.png inflating: /content/faces/thumbnails128x128/65902.png inflating: /content/faces/thumbnails128x128/65903.png inflating: /content/faces/thumbnails128x128/65904.png inflating: /content/faces/thumbnails128x128/65905.png inflating: /content/faces/thumbnails128x128/65906.png inflating: /content/faces/thumbnails128x128/65907.png inflating: /content/faces/thumbnails128x128/65908.png inflating: /content/faces/thumbnails128x128/65909.png inflating: /content/faces/thumbnails128x128/65910.png inflating: /content/faces/thumbnails128x128/65911.png inflating: /content/faces/thumbnails128x128/65912.png inflating: /content/faces/thumbnails128x128/65913.png inflating: /content/faces/thumbnails128x128/65914.png inflating: /content/faces/thumbnails128x128/65915.png inflating: /content/faces/thumbnails128x128/65916.png inflating: /content/faces/thumbnails128x128/65917.png inflating: /content/faces/thumbnails128x128/65918.png inflating: /content/faces/thumbnails128x128/65919.png inflating: /content/faces/thumbnails128x128/65920.png inflating: /content/faces/thumbnails128x128/65921.png inflating: /content/faces/thumbnails128x128/65922.png inflating: /content/faces/thumbnails128x128/65923.png inflating: /content/faces/thumbnails128x128/65924.png inflating: /content/faces/thumbnails128x128/65925.png inflating: /content/faces/thumbnails128x128/65926.png inflating: /content/faces/thumbnails128x128/65927.png inflating: /content/faces/thumbnails128x128/65928.png inflating: /content/faces/thumbnails128x128/65929.png inflating: /content/faces/thumbnails128x128/65930.png inflating: /content/faces/thumbnails128x128/65931.png inflating: /content/faces/thumbnails128x128/65932.png inflating: /content/faces/thumbnails128x128/65933.png inflating: /content/faces/thumbnails128x128/65934.png inflating: /content/faces/thumbnails128x128/65935.png inflating: /content/faces/thumbnails128x128/65936.png inflating: /content/faces/thumbnails128x128/65937.png inflating: /content/faces/thumbnails128x128/65938.png inflating: /content/faces/thumbnails128x128/65939.png inflating: /content/faces/thumbnails128x128/65940.png inflating: /content/faces/thumbnails128x128/65941.png inflating: /content/faces/thumbnails128x128/65942.png inflating: /content/faces/thumbnails128x128/65943.png inflating: /content/faces/thumbnails128x128/65944.png inflating: /content/faces/thumbnails128x128/65945.png inflating: /content/faces/thumbnails128x128/65946.png inflating: /content/faces/thumbnails128x128/65947.png inflating: /content/faces/thumbnails128x128/65948.png inflating: /content/faces/thumbnails128x128/65949.png inflating: /content/faces/thumbnails128x128/65950.png inflating: /content/faces/thumbnails128x128/65951.png inflating: /content/faces/thumbnails128x128/65952.png inflating: /content/faces/thumbnails128x128/65953.png inflating: /content/faces/thumbnails128x128/65954.png inflating: /content/faces/thumbnails128x128/65955.png inflating: /content/faces/thumbnails128x128/65956.png inflating: /content/faces/thumbnails128x128/65957.png inflating: /content/faces/thumbnails128x128/65958.png inflating: /content/faces/thumbnails128x128/65959.png inflating: /content/faces/thumbnails128x128/65960.png inflating: /content/faces/thumbnails128x128/65961.png inflating: /content/faces/thumbnails128x128/65962.png inflating: /content/faces/thumbnails128x128/65963.png inflating: /content/faces/thumbnails128x128/65964.png inflating: /content/faces/thumbnails128x128/65965.png inflating: /content/faces/thumbnails128x128/65966.png inflating: /content/faces/thumbnails128x128/65967.png inflating: /content/faces/thumbnails128x128/65968.png inflating: /content/faces/thumbnails128x128/65969.png inflating: /content/faces/thumbnails128x128/65970.png inflating: /content/faces/thumbnails128x128/65971.png inflating: /content/faces/thumbnails128x128/65972.png inflating: /content/faces/thumbnails128x128/65973.png inflating: /content/faces/thumbnails128x128/65974.png inflating: /content/faces/thumbnails128x128/65975.png inflating: /content/faces/thumbnails128x128/65976.png inflating: /content/faces/thumbnails128x128/65977.png inflating: /content/faces/thumbnails128x128/65978.png inflating: /content/faces/thumbnails128x128/65979.png inflating: /content/faces/thumbnails128x128/65980.png inflating: /content/faces/thumbnails128x128/65981.png inflating: /content/faces/thumbnails128x128/65982.png inflating: /content/faces/thumbnails128x128/65983.png inflating: /content/faces/thumbnails128x128/65984.png inflating: /content/faces/thumbnails128x128/65985.png inflating: /content/faces/thumbnails128x128/65986.png inflating: /content/faces/thumbnails128x128/65987.png inflating: /content/faces/thumbnails128x128/65988.png inflating: /content/faces/thumbnails128x128/65989.png inflating: /content/faces/thumbnails128x128/65990.png inflating: /content/faces/thumbnails128x128/65991.png inflating: /content/faces/thumbnails128x128/65992.png inflating: /content/faces/thumbnails128x128/65993.png inflating: /content/faces/thumbnails128x128/65994.png inflating: /content/faces/thumbnails128x128/65995.png inflating: /content/faces/thumbnails128x128/65996.png inflating: /content/faces/thumbnails128x128/65997.png inflating: /content/faces/thumbnails128x128/65998.png inflating: /content/faces/thumbnails128x128/65999.png inflating: /content/faces/thumbnails128x128/66000.png inflating: /content/faces/thumbnails128x128/66001.png inflating: /content/faces/thumbnails128x128/66002.png inflating: /content/faces/thumbnails128x128/66003.png inflating: /content/faces/thumbnails128x128/66004.png inflating: /content/faces/thumbnails128x128/66005.png inflating: /content/faces/thumbnails128x128/66006.png inflating: /content/faces/thumbnails128x128/66007.png inflating: /content/faces/thumbnails128x128/66008.png inflating: /content/faces/thumbnails128x128/66009.png inflating: /content/faces/thumbnails128x128/66010.png inflating: /content/faces/thumbnails128x128/66011.png inflating: /content/faces/thumbnails128x128/66012.png inflating: /content/faces/thumbnails128x128/66013.png inflating: /content/faces/thumbnails128x128/66014.png inflating: /content/faces/thumbnails128x128/66015.png inflating: /content/faces/thumbnails128x128/66016.png inflating: /content/faces/thumbnails128x128/66017.png inflating: /content/faces/thumbnails128x128/66018.png inflating: /content/faces/thumbnails128x128/66019.png inflating: /content/faces/thumbnails128x128/66020.png inflating: /content/faces/thumbnails128x128/66021.png inflating: /content/faces/thumbnails128x128/66022.png inflating: /content/faces/thumbnails128x128/66023.png inflating: /content/faces/thumbnails128x128/66024.png inflating: /content/faces/thumbnails128x128/66025.png inflating: /content/faces/thumbnails128x128/66026.png inflating: /content/faces/thumbnails128x128/66027.png inflating: /content/faces/thumbnails128x128/66028.png inflating: /content/faces/thumbnails128x128/66029.png inflating: /content/faces/thumbnails128x128/66030.png inflating: /content/faces/thumbnails128x128/66031.png inflating: /content/faces/thumbnails128x128/66032.png inflating: /content/faces/thumbnails128x128/66033.png inflating: /content/faces/thumbnails128x128/66034.png inflating: /content/faces/thumbnails128x128/66035.png inflating: /content/faces/thumbnails128x128/66036.png inflating: /content/faces/thumbnails128x128/66037.png inflating: /content/faces/thumbnails128x128/66038.png inflating: /content/faces/thumbnails128x128/66039.png inflating: /content/faces/thumbnails128x128/66040.png inflating: /content/faces/thumbnails128x128/66041.png inflating: /content/faces/thumbnails128x128/66042.png inflating: /content/faces/thumbnails128x128/66043.png inflating: /content/faces/thumbnails128x128/66044.png inflating: /content/faces/thumbnails128x128/66045.png inflating: /content/faces/thumbnails128x128/66046.png inflating: /content/faces/thumbnails128x128/66047.png inflating: /content/faces/thumbnails128x128/66048.png inflating: /content/faces/thumbnails128x128/66049.png inflating: /content/faces/thumbnails128x128/66050.png inflating: /content/faces/thumbnails128x128/66051.png inflating: /content/faces/thumbnails128x128/66052.png inflating: /content/faces/thumbnails128x128/66053.png inflating: /content/faces/thumbnails128x128/66054.png inflating: /content/faces/thumbnails128x128/66055.png inflating: /content/faces/thumbnails128x128/66056.png inflating: /content/faces/thumbnails128x128/66057.png inflating: /content/faces/thumbnails128x128/66058.png inflating: /content/faces/thumbnails128x128/66059.png inflating: /content/faces/thumbnails128x128/66060.png inflating: /content/faces/thumbnails128x128/66061.png inflating: /content/faces/thumbnails128x128/66062.png inflating: /content/faces/thumbnails128x128/66063.png inflating: /content/faces/thumbnails128x128/66064.png inflating: /content/faces/thumbnails128x128/66065.png inflating: /content/faces/thumbnails128x128/66066.png inflating: /content/faces/thumbnails128x128/66067.png inflating: /content/faces/thumbnails128x128/66068.png inflating: /content/faces/thumbnails128x128/66069.png inflating: /content/faces/thumbnails128x128/66070.png inflating: /content/faces/thumbnails128x128/66071.png inflating: /content/faces/thumbnails128x128/66072.png inflating: /content/faces/thumbnails128x128/66073.png inflating: /content/faces/thumbnails128x128/66074.png inflating: /content/faces/thumbnails128x128/66075.png inflating: /content/faces/thumbnails128x128/66076.png inflating: /content/faces/thumbnails128x128/66077.png inflating: /content/faces/thumbnails128x128/66078.png inflating: /content/faces/thumbnails128x128/66079.png inflating: /content/faces/thumbnails128x128/66080.png inflating: /content/faces/thumbnails128x128/66081.png inflating: /content/faces/thumbnails128x128/66082.png inflating: /content/faces/thumbnails128x128/66083.png inflating: /content/faces/thumbnails128x128/66084.png inflating: /content/faces/thumbnails128x128/66085.png inflating: /content/faces/thumbnails128x128/66086.png inflating: /content/faces/thumbnails128x128/66087.png inflating: /content/faces/thumbnails128x128/66088.png inflating: /content/faces/thumbnails128x128/66089.png inflating: /content/faces/thumbnails128x128/66090.png inflating: /content/faces/thumbnails128x128/66091.png inflating: /content/faces/thumbnails128x128/66092.png inflating: /content/faces/thumbnails128x128/66093.png inflating: /content/faces/thumbnails128x128/66094.png inflating: /content/faces/thumbnails128x128/66095.png inflating: /content/faces/thumbnails128x128/66096.png inflating: /content/faces/thumbnails128x128/66097.png inflating: /content/faces/thumbnails128x128/66098.png inflating: /content/faces/thumbnails128x128/66099.png inflating: /content/faces/thumbnails128x128/66100.png inflating: /content/faces/thumbnails128x128/66101.png inflating: /content/faces/thumbnails128x128/66102.png inflating: /content/faces/thumbnails128x128/66103.png inflating: /content/faces/thumbnails128x128/66104.png inflating: /content/faces/thumbnails128x128/66105.png inflating: /content/faces/thumbnails128x128/66106.png inflating: /content/faces/thumbnails128x128/66107.png inflating: /content/faces/thumbnails128x128/66108.png inflating: /content/faces/thumbnails128x128/66109.png inflating: /content/faces/thumbnails128x128/66110.png inflating: /content/faces/thumbnails128x128/66111.png inflating: /content/faces/thumbnails128x128/66112.png inflating: /content/faces/thumbnails128x128/66113.png inflating: /content/faces/thumbnails128x128/66114.png inflating: /content/faces/thumbnails128x128/66115.png inflating: /content/faces/thumbnails128x128/66116.png inflating: /content/faces/thumbnails128x128/66117.png inflating: /content/faces/thumbnails128x128/66118.png inflating: /content/faces/thumbnails128x128/66119.png inflating: /content/faces/thumbnails128x128/66120.png inflating: /content/faces/thumbnails128x128/66121.png inflating: /content/faces/thumbnails128x128/66122.png inflating: /content/faces/thumbnails128x128/66123.png inflating: /content/faces/thumbnails128x128/66124.png inflating: /content/faces/thumbnails128x128/66125.png inflating: /content/faces/thumbnails128x128/66126.png inflating: /content/faces/thumbnails128x128/66127.png inflating: /content/faces/thumbnails128x128/66128.png inflating: /content/faces/thumbnails128x128/66129.png inflating: /content/faces/thumbnails128x128/66130.png inflating: /content/faces/thumbnails128x128/66131.png inflating: /content/faces/thumbnails128x128/66132.png inflating: /content/faces/thumbnails128x128/66133.png inflating: /content/faces/thumbnails128x128/66134.png inflating: /content/faces/thumbnails128x128/66135.png inflating: /content/faces/thumbnails128x128/66136.png inflating: /content/faces/thumbnails128x128/66137.png inflating: /content/faces/thumbnails128x128/66138.png inflating: /content/faces/thumbnails128x128/66139.png inflating: /content/faces/thumbnails128x128/66140.png inflating: /content/faces/thumbnails128x128/66141.png inflating: /content/faces/thumbnails128x128/66142.png inflating: /content/faces/thumbnails128x128/66143.png inflating: /content/faces/thumbnails128x128/66144.png inflating: /content/faces/thumbnails128x128/66145.png inflating: /content/faces/thumbnails128x128/66146.png inflating: /content/faces/thumbnails128x128/66147.png inflating: /content/faces/thumbnails128x128/66148.png inflating: /content/faces/thumbnails128x128/66149.png inflating: /content/faces/thumbnails128x128/66150.png inflating: /content/faces/thumbnails128x128/66151.png inflating: /content/faces/thumbnails128x128/66152.png inflating: /content/faces/thumbnails128x128/66153.png inflating: /content/faces/thumbnails128x128/66154.png inflating: /content/faces/thumbnails128x128/66155.png inflating: /content/faces/thumbnails128x128/66156.png inflating: /content/faces/thumbnails128x128/66157.png inflating: /content/faces/thumbnails128x128/66158.png inflating: /content/faces/thumbnails128x128/66159.png inflating: /content/faces/thumbnails128x128/66160.png inflating: /content/faces/thumbnails128x128/66161.png inflating: /content/faces/thumbnails128x128/66162.png inflating: /content/faces/thumbnails128x128/66163.png inflating: /content/faces/thumbnails128x128/66164.png inflating: /content/faces/thumbnails128x128/66165.png inflating: /content/faces/thumbnails128x128/66166.png inflating: /content/faces/thumbnails128x128/66167.png inflating: /content/faces/thumbnails128x128/66168.png inflating: /content/faces/thumbnails128x128/66169.png inflating: /content/faces/thumbnails128x128/66170.png inflating: /content/faces/thumbnails128x128/66171.png inflating: /content/faces/thumbnails128x128/66172.png inflating: /content/faces/thumbnails128x128/66173.png inflating: /content/faces/thumbnails128x128/66174.png inflating: /content/faces/thumbnails128x128/66175.png inflating: /content/faces/thumbnails128x128/66176.png inflating: /content/faces/thumbnails128x128/66177.png inflating: /content/faces/thumbnails128x128/66178.png inflating: /content/faces/thumbnails128x128/66179.png inflating: /content/faces/thumbnails128x128/66180.png inflating: /content/faces/thumbnails128x128/66181.png inflating: /content/faces/thumbnails128x128/66182.png inflating: /content/faces/thumbnails128x128/66183.png inflating: /content/faces/thumbnails128x128/66184.png inflating: /content/faces/thumbnails128x128/66185.png inflating: /content/faces/thumbnails128x128/66186.png inflating: /content/faces/thumbnails128x128/66187.png inflating: /content/faces/thumbnails128x128/66188.png inflating: /content/faces/thumbnails128x128/66189.png inflating: /content/faces/thumbnails128x128/66190.png inflating: /content/faces/thumbnails128x128/66191.png inflating: /content/faces/thumbnails128x128/66192.png inflating: /content/faces/thumbnails128x128/66193.png inflating: /content/faces/thumbnails128x128/66194.png inflating: /content/faces/thumbnails128x128/66195.png inflating: /content/faces/thumbnails128x128/66196.png inflating: /content/faces/thumbnails128x128/66197.png inflating: /content/faces/thumbnails128x128/66198.png inflating: /content/faces/thumbnails128x128/66199.png inflating: /content/faces/thumbnails128x128/66200.png inflating: /content/faces/thumbnails128x128/66201.png inflating: /content/faces/thumbnails128x128/66202.png inflating: /content/faces/thumbnails128x128/66203.png inflating: /content/faces/thumbnails128x128/66204.png inflating: /content/faces/thumbnails128x128/66205.png inflating: /content/faces/thumbnails128x128/66206.png inflating: /content/faces/thumbnails128x128/66207.png inflating: /content/faces/thumbnails128x128/66208.png inflating: /content/faces/thumbnails128x128/66209.png inflating: /content/faces/thumbnails128x128/66210.png inflating: /content/faces/thumbnails128x128/66211.png inflating: /content/faces/thumbnails128x128/66212.png inflating: /content/faces/thumbnails128x128/66213.png inflating: /content/faces/thumbnails128x128/66214.png inflating: /content/faces/thumbnails128x128/66215.png inflating: /content/faces/thumbnails128x128/66216.png inflating: /content/faces/thumbnails128x128/66217.png inflating: /content/faces/thumbnails128x128/66218.png inflating: /content/faces/thumbnails128x128/66219.png inflating: /content/faces/thumbnails128x128/66220.png inflating: /content/faces/thumbnails128x128/66221.png inflating: /content/faces/thumbnails128x128/66222.png inflating: /content/faces/thumbnails128x128/66223.png inflating: /content/faces/thumbnails128x128/66224.png inflating: /content/faces/thumbnails128x128/66225.png inflating: /content/faces/thumbnails128x128/66226.png inflating: /content/faces/thumbnails128x128/66227.png inflating: /content/faces/thumbnails128x128/66228.png inflating: /content/faces/thumbnails128x128/66229.png inflating: /content/faces/thumbnails128x128/66230.png inflating: /content/faces/thumbnails128x128/66231.png inflating: /content/faces/thumbnails128x128/66232.png inflating: /content/faces/thumbnails128x128/66233.png inflating: /content/faces/thumbnails128x128/66234.png inflating: /content/faces/thumbnails128x128/66235.png inflating: /content/faces/thumbnails128x128/66236.png inflating: /content/faces/thumbnails128x128/66237.png inflating: /content/faces/thumbnails128x128/66238.png inflating: /content/faces/thumbnails128x128/66239.png inflating: /content/faces/thumbnails128x128/66240.png inflating: /content/faces/thumbnails128x128/66241.png inflating: /content/faces/thumbnails128x128/66242.png inflating: /content/faces/thumbnails128x128/66243.png inflating: /content/faces/thumbnails128x128/66244.png inflating: /content/faces/thumbnails128x128/66245.png inflating: /content/faces/thumbnails128x128/66246.png inflating: /content/faces/thumbnails128x128/66247.png inflating: /content/faces/thumbnails128x128/66248.png inflating: /content/faces/thumbnails128x128/66249.png inflating: /content/faces/thumbnails128x128/66250.png inflating: /content/faces/thumbnails128x128/66251.png inflating: /content/faces/thumbnails128x128/66252.png inflating: /content/faces/thumbnails128x128/66253.png inflating: /content/faces/thumbnails128x128/66254.png inflating: /content/faces/thumbnails128x128/66255.png inflating: /content/faces/thumbnails128x128/66256.png inflating: /content/faces/thumbnails128x128/66257.png inflating: /content/faces/thumbnails128x128/66258.png inflating: /content/faces/thumbnails128x128/66259.png inflating: /content/faces/thumbnails128x128/66260.png inflating: /content/faces/thumbnails128x128/66261.png inflating: /content/faces/thumbnails128x128/66262.png inflating: /content/faces/thumbnails128x128/66263.png inflating: /content/faces/thumbnails128x128/66264.png inflating: /content/faces/thumbnails128x128/66265.png inflating: /content/faces/thumbnails128x128/66266.png inflating: /content/faces/thumbnails128x128/66267.png inflating: /content/faces/thumbnails128x128/66268.png inflating: /content/faces/thumbnails128x128/66269.png inflating: /content/faces/thumbnails128x128/66270.png inflating: /content/faces/thumbnails128x128/66271.png inflating: /content/faces/thumbnails128x128/66272.png inflating: /content/faces/thumbnails128x128/66273.png inflating: /content/faces/thumbnails128x128/66274.png inflating: /content/faces/thumbnails128x128/66275.png inflating: /content/faces/thumbnails128x128/66276.png inflating: /content/faces/thumbnails128x128/66277.png inflating: /content/faces/thumbnails128x128/66278.png inflating: /content/faces/thumbnails128x128/66279.png inflating: /content/faces/thumbnails128x128/66280.png inflating: /content/faces/thumbnails128x128/66281.png inflating: /content/faces/thumbnails128x128/66282.png inflating: /content/faces/thumbnails128x128/66283.png inflating: /content/faces/thumbnails128x128/66284.png inflating: /content/faces/thumbnails128x128/66285.png inflating: /content/faces/thumbnails128x128/66286.png inflating: /content/faces/thumbnails128x128/66287.png inflating: /content/faces/thumbnails128x128/66288.png inflating: /content/faces/thumbnails128x128/66289.png inflating: /content/faces/thumbnails128x128/66290.png inflating: /content/faces/thumbnails128x128/66291.png inflating: /content/faces/thumbnails128x128/66292.png inflating: /content/faces/thumbnails128x128/66293.png inflating: /content/faces/thumbnails128x128/66294.png inflating: /content/faces/thumbnails128x128/66295.png inflating: /content/faces/thumbnails128x128/66296.png inflating: /content/faces/thumbnails128x128/66297.png inflating: /content/faces/thumbnails128x128/66298.png inflating: /content/faces/thumbnails128x128/66299.png inflating: /content/faces/thumbnails128x128/66300.png inflating: /content/faces/thumbnails128x128/66301.png inflating: /content/faces/thumbnails128x128/66302.png inflating: /content/faces/thumbnails128x128/66303.png inflating: /content/faces/thumbnails128x128/66304.png inflating: /content/faces/thumbnails128x128/66305.png inflating: /content/faces/thumbnails128x128/66306.png inflating: /content/faces/thumbnails128x128/66307.png inflating: /content/faces/thumbnails128x128/66308.png inflating: /content/faces/thumbnails128x128/66309.png inflating: /content/faces/thumbnails128x128/66310.png inflating: /content/faces/thumbnails128x128/66311.png inflating: /content/faces/thumbnails128x128/66312.png inflating: /content/faces/thumbnails128x128/66313.png inflating: /content/faces/thumbnails128x128/66314.png inflating: /content/faces/thumbnails128x128/66315.png inflating: /content/faces/thumbnails128x128/66316.png inflating: /content/faces/thumbnails128x128/66317.png inflating: /content/faces/thumbnails128x128/66318.png inflating: /content/faces/thumbnails128x128/66319.png inflating: /content/faces/thumbnails128x128/66320.png inflating: /content/faces/thumbnails128x128/66321.png inflating: /content/faces/thumbnails128x128/66322.png inflating: /content/faces/thumbnails128x128/66323.png inflating: /content/faces/thumbnails128x128/66324.png inflating: /content/faces/thumbnails128x128/66325.png inflating: /content/faces/thumbnails128x128/66326.png inflating: /content/faces/thumbnails128x128/66327.png inflating: /content/faces/thumbnails128x128/66328.png inflating: /content/faces/thumbnails128x128/66329.png inflating: /content/faces/thumbnails128x128/66330.png inflating: /content/faces/thumbnails128x128/66331.png inflating: /content/faces/thumbnails128x128/66332.png inflating: /content/faces/thumbnails128x128/66333.png inflating: /content/faces/thumbnails128x128/66334.png inflating: /content/faces/thumbnails128x128/66335.png inflating: /content/faces/thumbnails128x128/66336.png inflating: /content/faces/thumbnails128x128/66337.png inflating: /content/faces/thumbnails128x128/66338.png inflating: /content/faces/thumbnails128x128/66339.png inflating: /content/faces/thumbnails128x128/66340.png inflating: /content/faces/thumbnails128x128/66341.png inflating: /content/faces/thumbnails128x128/66342.png inflating: /content/faces/thumbnails128x128/66343.png inflating: /content/faces/thumbnails128x128/66344.png inflating: /content/faces/thumbnails128x128/66345.png inflating: /content/faces/thumbnails128x128/66346.png inflating: /content/faces/thumbnails128x128/66347.png inflating: /content/faces/thumbnails128x128/66348.png inflating: /content/faces/thumbnails128x128/66349.png inflating: /content/faces/thumbnails128x128/66350.png inflating: /content/faces/thumbnails128x128/66351.png inflating: /content/faces/thumbnails128x128/66352.png inflating: /content/faces/thumbnails128x128/66353.png inflating: /content/faces/thumbnails128x128/66354.png inflating: /content/faces/thumbnails128x128/66355.png inflating: /content/faces/thumbnails128x128/66356.png inflating: /content/faces/thumbnails128x128/66357.png inflating: /content/faces/thumbnails128x128/66358.png inflating: /content/faces/thumbnails128x128/66359.png inflating: /content/faces/thumbnails128x128/66360.png inflating: /content/faces/thumbnails128x128/66361.png inflating: /content/faces/thumbnails128x128/66362.png inflating: /content/faces/thumbnails128x128/66363.png inflating: /content/faces/thumbnails128x128/66364.png inflating: /content/faces/thumbnails128x128/66365.png inflating: /content/faces/thumbnails128x128/66366.png inflating: /content/faces/thumbnails128x128/66367.png inflating: /content/faces/thumbnails128x128/66368.png inflating: /content/faces/thumbnails128x128/66369.png inflating: /content/faces/thumbnails128x128/66370.png inflating: /content/faces/thumbnails128x128/66371.png inflating: /content/faces/thumbnails128x128/66372.png inflating: /content/faces/thumbnails128x128/66373.png inflating: /content/faces/thumbnails128x128/66374.png inflating: /content/faces/thumbnails128x128/66375.png inflating: /content/faces/thumbnails128x128/66376.png inflating: /content/faces/thumbnails128x128/66377.png inflating: /content/faces/thumbnails128x128/66378.png inflating: /content/faces/thumbnails128x128/66379.png inflating: /content/faces/thumbnails128x128/66380.png inflating: /content/faces/thumbnails128x128/66381.png inflating: /content/faces/thumbnails128x128/66382.png inflating: /content/faces/thumbnails128x128/66383.png inflating: /content/faces/thumbnails128x128/66384.png inflating: /content/faces/thumbnails128x128/66385.png inflating: /content/faces/thumbnails128x128/66386.png inflating: /content/faces/thumbnails128x128/66387.png inflating: /content/faces/thumbnails128x128/66388.png inflating: /content/faces/thumbnails128x128/66389.png inflating: /content/faces/thumbnails128x128/66390.png inflating: /content/faces/thumbnails128x128/66391.png inflating: /content/faces/thumbnails128x128/66392.png inflating: /content/faces/thumbnails128x128/66393.png inflating: /content/faces/thumbnails128x128/66394.png inflating: /content/faces/thumbnails128x128/66395.png inflating: /content/faces/thumbnails128x128/66396.png inflating: /content/faces/thumbnails128x128/66397.png inflating: /content/faces/thumbnails128x128/66398.png inflating: /content/faces/thumbnails128x128/66399.png inflating: /content/faces/thumbnails128x128/66400.png inflating: /content/faces/thumbnails128x128/66401.png inflating: /content/faces/thumbnails128x128/66402.png inflating: /content/faces/thumbnails128x128/66403.png inflating: /content/faces/thumbnails128x128/66404.png inflating: /content/faces/thumbnails128x128/66405.png inflating: /content/faces/thumbnails128x128/66406.png inflating: /content/faces/thumbnails128x128/66407.png inflating: /content/faces/thumbnails128x128/66408.png inflating: /content/faces/thumbnails128x128/66409.png inflating: /content/faces/thumbnails128x128/66410.png inflating: /content/faces/thumbnails128x128/66411.png inflating: /content/faces/thumbnails128x128/66412.png inflating: /content/faces/thumbnails128x128/66413.png inflating: /content/faces/thumbnails128x128/66414.png inflating: /content/faces/thumbnails128x128/66415.png inflating: /content/faces/thumbnails128x128/66416.png inflating: /content/faces/thumbnails128x128/66417.png inflating: /content/faces/thumbnails128x128/66418.png inflating: /content/faces/thumbnails128x128/66419.png inflating: /content/faces/thumbnails128x128/66420.png inflating: /content/faces/thumbnails128x128/66421.png inflating: /content/faces/thumbnails128x128/66422.png inflating: /content/faces/thumbnails128x128/66423.png inflating: /content/faces/thumbnails128x128/66424.png inflating: /content/faces/thumbnails128x128/66425.png inflating: /content/faces/thumbnails128x128/66426.png inflating: /content/faces/thumbnails128x128/66427.png inflating: /content/faces/thumbnails128x128/66428.png inflating: /content/faces/thumbnails128x128/66429.png inflating: /content/faces/thumbnails128x128/66430.png inflating: /content/faces/thumbnails128x128/66431.png inflating: /content/faces/thumbnails128x128/66432.png inflating: /content/faces/thumbnails128x128/66433.png inflating: /content/faces/thumbnails128x128/66434.png inflating: /content/faces/thumbnails128x128/66435.png inflating: /content/faces/thumbnails128x128/66436.png inflating: /content/faces/thumbnails128x128/66437.png inflating: /content/faces/thumbnails128x128/66438.png inflating: /content/faces/thumbnails128x128/66439.png inflating: /content/faces/thumbnails128x128/66440.png inflating: /content/faces/thumbnails128x128/66441.png inflating: /content/faces/thumbnails128x128/66442.png inflating: /content/faces/thumbnails128x128/66443.png inflating: /content/faces/thumbnails128x128/66444.png inflating: /content/faces/thumbnails128x128/66445.png inflating: /content/faces/thumbnails128x128/66446.png inflating: /content/faces/thumbnails128x128/66447.png inflating: /content/faces/thumbnails128x128/66448.png inflating: /content/faces/thumbnails128x128/66449.png inflating: /content/faces/thumbnails128x128/66450.png inflating: /content/faces/thumbnails128x128/66451.png inflating: /content/faces/thumbnails128x128/66452.png inflating: /content/faces/thumbnails128x128/66453.png inflating: /content/faces/thumbnails128x128/66454.png inflating: /content/faces/thumbnails128x128/66455.png inflating: /content/faces/thumbnails128x128/66456.png inflating: /content/faces/thumbnails128x128/66457.png inflating: /content/faces/thumbnails128x128/66458.png inflating: /content/faces/thumbnails128x128/66459.png inflating: /content/faces/thumbnails128x128/66460.png inflating: /content/faces/thumbnails128x128/66461.png inflating: /content/faces/thumbnails128x128/66462.png inflating: /content/faces/thumbnails128x128/66463.png inflating: /content/faces/thumbnails128x128/66464.png inflating: /content/faces/thumbnails128x128/66465.png inflating: /content/faces/thumbnails128x128/66466.png inflating: /content/faces/thumbnails128x128/66467.png inflating: /content/faces/thumbnails128x128/66468.png inflating: /content/faces/thumbnails128x128/66469.png inflating: /content/faces/thumbnails128x128/66470.png inflating: /content/faces/thumbnails128x128/66471.png inflating: /content/faces/thumbnails128x128/66472.png inflating: /content/faces/thumbnails128x128/66473.png inflating: /content/faces/thumbnails128x128/66474.png inflating: /content/faces/thumbnails128x128/66475.png inflating: /content/faces/thumbnails128x128/66476.png inflating: /content/faces/thumbnails128x128/66477.png inflating: /content/faces/thumbnails128x128/66478.png inflating: /content/faces/thumbnails128x128/66479.png inflating: /content/faces/thumbnails128x128/66480.png inflating: /content/faces/thumbnails128x128/66481.png inflating: /content/faces/thumbnails128x128/66482.png inflating: /content/faces/thumbnails128x128/66483.png inflating: /content/faces/thumbnails128x128/66484.png inflating: /content/faces/thumbnails128x128/66485.png inflating: /content/faces/thumbnails128x128/66486.png inflating: /content/faces/thumbnails128x128/66487.png inflating: /content/faces/thumbnails128x128/66488.png inflating: /content/faces/thumbnails128x128/66489.png inflating: /content/faces/thumbnails128x128/66490.png inflating: /content/faces/thumbnails128x128/66491.png inflating: /content/faces/thumbnails128x128/66492.png inflating: /content/faces/thumbnails128x128/66493.png inflating: /content/faces/thumbnails128x128/66494.png inflating: /content/faces/thumbnails128x128/66495.png inflating: /content/faces/thumbnails128x128/66496.png inflating: /content/faces/thumbnails128x128/66497.png inflating: /content/faces/thumbnails128x128/66498.png inflating: /content/faces/thumbnails128x128/66499.png inflating: /content/faces/thumbnails128x128/66500.png inflating: /content/faces/thumbnails128x128/66501.png inflating: /content/faces/thumbnails128x128/66502.png inflating: /content/faces/thumbnails128x128/66503.png inflating: /content/faces/thumbnails128x128/66504.png inflating: /content/faces/thumbnails128x128/66505.png inflating: /content/faces/thumbnails128x128/66506.png inflating: /content/faces/thumbnails128x128/66507.png inflating: /content/faces/thumbnails128x128/66508.png inflating: /content/faces/thumbnails128x128/66509.png inflating: /content/faces/thumbnails128x128/66510.png inflating: /content/faces/thumbnails128x128/66511.png inflating: /content/faces/thumbnails128x128/66512.png inflating: /content/faces/thumbnails128x128/66513.png inflating: /content/faces/thumbnails128x128/66514.png inflating: /content/faces/thumbnails128x128/66515.png inflating: /content/faces/thumbnails128x128/66516.png inflating: /content/faces/thumbnails128x128/66517.png inflating: /content/faces/thumbnails128x128/66518.png inflating: /content/faces/thumbnails128x128/66519.png inflating: /content/faces/thumbnails128x128/66520.png inflating: /content/faces/thumbnails128x128/66521.png inflating: /content/faces/thumbnails128x128/66522.png inflating: /content/faces/thumbnails128x128/66523.png inflating: /content/faces/thumbnails128x128/66524.png inflating: /content/faces/thumbnails128x128/66525.png inflating: /content/faces/thumbnails128x128/66526.png inflating: /content/faces/thumbnails128x128/66527.png inflating: /content/faces/thumbnails128x128/66528.png inflating: /content/faces/thumbnails128x128/66529.png inflating: /content/faces/thumbnails128x128/66530.png inflating: /content/faces/thumbnails128x128/66531.png inflating: /content/faces/thumbnails128x128/66532.png inflating: /content/faces/thumbnails128x128/66533.png inflating: /content/faces/thumbnails128x128/66534.png inflating: /content/faces/thumbnails128x128/66535.png inflating: /content/faces/thumbnails128x128/66536.png inflating: /content/faces/thumbnails128x128/66537.png inflating: /content/faces/thumbnails128x128/66538.png inflating: /content/faces/thumbnails128x128/66539.png inflating: /content/faces/thumbnails128x128/66540.png inflating: /content/faces/thumbnails128x128/66541.png inflating: /content/faces/thumbnails128x128/66542.png inflating: /content/faces/thumbnails128x128/66543.png inflating: /content/faces/thumbnails128x128/66544.png inflating: /content/faces/thumbnails128x128/66545.png inflating: /content/faces/thumbnails128x128/66546.png inflating: /content/faces/thumbnails128x128/66547.png inflating: /content/faces/thumbnails128x128/66548.png inflating: /content/faces/thumbnails128x128/66549.png inflating: /content/faces/thumbnails128x128/66550.png inflating: /content/faces/thumbnails128x128/66551.png inflating: /content/faces/thumbnails128x128/66552.png inflating: /content/faces/thumbnails128x128/66553.png inflating: /content/faces/thumbnails128x128/66554.png inflating: /content/faces/thumbnails128x128/66555.png inflating: /content/faces/thumbnails128x128/66556.png inflating: /content/faces/thumbnails128x128/66557.png inflating: /content/faces/thumbnails128x128/66558.png inflating: /content/faces/thumbnails128x128/66559.png inflating: /content/faces/thumbnails128x128/66560.png inflating: /content/faces/thumbnails128x128/66561.png inflating: /content/faces/thumbnails128x128/66562.png inflating: /content/faces/thumbnails128x128/66563.png inflating: /content/faces/thumbnails128x128/66564.png inflating: /content/faces/thumbnails128x128/66565.png inflating: /content/faces/thumbnails128x128/66566.png inflating: /content/faces/thumbnails128x128/66567.png inflating: /content/faces/thumbnails128x128/66568.png inflating: /content/faces/thumbnails128x128/66569.png inflating: /content/faces/thumbnails128x128/66570.png inflating: /content/faces/thumbnails128x128/66571.png inflating: /content/faces/thumbnails128x128/66572.png inflating: /content/faces/thumbnails128x128/66573.png inflating: /content/faces/thumbnails128x128/66574.png inflating: /content/faces/thumbnails128x128/66575.png inflating: /content/faces/thumbnails128x128/66576.png inflating: /content/faces/thumbnails128x128/66577.png inflating: /content/faces/thumbnails128x128/66578.png inflating: /content/faces/thumbnails128x128/66579.png inflating: /content/faces/thumbnails128x128/66580.png inflating: /content/faces/thumbnails128x128/66581.png inflating: /content/faces/thumbnails128x128/66582.png inflating: /content/faces/thumbnails128x128/66583.png inflating: /content/faces/thumbnails128x128/66584.png inflating: /content/faces/thumbnails128x128/66585.png inflating: /content/faces/thumbnails128x128/66586.png inflating: /content/faces/thumbnails128x128/66587.png inflating: /content/faces/thumbnails128x128/66588.png inflating: /content/faces/thumbnails128x128/66589.png inflating: /content/faces/thumbnails128x128/66590.png inflating: /content/faces/thumbnails128x128/66591.png inflating: /content/faces/thumbnails128x128/66592.png inflating: /content/faces/thumbnails128x128/66593.png inflating: /content/faces/thumbnails128x128/66594.png inflating: /content/faces/thumbnails128x128/66595.png inflating: /content/faces/thumbnails128x128/66596.png inflating: /content/faces/thumbnails128x128/66597.png inflating: /content/faces/thumbnails128x128/66598.png inflating: /content/faces/thumbnails128x128/66599.png inflating: /content/faces/thumbnails128x128/66600.png inflating: /content/faces/thumbnails128x128/66601.png inflating: /content/faces/thumbnails128x128/66602.png inflating: /content/faces/thumbnails128x128/66603.png inflating: /content/faces/thumbnails128x128/66604.png inflating: /content/faces/thumbnails128x128/66605.png inflating: /content/faces/thumbnails128x128/66606.png inflating: /content/faces/thumbnails128x128/66607.png inflating: /content/faces/thumbnails128x128/66608.png inflating: /content/faces/thumbnails128x128/66609.png inflating: /content/faces/thumbnails128x128/66610.png inflating: /content/faces/thumbnails128x128/66611.png inflating: /content/faces/thumbnails128x128/66612.png inflating: /content/faces/thumbnails128x128/66613.png inflating: /content/faces/thumbnails128x128/66614.png inflating: /content/faces/thumbnails128x128/66615.png inflating: /content/faces/thumbnails128x128/66616.png inflating: /content/faces/thumbnails128x128/66617.png inflating: /content/faces/thumbnails128x128/66618.png inflating: /content/faces/thumbnails128x128/66619.png inflating: /content/faces/thumbnails128x128/66620.png inflating: /content/faces/thumbnails128x128/66621.png inflating: /content/faces/thumbnails128x128/66622.png inflating: /content/faces/thumbnails128x128/66623.png inflating: /content/faces/thumbnails128x128/66624.png inflating: /content/faces/thumbnails128x128/66625.png inflating: /content/faces/thumbnails128x128/66626.png inflating: /content/faces/thumbnails128x128/66627.png inflating: /content/faces/thumbnails128x128/66628.png inflating: /content/faces/thumbnails128x128/66629.png inflating: /content/faces/thumbnails128x128/66630.png inflating: /content/faces/thumbnails128x128/66631.png inflating: /content/faces/thumbnails128x128/66632.png inflating: /content/faces/thumbnails128x128/66633.png inflating: /content/faces/thumbnails128x128/66634.png inflating: /content/faces/thumbnails128x128/66635.png inflating: /content/faces/thumbnails128x128/66636.png inflating: /content/faces/thumbnails128x128/66637.png inflating: /content/faces/thumbnails128x128/66638.png inflating: /content/faces/thumbnails128x128/66639.png inflating: /content/faces/thumbnails128x128/66640.png inflating: /content/faces/thumbnails128x128/66641.png inflating: /content/faces/thumbnails128x128/66642.png inflating: /content/faces/thumbnails128x128/66643.png inflating: /content/faces/thumbnails128x128/66644.png inflating: /content/faces/thumbnails128x128/66645.png inflating: /content/faces/thumbnails128x128/66646.png inflating: /content/faces/thumbnails128x128/66647.png inflating: /content/faces/thumbnails128x128/66648.png inflating: /content/faces/thumbnails128x128/66649.png inflating: /content/faces/thumbnails128x128/66650.png inflating: /content/faces/thumbnails128x128/66651.png inflating: /content/faces/thumbnails128x128/66652.png inflating: /content/faces/thumbnails128x128/66653.png inflating: /content/faces/thumbnails128x128/66654.png inflating: /content/faces/thumbnails128x128/66655.png inflating: /content/faces/thumbnails128x128/66656.png inflating: /content/faces/thumbnails128x128/66657.png inflating: /content/faces/thumbnails128x128/66658.png inflating: /content/faces/thumbnails128x128/66659.png inflating: /content/faces/thumbnails128x128/66660.png inflating: /content/faces/thumbnails128x128/66661.png inflating: /content/faces/thumbnails128x128/66662.png inflating: /content/faces/thumbnails128x128/66663.png inflating: /content/faces/thumbnails128x128/66664.png inflating: /content/faces/thumbnails128x128/66665.png inflating: /content/faces/thumbnails128x128/66666.png inflating: /content/faces/thumbnails128x128/66667.png inflating: /content/faces/thumbnails128x128/66668.png inflating: /content/faces/thumbnails128x128/66669.png inflating: /content/faces/thumbnails128x128/66670.png inflating: /content/faces/thumbnails128x128/66671.png inflating: /content/faces/thumbnails128x128/66672.png inflating: /content/faces/thumbnails128x128/66673.png inflating: /content/faces/thumbnails128x128/66674.png inflating: /content/faces/thumbnails128x128/66675.png inflating: /content/faces/thumbnails128x128/66676.png inflating: /content/faces/thumbnails128x128/66677.png inflating: /content/faces/thumbnails128x128/66678.png inflating: /content/faces/thumbnails128x128/66679.png inflating: /content/faces/thumbnails128x128/66680.png inflating: /content/faces/thumbnails128x128/66681.png inflating: /content/faces/thumbnails128x128/66682.png inflating: /content/faces/thumbnails128x128/66683.png inflating: /content/faces/thumbnails128x128/66684.png inflating: /content/faces/thumbnails128x128/66685.png inflating: /content/faces/thumbnails128x128/66686.png inflating: /content/faces/thumbnails128x128/66687.png inflating: /content/faces/thumbnails128x128/66688.png inflating: /content/faces/thumbnails128x128/66689.png inflating: /content/faces/thumbnails128x128/66690.png inflating: /content/faces/thumbnails128x128/66691.png inflating: /content/faces/thumbnails128x128/66692.png inflating: /content/faces/thumbnails128x128/66693.png inflating: /content/faces/thumbnails128x128/66694.png inflating: /content/faces/thumbnails128x128/66695.png inflating: /content/faces/thumbnails128x128/66696.png inflating: /content/faces/thumbnails128x128/66697.png inflating: /content/faces/thumbnails128x128/66698.png inflating: /content/faces/thumbnails128x128/66699.png inflating: /content/faces/thumbnails128x128/66700.png inflating: /content/faces/thumbnails128x128/66701.png inflating: /content/faces/thumbnails128x128/66702.png inflating: /content/faces/thumbnails128x128/66703.png inflating: /content/faces/thumbnails128x128/66704.png inflating: /content/faces/thumbnails128x128/66705.png inflating: /content/faces/thumbnails128x128/66706.png inflating: /content/faces/thumbnails128x128/66707.png inflating: /content/faces/thumbnails128x128/66708.png inflating: /content/faces/thumbnails128x128/66709.png inflating: /content/faces/thumbnails128x128/66710.png inflating: /content/faces/thumbnails128x128/66711.png inflating: /content/faces/thumbnails128x128/66712.png inflating: /content/faces/thumbnails128x128/66713.png inflating: /content/faces/thumbnails128x128/66714.png inflating: /content/faces/thumbnails128x128/66715.png inflating: /content/faces/thumbnails128x128/66716.png inflating: /content/faces/thumbnails128x128/66717.png inflating: /content/faces/thumbnails128x128/66718.png inflating: /content/faces/thumbnails128x128/66719.png inflating: /content/faces/thumbnails128x128/66720.png inflating: /content/faces/thumbnails128x128/66721.png inflating: /content/faces/thumbnails128x128/66722.png inflating: /content/faces/thumbnails128x128/66723.png inflating: /content/faces/thumbnails128x128/66724.png inflating: /content/faces/thumbnails128x128/66725.png inflating: /content/faces/thumbnails128x128/66726.png inflating: /content/faces/thumbnails128x128/66727.png inflating: /content/faces/thumbnails128x128/66728.png inflating: /content/faces/thumbnails128x128/66729.png inflating: /content/faces/thumbnails128x128/66730.png inflating: /content/faces/thumbnails128x128/66731.png inflating: /content/faces/thumbnails128x128/66732.png inflating: /content/faces/thumbnails128x128/66733.png inflating: /content/faces/thumbnails128x128/66734.png inflating: /content/faces/thumbnails128x128/66735.png inflating: /content/faces/thumbnails128x128/66736.png inflating: /content/faces/thumbnails128x128/66737.png inflating: /content/faces/thumbnails128x128/66738.png inflating: /content/faces/thumbnails128x128/66739.png inflating: /content/faces/thumbnails128x128/66740.png inflating: /content/faces/thumbnails128x128/66741.png inflating: /content/faces/thumbnails128x128/66742.png inflating: /content/faces/thumbnails128x128/66743.png inflating: /content/faces/thumbnails128x128/66744.png inflating: /content/faces/thumbnails128x128/66745.png inflating: /content/faces/thumbnails128x128/66746.png inflating: /content/faces/thumbnails128x128/66747.png inflating: /content/faces/thumbnails128x128/66748.png inflating: /content/faces/thumbnails128x128/66749.png inflating: /content/faces/thumbnails128x128/66750.png inflating: /content/faces/thumbnails128x128/66751.png inflating: /content/faces/thumbnails128x128/66752.png inflating: /content/faces/thumbnails128x128/66753.png inflating: /content/faces/thumbnails128x128/66754.png inflating: /content/faces/thumbnails128x128/66755.png inflating: /content/faces/thumbnails128x128/66756.png inflating: /content/faces/thumbnails128x128/66757.png inflating: /content/faces/thumbnails128x128/66758.png inflating: /content/faces/thumbnails128x128/66759.png inflating: /content/faces/thumbnails128x128/66760.png inflating: /content/faces/thumbnails128x128/66761.png inflating: /content/faces/thumbnails128x128/66762.png inflating: /content/faces/thumbnails128x128/66763.png inflating: /content/faces/thumbnails128x128/66764.png inflating: /content/faces/thumbnails128x128/66765.png inflating: /content/faces/thumbnails128x128/66766.png inflating: /content/faces/thumbnails128x128/66767.png inflating: /content/faces/thumbnails128x128/66768.png inflating: /content/faces/thumbnails128x128/66769.png inflating: /content/faces/thumbnails128x128/66770.png inflating: /content/faces/thumbnails128x128/66771.png inflating: /content/faces/thumbnails128x128/66772.png inflating: /content/faces/thumbnails128x128/66773.png inflating: /content/faces/thumbnails128x128/66774.png inflating: /content/faces/thumbnails128x128/66775.png inflating: /content/faces/thumbnails128x128/66776.png inflating: /content/faces/thumbnails128x128/66777.png inflating: /content/faces/thumbnails128x128/66778.png inflating: /content/faces/thumbnails128x128/66779.png inflating: /content/faces/thumbnails128x128/66780.png inflating: /content/faces/thumbnails128x128/66781.png inflating: /content/faces/thumbnails128x128/66782.png inflating: /content/faces/thumbnails128x128/66783.png inflating: /content/faces/thumbnails128x128/66784.png inflating: /content/faces/thumbnails128x128/66785.png inflating: /content/faces/thumbnails128x128/66786.png inflating: /content/faces/thumbnails128x128/66787.png inflating: /content/faces/thumbnails128x128/66788.png inflating: /content/faces/thumbnails128x128/66789.png inflating: /content/faces/thumbnails128x128/66790.png inflating: /content/faces/thumbnails128x128/66791.png inflating: /content/faces/thumbnails128x128/66792.png inflating: /content/faces/thumbnails128x128/66793.png inflating: /content/faces/thumbnails128x128/66794.png inflating: /content/faces/thumbnails128x128/66795.png inflating: /content/faces/thumbnails128x128/66796.png inflating: /content/faces/thumbnails128x128/66797.png inflating: /content/faces/thumbnails128x128/66798.png inflating: /content/faces/thumbnails128x128/66799.png inflating: /content/faces/thumbnails128x128/66800.png inflating: /content/faces/thumbnails128x128/66801.png inflating: /content/faces/thumbnails128x128/66802.png inflating: /content/faces/thumbnails128x128/66803.png inflating: /content/faces/thumbnails128x128/66804.png inflating: /content/faces/thumbnails128x128/66805.png inflating: /content/faces/thumbnails128x128/66806.png inflating: /content/faces/thumbnails128x128/66807.png inflating: /content/faces/thumbnails128x128/66808.png inflating: /content/faces/thumbnails128x128/66809.png inflating: /content/faces/thumbnails128x128/66810.png inflating: /content/faces/thumbnails128x128/66811.png inflating: /content/faces/thumbnails128x128/66812.png inflating: /content/faces/thumbnails128x128/66813.png inflating: /content/faces/thumbnails128x128/66814.png inflating: /content/faces/thumbnails128x128/66815.png inflating: /content/faces/thumbnails128x128/66816.png inflating: /content/faces/thumbnails128x128/66817.png inflating: /content/faces/thumbnails128x128/66818.png inflating: /content/faces/thumbnails128x128/66819.png inflating: /content/faces/thumbnails128x128/66820.png inflating: /content/faces/thumbnails128x128/66821.png inflating: /content/faces/thumbnails128x128/66822.png inflating: /content/faces/thumbnails128x128/66823.png inflating: /content/faces/thumbnails128x128/66824.png inflating: /content/faces/thumbnails128x128/66825.png inflating: /content/faces/thumbnails128x128/66826.png inflating: /content/faces/thumbnails128x128/66827.png inflating: /content/faces/thumbnails128x128/66828.png inflating: /content/faces/thumbnails128x128/66829.png inflating: /content/faces/thumbnails128x128/66830.png inflating: /content/faces/thumbnails128x128/66831.png inflating: /content/faces/thumbnails128x128/66832.png inflating: /content/faces/thumbnails128x128/66833.png inflating: /content/faces/thumbnails128x128/66834.png inflating: /content/faces/thumbnails128x128/66835.png inflating: /content/faces/thumbnails128x128/66836.png inflating: /content/faces/thumbnails128x128/66837.png inflating: /content/faces/thumbnails128x128/66838.png inflating: /content/faces/thumbnails128x128/66839.png inflating: /content/faces/thumbnails128x128/66840.png inflating: /content/faces/thumbnails128x128/66841.png inflating: /content/faces/thumbnails128x128/66842.png inflating: /content/faces/thumbnails128x128/66843.png inflating: /content/faces/thumbnails128x128/66844.png inflating: /content/faces/thumbnails128x128/66845.png inflating: /content/faces/thumbnails128x128/66846.png inflating: /content/faces/thumbnails128x128/66847.png inflating: /content/faces/thumbnails128x128/66848.png inflating: /content/faces/thumbnails128x128/66849.png inflating: /content/faces/thumbnails128x128/66850.png inflating: /content/faces/thumbnails128x128/66851.png inflating: /content/faces/thumbnails128x128/66852.png inflating: /content/faces/thumbnails128x128/66853.png inflating: /content/faces/thumbnails128x128/66854.png inflating: /content/faces/thumbnails128x128/66855.png inflating: /content/faces/thumbnails128x128/66856.png inflating: /content/faces/thumbnails128x128/66857.png inflating: /content/faces/thumbnails128x128/66858.png inflating: /content/faces/thumbnails128x128/66859.png inflating: /content/faces/thumbnails128x128/66860.png inflating: /content/faces/thumbnails128x128/66861.png inflating: /content/faces/thumbnails128x128/66862.png inflating: /content/faces/thumbnails128x128/66863.png inflating: /content/faces/thumbnails128x128/66864.png inflating: /content/faces/thumbnails128x128/66865.png inflating: /content/faces/thumbnails128x128/66866.png inflating: /content/faces/thumbnails128x128/66867.png inflating: /content/faces/thumbnails128x128/66868.png inflating: /content/faces/thumbnails128x128/66869.png inflating: /content/faces/thumbnails128x128/66870.png inflating: /content/faces/thumbnails128x128/66871.png inflating: /content/faces/thumbnails128x128/66872.png inflating: /content/faces/thumbnails128x128/66873.png inflating: /content/faces/thumbnails128x128/66874.png inflating: /content/faces/thumbnails128x128/66875.png inflating: /content/faces/thumbnails128x128/66876.png inflating: /content/faces/thumbnails128x128/66877.png inflating: /content/faces/thumbnails128x128/66878.png inflating: /content/faces/thumbnails128x128/66879.png inflating: /content/faces/thumbnails128x128/66880.png inflating: /content/faces/thumbnails128x128/66881.png inflating: /content/faces/thumbnails128x128/66882.png inflating: /content/faces/thumbnails128x128/66883.png inflating: /content/faces/thumbnails128x128/66884.png inflating: /content/faces/thumbnails128x128/66885.png inflating: /content/faces/thumbnails128x128/66886.png inflating: /content/faces/thumbnails128x128/66887.png inflating: /content/faces/thumbnails128x128/66888.png inflating: /content/faces/thumbnails128x128/66889.png inflating: /content/faces/thumbnails128x128/66890.png inflating: /content/faces/thumbnails128x128/66891.png inflating: /content/faces/thumbnails128x128/66892.png inflating: /content/faces/thumbnails128x128/66893.png inflating: /content/faces/thumbnails128x128/66894.png inflating: /content/faces/thumbnails128x128/66895.png inflating: /content/faces/thumbnails128x128/66896.png inflating: /content/faces/thumbnails128x128/66897.png inflating: /content/faces/thumbnails128x128/66898.png inflating: /content/faces/thumbnails128x128/66899.png inflating: /content/faces/thumbnails128x128/66900.png inflating: /content/faces/thumbnails128x128/66901.png inflating: /content/faces/thumbnails128x128/66902.png inflating: /content/faces/thumbnails128x128/66903.png inflating: /content/faces/thumbnails128x128/66904.png inflating: /content/faces/thumbnails128x128/66905.png inflating: /content/faces/thumbnails128x128/66906.png inflating: /content/faces/thumbnails128x128/66907.png inflating: /content/faces/thumbnails128x128/66908.png inflating: /content/faces/thumbnails128x128/66909.png inflating: /content/faces/thumbnails128x128/66910.png inflating: /content/faces/thumbnails128x128/66911.png inflating: /content/faces/thumbnails128x128/66912.png inflating: /content/faces/thumbnails128x128/66913.png inflating: /content/faces/thumbnails128x128/66914.png inflating: /content/faces/thumbnails128x128/66915.png inflating: /content/faces/thumbnails128x128/66916.png inflating: /content/faces/thumbnails128x128/66917.png inflating: /content/faces/thumbnails128x128/66918.png inflating: /content/faces/thumbnails128x128/66919.png inflating: /content/faces/thumbnails128x128/66920.png inflating: /content/faces/thumbnails128x128/66921.png inflating: /content/faces/thumbnails128x128/66922.png inflating: /content/faces/thumbnails128x128/66923.png inflating: /content/faces/thumbnails128x128/66924.png inflating: /content/faces/thumbnails128x128/66925.png inflating: /content/faces/thumbnails128x128/66926.png inflating: /content/faces/thumbnails128x128/66927.png inflating: /content/faces/thumbnails128x128/66928.png inflating: /content/faces/thumbnails128x128/66929.png inflating: /content/faces/thumbnails128x128/66930.png inflating: /content/faces/thumbnails128x128/66931.png inflating: /content/faces/thumbnails128x128/66932.png inflating: /content/faces/thumbnails128x128/66933.png inflating: /content/faces/thumbnails128x128/66934.png inflating: /content/faces/thumbnails128x128/66935.png inflating: /content/faces/thumbnails128x128/66936.png inflating: /content/faces/thumbnails128x128/66937.png inflating: /content/faces/thumbnails128x128/66938.png inflating: /content/faces/thumbnails128x128/66939.png inflating: /content/faces/thumbnails128x128/66940.png inflating: /content/faces/thumbnails128x128/66941.png inflating: /content/faces/thumbnails128x128/66942.png inflating: /content/faces/thumbnails128x128/66943.png inflating: /content/faces/thumbnails128x128/66944.png inflating: /content/faces/thumbnails128x128/66945.png inflating: /content/faces/thumbnails128x128/66946.png inflating: /content/faces/thumbnails128x128/66947.png inflating: /content/faces/thumbnails128x128/66948.png inflating: /content/faces/thumbnails128x128/66949.png inflating: /content/faces/thumbnails128x128/66950.png inflating: /content/faces/thumbnails128x128/66951.png inflating: /content/faces/thumbnails128x128/66952.png inflating: /content/faces/thumbnails128x128/66953.png inflating: /content/faces/thumbnails128x128/66954.png inflating: /content/faces/thumbnails128x128/66955.png inflating: /content/faces/thumbnails128x128/66956.png inflating: /content/faces/thumbnails128x128/66957.png inflating: /content/faces/thumbnails128x128/66958.png inflating: /content/faces/thumbnails128x128/66959.png inflating: /content/faces/thumbnails128x128/66960.png inflating: /content/faces/thumbnails128x128/66961.png inflating: /content/faces/thumbnails128x128/66962.png inflating: /content/faces/thumbnails128x128/66963.png inflating: /content/faces/thumbnails128x128/66964.png inflating: /content/faces/thumbnails128x128/66965.png inflating: /content/faces/thumbnails128x128/66966.png inflating: /content/faces/thumbnails128x128/66967.png inflating: /content/faces/thumbnails128x128/66968.png inflating: /content/faces/thumbnails128x128/66969.png inflating: /content/faces/thumbnails128x128/66970.png inflating: /content/faces/thumbnails128x128/66971.png inflating: /content/faces/thumbnails128x128/66972.png inflating: /content/faces/thumbnails128x128/66973.png inflating: /content/faces/thumbnails128x128/66974.png inflating: /content/faces/thumbnails128x128/66975.png inflating: /content/faces/thumbnails128x128/66976.png inflating: /content/faces/thumbnails128x128/66977.png inflating: /content/faces/thumbnails128x128/66978.png inflating: /content/faces/thumbnails128x128/66979.png inflating: /content/faces/thumbnails128x128/66980.png inflating: /content/faces/thumbnails128x128/66981.png inflating: /content/faces/thumbnails128x128/66982.png inflating: /content/faces/thumbnails128x128/66983.png inflating: /content/faces/thumbnails128x128/66984.png inflating: /content/faces/thumbnails128x128/66985.png inflating: /content/faces/thumbnails128x128/66986.png inflating: /content/faces/thumbnails128x128/66987.png inflating: /content/faces/thumbnails128x128/66988.png inflating: /content/faces/thumbnails128x128/66989.png inflating: /content/faces/thumbnails128x128/66990.png inflating: /content/faces/thumbnails128x128/66991.png inflating: /content/faces/thumbnails128x128/66992.png inflating: /content/faces/thumbnails128x128/66993.png inflating: /content/faces/thumbnails128x128/66994.png inflating: /content/faces/thumbnails128x128/66995.png inflating: /content/faces/thumbnails128x128/66996.png inflating: /content/faces/thumbnails128x128/66997.png inflating: /content/faces/thumbnails128x128/66998.png inflating: /content/faces/thumbnails128x128/66999.png inflating: /content/faces/thumbnails128x128/67000.png inflating: /content/faces/thumbnails128x128/67001.png inflating: /content/faces/thumbnails128x128/67002.png inflating: /content/faces/thumbnails128x128/67003.png inflating: /content/faces/thumbnails128x128/67004.png inflating: /content/faces/thumbnails128x128/67005.png inflating: /content/faces/thumbnails128x128/67006.png inflating: /content/faces/thumbnails128x128/67007.png inflating: /content/faces/thumbnails128x128/67008.png inflating: /content/faces/thumbnails128x128/67009.png inflating: /content/faces/thumbnails128x128/67010.png inflating: /content/faces/thumbnails128x128/67011.png inflating: /content/faces/thumbnails128x128/67012.png inflating: /content/faces/thumbnails128x128/67013.png inflating: /content/faces/thumbnails128x128/67014.png inflating: /content/faces/thumbnails128x128/67015.png inflating: /content/faces/thumbnails128x128/67016.png inflating: /content/faces/thumbnails128x128/67017.png inflating: /content/faces/thumbnails128x128/67018.png inflating: /content/faces/thumbnails128x128/67019.png inflating: /content/faces/thumbnails128x128/67020.png inflating: /content/faces/thumbnails128x128/67021.png inflating: /content/faces/thumbnails128x128/67022.png inflating: /content/faces/thumbnails128x128/67023.png inflating: /content/faces/thumbnails128x128/67024.png inflating: /content/faces/thumbnails128x128/67025.png inflating: /content/faces/thumbnails128x128/67026.png inflating: /content/faces/thumbnails128x128/67027.png inflating: /content/faces/thumbnails128x128/67028.png inflating: /content/faces/thumbnails128x128/67029.png inflating: /content/faces/thumbnails128x128/67030.png inflating: /content/faces/thumbnails128x128/67031.png inflating: /content/faces/thumbnails128x128/67032.png inflating: /content/faces/thumbnails128x128/67033.png inflating: /content/faces/thumbnails128x128/67034.png inflating: /content/faces/thumbnails128x128/67035.png inflating: /content/faces/thumbnails128x128/67036.png inflating: /content/faces/thumbnails128x128/67037.png inflating: /content/faces/thumbnails128x128/67038.png inflating: /content/faces/thumbnails128x128/67039.png inflating: /content/faces/thumbnails128x128/67040.png inflating: /content/faces/thumbnails128x128/67041.png inflating: /content/faces/thumbnails128x128/67042.png inflating: /content/faces/thumbnails128x128/67043.png inflating: /content/faces/thumbnails128x128/67044.png inflating: /content/faces/thumbnails128x128/67045.png inflating: /content/faces/thumbnails128x128/67046.png inflating: /content/faces/thumbnails128x128/67047.png inflating: /content/faces/thumbnails128x128/67048.png inflating: /content/faces/thumbnails128x128/67049.png inflating: /content/faces/thumbnails128x128/67050.png inflating: /content/faces/thumbnails128x128/67051.png inflating: /content/faces/thumbnails128x128/67052.png inflating: /content/faces/thumbnails128x128/67053.png inflating: /content/faces/thumbnails128x128/67054.png inflating: /content/faces/thumbnails128x128/67055.png inflating: /content/faces/thumbnails128x128/67056.png inflating: /content/faces/thumbnails128x128/67057.png inflating: /content/faces/thumbnails128x128/67058.png inflating: /content/faces/thumbnails128x128/67059.png inflating: /content/faces/thumbnails128x128/67060.png inflating: /content/faces/thumbnails128x128/67061.png inflating: /content/faces/thumbnails128x128/67062.png inflating: /content/faces/thumbnails128x128/67063.png inflating: /content/faces/thumbnails128x128/67064.png inflating: /content/faces/thumbnails128x128/67065.png inflating: /content/faces/thumbnails128x128/67066.png inflating: /content/faces/thumbnails128x128/67067.png inflating: /content/faces/thumbnails128x128/67068.png inflating: /content/faces/thumbnails128x128/67069.png inflating: /content/faces/thumbnails128x128/67070.png inflating: /content/faces/thumbnails128x128/67071.png inflating: /content/faces/thumbnails128x128/67072.png inflating: /content/faces/thumbnails128x128/67073.png inflating: /content/faces/thumbnails128x128/67074.png inflating: /content/faces/thumbnails128x128/67075.png inflating: /content/faces/thumbnails128x128/67076.png inflating: /content/faces/thumbnails128x128/67077.png inflating: /content/faces/thumbnails128x128/67078.png inflating: /content/faces/thumbnails128x128/67079.png inflating: /content/faces/thumbnails128x128/67080.png inflating: /content/faces/thumbnails128x128/67081.png inflating: /content/faces/thumbnails128x128/67082.png inflating: /content/faces/thumbnails128x128/67083.png inflating: /content/faces/thumbnails128x128/67084.png inflating: /content/faces/thumbnails128x128/67085.png inflating: /content/faces/thumbnails128x128/67086.png inflating: /content/faces/thumbnails128x128/67087.png inflating: /content/faces/thumbnails128x128/67088.png inflating: /content/faces/thumbnails128x128/67089.png inflating: /content/faces/thumbnails128x128/67090.png inflating: /content/faces/thumbnails128x128/67091.png inflating: /content/faces/thumbnails128x128/67092.png inflating: /content/faces/thumbnails128x128/67093.png inflating: /content/faces/thumbnails128x128/67094.png inflating: /content/faces/thumbnails128x128/67095.png inflating: /content/faces/thumbnails128x128/67096.png inflating: /content/faces/thumbnails128x128/67097.png inflating: /content/faces/thumbnails128x128/67098.png inflating: /content/faces/thumbnails128x128/67099.png inflating: /content/faces/thumbnails128x128/67100.png inflating: /content/faces/thumbnails128x128/67101.png inflating: /content/faces/thumbnails128x128/67102.png inflating: /content/faces/thumbnails128x128/67103.png inflating: /content/faces/thumbnails128x128/67104.png inflating: /content/faces/thumbnails128x128/67105.png inflating: /content/faces/thumbnails128x128/67106.png inflating: /content/faces/thumbnails128x128/67107.png inflating: /content/faces/thumbnails128x128/67108.png inflating: /content/faces/thumbnails128x128/67109.png inflating: /content/faces/thumbnails128x128/67110.png inflating: /content/faces/thumbnails128x128/67111.png inflating: /content/faces/thumbnails128x128/67112.png inflating: /content/faces/thumbnails128x128/67113.png inflating: /content/faces/thumbnails128x128/67114.png inflating: /content/faces/thumbnails128x128/67115.png inflating: /content/faces/thumbnails128x128/67116.png inflating: /content/faces/thumbnails128x128/67117.png inflating: /content/faces/thumbnails128x128/67118.png inflating: /content/faces/thumbnails128x128/67119.png inflating: /content/faces/thumbnails128x128/67120.png inflating: /content/faces/thumbnails128x128/67121.png inflating: /content/faces/thumbnails128x128/67122.png inflating: /content/faces/thumbnails128x128/67123.png inflating: /content/faces/thumbnails128x128/67124.png inflating: /content/faces/thumbnails128x128/67125.png inflating: /content/faces/thumbnails128x128/67126.png inflating: /content/faces/thumbnails128x128/67127.png inflating: /content/faces/thumbnails128x128/67128.png inflating: /content/faces/thumbnails128x128/67129.png inflating: /content/faces/thumbnails128x128/67130.png inflating: /content/faces/thumbnails128x128/67131.png inflating: /content/faces/thumbnails128x128/67132.png inflating: /content/faces/thumbnails128x128/67133.png inflating: /content/faces/thumbnails128x128/67134.png inflating: /content/faces/thumbnails128x128/67135.png inflating: /content/faces/thumbnails128x128/67136.png inflating: /content/faces/thumbnails128x128/67137.png inflating: /content/faces/thumbnails128x128/67138.png inflating: /content/faces/thumbnails128x128/67139.png inflating: /content/faces/thumbnails128x128/67140.png inflating: /content/faces/thumbnails128x128/67141.png inflating: /content/faces/thumbnails128x128/67142.png inflating: /content/faces/thumbnails128x128/67143.png inflating: /content/faces/thumbnails128x128/67144.png inflating: /content/faces/thumbnails128x128/67145.png inflating: /content/faces/thumbnails128x128/67146.png inflating: /content/faces/thumbnails128x128/67147.png inflating: /content/faces/thumbnails128x128/67148.png inflating: /content/faces/thumbnails128x128/67149.png inflating: /content/faces/thumbnails128x128/67150.png inflating: /content/faces/thumbnails128x128/67151.png inflating: /content/faces/thumbnails128x128/67152.png inflating: /content/faces/thumbnails128x128/67153.png inflating: /content/faces/thumbnails128x128/67154.png inflating: /content/faces/thumbnails128x128/67155.png inflating: /content/faces/thumbnails128x128/67156.png inflating: /content/faces/thumbnails128x128/67157.png inflating: /content/faces/thumbnails128x128/67158.png inflating: /content/faces/thumbnails128x128/67159.png inflating: /content/faces/thumbnails128x128/67160.png inflating: /content/faces/thumbnails128x128/67161.png inflating: /content/faces/thumbnails128x128/67162.png inflating: /content/faces/thumbnails128x128/67163.png inflating: /content/faces/thumbnails128x128/67164.png inflating: /content/faces/thumbnails128x128/67165.png inflating: /content/faces/thumbnails128x128/67166.png inflating: /content/faces/thumbnails128x128/67167.png inflating: /content/faces/thumbnails128x128/67168.png inflating: /content/faces/thumbnails128x128/67169.png inflating: /content/faces/thumbnails128x128/67170.png inflating: /content/faces/thumbnails128x128/67171.png inflating: /content/faces/thumbnails128x128/67172.png inflating: /content/faces/thumbnails128x128/67173.png inflating: /content/faces/thumbnails128x128/67174.png inflating: /content/faces/thumbnails128x128/67175.png inflating: /content/faces/thumbnails128x128/67176.png inflating: /content/faces/thumbnails128x128/67177.png inflating: /content/faces/thumbnails128x128/67178.png inflating: /content/faces/thumbnails128x128/67179.png inflating: /content/faces/thumbnails128x128/67180.png inflating: /content/faces/thumbnails128x128/67181.png inflating: /content/faces/thumbnails128x128/67182.png inflating: /content/faces/thumbnails128x128/67183.png inflating: /content/faces/thumbnails128x128/67184.png inflating: /content/faces/thumbnails128x128/67185.png inflating: /content/faces/thumbnails128x128/67186.png inflating: /content/faces/thumbnails128x128/67187.png inflating: /content/faces/thumbnails128x128/67188.png inflating: /content/faces/thumbnails128x128/67189.png inflating: /content/faces/thumbnails128x128/67190.png inflating: /content/faces/thumbnails128x128/67191.png inflating: /content/faces/thumbnails128x128/67192.png inflating: /content/faces/thumbnails128x128/67193.png inflating: /content/faces/thumbnails128x128/67194.png inflating: /content/faces/thumbnails128x128/67195.png inflating: /content/faces/thumbnails128x128/67196.png inflating: /content/faces/thumbnails128x128/67197.png inflating: /content/faces/thumbnails128x128/67198.png inflating: /content/faces/thumbnails128x128/67199.png inflating: /content/faces/thumbnails128x128/67200.png inflating: /content/faces/thumbnails128x128/67201.png inflating: /content/faces/thumbnails128x128/67202.png inflating: /content/faces/thumbnails128x128/67203.png inflating: /content/faces/thumbnails128x128/67204.png inflating: /content/faces/thumbnails128x128/67205.png inflating: /content/faces/thumbnails128x128/67206.png inflating: /content/faces/thumbnails128x128/67207.png inflating: /content/faces/thumbnails128x128/67208.png inflating: /content/faces/thumbnails128x128/67209.png inflating: /content/faces/thumbnails128x128/67210.png inflating: /content/faces/thumbnails128x128/67211.png inflating: /content/faces/thumbnails128x128/67212.png inflating: /content/faces/thumbnails128x128/67213.png inflating: /content/faces/thumbnails128x128/67214.png inflating: /content/faces/thumbnails128x128/67215.png inflating: /content/faces/thumbnails128x128/67216.png inflating: /content/faces/thumbnails128x128/67217.png inflating: /content/faces/thumbnails128x128/67218.png inflating: /content/faces/thumbnails128x128/67219.png inflating: /content/faces/thumbnails128x128/67220.png inflating: /content/faces/thumbnails128x128/67221.png inflating: /content/faces/thumbnails128x128/67222.png inflating: /content/faces/thumbnails128x128/67223.png inflating: /content/faces/thumbnails128x128/67224.png inflating: /content/faces/thumbnails128x128/67225.png inflating: /content/faces/thumbnails128x128/67226.png inflating: /content/faces/thumbnails128x128/67227.png inflating: /content/faces/thumbnails128x128/67228.png inflating: /content/faces/thumbnails128x128/67229.png inflating: /content/faces/thumbnails128x128/67230.png inflating: /content/faces/thumbnails128x128/67231.png inflating: /content/faces/thumbnails128x128/67232.png inflating: /content/faces/thumbnails128x128/67233.png inflating: /content/faces/thumbnails128x128/67234.png inflating: /content/faces/thumbnails128x128/67235.png inflating: /content/faces/thumbnails128x128/67236.png inflating: /content/faces/thumbnails128x128/67237.png inflating: /content/faces/thumbnails128x128/67238.png inflating: /content/faces/thumbnails128x128/67239.png inflating: /content/faces/thumbnails128x128/67240.png inflating: /content/faces/thumbnails128x128/67241.png inflating: /content/faces/thumbnails128x128/67242.png inflating: /content/faces/thumbnails128x128/67243.png inflating: /content/faces/thumbnails128x128/67244.png inflating: /content/faces/thumbnails128x128/67245.png inflating: /content/faces/thumbnails128x128/67246.png inflating: /content/faces/thumbnails128x128/67247.png inflating: /content/faces/thumbnails128x128/67248.png inflating: /content/faces/thumbnails128x128/67249.png inflating: /content/faces/thumbnails128x128/67250.png inflating: /content/faces/thumbnails128x128/67251.png inflating: /content/faces/thumbnails128x128/67252.png inflating: /content/faces/thumbnails128x128/67253.png inflating: /content/faces/thumbnails128x128/67254.png inflating: /content/faces/thumbnails128x128/67255.png inflating: /content/faces/thumbnails128x128/67256.png inflating: /content/faces/thumbnails128x128/67257.png inflating: /content/faces/thumbnails128x128/67258.png inflating: /content/faces/thumbnails128x128/67259.png inflating: /content/faces/thumbnails128x128/67260.png inflating: /content/faces/thumbnails128x128/67261.png inflating: /content/faces/thumbnails128x128/67262.png inflating: /content/faces/thumbnails128x128/67263.png inflating: /content/faces/thumbnails128x128/67264.png inflating: /content/faces/thumbnails128x128/67265.png inflating: /content/faces/thumbnails128x128/67266.png inflating: /content/faces/thumbnails128x128/67267.png inflating: /content/faces/thumbnails128x128/67268.png inflating: /content/faces/thumbnails128x128/67269.png inflating: /content/faces/thumbnails128x128/67270.png inflating: /content/faces/thumbnails128x128/67271.png inflating: /content/faces/thumbnails128x128/67272.png inflating: /content/faces/thumbnails128x128/67273.png inflating: /content/faces/thumbnails128x128/67274.png inflating: /content/faces/thumbnails128x128/67275.png inflating: /content/faces/thumbnails128x128/67276.png inflating: /content/faces/thumbnails128x128/67277.png inflating: /content/faces/thumbnails128x128/67278.png inflating: /content/faces/thumbnails128x128/67279.png inflating: /content/faces/thumbnails128x128/67280.png inflating: /content/faces/thumbnails128x128/67281.png inflating: /content/faces/thumbnails128x128/67282.png inflating: /content/faces/thumbnails128x128/67283.png inflating: /content/faces/thumbnails128x128/67284.png inflating: /content/faces/thumbnails128x128/67285.png inflating: /content/faces/thumbnails128x128/67286.png inflating: /content/faces/thumbnails128x128/67287.png inflating: /content/faces/thumbnails128x128/67288.png inflating: /content/faces/thumbnails128x128/67289.png inflating: /content/faces/thumbnails128x128/67290.png inflating: /content/faces/thumbnails128x128/67291.png inflating: /content/faces/thumbnails128x128/67292.png inflating: /content/faces/thumbnails128x128/67293.png inflating: /content/faces/thumbnails128x128/67294.png inflating: /content/faces/thumbnails128x128/67295.png inflating: /content/faces/thumbnails128x128/67296.png inflating: /content/faces/thumbnails128x128/67297.png inflating: /content/faces/thumbnails128x128/67298.png inflating: /content/faces/thumbnails128x128/67299.png inflating: /content/faces/thumbnails128x128/67300.png inflating: /content/faces/thumbnails128x128/67301.png inflating: /content/faces/thumbnails128x128/67302.png inflating: /content/faces/thumbnails128x128/67303.png inflating: /content/faces/thumbnails128x128/67304.png inflating: /content/faces/thumbnails128x128/67305.png inflating: /content/faces/thumbnails128x128/67306.png inflating: /content/faces/thumbnails128x128/67307.png inflating: /content/faces/thumbnails128x128/67308.png inflating: /content/faces/thumbnails128x128/67309.png inflating: /content/faces/thumbnails128x128/67310.png inflating: /content/faces/thumbnails128x128/67311.png inflating: /content/faces/thumbnails128x128/67312.png inflating: /content/faces/thumbnails128x128/67313.png inflating: /content/faces/thumbnails128x128/67314.png inflating: /content/faces/thumbnails128x128/67315.png inflating: /content/faces/thumbnails128x128/67316.png inflating: /content/faces/thumbnails128x128/67317.png inflating: /content/faces/thumbnails128x128/67318.png inflating: /content/faces/thumbnails128x128/67319.png inflating: /content/faces/thumbnails128x128/67320.png inflating: /content/faces/thumbnails128x128/67321.png inflating: /content/faces/thumbnails128x128/67322.png inflating: /content/faces/thumbnails128x128/67323.png inflating: /content/faces/thumbnails128x128/67324.png inflating: /content/faces/thumbnails128x128/67325.png inflating: /content/faces/thumbnails128x128/67326.png inflating: /content/faces/thumbnails128x128/67327.png inflating: /content/faces/thumbnails128x128/67328.png inflating: /content/faces/thumbnails128x128/67329.png inflating: /content/faces/thumbnails128x128/67330.png inflating: /content/faces/thumbnails128x128/67331.png inflating: /content/faces/thumbnails128x128/67332.png inflating: /content/faces/thumbnails128x128/67333.png inflating: /content/faces/thumbnails128x128/67334.png inflating: /content/faces/thumbnails128x128/67335.png inflating: /content/faces/thumbnails128x128/67336.png inflating: /content/faces/thumbnails128x128/67337.png inflating: /content/faces/thumbnails128x128/67338.png inflating: /content/faces/thumbnails128x128/67339.png inflating: /content/faces/thumbnails128x128/67340.png inflating: /content/faces/thumbnails128x128/67341.png inflating: /content/faces/thumbnails128x128/67342.png inflating: /content/faces/thumbnails128x128/67343.png inflating: /content/faces/thumbnails128x128/67344.png inflating: /content/faces/thumbnails128x128/67345.png inflating: /content/faces/thumbnails128x128/67346.png inflating: /content/faces/thumbnails128x128/67347.png inflating: /content/faces/thumbnails128x128/67348.png inflating: /content/faces/thumbnails128x128/67349.png inflating: /content/faces/thumbnails128x128/67350.png inflating: /content/faces/thumbnails128x128/67351.png inflating: /content/faces/thumbnails128x128/67352.png inflating: /content/faces/thumbnails128x128/67353.png inflating: /content/faces/thumbnails128x128/67354.png inflating: /content/faces/thumbnails128x128/67355.png inflating: /content/faces/thumbnails128x128/67356.png inflating: /content/faces/thumbnails128x128/67357.png inflating: /content/faces/thumbnails128x128/67358.png inflating: /content/faces/thumbnails128x128/67359.png inflating: /content/faces/thumbnails128x128/67360.png inflating: /content/faces/thumbnails128x128/67361.png inflating: /content/faces/thumbnails128x128/67362.png inflating: /content/faces/thumbnails128x128/67363.png inflating: /content/faces/thumbnails128x128/67364.png inflating: /content/faces/thumbnails128x128/67365.png inflating: /content/faces/thumbnails128x128/67366.png inflating: /content/faces/thumbnails128x128/67367.png inflating: /content/faces/thumbnails128x128/67368.png inflating: /content/faces/thumbnails128x128/67369.png inflating: /content/faces/thumbnails128x128/67370.png inflating: /content/faces/thumbnails128x128/67371.png inflating: /content/faces/thumbnails128x128/67372.png inflating: /content/faces/thumbnails128x128/67373.png inflating: /content/faces/thumbnails128x128/67374.png inflating: /content/faces/thumbnails128x128/67375.png inflating: /content/faces/thumbnails128x128/67376.png inflating: /content/faces/thumbnails128x128/67377.png inflating: /content/faces/thumbnails128x128/67378.png inflating: /content/faces/thumbnails128x128/67379.png inflating: /content/faces/thumbnails128x128/67380.png inflating: /content/faces/thumbnails128x128/67381.png inflating: /content/faces/thumbnails128x128/67382.png inflating: /content/faces/thumbnails128x128/67383.png inflating: /content/faces/thumbnails128x128/67384.png inflating: /content/faces/thumbnails128x128/67385.png inflating: /content/faces/thumbnails128x128/67386.png inflating: /content/faces/thumbnails128x128/67387.png inflating: /content/faces/thumbnails128x128/67388.png inflating: /content/faces/thumbnails128x128/67389.png inflating: /content/faces/thumbnails128x128/67390.png inflating: /content/faces/thumbnails128x128/67391.png inflating: /content/faces/thumbnails128x128/67392.png inflating: /content/faces/thumbnails128x128/67393.png inflating: /content/faces/thumbnails128x128/67394.png inflating: /content/faces/thumbnails128x128/67395.png inflating: /content/faces/thumbnails128x128/67396.png inflating: /content/faces/thumbnails128x128/67397.png inflating: /content/faces/thumbnails128x128/67398.png inflating: /content/faces/thumbnails128x128/67399.png inflating: /content/faces/thumbnails128x128/67400.png inflating: /content/faces/thumbnails128x128/67401.png inflating: /content/faces/thumbnails128x128/67402.png inflating: /content/faces/thumbnails128x128/67403.png inflating: /content/faces/thumbnails128x128/67404.png inflating: /content/faces/thumbnails128x128/67405.png inflating: /content/faces/thumbnails128x128/67406.png inflating: /content/faces/thumbnails128x128/67407.png inflating: /content/faces/thumbnails128x128/67408.png inflating: /content/faces/thumbnails128x128/67409.png inflating: /content/faces/thumbnails128x128/67410.png inflating: /content/faces/thumbnails128x128/67411.png inflating: /content/faces/thumbnails128x128/67412.png inflating: /content/faces/thumbnails128x128/67413.png inflating: /content/faces/thumbnails128x128/67414.png inflating: /content/faces/thumbnails128x128/67415.png inflating: /content/faces/thumbnails128x128/67416.png inflating: /content/faces/thumbnails128x128/67417.png inflating: /content/faces/thumbnails128x128/67418.png inflating: /content/faces/thumbnails128x128/67419.png inflating: /content/faces/thumbnails128x128/67420.png inflating: /content/faces/thumbnails128x128/67421.png inflating: /content/faces/thumbnails128x128/67422.png inflating: /content/faces/thumbnails128x128/67423.png inflating: /content/faces/thumbnails128x128/67424.png inflating: /content/faces/thumbnails128x128/67425.png inflating: /content/faces/thumbnails128x128/67426.png inflating: /content/faces/thumbnails128x128/67427.png inflating: /content/faces/thumbnails128x128/67428.png inflating: /content/faces/thumbnails128x128/67429.png inflating: /content/faces/thumbnails128x128/67430.png inflating: /content/faces/thumbnails128x128/67431.png inflating: /content/faces/thumbnails128x128/67432.png inflating: /content/faces/thumbnails128x128/67433.png inflating: /content/faces/thumbnails128x128/67434.png inflating: /content/faces/thumbnails128x128/67435.png inflating: /content/faces/thumbnails128x128/67436.png inflating: /content/faces/thumbnails128x128/67437.png inflating: /content/faces/thumbnails128x128/67438.png inflating: /content/faces/thumbnails128x128/67439.png inflating: /content/faces/thumbnails128x128/67440.png inflating: /content/faces/thumbnails128x128/67441.png inflating: /content/faces/thumbnails128x128/67442.png inflating: /content/faces/thumbnails128x128/67443.png inflating: /content/faces/thumbnails128x128/67444.png inflating: /content/faces/thumbnails128x128/67445.png inflating: /content/faces/thumbnails128x128/67446.png inflating: /content/faces/thumbnails128x128/67447.png inflating: /content/faces/thumbnails128x128/67448.png inflating: /content/faces/thumbnails128x128/67449.png inflating: /content/faces/thumbnails128x128/67450.png inflating: /content/faces/thumbnails128x128/67451.png inflating: /content/faces/thumbnails128x128/67452.png inflating: /content/faces/thumbnails128x128/67453.png inflating: /content/faces/thumbnails128x128/67454.png inflating: /content/faces/thumbnails128x128/67455.png inflating: /content/faces/thumbnails128x128/67456.png inflating: /content/faces/thumbnails128x128/67457.png inflating: /content/faces/thumbnails128x128/67458.png inflating: /content/faces/thumbnails128x128/67459.png inflating: /content/faces/thumbnails128x128/67460.png inflating: /content/faces/thumbnails128x128/67461.png inflating: /content/faces/thumbnails128x128/67462.png inflating: /content/faces/thumbnails128x128/67463.png inflating: /content/faces/thumbnails128x128/67464.png inflating: /content/faces/thumbnails128x128/67465.png inflating: /content/faces/thumbnails128x128/67466.png inflating: /content/faces/thumbnails128x128/67467.png inflating: /content/faces/thumbnails128x128/67468.png inflating: /content/faces/thumbnails128x128/67469.png inflating: /content/faces/thumbnails128x128/67470.png inflating: /content/faces/thumbnails128x128/67471.png inflating: /content/faces/thumbnails128x128/67472.png inflating: /content/faces/thumbnails128x128/67473.png inflating: /content/faces/thumbnails128x128/67474.png inflating: /content/faces/thumbnails128x128/67475.png inflating: /content/faces/thumbnails128x128/67476.png inflating: /content/faces/thumbnails128x128/67477.png inflating: /content/faces/thumbnails128x128/67478.png inflating: /content/faces/thumbnails128x128/67479.png inflating: /content/faces/thumbnails128x128/67480.png inflating: /content/faces/thumbnails128x128/67481.png inflating: /content/faces/thumbnails128x128/67482.png inflating: /content/faces/thumbnails128x128/67483.png inflating: /content/faces/thumbnails128x128/67484.png inflating: /content/faces/thumbnails128x128/67485.png inflating: /content/faces/thumbnails128x128/67486.png inflating: /content/faces/thumbnails128x128/67487.png inflating: /content/faces/thumbnails128x128/67488.png inflating: /content/faces/thumbnails128x128/67489.png inflating: /content/faces/thumbnails128x128/67490.png inflating: /content/faces/thumbnails128x128/67491.png inflating: /content/faces/thumbnails128x128/67492.png inflating: /content/faces/thumbnails128x128/67493.png inflating: /content/faces/thumbnails128x128/67494.png inflating: /content/faces/thumbnails128x128/67495.png inflating: /content/faces/thumbnails128x128/67496.png inflating: /content/faces/thumbnails128x128/67497.png inflating: /content/faces/thumbnails128x128/67498.png inflating: /content/faces/thumbnails128x128/67499.png inflating: /content/faces/thumbnails128x128/67500.png inflating: /content/faces/thumbnails128x128/67501.png inflating: /content/faces/thumbnails128x128/67502.png inflating: /content/faces/thumbnails128x128/67503.png inflating: /content/faces/thumbnails128x128/67504.png inflating: /content/faces/thumbnails128x128/67505.png inflating: /content/faces/thumbnails128x128/67506.png inflating: /content/faces/thumbnails128x128/67507.png inflating: /content/faces/thumbnails128x128/67508.png inflating: /content/faces/thumbnails128x128/67509.png inflating: /content/faces/thumbnails128x128/67510.png inflating: /content/faces/thumbnails128x128/67511.png inflating: /content/faces/thumbnails128x128/67512.png inflating: /content/faces/thumbnails128x128/67513.png inflating: /content/faces/thumbnails128x128/67514.png inflating: /content/faces/thumbnails128x128/67515.png inflating: /content/faces/thumbnails128x128/67516.png inflating: /content/faces/thumbnails128x128/67517.png inflating: /content/faces/thumbnails128x128/67518.png inflating: /content/faces/thumbnails128x128/67519.png inflating: /content/faces/thumbnails128x128/67520.png inflating: /content/faces/thumbnails128x128/67521.png inflating: /content/faces/thumbnails128x128/67522.png inflating: /content/faces/thumbnails128x128/67523.png inflating: /content/faces/thumbnails128x128/67524.png inflating: /content/faces/thumbnails128x128/67525.png inflating: /content/faces/thumbnails128x128/67526.png inflating: /content/faces/thumbnails128x128/67527.png inflating: /content/faces/thumbnails128x128/67528.png inflating: /content/faces/thumbnails128x128/67529.png inflating: /content/faces/thumbnails128x128/67530.png inflating: /content/faces/thumbnails128x128/67531.png inflating: /content/faces/thumbnails128x128/67532.png inflating: /content/faces/thumbnails128x128/67533.png inflating: /content/faces/thumbnails128x128/67534.png inflating: /content/faces/thumbnails128x128/67535.png inflating: /content/faces/thumbnails128x128/67536.png inflating: /content/faces/thumbnails128x128/67537.png inflating: /content/faces/thumbnails128x128/67538.png inflating: /content/faces/thumbnails128x128/67539.png inflating: /content/faces/thumbnails128x128/67540.png inflating: /content/faces/thumbnails128x128/67541.png inflating: /content/faces/thumbnails128x128/67542.png inflating: /content/faces/thumbnails128x128/67543.png inflating: /content/faces/thumbnails128x128/67544.png inflating: /content/faces/thumbnails128x128/67545.png inflating: /content/faces/thumbnails128x128/67546.png inflating: /content/faces/thumbnails128x128/67547.png inflating: /content/faces/thumbnails128x128/67548.png inflating: /content/faces/thumbnails128x128/67549.png inflating: /content/faces/thumbnails128x128/67550.png inflating: /content/faces/thumbnails128x128/67551.png inflating: /content/faces/thumbnails128x128/67552.png inflating: /content/faces/thumbnails128x128/67553.png inflating: /content/faces/thumbnails128x128/67554.png inflating: /content/faces/thumbnails128x128/67555.png inflating: /content/faces/thumbnails128x128/67556.png inflating: /content/faces/thumbnails128x128/67557.png inflating: /content/faces/thumbnails128x128/67558.png inflating: /content/faces/thumbnails128x128/67559.png inflating: /content/faces/thumbnails128x128/67560.png inflating: /content/faces/thumbnails128x128/67561.png inflating: /content/faces/thumbnails128x128/67562.png inflating: /content/faces/thumbnails128x128/67563.png inflating: /content/faces/thumbnails128x128/67564.png inflating: /content/faces/thumbnails128x128/67565.png inflating: /content/faces/thumbnails128x128/67566.png inflating: /content/faces/thumbnails128x128/67567.png inflating: /content/faces/thumbnails128x128/67568.png inflating: /content/faces/thumbnails128x128/67569.png inflating: /content/faces/thumbnails128x128/67570.png inflating: /content/faces/thumbnails128x128/67571.png inflating: /content/faces/thumbnails128x128/67572.png inflating: /content/faces/thumbnails128x128/67573.png inflating: /content/faces/thumbnails128x128/67574.png inflating: /content/faces/thumbnails128x128/67575.png inflating: /content/faces/thumbnails128x128/67576.png inflating: /content/faces/thumbnails128x128/67577.png inflating: /content/faces/thumbnails128x128/67578.png inflating: /content/faces/thumbnails128x128/67579.png inflating: /content/faces/thumbnails128x128/67580.png inflating: /content/faces/thumbnails128x128/67581.png inflating: /content/faces/thumbnails128x128/67582.png inflating: /content/faces/thumbnails128x128/67583.png inflating: /content/faces/thumbnails128x128/67584.png inflating: /content/faces/thumbnails128x128/67585.png inflating: /content/faces/thumbnails128x128/67586.png inflating: /content/faces/thumbnails128x128/67587.png inflating: /content/faces/thumbnails128x128/67588.png inflating: /content/faces/thumbnails128x128/67589.png inflating: /content/faces/thumbnails128x128/67590.png inflating: /content/faces/thumbnails128x128/67591.png inflating: /content/faces/thumbnails128x128/67592.png inflating: /content/faces/thumbnails128x128/67593.png inflating: /content/faces/thumbnails128x128/67594.png inflating: /content/faces/thumbnails128x128/67595.png inflating: /content/faces/thumbnails128x128/67596.png inflating: /content/faces/thumbnails128x128/67597.png inflating: /content/faces/thumbnails128x128/67598.png inflating: /content/faces/thumbnails128x128/67599.png inflating: /content/faces/thumbnails128x128/67600.png inflating: /content/faces/thumbnails128x128/67601.png inflating: /content/faces/thumbnails128x128/67602.png inflating: /content/faces/thumbnails128x128/67603.png inflating: /content/faces/thumbnails128x128/67604.png inflating: /content/faces/thumbnails128x128/67605.png inflating: /content/faces/thumbnails128x128/67606.png inflating: /content/faces/thumbnails128x128/67607.png inflating: /content/faces/thumbnails128x128/67608.png inflating: /content/faces/thumbnails128x128/67609.png inflating: /content/faces/thumbnails128x128/67610.png inflating: /content/faces/thumbnails128x128/67611.png inflating: /content/faces/thumbnails128x128/67612.png inflating: /content/faces/thumbnails128x128/67613.png inflating: /content/faces/thumbnails128x128/67614.png inflating: /content/faces/thumbnails128x128/67615.png inflating: /content/faces/thumbnails128x128/67616.png inflating: /content/faces/thumbnails128x128/67617.png inflating: /content/faces/thumbnails128x128/67618.png inflating: /content/faces/thumbnails128x128/67619.png inflating: /content/faces/thumbnails128x128/67620.png inflating: /content/faces/thumbnails128x128/67621.png inflating: /content/faces/thumbnails128x128/67622.png inflating: /content/faces/thumbnails128x128/67623.png inflating: /content/faces/thumbnails128x128/67624.png inflating: /content/faces/thumbnails128x128/67625.png inflating: /content/faces/thumbnails128x128/67626.png inflating: /content/faces/thumbnails128x128/67627.png inflating: /content/faces/thumbnails128x128/67628.png inflating: /content/faces/thumbnails128x128/67629.png inflating: /content/faces/thumbnails128x128/67630.png inflating: /content/faces/thumbnails128x128/67631.png inflating: /content/faces/thumbnails128x128/67632.png inflating: /content/faces/thumbnails128x128/67633.png inflating: /content/faces/thumbnails128x128/67634.png inflating: /content/faces/thumbnails128x128/67635.png inflating: /content/faces/thumbnails128x128/67636.png inflating: /content/faces/thumbnails128x128/67637.png inflating: /content/faces/thumbnails128x128/67638.png inflating: /content/faces/thumbnails128x128/67639.png inflating: /content/faces/thumbnails128x128/67640.png inflating: /content/faces/thumbnails128x128/67641.png inflating: /content/faces/thumbnails128x128/67642.png inflating: /content/faces/thumbnails128x128/67643.png inflating: /content/faces/thumbnails128x128/67644.png inflating: /content/faces/thumbnails128x128/67645.png inflating: /content/faces/thumbnails128x128/67646.png inflating: /content/faces/thumbnails128x128/67647.png inflating: /content/faces/thumbnails128x128/67648.png inflating: /content/faces/thumbnails128x128/67649.png inflating: /content/faces/thumbnails128x128/67650.png inflating: /content/faces/thumbnails128x128/67651.png inflating: /content/faces/thumbnails128x128/67652.png inflating: /content/faces/thumbnails128x128/67653.png inflating: /content/faces/thumbnails128x128/67654.png inflating: /content/faces/thumbnails128x128/67655.png inflating: /content/faces/thumbnails128x128/67656.png inflating: /content/faces/thumbnails128x128/67657.png inflating: /content/faces/thumbnails128x128/67658.png inflating: /content/faces/thumbnails128x128/67659.png inflating: /content/faces/thumbnails128x128/67660.png inflating: /content/faces/thumbnails128x128/67661.png inflating: /content/faces/thumbnails128x128/67662.png inflating: /content/faces/thumbnails128x128/67663.png inflating: /content/faces/thumbnails128x128/67664.png inflating: /content/faces/thumbnails128x128/67665.png inflating: /content/faces/thumbnails128x128/67666.png inflating: /content/faces/thumbnails128x128/67667.png inflating: /content/faces/thumbnails128x128/67668.png inflating: /content/faces/thumbnails128x128/67669.png inflating: /content/faces/thumbnails128x128/67670.png inflating: /content/faces/thumbnails128x128/67671.png inflating: /content/faces/thumbnails128x128/67672.png inflating: /content/faces/thumbnails128x128/67673.png inflating: /content/faces/thumbnails128x128/67674.png inflating: /content/faces/thumbnails128x128/67675.png inflating: /content/faces/thumbnails128x128/67676.png inflating: /content/faces/thumbnails128x128/67677.png inflating: /content/faces/thumbnails128x128/67678.png inflating: /content/faces/thumbnails128x128/67679.png inflating: /content/faces/thumbnails128x128/67680.png inflating: /content/faces/thumbnails128x128/67681.png inflating: /content/faces/thumbnails128x128/67682.png inflating: /content/faces/thumbnails128x128/67683.png inflating: /content/faces/thumbnails128x128/67684.png inflating: /content/faces/thumbnails128x128/67685.png inflating: /content/faces/thumbnails128x128/67686.png inflating: /content/faces/thumbnails128x128/67687.png inflating: /content/faces/thumbnails128x128/67688.png inflating: /content/faces/thumbnails128x128/67689.png inflating: /content/faces/thumbnails128x128/67690.png inflating: /content/faces/thumbnails128x128/67691.png inflating: /content/faces/thumbnails128x128/67692.png inflating: /content/faces/thumbnails128x128/67693.png inflating: /content/faces/thumbnails128x128/67694.png inflating: /content/faces/thumbnails128x128/67695.png inflating: /content/faces/thumbnails128x128/67696.png inflating: /content/faces/thumbnails128x128/67697.png inflating: /content/faces/thumbnails128x128/67698.png inflating: /content/faces/thumbnails128x128/67699.png inflating: /content/faces/thumbnails128x128/67700.png inflating: /content/faces/thumbnails128x128/67701.png inflating: /content/faces/thumbnails128x128/67702.png inflating: /content/faces/thumbnails128x128/67703.png inflating: /content/faces/thumbnails128x128/67704.png inflating: /content/faces/thumbnails128x128/67705.png inflating: /content/faces/thumbnails128x128/67706.png inflating: /content/faces/thumbnails128x128/67707.png inflating: /content/faces/thumbnails128x128/67708.png inflating: /content/faces/thumbnails128x128/67709.png inflating: /content/faces/thumbnails128x128/67710.png inflating: /content/faces/thumbnails128x128/67711.png inflating: /content/faces/thumbnails128x128/67712.png inflating: /content/faces/thumbnails128x128/67713.png inflating: /content/faces/thumbnails128x128/67714.png inflating: /content/faces/thumbnails128x128/67715.png inflating: /content/faces/thumbnails128x128/67716.png inflating: /content/faces/thumbnails128x128/67717.png inflating: /content/faces/thumbnails128x128/67718.png inflating: /content/faces/thumbnails128x128/67719.png inflating: /content/faces/thumbnails128x128/67720.png inflating: /content/faces/thumbnails128x128/67721.png inflating: /content/faces/thumbnails128x128/67722.png inflating: /content/faces/thumbnails128x128/67723.png inflating: /content/faces/thumbnails128x128/67724.png inflating: /content/faces/thumbnails128x128/67725.png inflating: /content/faces/thumbnails128x128/67726.png inflating: /content/faces/thumbnails128x128/67727.png inflating: /content/faces/thumbnails128x128/67728.png inflating: /content/faces/thumbnails128x128/67729.png inflating: /content/faces/thumbnails128x128/67730.png inflating: /content/faces/thumbnails128x128/67731.png inflating: /content/faces/thumbnails128x128/67732.png inflating: /content/faces/thumbnails128x128/67733.png inflating: /content/faces/thumbnails128x128/67734.png inflating: /content/faces/thumbnails128x128/67735.png inflating: /content/faces/thumbnails128x128/67736.png inflating: /content/faces/thumbnails128x128/67737.png inflating: /content/faces/thumbnails128x128/67738.png inflating: /content/faces/thumbnails128x128/67739.png inflating: /content/faces/thumbnails128x128/67740.png inflating: /content/faces/thumbnails128x128/67741.png inflating: /content/faces/thumbnails128x128/67742.png inflating: /content/faces/thumbnails128x128/67743.png inflating: /content/faces/thumbnails128x128/67744.png inflating: /content/faces/thumbnails128x128/67745.png inflating: /content/faces/thumbnails128x128/67746.png inflating: /content/faces/thumbnails128x128/67747.png inflating: /content/faces/thumbnails128x128/67748.png inflating: /content/faces/thumbnails128x128/67749.png inflating: /content/faces/thumbnails128x128/67750.png inflating: /content/faces/thumbnails128x128/67751.png inflating: /content/faces/thumbnails128x128/67752.png inflating: /content/faces/thumbnails128x128/67753.png inflating: /content/faces/thumbnails128x128/67754.png inflating: /content/faces/thumbnails128x128/67755.png inflating: /content/faces/thumbnails128x128/67756.png inflating: /content/faces/thumbnails128x128/67757.png inflating: /content/faces/thumbnails128x128/67758.png inflating: /content/faces/thumbnails128x128/67759.png inflating: /content/faces/thumbnails128x128/67760.png inflating: /content/faces/thumbnails128x128/67761.png inflating: /content/faces/thumbnails128x128/67762.png inflating: /content/faces/thumbnails128x128/67763.png inflating: /content/faces/thumbnails128x128/67764.png inflating: /content/faces/thumbnails128x128/67765.png inflating: /content/faces/thumbnails128x128/67766.png inflating: /content/faces/thumbnails128x128/67767.png inflating: /content/faces/thumbnails128x128/67768.png inflating: /content/faces/thumbnails128x128/67769.png inflating: /content/faces/thumbnails128x128/67770.png inflating: /content/faces/thumbnails128x128/67771.png inflating: /content/faces/thumbnails128x128/67772.png inflating: /content/faces/thumbnails128x128/67773.png inflating: /content/faces/thumbnails128x128/67774.png inflating: /content/faces/thumbnails128x128/67775.png inflating: /content/faces/thumbnails128x128/67776.png inflating: /content/faces/thumbnails128x128/67777.png inflating: /content/faces/thumbnails128x128/67778.png inflating: /content/faces/thumbnails128x128/67779.png inflating: /content/faces/thumbnails128x128/67780.png inflating: /content/faces/thumbnails128x128/67781.png inflating: /content/faces/thumbnails128x128/67782.png inflating: /content/faces/thumbnails128x128/67783.png inflating: /content/faces/thumbnails128x128/67784.png inflating: /content/faces/thumbnails128x128/67785.png inflating: /content/faces/thumbnails128x128/67786.png inflating: /content/faces/thumbnails128x128/67787.png inflating: /content/faces/thumbnails128x128/67788.png inflating: /content/faces/thumbnails128x128/67789.png inflating: /content/faces/thumbnails128x128/67790.png inflating: /content/faces/thumbnails128x128/67791.png inflating: /content/faces/thumbnails128x128/67792.png inflating: /content/faces/thumbnails128x128/67793.png inflating: /content/faces/thumbnails128x128/67794.png inflating: /content/faces/thumbnails128x128/67795.png inflating: /content/faces/thumbnails128x128/67796.png inflating: /content/faces/thumbnails128x128/67797.png inflating: /content/faces/thumbnails128x128/67798.png inflating: /content/faces/thumbnails128x128/67799.png inflating: /content/faces/thumbnails128x128/67800.png inflating: /content/faces/thumbnails128x128/67801.png inflating: /content/faces/thumbnails128x128/67802.png inflating: /content/faces/thumbnails128x128/67803.png inflating: /content/faces/thumbnails128x128/67804.png inflating: /content/faces/thumbnails128x128/67805.png inflating: /content/faces/thumbnails128x128/67806.png inflating: /content/faces/thumbnails128x128/67807.png inflating: /content/faces/thumbnails128x128/67808.png inflating: /content/faces/thumbnails128x128/67809.png inflating: /content/faces/thumbnails128x128/67810.png inflating: /content/faces/thumbnails128x128/67811.png inflating: /content/faces/thumbnails128x128/67812.png inflating: /content/faces/thumbnails128x128/67813.png inflating: /content/faces/thumbnails128x128/67814.png inflating: /content/faces/thumbnails128x128/67815.png inflating: /content/faces/thumbnails128x128/67816.png inflating: /content/faces/thumbnails128x128/67817.png inflating: /content/faces/thumbnails128x128/67818.png inflating: /content/faces/thumbnails128x128/67819.png inflating: /content/faces/thumbnails128x128/67820.png inflating: /content/faces/thumbnails128x128/67821.png inflating: /content/faces/thumbnails128x128/67822.png inflating: /content/faces/thumbnails128x128/67823.png inflating: /content/faces/thumbnails128x128/67824.png inflating: /content/faces/thumbnails128x128/67825.png inflating: /content/faces/thumbnails128x128/67826.png inflating: /content/faces/thumbnails128x128/67827.png inflating: /content/faces/thumbnails128x128/67828.png inflating: /content/faces/thumbnails128x128/67829.png inflating: /content/faces/thumbnails128x128/67830.png inflating: /content/faces/thumbnails128x128/67831.png inflating: /content/faces/thumbnails128x128/67832.png inflating: /content/faces/thumbnails128x128/67833.png inflating: /content/faces/thumbnails128x128/67834.png inflating: /content/faces/thumbnails128x128/67835.png inflating: /content/faces/thumbnails128x128/67836.png inflating: /content/faces/thumbnails128x128/67837.png inflating: /content/faces/thumbnails128x128/67838.png inflating: /content/faces/thumbnails128x128/67839.png inflating: /content/faces/thumbnails128x128/67840.png inflating: /content/faces/thumbnails128x128/67841.png inflating: /content/faces/thumbnails128x128/67842.png inflating: /content/faces/thumbnails128x128/67843.png inflating: /content/faces/thumbnails128x128/67844.png inflating: /content/faces/thumbnails128x128/67845.png inflating: /content/faces/thumbnails128x128/67846.png inflating: /content/faces/thumbnails128x128/67847.png inflating: /content/faces/thumbnails128x128/67848.png inflating: /content/faces/thumbnails128x128/67849.png inflating: /content/faces/thumbnails128x128/67850.png inflating: /content/faces/thumbnails128x128/67851.png inflating: /content/faces/thumbnails128x128/67852.png inflating: /content/faces/thumbnails128x128/67853.png inflating: /content/faces/thumbnails128x128/67854.png inflating: /content/faces/thumbnails128x128/67855.png inflating: /content/faces/thumbnails128x128/67856.png inflating: /content/faces/thumbnails128x128/67857.png inflating: /content/faces/thumbnails128x128/67858.png inflating: /content/faces/thumbnails128x128/67859.png inflating: /content/faces/thumbnails128x128/67860.png inflating: /content/faces/thumbnails128x128/67861.png inflating: /content/faces/thumbnails128x128/67862.png inflating: /content/faces/thumbnails128x128/67863.png inflating: /content/faces/thumbnails128x128/67864.png inflating: /content/faces/thumbnails128x128/67865.png inflating: /content/faces/thumbnails128x128/67866.png inflating: /content/faces/thumbnails128x128/67867.png inflating: /content/faces/thumbnails128x128/67868.png inflating: /content/faces/thumbnails128x128/67869.png inflating: /content/faces/thumbnails128x128/67870.png inflating: /content/faces/thumbnails128x128/67871.png inflating: /content/faces/thumbnails128x128/67872.png inflating: /content/faces/thumbnails128x128/67873.png inflating: /content/faces/thumbnails128x128/67874.png inflating: /content/faces/thumbnails128x128/67875.png inflating: /content/faces/thumbnails128x128/67876.png inflating: /content/faces/thumbnails128x128/67877.png inflating: /content/faces/thumbnails128x128/67878.png inflating: /content/faces/thumbnails128x128/67879.png inflating: /content/faces/thumbnails128x128/67880.png inflating: /content/faces/thumbnails128x128/67881.png inflating: /content/faces/thumbnails128x128/67882.png inflating: /content/faces/thumbnails128x128/67883.png inflating: /content/faces/thumbnails128x128/67884.png inflating: /content/faces/thumbnails128x128/67885.png inflating: /content/faces/thumbnails128x128/67886.png inflating: /content/faces/thumbnails128x128/67887.png inflating: /content/faces/thumbnails128x128/67888.png inflating: /content/faces/thumbnails128x128/67889.png inflating: /content/faces/thumbnails128x128/67890.png inflating: /content/faces/thumbnails128x128/67891.png inflating: /content/faces/thumbnails128x128/67892.png inflating: /content/faces/thumbnails128x128/67893.png inflating: /content/faces/thumbnails128x128/67894.png inflating: /content/faces/thumbnails128x128/67895.png inflating: /content/faces/thumbnails128x128/67896.png inflating: /content/faces/thumbnails128x128/67897.png inflating: /content/faces/thumbnails128x128/67898.png inflating: /content/faces/thumbnails128x128/67899.png inflating: /content/faces/thumbnails128x128/67900.png inflating: /content/faces/thumbnails128x128/67901.png inflating: /content/faces/thumbnails128x128/67902.png inflating: /content/faces/thumbnails128x128/67903.png inflating: /content/faces/thumbnails128x128/67904.png inflating: /content/faces/thumbnails128x128/67905.png inflating: /content/faces/thumbnails128x128/67906.png inflating: /content/faces/thumbnails128x128/67907.png inflating: /content/faces/thumbnails128x128/67908.png inflating: /content/faces/thumbnails128x128/67909.png inflating: /content/faces/thumbnails128x128/67910.png inflating: /content/faces/thumbnails128x128/67911.png inflating: /content/faces/thumbnails128x128/67912.png inflating: /content/faces/thumbnails128x128/67913.png inflating: /content/faces/thumbnails128x128/67914.png inflating: /content/faces/thumbnails128x128/67915.png inflating: /content/faces/thumbnails128x128/67916.png inflating: /content/faces/thumbnails128x128/67917.png inflating: /content/faces/thumbnails128x128/67918.png inflating: /content/faces/thumbnails128x128/67919.png inflating: /content/faces/thumbnails128x128/67920.png inflating: /content/faces/thumbnails128x128/67921.png inflating: /content/faces/thumbnails128x128/67922.png inflating: /content/faces/thumbnails128x128/67923.png inflating: /content/faces/thumbnails128x128/67924.png inflating: /content/faces/thumbnails128x128/67925.png inflating: /content/faces/thumbnails128x128/67926.png inflating: /content/faces/thumbnails128x128/67927.png inflating: /content/faces/thumbnails128x128/67928.png inflating: /content/faces/thumbnails128x128/67929.png inflating: /content/faces/thumbnails128x128/67930.png inflating: /content/faces/thumbnails128x128/67931.png inflating: /content/faces/thumbnails128x128/67932.png inflating: /content/faces/thumbnails128x128/67933.png inflating: /content/faces/thumbnails128x128/67934.png inflating: /content/faces/thumbnails128x128/67935.png inflating: /content/faces/thumbnails128x128/67936.png inflating: /content/faces/thumbnails128x128/67937.png inflating: /content/faces/thumbnails128x128/67938.png inflating: /content/faces/thumbnails128x128/67939.png inflating: /content/faces/thumbnails128x128/67940.png inflating: /content/faces/thumbnails128x128/67941.png inflating: /content/faces/thumbnails128x128/67942.png inflating: /content/faces/thumbnails128x128/67943.png inflating: /content/faces/thumbnails128x128/67944.png inflating: /content/faces/thumbnails128x128/67945.png inflating: /content/faces/thumbnails128x128/67946.png inflating: /content/faces/thumbnails128x128/67947.png inflating: /content/faces/thumbnails128x128/67948.png inflating: /content/faces/thumbnails128x128/67949.png inflating: /content/faces/thumbnails128x128/67950.png inflating: /content/faces/thumbnails128x128/67951.png inflating: /content/faces/thumbnails128x128/67952.png inflating: /content/faces/thumbnails128x128/67953.png inflating: /content/faces/thumbnails128x128/67954.png inflating: /content/faces/thumbnails128x128/67955.png inflating: /content/faces/thumbnails128x128/67956.png inflating: /content/faces/thumbnails128x128/67957.png inflating: /content/faces/thumbnails128x128/67958.png inflating: /content/faces/thumbnails128x128/67959.png inflating: /content/faces/thumbnails128x128/67960.png inflating: /content/faces/thumbnails128x128/67961.png inflating: /content/faces/thumbnails128x128/67962.png inflating: /content/faces/thumbnails128x128/67963.png inflating: /content/faces/thumbnails128x128/67964.png inflating: /content/faces/thumbnails128x128/67965.png inflating: /content/faces/thumbnails128x128/67966.png inflating: /content/faces/thumbnails128x128/67967.png inflating: /content/faces/thumbnails128x128/67968.png inflating: /content/faces/thumbnails128x128/67969.png inflating: /content/faces/thumbnails128x128/67970.png inflating: /content/faces/thumbnails128x128/67971.png inflating: /content/faces/thumbnails128x128/67972.png inflating: /content/faces/thumbnails128x128/67973.png inflating: /content/faces/thumbnails128x128/67974.png inflating: /content/faces/thumbnails128x128/67975.png inflating: /content/faces/thumbnails128x128/67976.png inflating: /content/faces/thumbnails128x128/67977.png inflating: /content/faces/thumbnails128x128/67978.png inflating: /content/faces/thumbnails128x128/67979.png inflating: /content/faces/thumbnails128x128/67980.png inflating: /content/faces/thumbnails128x128/67981.png inflating: /content/faces/thumbnails128x128/67982.png inflating: /content/faces/thumbnails128x128/67983.png inflating: /content/faces/thumbnails128x128/67984.png inflating: /content/faces/thumbnails128x128/67985.png inflating: /content/faces/thumbnails128x128/67986.png inflating: /content/faces/thumbnails128x128/67987.png inflating: /content/faces/thumbnails128x128/67988.png inflating: /content/faces/thumbnails128x128/67989.png inflating: /content/faces/thumbnails128x128/67990.png inflating: /content/faces/thumbnails128x128/67991.png inflating: /content/faces/thumbnails128x128/67992.png inflating: /content/faces/thumbnails128x128/67993.png inflating: /content/faces/thumbnails128x128/67994.png inflating: /content/faces/thumbnails128x128/67995.png inflating: /content/faces/thumbnails128x128/67996.png inflating: /content/faces/thumbnails128x128/67997.png inflating: /content/faces/thumbnails128x128/67998.png inflating: /content/faces/thumbnails128x128/67999.png inflating: /content/faces/thumbnails128x128/68000.png inflating: /content/faces/thumbnails128x128/68001.png inflating: /content/faces/thumbnails128x128/68002.png inflating: /content/faces/thumbnails128x128/68003.png inflating: /content/faces/thumbnails128x128/68004.png inflating: /content/faces/thumbnails128x128/68005.png inflating: /content/faces/thumbnails128x128/68006.png inflating: /content/faces/thumbnails128x128/68007.png inflating: /content/faces/thumbnails128x128/68008.png inflating: /content/faces/thumbnails128x128/68009.png inflating: /content/faces/thumbnails128x128/68010.png inflating: /content/faces/thumbnails128x128/68011.png inflating: /content/faces/thumbnails128x128/68012.png inflating: /content/faces/thumbnails128x128/68013.png inflating: /content/faces/thumbnails128x128/68014.png inflating: /content/faces/thumbnails128x128/68015.png inflating: /content/faces/thumbnails128x128/68016.png inflating: /content/faces/thumbnails128x128/68017.png inflating: /content/faces/thumbnails128x128/68018.png inflating: /content/faces/thumbnails128x128/68019.png inflating: /content/faces/thumbnails128x128/68020.png inflating: /content/faces/thumbnails128x128/68021.png inflating: /content/faces/thumbnails128x128/68022.png inflating: /content/faces/thumbnails128x128/68023.png inflating: /content/faces/thumbnails128x128/68024.png inflating: /content/faces/thumbnails128x128/68025.png inflating: /content/faces/thumbnails128x128/68026.png inflating: /content/faces/thumbnails128x128/68027.png inflating: /content/faces/thumbnails128x128/68028.png inflating: /content/faces/thumbnails128x128/68029.png inflating: /content/faces/thumbnails128x128/68030.png inflating: /content/faces/thumbnails128x128/68031.png inflating: /content/faces/thumbnails128x128/68032.png inflating: /content/faces/thumbnails128x128/68033.png inflating: /content/faces/thumbnails128x128/68034.png inflating: /content/faces/thumbnails128x128/68035.png inflating: /content/faces/thumbnails128x128/68036.png inflating: /content/faces/thumbnails128x128/68037.png inflating: /content/faces/thumbnails128x128/68038.png inflating: /content/faces/thumbnails128x128/68039.png inflating: /content/faces/thumbnails128x128/68040.png inflating: /content/faces/thumbnails128x128/68041.png inflating: /content/faces/thumbnails128x128/68042.png inflating: /content/faces/thumbnails128x128/68043.png inflating: /content/faces/thumbnails128x128/68044.png inflating: /content/faces/thumbnails128x128/68045.png inflating: /content/faces/thumbnails128x128/68046.png inflating: /content/faces/thumbnails128x128/68047.png inflating: /content/faces/thumbnails128x128/68048.png inflating: /content/faces/thumbnails128x128/68049.png inflating: /content/faces/thumbnails128x128/68050.png inflating: /content/faces/thumbnails128x128/68051.png inflating: /content/faces/thumbnails128x128/68052.png inflating: /content/faces/thumbnails128x128/68053.png inflating: /content/faces/thumbnails128x128/68054.png inflating: /content/faces/thumbnails128x128/68055.png inflating: /content/faces/thumbnails128x128/68056.png inflating: /content/faces/thumbnails128x128/68057.png inflating: /content/faces/thumbnails128x128/68058.png inflating: /content/faces/thumbnails128x128/68059.png inflating: /content/faces/thumbnails128x128/68060.png inflating: /content/faces/thumbnails128x128/68061.png inflating: /content/faces/thumbnails128x128/68062.png inflating: /content/faces/thumbnails128x128/68063.png inflating: /content/faces/thumbnails128x128/68064.png inflating: /content/faces/thumbnails128x128/68065.png inflating: /content/faces/thumbnails128x128/68066.png inflating: /content/faces/thumbnails128x128/68067.png inflating: /content/faces/thumbnails128x128/68068.png inflating: /content/faces/thumbnails128x128/68069.png inflating: /content/faces/thumbnails128x128/68070.png inflating: /content/faces/thumbnails128x128/68071.png inflating: /content/faces/thumbnails128x128/68072.png inflating: /content/faces/thumbnails128x128/68073.png inflating: /content/faces/thumbnails128x128/68074.png inflating: /content/faces/thumbnails128x128/68075.png inflating: /content/faces/thumbnails128x128/68076.png inflating: /content/faces/thumbnails128x128/68077.png inflating: /content/faces/thumbnails128x128/68078.png inflating: /content/faces/thumbnails128x128/68079.png inflating: /content/faces/thumbnails128x128/68080.png inflating: /content/faces/thumbnails128x128/68081.png inflating: /content/faces/thumbnails128x128/68082.png inflating: /content/faces/thumbnails128x128/68083.png inflating: /content/faces/thumbnails128x128/68084.png inflating: /content/faces/thumbnails128x128/68085.png inflating: /content/faces/thumbnails128x128/68086.png inflating: /content/faces/thumbnails128x128/68087.png inflating: /content/faces/thumbnails128x128/68088.png inflating: /content/faces/thumbnails128x128/68089.png inflating: /content/faces/thumbnails128x128/68090.png inflating: /content/faces/thumbnails128x128/68091.png inflating: /content/faces/thumbnails128x128/68092.png inflating: /content/faces/thumbnails128x128/68093.png inflating: /content/faces/thumbnails128x128/68094.png inflating: /content/faces/thumbnails128x128/68095.png inflating: /content/faces/thumbnails128x128/68096.png inflating: /content/faces/thumbnails128x128/68097.png inflating: /content/faces/thumbnails128x128/68098.png inflating: /content/faces/thumbnails128x128/68099.png inflating: /content/faces/thumbnails128x128/68100.png inflating: /content/faces/thumbnails128x128/68101.png inflating: /content/faces/thumbnails128x128/68102.png inflating: /content/faces/thumbnails128x128/68103.png inflating: /content/faces/thumbnails128x128/68104.png inflating: /content/faces/thumbnails128x128/68105.png inflating: /content/faces/thumbnails128x128/68106.png inflating: /content/faces/thumbnails128x128/68107.png inflating: /content/faces/thumbnails128x128/68108.png inflating: /content/faces/thumbnails128x128/68109.png inflating: /content/faces/thumbnails128x128/68110.png inflating: /content/faces/thumbnails128x128/68111.png inflating: /content/faces/thumbnails128x128/68112.png inflating: /content/faces/thumbnails128x128/68113.png inflating: /content/faces/thumbnails128x128/68114.png inflating: /content/faces/thumbnails128x128/68115.png inflating: /content/faces/thumbnails128x128/68116.png inflating: /content/faces/thumbnails128x128/68117.png inflating: /content/faces/thumbnails128x128/68118.png inflating: /content/faces/thumbnails128x128/68119.png inflating: /content/faces/thumbnails128x128/68120.png inflating: /content/faces/thumbnails128x128/68121.png inflating: /content/faces/thumbnails128x128/68122.png inflating: /content/faces/thumbnails128x128/68123.png inflating: /content/faces/thumbnails128x128/68124.png inflating: /content/faces/thumbnails128x128/68125.png inflating: /content/faces/thumbnails128x128/68126.png inflating: /content/faces/thumbnails128x128/68127.png inflating: /content/faces/thumbnails128x128/68128.png inflating: /content/faces/thumbnails128x128/68129.png inflating: /content/faces/thumbnails128x128/68130.png inflating: /content/faces/thumbnails128x128/68131.png inflating: /content/faces/thumbnails128x128/68132.png inflating: /content/faces/thumbnails128x128/68133.png inflating: /content/faces/thumbnails128x128/68134.png inflating: /content/faces/thumbnails128x128/68135.png inflating: /content/faces/thumbnails128x128/68136.png inflating: /content/faces/thumbnails128x128/68137.png inflating: /content/faces/thumbnails128x128/68138.png inflating: /content/faces/thumbnails128x128/68139.png inflating: /content/faces/thumbnails128x128/68140.png inflating: /content/faces/thumbnails128x128/68141.png inflating: /content/faces/thumbnails128x128/68142.png inflating: /content/faces/thumbnails128x128/68143.png inflating: /content/faces/thumbnails128x128/68144.png inflating: /content/faces/thumbnails128x128/68145.png inflating: /content/faces/thumbnails128x128/68146.png inflating: /content/faces/thumbnails128x128/68147.png inflating: /content/faces/thumbnails128x128/68148.png inflating: /content/faces/thumbnails128x128/68149.png inflating: /content/faces/thumbnails128x128/68150.png inflating: /content/faces/thumbnails128x128/68151.png inflating: /content/faces/thumbnails128x128/68152.png inflating: /content/faces/thumbnails128x128/68153.png inflating: /content/faces/thumbnails128x128/68154.png inflating: /content/faces/thumbnails128x128/68155.png inflating: /content/faces/thumbnails128x128/68156.png inflating: /content/faces/thumbnails128x128/68157.png inflating: /content/faces/thumbnails128x128/68158.png inflating: /content/faces/thumbnails128x128/68159.png inflating: /content/faces/thumbnails128x128/68160.png inflating: /content/faces/thumbnails128x128/68161.png inflating: /content/faces/thumbnails128x128/68162.png inflating: /content/faces/thumbnails128x128/68163.png inflating: /content/faces/thumbnails128x128/68164.png inflating: /content/faces/thumbnails128x128/68165.png inflating: /content/faces/thumbnails128x128/68166.png inflating: /content/faces/thumbnails128x128/68167.png inflating: /content/faces/thumbnails128x128/68168.png inflating: /content/faces/thumbnails128x128/68169.png inflating: /content/faces/thumbnails128x128/68170.png inflating: /content/faces/thumbnails128x128/68171.png inflating: /content/faces/thumbnails128x128/68172.png inflating: /content/faces/thumbnails128x128/68173.png inflating: /content/faces/thumbnails128x128/68174.png inflating: /content/faces/thumbnails128x128/68175.png inflating: /content/faces/thumbnails128x128/68176.png inflating: /content/faces/thumbnails128x128/68177.png inflating: /content/faces/thumbnails128x128/68178.png inflating: /content/faces/thumbnails128x128/68179.png inflating: /content/faces/thumbnails128x128/68180.png inflating: /content/faces/thumbnails128x128/68181.png inflating: /content/faces/thumbnails128x128/68182.png inflating: /content/faces/thumbnails128x128/68183.png inflating: /content/faces/thumbnails128x128/68184.png inflating: /content/faces/thumbnails128x128/68185.png inflating: /content/faces/thumbnails128x128/68186.png inflating: /content/faces/thumbnails128x128/68187.png inflating: /content/faces/thumbnails128x128/68188.png inflating: /content/faces/thumbnails128x128/68189.png inflating: /content/faces/thumbnails128x128/68190.png inflating: /content/faces/thumbnails128x128/68191.png inflating: /content/faces/thumbnails128x128/68192.png inflating: /content/faces/thumbnails128x128/68193.png inflating: /content/faces/thumbnails128x128/68194.png inflating: /content/faces/thumbnails128x128/68195.png inflating: /content/faces/thumbnails128x128/68196.png inflating: /content/faces/thumbnails128x128/68197.png inflating: /content/faces/thumbnails128x128/68198.png inflating: /content/faces/thumbnails128x128/68199.png inflating: /content/faces/thumbnails128x128/68200.png inflating: /content/faces/thumbnails128x128/68201.png inflating: /content/faces/thumbnails128x128/68202.png inflating: /content/faces/thumbnails128x128/68203.png inflating: /content/faces/thumbnails128x128/68204.png inflating: /content/faces/thumbnails128x128/68205.png inflating: /content/faces/thumbnails128x128/68206.png inflating: /content/faces/thumbnails128x128/68207.png inflating: /content/faces/thumbnails128x128/68208.png inflating: /content/faces/thumbnails128x128/68209.png inflating: /content/faces/thumbnails128x128/68210.png inflating: /content/faces/thumbnails128x128/68211.png inflating: /content/faces/thumbnails128x128/68212.png inflating: /content/faces/thumbnails128x128/68213.png inflating: /content/faces/thumbnails128x128/68214.png inflating: /content/faces/thumbnails128x128/68215.png inflating: /content/faces/thumbnails128x128/68216.png inflating: /content/faces/thumbnails128x128/68217.png inflating: /content/faces/thumbnails128x128/68218.png inflating: /content/faces/thumbnails128x128/68219.png inflating: /content/faces/thumbnails128x128/68220.png inflating: /content/faces/thumbnails128x128/68221.png inflating: /content/faces/thumbnails128x128/68222.png inflating: /content/faces/thumbnails128x128/68223.png inflating: /content/faces/thumbnails128x128/68224.png inflating: /content/faces/thumbnails128x128/68225.png inflating: /content/faces/thumbnails128x128/68226.png inflating: /content/faces/thumbnails128x128/68227.png inflating: /content/faces/thumbnails128x128/68228.png inflating: /content/faces/thumbnails128x128/68229.png inflating: /content/faces/thumbnails128x128/68230.png inflating: /content/faces/thumbnails128x128/68231.png inflating: /content/faces/thumbnails128x128/68232.png inflating: /content/faces/thumbnails128x128/68233.png inflating: /content/faces/thumbnails128x128/68234.png inflating: /content/faces/thumbnails128x128/68235.png inflating: /content/faces/thumbnails128x128/68236.png inflating: /content/faces/thumbnails128x128/68237.png inflating: /content/faces/thumbnails128x128/68238.png inflating: /content/faces/thumbnails128x128/68239.png inflating: /content/faces/thumbnails128x128/68240.png inflating: /content/faces/thumbnails128x128/68241.png inflating: /content/faces/thumbnails128x128/68242.png inflating: /content/faces/thumbnails128x128/68243.png inflating: /content/faces/thumbnails128x128/68244.png inflating: /content/faces/thumbnails128x128/68245.png inflating: /content/faces/thumbnails128x128/68246.png inflating: /content/faces/thumbnails128x128/68247.png inflating: /content/faces/thumbnails128x128/68248.png inflating: /content/faces/thumbnails128x128/68249.png inflating: /content/faces/thumbnails128x128/68250.png inflating: /content/faces/thumbnails128x128/68251.png inflating: /content/faces/thumbnails128x128/68252.png inflating: /content/faces/thumbnails128x128/68253.png inflating: /content/faces/thumbnails128x128/68254.png inflating: /content/faces/thumbnails128x128/68255.png inflating: /content/faces/thumbnails128x128/68256.png inflating: /content/faces/thumbnails128x128/68257.png inflating: /content/faces/thumbnails128x128/68258.png inflating: /content/faces/thumbnails128x128/68259.png inflating: /content/faces/thumbnails128x128/68260.png inflating: /content/faces/thumbnails128x128/68261.png inflating: /content/faces/thumbnails128x128/68262.png inflating: /content/faces/thumbnails128x128/68263.png inflating: /content/faces/thumbnails128x128/68264.png inflating: /content/faces/thumbnails128x128/68265.png inflating: /content/faces/thumbnails128x128/68266.png inflating: /content/faces/thumbnails128x128/68267.png inflating: /content/faces/thumbnails128x128/68268.png inflating: /content/faces/thumbnails128x128/68269.png inflating: /content/faces/thumbnails128x128/68270.png inflating: /content/faces/thumbnails128x128/68271.png inflating: /content/faces/thumbnails128x128/68272.png inflating: /content/faces/thumbnails128x128/68273.png inflating: /content/faces/thumbnails128x128/68274.png inflating: /content/faces/thumbnails128x128/68275.png inflating: /content/faces/thumbnails128x128/68276.png inflating: /content/faces/thumbnails128x128/68277.png inflating: /content/faces/thumbnails128x128/68278.png inflating: /content/faces/thumbnails128x128/68279.png inflating: /content/faces/thumbnails128x128/68280.png inflating: /content/faces/thumbnails128x128/68281.png inflating: /content/faces/thumbnails128x128/68282.png inflating: /content/faces/thumbnails128x128/68283.png inflating: /content/faces/thumbnails128x128/68284.png inflating: /content/faces/thumbnails128x128/68285.png inflating: /content/faces/thumbnails128x128/68286.png inflating: /content/faces/thumbnails128x128/68287.png inflating: /content/faces/thumbnails128x128/68288.png inflating: /content/faces/thumbnails128x128/68289.png inflating: /content/faces/thumbnails128x128/68290.png inflating: /content/faces/thumbnails128x128/68291.png inflating: /content/faces/thumbnails128x128/68292.png inflating: /content/faces/thumbnails128x128/68293.png inflating: /content/faces/thumbnails128x128/68294.png inflating: /content/faces/thumbnails128x128/68295.png inflating: /content/faces/thumbnails128x128/68296.png inflating: /content/faces/thumbnails128x128/68297.png inflating: /content/faces/thumbnails128x128/68298.png inflating: /content/faces/thumbnails128x128/68299.png inflating: /content/faces/thumbnails128x128/68300.png inflating: /content/faces/thumbnails128x128/68301.png inflating: /content/faces/thumbnails128x128/68302.png inflating: /content/faces/thumbnails128x128/68303.png inflating: /content/faces/thumbnails128x128/68304.png inflating: /content/faces/thumbnails128x128/68305.png inflating: /content/faces/thumbnails128x128/68306.png inflating: /content/faces/thumbnails128x128/68307.png inflating: /content/faces/thumbnails128x128/68308.png inflating: /content/faces/thumbnails128x128/68309.png inflating: /content/faces/thumbnails128x128/68310.png inflating: /content/faces/thumbnails128x128/68311.png inflating: /content/faces/thumbnails128x128/68312.png inflating: /content/faces/thumbnails128x128/68313.png inflating: /content/faces/thumbnails128x128/68314.png inflating: /content/faces/thumbnails128x128/68315.png inflating: /content/faces/thumbnails128x128/68316.png inflating: /content/faces/thumbnails128x128/68317.png inflating: /content/faces/thumbnails128x128/68318.png inflating: /content/faces/thumbnails128x128/68319.png inflating: /content/faces/thumbnails128x128/68320.png inflating: /content/faces/thumbnails128x128/68321.png inflating: /content/faces/thumbnails128x128/68322.png inflating: /content/faces/thumbnails128x128/68323.png inflating: /content/faces/thumbnails128x128/68324.png inflating: /content/faces/thumbnails128x128/68325.png inflating: /content/faces/thumbnails128x128/68326.png inflating: /content/faces/thumbnails128x128/68327.png inflating: /content/faces/thumbnails128x128/68328.png inflating: /content/faces/thumbnails128x128/68329.png inflating: /content/faces/thumbnails128x128/68330.png inflating: /content/faces/thumbnails128x128/68331.png inflating: /content/faces/thumbnails128x128/68332.png inflating: /content/faces/thumbnails128x128/68333.png inflating: /content/faces/thumbnails128x128/68334.png inflating: /content/faces/thumbnails128x128/68335.png inflating: /content/faces/thumbnails128x128/68336.png inflating: /content/faces/thumbnails128x128/68337.png inflating: /content/faces/thumbnails128x128/68338.png inflating: /content/faces/thumbnails128x128/68339.png inflating: /content/faces/thumbnails128x128/68340.png inflating: /content/faces/thumbnails128x128/68341.png inflating: /content/faces/thumbnails128x128/68342.png inflating: /content/faces/thumbnails128x128/68343.png inflating: /content/faces/thumbnails128x128/68344.png inflating: /content/faces/thumbnails128x128/68345.png inflating: /content/faces/thumbnails128x128/68346.png inflating: /content/faces/thumbnails128x128/68347.png inflating: /content/faces/thumbnails128x128/68348.png inflating: /content/faces/thumbnails128x128/68349.png inflating: /content/faces/thumbnails128x128/68350.png inflating: /content/faces/thumbnails128x128/68351.png inflating: /content/faces/thumbnails128x128/68352.png inflating: /content/faces/thumbnails128x128/68353.png inflating: /content/faces/thumbnails128x128/68354.png inflating: /content/faces/thumbnails128x128/68355.png inflating: /content/faces/thumbnails128x128/68356.png inflating: /content/faces/thumbnails128x128/68357.png inflating: /content/faces/thumbnails128x128/68358.png inflating: /content/faces/thumbnails128x128/68359.png inflating: /content/faces/thumbnails128x128/68360.png inflating: /content/faces/thumbnails128x128/68361.png inflating: /content/faces/thumbnails128x128/68362.png inflating: /content/faces/thumbnails128x128/68363.png inflating: /content/faces/thumbnails128x128/68364.png inflating: /content/faces/thumbnails128x128/68365.png inflating: /content/faces/thumbnails128x128/68366.png inflating: /content/faces/thumbnails128x128/68367.png inflating: /content/faces/thumbnails128x128/68368.png inflating: /content/faces/thumbnails128x128/68369.png inflating: /content/faces/thumbnails128x128/68370.png inflating: /content/faces/thumbnails128x128/68371.png inflating: /content/faces/thumbnails128x128/68372.png inflating: /content/faces/thumbnails128x128/68373.png inflating: /content/faces/thumbnails128x128/68374.png inflating: /content/faces/thumbnails128x128/68375.png inflating: /content/faces/thumbnails128x128/68376.png inflating: /content/faces/thumbnails128x128/68377.png inflating: /content/faces/thumbnails128x128/68378.png inflating: /content/faces/thumbnails128x128/68379.png inflating: /content/faces/thumbnails128x128/68380.png inflating: /content/faces/thumbnails128x128/68381.png inflating: /content/faces/thumbnails128x128/68382.png inflating: /content/faces/thumbnails128x128/68383.png inflating: /content/faces/thumbnails128x128/68384.png inflating: /content/faces/thumbnails128x128/68385.png inflating: /content/faces/thumbnails128x128/68386.png inflating: /content/faces/thumbnails128x128/68387.png inflating: /content/faces/thumbnails128x128/68388.png inflating: /content/faces/thumbnails128x128/68389.png inflating: /content/faces/thumbnails128x128/68390.png inflating: /content/faces/thumbnails128x128/68391.png inflating: /content/faces/thumbnails128x128/68392.png inflating: /content/faces/thumbnails128x128/68393.png inflating: /content/faces/thumbnails128x128/68394.png inflating: /content/faces/thumbnails128x128/68395.png inflating: /content/faces/thumbnails128x128/68396.png inflating: /content/faces/thumbnails128x128/68397.png inflating: /content/faces/thumbnails128x128/68398.png inflating: /content/faces/thumbnails128x128/68399.png inflating: /content/faces/thumbnails128x128/68400.png inflating: /content/faces/thumbnails128x128/68401.png inflating: /content/faces/thumbnails128x128/68402.png inflating: /content/faces/thumbnails128x128/68403.png inflating: /content/faces/thumbnails128x128/68404.png inflating: /content/faces/thumbnails128x128/68405.png inflating: /content/faces/thumbnails128x128/68406.png inflating: /content/faces/thumbnails128x128/68407.png inflating: /content/faces/thumbnails128x128/68408.png inflating: /content/faces/thumbnails128x128/68409.png inflating: /content/faces/thumbnails128x128/68410.png inflating: /content/faces/thumbnails128x128/68411.png inflating: /content/faces/thumbnails128x128/68412.png inflating: /content/faces/thumbnails128x128/68413.png inflating: /content/faces/thumbnails128x128/68414.png inflating: /content/faces/thumbnails128x128/68415.png inflating: /content/faces/thumbnails128x128/68416.png inflating: /content/faces/thumbnails128x128/68417.png inflating: /content/faces/thumbnails128x128/68418.png inflating: /content/faces/thumbnails128x128/68419.png inflating: /content/faces/thumbnails128x128/68420.png inflating: /content/faces/thumbnails128x128/68421.png inflating: /content/faces/thumbnails128x128/68422.png inflating: /content/faces/thumbnails128x128/68423.png inflating: /content/faces/thumbnails128x128/68424.png inflating: /content/faces/thumbnails128x128/68425.png inflating: /content/faces/thumbnails128x128/68426.png inflating: /content/faces/thumbnails128x128/68427.png inflating: /content/faces/thumbnails128x128/68428.png inflating: /content/faces/thumbnails128x128/68429.png inflating: /content/faces/thumbnails128x128/68430.png inflating: /content/faces/thumbnails128x128/68431.png inflating: /content/faces/thumbnails128x128/68432.png inflating: /content/faces/thumbnails128x128/68433.png inflating: /content/faces/thumbnails128x128/68434.png inflating: /content/faces/thumbnails128x128/68435.png inflating: /content/faces/thumbnails128x128/68436.png inflating: /content/faces/thumbnails128x128/68437.png inflating: /content/faces/thumbnails128x128/68438.png inflating: /content/faces/thumbnails128x128/68439.png inflating: /content/faces/thumbnails128x128/68440.png inflating: /content/faces/thumbnails128x128/68441.png inflating: /content/faces/thumbnails128x128/68442.png inflating: /content/faces/thumbnails128x128/68443.png inflating: /content/faces/thumbnails128x128/68444.png inflating: /content/faces/thumbnails128x128/68445.png inflating: /content/faces/thumbnails128x128/68446.png inflating: /content/faces/thumbnails128x128/68447.png inflating: /content/faces/thumbnails128x128/68448.png inflating: /content/faces/thumbnails128x128/68449.png inflating: /content/faces/thumbnails128x128/68450.png inflating: /content/faces/thumbnails128x128/68451.png inflating: /content/faces/thumbnails128x128/68452.png inflating: /content/faces/thumbnails128x128/68453.png inflating: /content/faces/thumbnails128x128/68454.png inflating: /content/faces/thumbnails128x128/68455.png inflating: /content/faces/thumbnails128x128/68456.png inflating: /content/faces/thumbnails128x128/68457.png inflating: /content/faces/thumbnails128x128/68458.png inflating: /content/faces/thumbnails128x128/68459.png inflating: /content/faces/thumbnails128x128/68460.png inflating: /content/faces/thumbnails128x128/68461.png inflating: /content/faces/thumbnails128x128/68462.png inflating: /content/faces/thumbnails128x128/68463.png inflating: /content/faces/thumbnails128x128/68464.png inflating: /content/faces/thumbnails128x128/68465.png inflating: /content/faces/thumbnails128x128/68466.png inflating: /content/faces/thumbnails128x128/68467.png inflating: /content/faces/thumbnails128x128/68468.png inflating: /content/faces/thumbnails128x128/68469.png inflating: /content/faces/thumbnails128x128/68470.png inflating: /content/faces/thumbnails128x128/68471.png inflating: /content/faces/thumbnails128x128/68472.png inflating: /content/faces/thumbnails128x128/68473.png inflating: /content/faces/thumbnails128x128/68474.png inflating: /content/faces/thumbnails128x128/68475.png inflating: /content/faces/thumbnails128x128/68476.png inflating: /content/faces/thumbnails128x128/68477.png inflating: /content/faces/thumbnails128x128/68478.png inflating: /content/faces/thumbnails128x128/68479.png inflating: /content/faces/thumbnails128x128/68480.png inflating: /content/faces/thumbnails128x128/68481.png inflating: /content/faces/thumbnails128x128/68482.png inflating: /content/faces/thumbnails128x128/68483.png inflating: /content/faces/thumbnails128x128/68484.png inflating: /content/faces/thumbnails128x128/68485.png inflating: /content/faces/thumbnails128x128/68486.png inflating: /content/faces/thumbnails128x128/68487.png inflating: /content/faces/thumbnails128x128/68488.png inflating: /content/faces/thumbnails128x128/68489.png inflating: /content/faces/thumbnails128x128/68490.png inflating: /content/faces/thumbnails128x128/68491.png inflating: /content/faces/thumbnails128x128/68492.png inflating: /content/faces/thumbnails128x128/68493.png inflating: /content/faces/thumbnails128x128/68494.png inflating: /content/faces/thumbnails128x128/68495.png inflating: /content/faces/thumbnails128x128/68496.png inflating: /content/faces/thumbnails128x128/68497.png inflating: /content/faces/thumbnails128x128/68498.png inflating: /content/faces/thumbnails128x128/68499.png inflating: /content/faces/thumbnails128x128/68500.png inflating: /content/faces/thumbnails128x128/68501.png inflating: /content/faces/thumbnails128x128/68502.png inflating: /content/faces/thumbnails128x128/68503.png inflating: /content/faces/thumbnails128x128/68504.png inflating: /content/faces/thumbnails128x128/68505.png inflating: /content/faces/thumbnails128x128/68506.png inflating: /content/faces/thumbnails128x128/68507.png inflating: /content/faces/thumbnails128x128/68508.png inflating: /content/faces/thumbnails128x128/68509.png inflating: /content/faces/thumbnails128x128/68510.png inflating: /content/faces/thumbnails128x128/68511.png inflating: /content/faces/thumbnails128x128/68512.png inflating: /content/faces/thumbnails128x128/68513.png inflating: /content/faces/thumbnails128x128/68514.png inflating: /content/faces/thumbnails128x128/68515.png inflating: /content/faces/thumbnails128x128/68516.png inflating: /content/faces/thumbnails128x128/68517.png inflating: /content/faces/thumbnails128x128/68518.png inflating: /content/faces/thumbnails128x128/68519.png inflating: /content/faces/thumbnails128x128/68520.png inflating: /content/faces/thumbnails128x128/68521.png inflating: /content/faces/thumbnails128x128/68522.png inflating: /content/faces/thumbnails128x128/68523.png inflating: /content/faces/thumbnails128x128/68524.png inflating: /content/faces/thumbnails128x128/68525.png inflating: /content/faces/thumbnails128x128/68526.png inflating: /content/faces/thumbnails128x128/68527.png inflating: /content/faces/thumbnails128x128/68528.png inflating: /content/faces/thumbnails128x128/68529.png inflating: /content/faces/thumbnails128x128/68530.png inflating: /content/faces/thumbnails128x128/68531.png inflating: /content/faces/thumbnails128x128/68532.png inflating: /content/faces/thumbnails128x128/68533.png inflating: /content/faces/thumbnails128x128/68534.png inflating: /content/faces/thumbnails128x128/68535.png inflating: /content/faces/thumbnails128x128/68536.png inflating: /content/faces/thumbnails128x128/68537.png inflating: /content/faces/thumbnails128x128/68538.png inflating: /content/faces/thumbnails128x128/68539.png inflating: /content/faces/thumbnails128x128/68540.png inflating: /content/faces/thumbnails128x128/68541.png inflating: /content/faces/thumbnails128x128/68542.png inflating: /content/faces/thumbnails128x128/68543.png inflating: /content/faces/thumbnails128x128/68544.png inflating: /content/faces/thumbnails128x128/68545.png inflating: /content/faces/thumbnails128x128/68546.png inflating: /content/faces/thumbnails128x128/68547.png inflating: /content/faces/thumbnails128x128/68548.png inflating: /content/faces/thumbnails128x128/68549.png inflating: /content/faces/thumbnails128x128/68550.png inflating: /content/faces/thumbnails128x128/68551.png inflating: /content/faces/thumbnails128x128/68552.png inflating: /content/faces/thumbnails128x128/68553.png inflating: /content/faces/thumbnails128x128/68554.png inflating: /content/faces/thumbnails128x128/68555.png inflating: /content/faces/thumbnails128x128/68556.png inflating: /content/faces/thumbnails128x128/68557.png inflating: /content/faces/thumbnails128x128/68558.png inflating: /content/faces/thumbnails128x128/68559.png inflating: /content/faces/thumbnails128x128/68560.png inflating: /content/faces/thumbnails128x128/68561.png inflating: /content/faces/thumbnails128x128/68562.png inflating: /content/faces/thumbnails128x128/68563.png inflating: /content/faces/thumbnails128x128/68564.png inflating: /content/faces/thumbnails128x128/68565.png inflating: /content/faces/thumbnails128x128/68566.png inflating: /content/faces/thumbnails128x128/68567.png inflating: /content/faces/thumbnails128x128/68568.png inflating: /content/faces/thumbnails128x128/68569.png inflating: /content/faces/thumbnails128x128/68570.png inflating: /content/faces/thumbnails128x128/68571.png inflating: /content/faces/thumbnails128x128/68572.png inflating: /content/faces/thumbnails128x128/68573.png inflating: /content/faces/thumbnails128x128/68574.png inflating: /content/faces/thumbnails128x128/68575.png inflating: /content/faces/thumbnails128x128/68576.png inflating: /content/faces/thumbnails128x128/68577.png inflating: /content/faces/thumbnails128x128/68578.png inflating: /content/faces/thumbnails128x128/68579.png inflating: /content/faces/thumbnails128x128/68580.png inflating: /content/faces/thumbnails128x128/68581.png inflating: /content/faces/thumbnails128x128/68582.png inflating: /content/faces/thumbnails128x128/68583.png inflating: /content/faces/thumbnails128x128/68584.png inflating: /content/faces/thumbnails128x128/68585.png inflating: /content/faces/thumbnails128x128/68586.png inflating: /content/faces/thumbnails128x128/68587.png inflating: /content/faces/thumbnails128x128/68588.png inflating: /content/faces/thumbnails128x128/68589.png inflating: /content/faces/thumbnails128x128/68590.png inflating: /content/faces/thumbnails128x128/68591.png inflating: /content/faces/thumbnails128x128/68592.png inflating: /content/faces/thumbnails128x128/68593.png inflating: /content/faces/thumbnails128x128/68594.png inflating: /content/faces/thumbnails128x128/68595.png inflating: /content/faces/thumbnails128x128/68596.png inflating: /content/faces/thumbnails128x128/68597.png inflating: /content/faces/thumbnails128x128/68598.png inflating: /content/faces/thumbnails128x128/68599.png inflating: /content/faces/thumbnails128x128/68600.png inflating: /content/faces/thumbnails128x128/68601.png inflating: /content/faces/thumbnails128x128/68602.png inflating: /content/faces/thumbnails128x128/68603.png inflating: /content/faces/thumbnails128x128/68604.png inflating: /content/faces/thumbnails128x128/68605.png inflating: /content/faces/thumbnails128x128/68606.png inflating: /content/faces/thumbnails128x128/68607.png inflating: /content/faces/thumbnails128x128/68608.png inflating: /content/faces/thumbnails128x128/68609.png inflating: /content/faces/thumbnails128x128/68610.png inflating: /content/faces/thumbnails128x128/68611.png inflating: /content/faces/thumbnails128x128/68612.png inflating: /content/faces/thumbnails128x128/68613.png inflating: /content/faces/thumbnails128x128/68614.png inflating: /content/faces/thumbnails128x128/68615.png inflating: /content/faces/thumbnails128x128/68616.png inflating: /content/faces/thumbnails128x128/68617.png inflating: /content/faces/thumbnails128x128/68618.png inflating: /content/faces/thumbnails128x128/68619.png inflating: /content/faces/thumbnails128x128/68620.png inflating: /content/faces/thumbnails128x128/68621.png inflating: /content/faces/thumbnails128x128/68622.png inflating: /content/faces/thumbnails128x128/68623.png inflating: /content/faces/thumbnails128x128/68624.png inflating: /content/faces/thumbnails128x128/68625.png inflating: /content/faces/thumbnails128x128/68626.png inflating: /content/faces/thumbnails128x128/68627.png inflating: /content/faces/thumbnails128x128/68628.png inflating: /content/faces/thumbnails128x128/68629.png inflating: /content/faces/thumbnails128x128/68630.png inflating: /content/faces/thumbnails128x128/68631.png inflating: /content/faces/thumbnails128x128/68632.png inflating: /content/faces/thumbnails128x128/68633.png inflating: /content/faces/thumbnails128x128/68634.png inflating: /content/faces/thumbnails128x128/68635.png inflating: /content/faces/thumbnails128x128/68636.png inflating: /content/faces/thumbnails128x128/68637.png inflating: /content/faces/thumbnails128x128/68638.png inflating: /content/faces/thumbnails128x128/68639.png inflating: /content/faces/thumbnails128x128/68640.png inflating: /content/faces/thumbnails128x128/68641.png inflating: /content/faces/thumbnails128x128/68642.png inflating: /content/faces/thumbnails128x128/68643.png inflating: /content/faces/thumbnails128x128/68644.png inflating: /content/faces/thumbnails128x128/68645.png inflating: /content/faces/thumbnails128x128/68646.png inflating: /content/faces/thumbnails128x128/68647.png inflating: /content/faces/thumbnails128x128/68648.png inflating: /content/faces/thumbnails128x128/68649.png inflating: /content/faces/thumbnails128x128/68650.png inflating: /content/faces/thumbnails128x128/68651.png inflating: /content/faces/thumbnails128x128/68652.png inflating: /content/faces/thumbnails128x128/68653.png inflating: /content/faces/thumbnails128x128/68654.png inflating: /content/faces/thumbnails128x128/68655.png inflating: /content/faces/thumbnails128x128/68656.png inflating: /content/faces/thumbnails128x128/68657.png inflating: /content/faces/thumbnails128x128/68658.png inflating: /content/faces/thumbnails128x128/68659.png inflating: /content/faces/thumbnails128x128/68660.png inflating: /content/faces/thumbnails128x128/68661.png inflating: /content/faces/thumbnails128x128/68662.png inflating: /content/faces/thumbnails128x128/68663.png inflating: /content/faces/thumbnails128x128/68664.png inflating: /content/faces/thumbnails128x128/68665.png inflating: /content/faces/thumbnails128x128/68666.png inflating: /content/faces/thumbnails128x128/68667.png inflating: /content/faces/thumbnails128x128/68668.png inflating: /content/faces/thumbnails128x128/68669.png inflating: /content/faces/thumbnails128x128/68670.png inflating: /content/faces/thumbnails128x128/68671.png inflating: /content/faces/thumbnails128x128/68672.png inflating: /content/faces/thumbnails128x128/68673.png inflating: /content/faces/thumbnails128x128/68674.png inflating: /content/faces/thumbnails128x128/68675.png inflating: /content/faces/thumbnails128x128/68676.png inflating: /content/faces/thumbnails128x128/68677.png inflating: /content/faces/thumbnails128x128/68678.png inflating: /content/faces/thumbnails128x128/68679.png inflating: /content/faces/thumbnails128x128/68680.png inflating: /content/faces/thumbnails128x128/68681.png inflating: /content/faces/thumbnails128x128/68682.png inflating: /content/faces/thumbnails128x128/68683.png inflating: /content/faces/thumbnails128x128/68684.png inflating: /content/faces/thumbnails128x128/68685.png inflating: /content/faces/thumbnails128x128/68686.png inflating: /content/faces/thumbnails128x128/68687.png inflating: /content/faces/thumbnails128x128/68688.png inflating: /content/faces/thumbnails128x128/68689.png inflating: /content/faces/thumbnails128x128/68690.png inflating: /content/faces/thumbnails128x128/68691.png inflating: /content/faces/thumbnails128x128/68692.png inflating: /content/faces/thumbnails128x128/68693.png inflating: /content/faces/thumbnails128x128/68694.png inflating: /content/faces/thumbnails128x128/68695.png inflating: /content/faces/thumbnails128x128/68696.png inflating: /content/faces/thumbnails128x128/68697.png inflating: /content/faces/thumbnails128x128/68698.png inflating: /content/faces/thumbnails128x128/68699.png inflating: /content/faces/thumbnails128x128/68700.png inflating: /content/faces/thumbnails128x128/68701.png inflating: /content/faces/thumbnails128x128/68702.png inflating: /content/faces/thumbnails128x128/68703.png inflating: /content/faces/thumbnails128x128/68704.png inflating: /content/faces/thumbnails128x128/68705.png inflating: /content/faces/thumbnails128x128/68706.png inflating: /content/faces/thumbnails128x128/68707.png inflating: /content/faces/thumbnails128x128/68708.png inflating: /content/faces/thumbnails128x128/68709.png inflating: /content/faces/thumbnails128x128/68710.png inflating: /content/faces/thumbnails128x128/68711.png inflating: /content/faces/thumbnails128x128/68712.png inflating: /content/faces/thumbnails128x128/68713.png inflating: /content/faces/thumbnails128x128/68714.png inflating: /content/faces/thumbnails128x128/68715.png inflating: /content/faces/thumbnails128x128/68716.png inflating: /content/faces/thumbnails128x128/68717.png inflating: /content/faces/thumbnails128x128/68718.png inflating: /content/faces/thumbnails128x128/68719.png inflating: /content/faces/thumbnails128x128/68720.png inflating: /content/faces/thumbnails128x128/68721.png inflating: /content/faces/thumbnails128x128/68722.png inflating: /content/faces/thumbnails128x128/68723.png inflating: /content/faces/thumbnails128x128/68724.png inflating: /content/faces/thumbnails128x128/68725.png inflating: /content/faces/thumbnails128x128/68726.png inflating: /content/faces/thumbnails128x128/68727.png inflating: /content/faces/thumbnails128x128/68728.png inflating: /content/faces/thumbnails128x128/68729.png inflating: /content/faces/thumbnails128x128/68730.png inflating: /content/faces/thumbnails128x128/68731.png inflating: /content/faces/thumbnails128x128/68732.png inflating: /content/faces/thumbnails128x128/68733.png inflating: /content/faces/thumbnails128x128/68734.png inflating: /content/faces/thumbnails128x128/68735.png inflating: /content/faces/thumbnails128x128/68736.png inflating: /content/faces/thumbnails128x128/68737.png inflating: /content/faces/thumbnails128x128/68738.png inflating: /content/faces/thumbnails128x128/68739.png inflating: /content/faces/thumbnails128x128/68740.png inflating: /content/faces/thumbnails128x128/68741.png inflating: /content/faces/thumbnails128x128/68742.png inflating: /content/faces/thumbnails128x128/68743.png inflating: /content/faces/thumbnails128x128/68744.png inflating: /content/faces/thumbnails128x128/68745.png inflating: /content/faces/thumbnails128x128/68746.png inflating: /content/faces/thumbnails128x128/68747.png inflating: /content/faces/thumbnails128x128/68748.png inflating: /content/faces/thumbnails128x128/68749.png inflating: /content/faces/thumbnails128x128/68750.png inflating: /content/faces/thumbnails128x128/68751.png inflating: /content/faces/thumbnails128x128/68752.png inflating: /content/faces/thumbnails128x128/68753.png inflating: /content/faces/thumbnails128x128/68754.png inflating: /content/faces/thumbnails128x128/68755.png inflating: /content/faces/thumbnails128x128/68756.png inflating: /content/faces/thumbnails128x128/68757.png inflating: /content/faces/thumbnails128x128/68758.png inflating: /content/faces/thumbnails128x128/68759.png inflating: /content/faces/thumbnails128x128/68760.png inflating: /content/faces/thumbnails128x128/68761.png inflating: /content/faces/thumbnails128x128/68762.png inflating: /content/faces/thumbnails128x128/68763.png inflating: /content/faces/thumbnails128x128/68764.png inflating: /content/faces/thumbnails128x128/68765.png inflating: /content/faces/thumbnails128x128/68766.png inflating: /content/faces/thumbnails128x128/68767.png inflating: /content/faces/thumbnails128x128/68768.png inflating: /content/faces/thumbnails128x128/68769.png inflating: /content/faces/thumbnails128x128/68770.png inflating: /content/faces/thumbnails128x128/68771.png inflating: /content/faces/thumbnails128x128/68772.png inflating: /content/faces/thumbnails128x128/68773.png inflating: /content/faces/thumbnails128x128/68774.png inflating: /content/faces/thumbnails128x128/68775.png inflating: /content/faces/thumbnails128x128/68776.png inflating: /content/faces/thumbnails128x128/68777.png inflating: /content/faces/thumbnails128x128/68778.png inflating: /content/faces/thumbnails128x128/68779.png inflating: /content/faces/thumbnails128x128/68780.png inflating: /content/faces/thumbnails128x128/68781.png inflating: /content/faces/thumbnails128x128/68782.png inflating: /content/faces/thumbnails128x128/68783.png inflating: /content/faces/thumbnails128x128/68784.png inflating: /content/faces/thumbnails128x128/68785.png inflating: /content/faces/thumbnails128x128/68786.png inflating: /content/faces/thumbnails128x128/68787.png inflating: /content/faces/thumbnails128x128/68788.png inflating: /content/faces/thumbnails128x128/68789.png inflating: /content/faces/thumbnails128x128/68790.png inflating: /content/faces/thumbnails128x128/68791.png inflating: /content/faces/thumbnails128x128/68792.png inflating: /content/faces/thumbnails128x128/68793.png inflating: /content/faces/thumbnails128x128/68794.png inflating: /content/faces/thumbnails128x128/68795.png inflating: /content/faces/thumbnails128x128/68796.png inflating: /content/faces/thumbnails128x128/68797.png inflating: /content/faces/thumbnails128x128/68798.png inflating: /content/faces/thumbnails128x128/68799.png inflating: /content/faces/thumbnails128x128/68800.png inflating: /content/faces/thumbnails128x128/68801.png inflating: /content/faces/thumbnails128x128/68802.png inflating: /content/faces/thumbnails128x128/68803.png inflating: /content/faces/thumbnails128x128/68804.png inflating: /content/faces/thumbnails128x128/68805.png inflating: /content/faces/thumbnails128x128/68806.png inflating: /content/faces/thumbnails128x128/68807.png inflating: /content/faces/thumbnails128x128/68808.png inflating: /content/faces/thumbnails128x128/68809.png inflating: /content/faces/thumbnails128x128/68810.png inflating: /content/faces/thumbnails128x128/68811.png inflating: /content/faces/thumbnails128x128/68812.png inflating: /content/faces/thumbnails128x128/68813.png inflating: /content/faces/thumbnails128x128/68814.png inflating: /content/faces/thumbnails128x128/68815.png inflating: /content/faces/thumbnails128x128/68816.png inflating: /content/faces/thumbnails128x128/68817.png inflating: /content/faces/thumbnails128x128/68818.png inflating: /content/faces/thumbnails128x128/68819.png inflating: /content/faces/thumbnails128x128/68820.png inflating: /content/faces/thumbnails128x128/68821.png inflating: /content/faces/thumbnails128x128/68822.png inflating: /content/faces/thumbnails128x128/68823.png inflating: /content/faces/thumbnails128x128/68824.png inflating: /content/faces/thumbnails128x128/68825.png inflating: /content/faces/thumbnails128x128/68826.png inflating: /content/faces/thumbnails128x128/68827.png inflating: /content/faces/thumbnails128x128/68828.png inflating: /content/faces/thumbnails128x128/68829.png inflating: /content/faces/thumbnails128x128/68830.png inflating: /content/faces/thumbnails128x128/68831.png inflating: /content/faces/thumbnails128x128/68832.png inflating: /content/faces/thumbnails128x128/68833.png inflating: /content/faces/thumbnails128x128/68834.png inflating: /content/faces/thumbnails128x128/68835.png inflating: /content/faces/thumbnails128x128/68836.png inflating: /content/faces/thumbnails128x128/68837.png inflating: /content/faces/thumbnails128x128/68838.png inflating: /content/faces/thumbnails128x128/68839.png inflating: /content/faces/thumbnails128x128/68840.png inflating: /content/faces/thumbnails128x128/68841.png inflating: /content/faces/thumbnails128x128/68842.png inflating: /content/faces/thumbnails128x128/68843.png inflating: /content/faces/thumbnails128x128/68844.png inflating: /content/faces/thumbnails128x128/68845.png inflating: /content/faces/thumbnails128x128/68846.png inflating: /content/faces/thumbnails128x128/68847.png inflating: /content/faces/thumbnails128x128/68848.png inflating: /content/faces/thumbnails128x128/68849.png inflating: /content/faces/thumbnails128x128/68850.png inflating: /content/faces/thumbnails128x128/68851.png inflating: /content/faces/thumbnails128x128/68852.png inflating: /content/faces/thumbnails128x128/68853.png inflating: /content/faces/thumbnails128x128/68854.png inflating: /content/faces/thumbnails128x128/68855.png inflating: /content/faces/thumbnails128x128/68856.png inflating: /content/faces/thumbnails128x128/68857.png inflating: /content/faces/thumbnails128x128/68858.png inflating: /content/faces/thumbnails128x128/68859.png inflating: /content/faces/thumbnails128x128/68860.png inflating: /content/faces/thumbnails128x128/68861.png inflating: /content/faces/thumbnails128x128/68862.png inflating: /content/faces/thumbnails128x128/68863.png inflating: /content/faces/thumbnails128x128/68864.png inflating: /content/faces/thumbnails128x128/68865.png inflating: /content/faces/thumbnails128x128/68866.png inflating: /content/faces/thumbnails128x128/68867.png inflating: /content/faces/thumbnails128x128/68868.png inflating: /content/faces/thumbnails128x128/68869.png inflating: /content/faces/thumbnails128x128/68870.png inflating: /content/faces/thumbnails128x128/68871.png inflating: /content/faces/thumbnails128x128/68872.png inflating: /content/faces/thumbnails128x128/68873.png inflating: /content/faces/thumbnails128x128/68874.png inflating: /content/faces/thumbnails128x128/68875.png inflating: /content/faces/thumbnails128x128/68876.png inflating: /content/faces/thumbnails128x128/68877.png inflating: /content/faces/thumbnails128x128/68878.png inflating: /content/faces/thumbnails128x128/68879.png inflating: /content/faces/thumbnails128x128/68880.png inflating: /content/faces/thumbnails128x128/68881.png inflating: /content/faces/thumbnails128x128/68882.png inflating: /content/faces/thumbnails128x128/68883.png inflating: /content/faces/thumbnails128x128/68884.png inflating: /content/faces/thumbnails128x128/68885.png inflating: /content/faces/thumbnails128x128/68886.png inflating: /content/faces/thumbnails128x128/68887.png inflating: /content/faces/thumbnails128x128/68888.png inflating: /content/faces/thumbnails128x128/68889.png inflating: /content/faces/thumbnails128x128/68890.png inflating: /content/faces/thumbnails128x128/68891.png inflating: /content/faces/thumbnails128x128/68892.png inflating: /content/faces/thumbnails128x128/68893.png inflating: /content/faces/thumbnails128x128/68894.png inflating: /content/faces/thumbnails128x128/68895.png inflating: /content/faces/thumbnails128x128/68896.png inflating: /content/faces/thumbnails128x128/68897.png inflating: /content/faces/thumbnails128x128/68898.png inflating: /content/faces/thumbnails128x128/68899.png inflating: /content/faces/thumbnails128x128/68900.png inflating: /content/faces/thumbnails128x128/68901.png inflating: /content/faces/thumbnails128x128/68902.png inflating: /content/faces/thumbnails128x128/68903.png inflating: /content/faces/thumbnails128x128/68904.png inflating: /content/faces/thumbnails128x128/68905.png inflating: /content/faces/thumbnails128x128/68906.png inflating: /content/faces/thumbnails128x128/68907.png inflating: /content/faces/thumbnails128x128/68908.png inflating: /content/faces/thumbnails128x128/68909.png inflating: /content/faces/thumbnails128x128/68910.png inflating: /content/faces/thumbnails128x128/68911.png inflating: /content/faces/thumbnails128x128/68912.png inflating: /content/faces/thumbnails128x128/68913.png inflating: /content/faces/thumbnails128x128/68914.png inflating: /content/faces/thumbnails128x128/68915.png inflating: /content/faces/thumbnails128x128/68916.png inflating: /content/faces/thumbnails128x128/68917.png inflating: /content/faces/thumbnails128x128/68918.png inflating: /content/faces/thumbnails128x128/68919.png inflating: /content/faces/thumbnails128x128/68920.png inflating: /content/faces/thumbnails128x128/68921.png inflating: /content/faces/thumbnails128x128/68922.png inflating: /content/faces/thumbnails128x128/68923.png inflating: /content/faces/thumbnails128x128/68924.png inflating: /content/faces/thumbnails128x128/68925.png inflating: /content/faces/thumbnails128x128/68926.png inflating: /content/faces/thumbnails128x128/68927.png inflating: /content/faces/thumbnails128x128/68928.png inflating: /content/faces/thumbnails128x128/68929.png inflating: /content/faces/thumbnails128x128/68930.png inflating: /content/faces/thumbnails128x128/68931.png inflating: /content/faces/thumbnails128x128/68932.png inflating: /content/faces/thumbnails128x128/68933.png inflating: /content/faces/thumbnails128x128/68934.png inflating: /content/faces/thumbnails128x128/68935.png inflating: /content/faces/thumbnails128x128/68936.png inflating: /content/faces/thumbnails128x128/68937.png inflating: /content/faces/thumbnails128x128/68938.png inflating: /content/faces/thumbnails128x128/68939.png inflating: /content/faces/thumbnails128x128/68940.png inflating: /content/faces/thumbnails128x128/68941.png inflating: /content/faces/thumbnails128x128/68942.png inflating: /content/faces/thumbnails128x128/68943.png inflating: /content/faces/thumbnails128x128/68944.png inflating: /content/faces/thumbnails128x128/68945.png inflating: /content/faces/thumbnails128x128/68946.png inflating: /content/faces/thumbnails128x128/68947.png inflating: /content/faces/thumbnails128x128/68948.png inflating: /content/faces/thumbnails128x128/68949.png inflating: /content/faces/thumbnails128x128/68950.png inflating: /content/faces/thumbnails128x128/68951.png inflating: /content/faces/thumbnails128x128/68952.png inflating: /content/faces/thumbnails128x128/68953.png inflating: /content/faces/thumbnails128x128/68954.png inflating: /content/faces/thumbnails128x128/68955.png inflating: /content/faces/thumbnails128x128/68956.png inflating: /content/faces/thumbnails128x128/68957.png inflating: /content/faces/thumbnails128x128/68958.png inflating: /content/faces/thumbnails128x128/68959.png inflating: /content/faces/thumbnails128x128/68960.png inflating: /content/faces/thumbnails128x128/68961.png inflating: /content/faces/thumbnails128x128/68962.png inflating: /content/faces/thumbnails128x128/68963.png inflating: /content/faces/thumbnails128x128/68964.png inflating: /content/faces/thumbnails128x128/68965.png inflating: /content/faces/thumbnails128x128/68966.png inflating: /content/faces/thumbnails128x128/68967.png inflating: /content/faces/thumbnails128x128/68968.png inflating: /content/faces/thumbnails128x128/68969.png inflating: /content/faces/thumbnails128x128/68970.png inflating: /content/faces/thumbnails128x128/68971.png inflating: /content/faces/thumbnails128x128/68972.png inflating: /content/faces/thumbnails128x128/68973.png inflating: /content/faces/thumbnails128x128/68974.png inflating: /content/faces/thumbnails128x128/68975.png inflating: /content/faces/thumbnails128x128/68976.png inflating: /content/faces/thumbnails128x128/68977.png inflating: /content/faces/thumbnails128x128/68978.png inflating: /content/faces/thumbnails128x128/68979.png inflating: /content/faces/thumbnails128x128/68980.png inflating: /content/faces/thumbnails128x128/68981.png inflating: /content/faces/thumbnails128x128/68982.png inflating: /content/faces/thumbnails128x128/68983.png inflating: /content/faces/thumbnails128x128/68984.png inflating: /content/faces/thumbnails128x128/68985.png inflating: /content/faces/thumbnails128x128/68986.png inflating: /content/faces/thumbnails128x128/68987.png inflating: /content/faces/thumbnails128x128/68988.png inflating: /content/faces/thumbnails128x128/68989.png inflating: /content/faces/thumbnails128x128/68990.png inflating: /content/faces/thumbnails128x128/68991.png inflating: /content/faces/thumbnails128x128/68992.png inflating: /content/faces/thumbnails128x128/68993.png inflating: /content/faces/thumbnails128x128/68994.png inflating: /content/faces/thumbnails128x128/68995.png inflating: /content/faces/thumbnails128x128/68996.png inflating: /content/faces/thumbnails128x128/68997.png inflating: /content/faces/thumbnails128x128/68998.png inflating: /content/faces/thumbnails128x128/68999.png inflating: /content/faces/thumbnails128x128/69000.png inflating: /content/faces/thumbnails128x128/69001.png inflating: /content/faces/thumbnails128x128/69002.png inflating: /content/faces/thumbnails128x128/69003.png inflating: /content/faces/thumbnails128x128/69004.png inflating: /content/faces/thumbnails128x128/69005.png inflating: /content/faces/thumbnails128x128/69006.png inflating: /content/faces/thumbnails128x128/69007.png inflating: /content/faces/thumbnails128x128/69008.png inflating: /content/faces/thumbnails128x128/69009.png inflating: /content/faces/thumbnails128x128/69010.png inflating: /content/faces/thumbnails128x128/69011.png inflating: /content/faces/thumbnails128x128/69012.png inflating: /content/faces/thumbnails128x128/69013.png inflating: /content/faces/thumbnails128x128/69014.png inflating: /content/faces/thumbnails128x128/69015.png inflating: /content/faces/thumbnails128x128/69016.png inflating: /content/faces/thumbnails128x128/69017.png inflating: /content/faces/thumbnails128x128/69018.png inflating: /content/faces/thumbnails128x128/69019.png inflating: /content/faces/thumbnails128x128/69020.png inflating: /content/faces/thumbnails128x128/69021.png inflating: /content/faces/thumbnails128x128/69022.png inflating: /content/faces/thumbnails128x128/69023.png inflating: /content/faces/thumbnails128x128/69024.png inflating: /content/faces/thumbnails128x128/69025.png inflating: /content/faces/thumbnails128x128/69026.png inflating: /content/faces/thumbnails128x128/69027.png inflating: /content/faces/thumbnails128x128/69028.png inflating: /content/faces/thumbnails128x128/69029.png inflating: /content/faces/thumbnails128x128/69030.png inflating: /content/faces/thumbnails128x128/69031.png inflating: /content/faces/thumbnails128x128/69032.png inflating: /content/faces/thumbnails128x128/69033.png inflating: /content/faces/thumbnails128x128/69034.png inflating: /content/faces/thumbnails128x128/69035.png inflating: /content/faces/thumbnails128x128/69036.png inflating: /content/faces/thumbnails128x128/69037.png inflating: /content/faces/thumbnails128x128/69038.png inflating: /content/faces/thumbnails128x128/69039.png inflating: /content/faces/thumbnails128x128/69040.png inflating: /content/faces/thumbnails128x128/69041.png inflating: /content/faces/thumbnails128x128/69042.png inflating: /content/faces/thumbnails128x128/69043.png inflating: /content/faces/thumbnails128x128/69044.png inflating: /content/faces/thumbnails128x128/69045.png inflating: /content/faces/thumbnails128x128/69046.png inflating: /content/faces/thumbnails128x128/69047.png inflating: /content/faces/thumbnails128x128/69048.png inflating: /content/faces/thumbnails128x128/69049.png inflating: /content/faces/thumbnails128x128/69050.png inflating: /content/faces/thumbnails128x128/69051.png inflating: /content/faces/thumbnails128x128/69052.png inflating: /content/faces/thumbnails128x128/69053.png inflating: /content/faces/thumbnails128x128/69054.png inflating: /content/faces/thumbnails128x128/69055.png inflating: /content/faces/thumbnails128x128/69056.png inflating: /content/faces/thumbnails128x128/69057.png inflating: /content/faces/thumbnails128x128/69058.png inflating: /content/faces/thumbnails128x128/69059.png inflating: /content/faces/thumbnails128x128/69060.png inflating: /content/faces/thumbnails128x128/69061.png inflating: /content/faces/thumbnails128x128/69062.png inflating: /content/faces/thumbnails128x128/69063.png inflating: /content/faces/thumbnails128x128/69064.png inflating: /content/faces/thumbnails128x128/69065.png inflating: /content/faces/thumbnails128x128/69066.png inflating: /content/faces/thumbnails128x128/69067.png inflating: /content/faces/thumbnails128x128/69068.png inflating: /content/faces/thumbnails128x128/69069.png inflating: /content/faces/thumbnails128x128/69070.png inflating: /content/faces/thumbnails128x128/69071.png inflating: /content/faces/thumbnails128x128/69072.png inflating: /content/faces/thumbnails128x128/69073.png inflating: /content/faces/thumbnails128x128/69074.png inflating: /content/faces/thumbnails128x128/69075.png inflating: /content/faces/thumbnails128x128/69076.png inflating: /content/faces/thumbnails128x128/69077.png inflating: /content/faces/thumbnails128x128/69078.png inflating: /content/faces/thumbnails128x128/69079.png inflating: /content/faces/thumbnails128x128/69080.png inflating: /content/faces/thumbnails128x128/69081.png inflating: /content/faces/thumbnails128x128/69082.png inflating: /content/faces/thumbnails128x128/69083.png inflating: /content/faces/thumbnails128x128/69084.png inflating: /content/faces/thumbnails128x128/69085.png inflating: /content/faces/thumbnails128x128/69086.png inflating: /content/faces/thumbnails128x128/69087.png inflating: /content/faces/thumbnails128x128/69088.png inflating: /content/faces/thumbnails128x128/69089.png inflating: /content/faces/thumbnails128x128/69090.png inflating: /content/faces/thumbnails128x128/69091.png inflating: /content/faces/thumbnails128x128/69092.png inflating: /content/faces/thumbnails128x128/69093.png inflating: /content/faces/thumbnails128x128/69094.png inflating: /content/faces/thumbnails128x128/69095.png inflating: /content/faces/thumbnails128x128/69096.png inflating: /content/faces/thumbnails128x128/69097.png inflating: /content/faces/thumbnails128x128/69098.png inflating: /content/faces/thumbnails128x128/69099.png inflating: /content/faces/thumbnails128x128/69100.png inflating: /content/faces/thumbnails128x128/69101.png inflating: /content/faces/thumbnails128x128/69102.png inflating: /content/faces/thumbnails128x128/69103.png inflating: /content/faces/thumbnails128x128/69104.png inflating: /content/faces/thumbnails128x128/69105.png inflating: /content/faces/thumbnails128x128/69106.png inflating: /content/faces/thumbnails128x128/69107.png inflating: /content/faces/thumbnails128x128/69108.png inflating: /content/faces/thumbnails128x128/69109.png inflating: /content/faces/thumbnails128x128/69110.png inflating: /content/faces/thumbnails128x128/69111.png inflating: /content/faces/thumbnails128x128/69112.png inflating: /content/faces/thumbnails128x128/69113.png inflating: /content/faces/thumbnails128x128/69114.png inflating: /content/faces/thumbnails128x128/69115.png inflating: /content/faces/thumbnails128x128/69116.png inflating: /content/faces/thumbnails128x128/69117.png inflating: /content/faces/thumbnails128x128/69118.png inflating: /content/faces/thumbnails128x128/69119.png inflating: /content/faces/thumbnails128x128/69120.png inflating: /content/faces/thumbnails128x128/69121.png inflating: /content/faces/thumbnails128x128/69122.png inflating: /content/faces/thumbnails128x128/69123.png inflating: /content/faces/thumbnails128x128/69124.png inflating: /content/faces/thumbnails128x128/69125.png inflating: /content/faces/thumbnails128x128/69126.png inflating: /content/faces/thumbnails128x128/69127.png inflating: /content/faces/thumbnails128x128/69128.png inflating: /content/faces/thumbnails128x128/69129.png inflating: /content/faces/thumbnails128x128/69130.png inflating: /content/faces/thumbnails128x128/69131.png inflating: /content/faces/thumbnails128x128/69132.png inflating: /content/faces/thumbnails128x128/69133.png inflating: /content/faces/thumbnails128x128/69134.png inflating: /content/faces/thumbnails128x128/69135.png inflating: /content/faces/thumbnails128x128/69136.png inflating: /content/faces/thumbnails128x128/69137.png inflating: /content/faces/thumbnails128x128/69138.png inflating: /content/faces/thumbnails128x128/69139.png inflating: /content/faces/thumbnails128x128/69140.png inflating: /content/faces/thumbnails128x128/69141.png inflating: /content/faces/thumbnails128x128/69142.png inflating: /content/faces/thumbnails128x128/69143.png inflating: /content/faces/thumbnails128x128/69144.png inflating: /content/faces/thumbnails128x128/69145.png inflating: /content/faces/thumbnails128x128/69146.png inflating: /content/faces/thumbnails128x128/69147.png inflating: /content/faces/thumbnails128x128/69148.png inflating: /content/faces/thumbnails128x128/69149.png inflating: /content/faces/thumbnails128x128/69150.png inflating: /content/faces/thumbnails128x128/69151.png inflating: /content/faces/thumbnails128x128/69152.png inflating: /content/faces/thumbnails128x128/69153.png inflating: /content/faces/thumbnails128x128/69154.png inflating: /content/faces/thumbnails128x128/69155.png inflating: /content/faces/thumbnails128x128/69156.png inflating: /content/faces/thumbnails128x128/69157.png inflating: /content/faces/thumbnails128x128/69158.png inflating: /content/faces/thumbnails128x128/69159.png inflating: /content/faces/thumbnails128x128/69160.png inflating: /content/faces/thumbnails128x128/69161.png inflating: /content/faces/thumbnails128x128/69162.png inflating: /content/faces/thumbnails128x128/69163.png inflating: /content/faces/thumbnails128x128/69164.png inflating: /content/faces/thumbnails128x128/69165.png inflating: /content/faces/thumbnails128x128/69166.png inflating: /content/faces/thumbnails128x128/69167.png inflating: /content/faces/thumbnails128x128/69168.png inflating: /content/faces/thumbnails128x128/69169.png inflating: /content/faces/thumbnails128x128/69170.png inflating: /content/faces/thumbnails128x128/69171.png inflating: /content/faces/thumbnails128x128/69172.png inflating: /content/faces/thumbnails128x128/69173.png inflating: /content/faces/thumbnails128x128/69174.png inflating: /content/faces/thumbnails128x128/69175.png inflating: /content/faces/thumbnails128x128/69176.png inflating: /content/faces/thumbnails128x128/69177.png inflating: /content/faces/thumbnails128x128/69178.png inflating: /content/faces/thumbnails128x128/69179.png inflating: /content/faces/thumbnails128x128/69180.png inflating: /content/faces/thumbnails128x128/69181.png inflating: /content/faces/thumbnails128x128/69182.png inflating: /content/faces/thumbnails128x128/69183.png inflating: /content/faces/thumbnails128x128/69184.png inflating: /content/faces/thumbnails128x128/69185.png inflating: /content/faces/thumbnails128x128/69186.png inflating: /content/faces/thumbnails128x128/69187.png inflating: /content/faces/thumbnails128x128/69188.png inflating: /content/faces/thumbnails128x128/69189.png inflating: /content/faces/thumbnails128x128/69190.png inflating: /content/faces/thumbnails128x128/69191.png inflating: /content/faces/thumbnails128x128/69192.png inflating: /content/faces/thumbnails128x128/69193.png inflating: /content/faces/thumbnails128x128/69194.png inflating: /content/faces/thumbnails128x128/69195.png inflating: /content/faces/thumbnails128x128/69196.png inflating: /content/faces/thumbnails128x128/69197.png inflating: /content/faces/thumbnails128x128/69198.png inflating: /content/faces/thumbnails128x128/69199.png inflating: /content/faces/thumbnails128x128/69200.png inflating: /content/faces/thumbnails128x128/69201.png inflating: /content/faces/thumbnails128x128/69202.png inflating: /content/faces/thumbnails128x128/69203.png inflating: /content/faces/thumbnails128x128/69204.png inflating: /content/faces/thumbnails128x128/69205.png inflating: /content/faces/thumbnails128x128/69206.png inflating: /content/faces/thumbnails128x128/69207.png inflating: /content/faces/thumbnails128x128/69208.png inflating: /content/faces/thumbnails128x128/69209.png inflating: /content/faces/thumbnails128x128/69210.png inflating: /content/faces/thumbnails128x128/69211.png inflating: /content/faces/thumbnails128x128/69212.png inflating: /content/faces/thumbnails128x128/69213.png inflating: /content/faces/thumbnails128x128/69214.png inflating: /content/faces/thumbnails128x128/69215.png inflating: /content/faces/thumbnails128x128/69216.png inflating: /content/faces/thumbnails128x128/69217.png inflating: /content/faces/thumbnails128x128/69218.png inflating: /content/faces/thumbnails128x128/69219.png inflating: /content/faces/thumbnails128x128/69220.png inflating: /content/faces/thumbnails128x128/69221.png inflating: /content/faces/thumbnails128x128/69222.png inflating: /content/faces/thumbnails128x128/69223.png inflating: /content/faces/thumbnails128x128/69224.png inflating: /content/faces/thumbnails128x128/69225.png inflating: /content/faces/thumbnails128x128/69226.png inflating: /content/faces/thumbnails128x128/69227.png inflating: /content/faces/thumbnails128x128/69228.png inflating: /content/faces/thumbnails128x128/69229.png inflating: /content/faces/thumbnails128x128/69230.png inflating: /content/faces/thumbnails128x128/69231.png inflating: /content/faces/thumbnails128x128/69232.png inflating: /content/faces/thumbnails128x128/69233.png inflating: /content/faces/thumbnails128x128/69234.png inflating: /content/faces/thumbnails128x128/69235.png inflating: /content/faces/thumbnails128x128/69236.png inflating: /content/faces/thumbnails128x128/69237.png inflating: /content/faces/thumbnails128x128/69238.png inflating: /content/faces/thumbnails128x128/69239.png inflating: /content/faces/thumbnails128x128/69240.png inflating: /content/faces/thumbnails128x128/69241.png inflating: /content/faces/thumbnails128x128/69242.png inflating: /content/faces/thumbnails128x128/69243.png inflating: /content/faces/thumbnails128x128/69244.png inflating: /content/faces/thumbnails128x128/69245.png inflating: /content/faces/thumbnails128x128/69246.png inflating: /content/faces/thumbnails128x128/69247.png inflating: /content/faces/thumbnails128x128/69248.png inflating: /content/faces/thumbnails128x128/69249.png inflating: /content/faces/thumbnails128x128/69250.png inflating: /content/faces/thumbnails128x128/69251.png inflating: /content/faces/thumbnails128x128/69252.png inflating: /content/faces/thumbnails128x128/69253.png inflating: /content/faces/thumbnails128x128/69254.png inflating: /content/faces/thumbnails128x128/69255.png inflating: /content/faces/thumbnails128x128/69256.png inflating: /content/faces/thumbnails128x128/69257.png inflating: /content/faces/thumbnails128x128/69258.png inflating: /content/faces/thumbnails128x128/69259.png inflating: /content/faces/thumbnails128x128/69260.png inflating: /content/faces/thumbnails128x128/69261.png inflating: /content/faces/thumbnails128x128/69262.png inflating: /content/faces/thumbnails128x128/69263.png inflating: /content/faces/thumbnails128x128/69264.png inflating: /content/faces/thumbnails128x128/69265.png inflating: /content/faces/thumbnails128x128/69266.png inflating: /content/faces/thumbnails128x128/69267.png inflating: /content/faces/thumbnails128x128/69268.png inflating: /content/faces/thumbnails128x128/69269.png inflating: /content/faces/thumbnails128x128/69270.png inflating: /content/faces/thumbnails128x128/69271.png inflating: /content/faces/thumbnails128x128/69272.png inflating: /content/faces/thumbnails128x128/69273.png inflating: /content/faces/thumbnails128x128/69274.png inflating: /content/faces/thumbnails128x128/69275.png inflating: /content/faces/thumbnails128x128/69276.png inflating: /content/faces/thumbnails128x128/69277.png inflating: /content/faces/thumbnails128x128/69278.png inflating: /content/faces/thumbnails128x128/69279.png inflating: /content/faces/thumbnails128x128/69280.png inflating: /content/faces/thumbnails128x128/69281.png inflating: /content/faces/thumbnails128x128/69282.png inflating: /content/faces/thumbnails128x128/69283.png inflating: /content/faces/thumbnails128x128/69284.png inflating: /content/faces/thumbnails128x128/69285.png inflating: /content/faces/thumbnails128x128/69286.png inflating: /content/faces/thumbnails128x128/69287.png inflating: /content/faces/thumbnails128x128/69288.png inflating: /content/faces/thumbnails128x128/69289.png inflating: /content/faces/thumbnails128x128/69290.png inflating: /content/faces/thumbnails128x128/69291.png inflating: /content/faces/thumbnails128x128/69292.png inflating: /content/faces/thumbnails128x128/69293.png inflating: /content/faces/thumbnails128x128/69294.png inflating: /content/faces/thumbnails128x128/69295.png inflating: /content/faces/thumbnails128x128/69296.png inflating: /content/faces/thumbnails128x128/69297.png inflating: /content/faces/thumbnails128x128/69298.png inflating: /content/faces/thumbnails128x128/69299.png inflating: /content/faces/thumbnails128x128/69300.png inflating: /content/faces/thumbnails128x128/69301.png inflating: /content/faces/thumbnails128x128/69302.png inflating: /content/faces/thumbnails128x128/69303.png inflating: /content/faces/thumbnails128x128/69304.png inflating: /content/faces/thumbnails128x128/69305.png inflating: /content/faces/thumbnails128x128/69306.png inflating: /content/faces/thumbnails128x128/69307.png inflating: /content/faces/thumbnails128x128/69308.png inflating: /content/faces/thumbnails128x128/69309.png inflating: /content/faces/thumbnails128x128/69310.png inflating: /content/faces/thumbnails128x128/69311.png inflating: /content/faces/thumbnails128x128/69312.png inflating: /content/faces/thumbnails128x128/69313.png inflating: /content/faces/thumbnails128x128/69314.png inflating: /content/faces/thumbnails128x128/69315.png inflating: /content/faces/thumbnails128x128/69316.png inflating: /content/faces/thumbnails128x128/69317.png inflating: /content/faces/thumbnails128x128/69318.png inflating: /content/faces/thumbnails128x128/69319.png inflating: /content/faces/thumbnails128x128/69320.png inflating: /content/faces/thumbnails128x128/69321.png inflating: /content/faces/thumbnails128x128/69322.png inflating: /content/faces/thumbnails128x128/69323.png inflating: /content/faces/thumbnails128x128/69324.png inflating: /content/faces/thumbnails128x128/69325.png inflating: /content/faces/thumbnails128x128/69326.png inflating: /content/faces/thumbnails128x128/69327.png inflating: /content/faces/thumbnails128x128/69328.png inflating: /content/faces/thumbnails128x128/69329.png inflating: /content/faces/thumbnails128x128/69330.png inflating: /content/faces/thumbnails128x128/69331.png inflating: /content/faces/thumbnails128x128/69332.png inflating: /content/faces/thumbnails128x128/69333.png inflating: /content/faces/thumbnails128x128/69334.png inflating: /content/faces/thumbnails128x128/69335.png inflating: /content/faces/thumbnails128x128/69336.png inflating: /content/faces/thumbnails128x128/69337.png inflating: /content/faces/thumbnails128x128/69338.png inflating: /content/faces/thumbnails128x128/69339.png inflating: /content/faces/thumbnails128x128/69340.png inflating: /content/faces/thumbnails128x128/69341.png inflating: /content/faces/thumbnails128x128/69342.png inflating: /content/faces/thumbnails128x128/69343.png inflating: /content/faces/thumbnails128x128/69344.png inflating: /content/faces/thumbnails128x128/69345.png inflating: /content/faces/thumbnails128x128/69346.png inflating: /content/faces/thumbnails128x128/69347.png inflating: /content/faces/thumbnails128x128/69348.png inflating: /content/faces/thumbnails128x128/69349.png inflating: /content/faces/thumbnails128x128/69350.png inflating: /content/faces/thumbnails128x128/69351.png inflating: /content/faces/thumbnails128x128/69352.png inflating: /content/faces/thumbnails128x128/69353.png inflating: /content/faces/thumbnails128x128/69354.png inflating: /content/faces/thumbnails128x128/69355.png inflating: /content/faces/thumbnails128x128/69356.png inflating: /content/faces/thumbnails128x128/69357.png inflating: /content/faces/thumbnails128x128/69358.png inflating: /content/faces/thumbnails128x128/69359.png inflating: /content/faces/thumbnails128x128/69360.png inflating: /content/faces/thumbnails128x128/69361.png inflating: /content/faces/thumbnails128x128/69362.png inflating: /content/faces/thumbnails128x128/69363.png inflating: /content/faces/thumbnails128x128/69364.png inflating: /content/faces/thumbnails128x128/69365.png inflating: /content/faces/thumbnails128x128/69366.png inflating: /content/faces/thumbnails128x128/69367.png inflating: /content/faces/thumbnails128x128/69368.png inflating: /content/faces/thumbnails128x128/69369.png inflating: /content/faces/thumbnails128x128/69370.png inflating: /content/faces/thumbnails128x128/69371.png inflating: /content/faces/thumbnails128x128/69372.png inflating: /content/faces/thumbnails128x128/69373.png inflating: /content/faces/thumbnails128x128/69374.png inflating: /content/faces/thumbnails128x128/69375.png inflating: /content/faces/thumbnails128x128/69376.png inflating: /content/faces/thumbnails128x128/69377.png inflating: /content/faces/thumbnails128x128/69378.png inflating: /content/faces/thumbnails128x128/69379.png inflating: /content/faces/thumbnails128x128/69380.png inflating: /content/faces/thumbnails128x128/69381.png inflating: /content/faces/thumbnails128x128/69382.png inflating: /content/faces/thumbnails128x128/69383.png inflating: /content/faces/thumbnails128x128/69384.png inflating: /content/faces/thumbnails128x128/69385.png inflating: /content/faces/thumbnails128x128/69386.png inflating: /content/faces/thumbnails128x128/69387.png inflating: /content/faces/thumbnails128x128/69388.png inflating: /content/faces/thumbnails128x128/69389.png inflating: /content/faces/thumbnails128x128/69390.png inflating: /content/faces/thumbnails128x128/69391.png inflating: /content/faces/thumbnails128x128/69392.png inflating: /content/faces/thumbnails128x128/69393.png inflating: /content/faces/thumbnails128x128/69394.png inflating: /content/faces/thumbnails128x128/69395.png inflating: /content/faces/thumbnails128x128/69396.png inflating: /content/faces/thumbnails128x128/69397.png inflating: /content/faces/thumbnails128x128/69398.png inflating: /content/faces/thumbnails128x128/69399.png inflating: /content/faces/thumbnails128x128/69400.png inflating: /content/faces/thumbnails128x128/69401.png inflating: /content/faces/thumbnails128x128/69402.png inflating: /content/faces/thumbnails128x128/69403.png inflating: /content/faces/thumbnails128x128/69404.png inflating: /content/faces/thumbnails128x128/69405.png inflating: /content/faces/thumbnails128x128/69406.png inflating: /content/faces/thumbnails128x128/69407.png inflating: /content/faces/thumbnails128x128/69408.png inflating: /content/faces/thumbnails128x128/69409.png inflating: /content/faces/thumbnails128x128/69410.png inflating: /content/faces/thumbnails128x128/69411.png inflating: /content/faces/thumbnails128x128/69412.png inflating: /content/faces/thumbnails128x128/69413.png inflating: /content/faces/thumbnails128x128/69414.png inflating: /content/faces/thumbnails128x128/69415.png inflating: /content/faces/thumbnails128x128/69416.png inflating: /content/faces/thumbnails128x128/69417.png inflating: /content/faces/thumbnails128x128/69418.png inflating: /content/faces/thumbnails128x128/69419.png inflating: /content/faces/thumbnails128x128/69420.png inflating: /content/faces/thumbnails128x128/69421.png inflating: /content/faces/thumbnails128x128/69422.png inflating: /content/faces/thumbnails128x128/69423.png inflating: /content/faces/thumbnails128x128/69424.png inflating: /content/faces/thumbnails128x128/69425.png inflating: /content/faces/thumbnails128x128/69426.png inflating: /content/faces/thumbnails128x128/69427.png inflating: /content/faces/thumbnails128x128/69428.png inflating: /content/faces/thumbnails128x128/69429.png inflating: /content/faces/thumbnails128x128/69430.png inflating: /content/faces/thumbnails128x128/69431.png inflating: /content/faces/thumbnails128x128/69432.png inflating: /content/faces/thumbnails128x128/69433.png inflating: /content/faces/thumbnails128x128/69434.png inflating: /content/faces/thumbnails128x128/69435.png inflating: /content/faces/thumbnails128x128/69436.png inflating: /content/faces/thumbnails128x128/69437.png inflating: /content/faces/thumbnails128x128/69438.png inflating: /content/faces/thumbnails128x128/69439.png inflating: /content/faces/thumbnails128x128/69440.png inflating: /content/faces/thumbnails128x128/69441.png inflating: /content/faces/thumbnails128x128/69442.png inflating: /content/faces/thumbnails128x128/69443.png inflating: /content/faces/thumbnails128x128/69444.png inflating: /content/faces/thumbnails128x128/69445.png inflating: /content/faces/thumbnails128x128/69446.png inflating: /content/faces/thumbnails128x128/69447.png inflating: /content/faces/thumbnails128x128/69448.png inflating: /content/faces/thumbnails128x128/69449.png inflating: /content/faces/thumbnails128x128/69450.png inflating: /content/faces/thumbnails128x128/69451.png inflating: /content/faces/thumbnails128x128/69452.png inflating: /content/faces/thumbnails128x128/69453.png inflating: /content/faces/thumbnails128x128/69454.png inflating: /content/faces/thumbnails128x128/69455.png inflating: /content/faces/thumbnails128x128/69456.png inflating: /content/faces/thumbnails128x128/69457.png inflating: /content/faces/thumbnails128x128/69458.png inflating: /content/faces/thumbnails128x128/69459.png inflating: /content/faces/thumbnails128x128/69460.png inflating: /content/faces/thumbnails128x128/69461.png inflating: /content/faces/thumbnails128x128/69462.png inflating: /content/faces/thumbnails128x128/69463.png inflating: /content/faces/thumbnails128x128/69464.png inflating: /content/faces/thumbnails128x128/69465.png inflating: /content/faces/thumbnails128x128/69466.png inflating: /content/faces/thumbnails128x128/69467.png inflating: /content/faces/thumbnails128x128/69468.png inflating: /content/faces/thumbnails128x128/69469.png inflating: /content/faces/thumbnails128x128/69470.png inflating: /content/faces/thumbnails128x128/69471.png inflating: /content/faces/thumbnails128x128/69472.png inflating: /content/faces/thumbnails128x128/69473.png inflating: /content/faces/thumbnails128x128/69474.png inflating: /content/faces/thumbnails128x128/69475.png inflating: /content/faces/thumbnails128x128/69476.png inflating: /content/faces/thumbnails128x128/69477.png inflating: /content/faces/thumbnails128x128/69478.png inflating: /content/faces/thumbnails128x128/69479.png inflating: /content/faces/thumbnails128x128/69480.png inflating: /content/faces/thumbnails128x128/69481.png inflating: /content/faces/thumbnails128x128/69482.png inflating: /content/faces/thumbnails128x128/69483.png inflating: /content/faces/thumbnails128x128/69484.png inflating: /content/faces/thumbnails128x128/69485.png inflating: /content/faces/thumbnails128x128/69486.png inflating: /content/faces/thumbnails128x128/69487.png inflating: /content/faces/thumbnails128x128/69488.png inflating: /content/faces/thumbnails128x128/69489.png inflating: /content/faces/thumbnails128x128/69490.png inflating: /content/faces/thumbnails128x128/69491.png inflating: /content/faces/thumbnails128x128/69492.png inflating: /content/faces/thumbnails128x128/69493.png inflating: /content/faces/thumbnails128x128/69494.png inflating: /content/faces/thumbnails128x128/69495.png inflating: /content/faces/thumbnails128x128/69496.png inflating: /content/faces/thumbnails128x128/69497.png inflating: /content/faces/thumbnails128x128/69498.png inflating: /content/faces/thumbnails128x128/69499.png inflating: /content/faces/thumbnails128x128/69500.png inflating: /content/faces/thumbnails128x128/69501.png inflating: /content/faces/thumbnails128x128/69502.png inflating: /content/faces/thumbnails128x128/69503.png inflating: /content/faces/thumbnails128x128/69504.png inflating: /content/faces/thumbnails128x128/69505.png inflating: /content/faces/thumbnails128x128/69506.png inflating: /content/faces/thumbnails128x128/69507.png inflating: /content/faces/thumbnails128x128/69508.png inflating: /content/faces/thumbnails128x128/69509.png inflating: /content/faces/thumbnails128x128/69510.png inflating: /content/faces/thumbnails128x128/69511.png inflating: /content/faces/thumbnails128x128/69512.png inflating: /content/faces/thumbnails128x128/69513.png inflating: /content/faces/thumbnails128x128/69514.png inflating: /content/faces/thumbnails128x128/69515.png inflating: /content/faces/thumbnails128x128/69516.png inflating: /content/faces/thumbnails128x128/69517.png inflating: /content/faces/thumbnails128x128/69518.png inflating: /content/faces/thumbnails128x128/69519.png inflating: /content/faces/thumbnails128x128/69520.png inflating: /content/faces/thumbnails128x128/69521.png inflating: /content/faces/thumbnails128x128/69522.png inflating: /content/faces/thumbnails128x128/69523.png inflating: /content/faces/thumbnails128x128/69524.png inflating: /content/faces/thumbnails128x128/69525.png inflating: /content/faces/thumbnails128x128/69526.png inflating: /content/faces/thumbnails128x128/69527.png inflating: /content/faces/thumbnails128x128/69528.png inflating: /content/faces/thumbnails128x128/69529.png inflating: /content/faces/thumbnails128x128/69530.png inflating: /content/faces/thumbnails128x128/69531.png inflating: /content/faces/thumbnails128x128/69532.png inflating: /content/faces/thumbnails128x128/69533.png inflating: /content/faces/thumbnails128x128/69534.png inflating: /content/faces/thumbnails128x128/69535.png inflating: /content/faces/thumbnails128x128/69536.png inflating: /content/faces/thumbnails128x128/69537.png inflating: /content/faces/thumbnails128x128/69538.png inflating: /content/faces/thumbnails128x128/69539.png inflating: /content/faces/thumbnails128x128/69540.png inflating: /content/faces/thumbnails128x128/69541.png inflating: /content/faces/thumbnails128x128/69542.png inflating: /content/faces/thumbnails128x128/69543.png inflating: /content/faces/thumbnails128x128/69544.png inflating: /content/faces/thumbnails128x128/69545.png inflating: /content/faces/thumbnails128x128/69546.png inflating: /content/faces/thumbnails128x128/69547.png inflating: /content/faces/thumbnails128x128/69548.png inflating: /content/faces/thumbnails128x128/69549.png inflating: /content/faces/thumbnails128x128/69550.png inflating: /content/faces/thumbnails128x128/69551.png inflating: /content/faces/thumbnails128x128/69552.png inflating: /content/faces/thumbnails128x128/69553.png inflating: /content/faces/thumbnails128x128/69554.png inflating: /content/faces/thumbnails128x128/69555.png inflating: /content/faces/thumbnails128x128/69556.png inflating: /content/faces/thumbnails128x128/69557.png inflating: /content/faces/thumbnails128x128/69558.png inflating: /content/faces/thumbnails128x128/69559.png inflating: /content/faces/thumbnails128x128/69560.png inflating: /content/faces/thumbnails128x128/69561.png inflating: /content/faces/thumbnails128x128/69562.png inflating: /content/faces/thumbnails128x128/69563.png inflating: /content/faces/thumbnails128x128/69564.png inflating: /content/faces/thumbnails128x128/69565.png inflating: /content/faces/thumbnails128x128/69566.png inflating: /content/faces/thumbnails128x128/69567.png inflating: /content/faces/thumbnails128x128/69568.png inflating: /content/faces/thumbnails128x128/69569.png inflating: /content/faces/thumbnails128x128/69570.png inflating: /content/faces/thumbnails128x128/69571.png inflating: /content/faces/thumbnails128x128/69572.png inflating: /content/faces/thumbnails128x128/69573.png inflating: /content/faces/thumbnails128x128/69574.png inflating: /content/faces/thumbnails128x128/69575.png inflating: /content/faces/thumbnails128x128/69576.png inflating: /content/faces/thumbnails128x128/69577.png inflating: /content/faces/thumbnails128x128/69578.png inflating: /content/faces/thumbnails128x128/69579.png inflating: /content/faces/thumbnails128x128/69580.png inflating: /content/faces/thumbnails128x128/69581.png inflating: /content/faces/thumbnails128x128/69582.png inflating: /content/faces/thumbnails128x128/69583.png inflating: /content/faces/thumbnails128x128/69584.png inflating: /content/faces/thumbnails128x128/69585.png inflating: /content/faces/thumbnails128x128/69586.png inflating: /content/faces/thumbnails128x128/69587.png inflating: /content/faces/thumbnails128x128/69588.png inflating: /content/faces/thumbnails128x128/69589.png inflating: /content/faces/thumbnails128x128/69590.png inflating: /content/faces/thumbnails128x128/69591.png inflating: /content/faces/thumbnails128x128/69592.png inflating: /content/faces/thumbnails128x128/69593.png inflating: /content/faces/thumbnails128x128/69594.png inflating: /content/faces/thumbnails128x128/69595.png inflating: /content/faces/thumbnails128x128/69596.png inflating: /content/faces/thumbnails128x128/69597.png inflating: /content/faces/thumbnails128x128/69598.png inflating: /content/faces/thumbnails128x128/69599.png inflating: /content/faces/thumbnails128x128/69600.png inflating: /content/faces/thumbnails128x128/69601.png inflating: /content/faces/thumbnails128x128/69602.png inflating: /content/faces/thumbnails128x128/69603.png inflating: /content/faces/thumbnails128x128/69604.png inflating: /content/faces/thumbnails128x128/69605.png inflating: /content/faces/thumbnails128x128/69606.png inflating: /content/faces/thumbnails128x128/69607.png inflating: /content/faces/thumbnails128x128/69608.png inflating: /content/faces/thumbnails128x128/69609.png inflating: /content/faces/thumbnails128x128/69610.png inflating: /content/faces/thumbnails128x128/69611.png inflating: /content/faces/thumbnails128x128/69612.png inflating: /content/faces/thumbnails128x128/69613.png inflating: /content/faces/thumbnails128x128/69614.png inflating: /content/faces/thumbnails128x128/69615.png inflating: /content/faces/thumbnails128x128/69616.png inflating: /content/faces/thumbnails128x128/69617.png inflating: /content/faces/thumbnails128x128/69618.png inflating: /content/faces/thumbnails128x128/69619.png inflating: /content/faces/thumbnails128x128/69620.png inflating: /content/faces/thumbnails128x128/69621.png inflating: /content/faces/thumbnails128x128/69622.png inflating: /content/faces/thumbnails128x128/69623.png inflating: /content/faces/thumbnails128x128/69624.png inflating: /content/faces/thumbnails128x128/69625.png inflating: /content/faces/thumbnails128x128/69626.png inflating: /content/faces/thumbnails128x128/69627.png inflating: /content/faces/thumbnails128x128/69628.png inflating: /content/faces/thumbnails128x128/69629.png inflating: /content/faces/thumbnails128x128/69630.png inflating: /content/faces/thumbnails128x128/69631.png inflating: /content/faces/thumbnails128x128/69632.png inflating: /content/faces/thumbnails128x128/69633.png inflating: /content/faces/thumbnails128x128/69634.png inflating: /content/faces/thumbnails128x128/69635.png inflating: /content/faces/thumbnails128x128/69636.png inflating: /content/faces/thumbnails128x128/69637.png inflating: /content/faces/thumbnails128x128/69638.png inflating: /content/faces/thumbnails128x128/69639.png inflating: /content/faces/thumbnails128x128/69640.png inflating: /content/faces/thumbnails128x128/69641.png inflating: /content/faces/thumbnails128x128/69642.png inflating: /content/faces/thumbnails128x128/69643.png inflating: /content/faces/thumbnails128x128/69644.png inflating: /content/faces/thumbnails128x128/69645.png inflating: /content/faces/thumbnails128x128/69646.png inflating: /content/faces/thumbnails128x128/69647.png inflating: /content/faces/thumbnails128x128/69648.png inflating: /content/faces/thumbnails128x128/69649.png inflating: /content/faces/thumbnails128x128/69650.png inflating: /content/faces/thumbnails128x128/69651.png inflating: /content/faces/thumbnails128x128/69652.png inflating: /content/faces/thumbnails128x128/69653.png inflating: /content/faces/thumbnails128x128/69654.png inflating: /content/faces/thumbnails128x128/69655.png inflating: /content/faces/thumbnails128x128/69656.png inflating: /content/faces/thumbnails128x128/69657.png inflating: /content/faces/thumbnails128x128/69658.png inflating: /content/faces/thumbnails128x128/69659.png inflating: /content/faces/thumbnails128x128/69660.png inflating: /content/faces/thumbnails128x128/69661.png inflating: /content/faces/thumbnails128x128/69662.png inflating: /content/faces/thumbnails128x128/69663.png inflating: /content/faces/thumbnails128x128/69664.png inflating: /content/faces/thumbnails128x128/69665.png inflating: /content/faces/thumbnails128x128/69666.png inflating: /content/faces/thumbnails128x128/69667.png inflating: /content/faces/thumbnails128x128/69668.png inflating: /content/faces/thumbnails128x128/69669.png inflating: /content/faces/thumbnails128x128/69670.png inflating: /content/faces/thumbnails128x128/69671.png inflating: /content/faces/thumbnails128x128/69672.png inflating: /content/faces/thumbnails128x128/69673.png inflating: /content/faces/thumbnails128x128/69674.png inflating: /content/faces/thumbnails128x128/69675.png inflating: /content/faces/thumbnails128x128/69676.png inflating: /content/faces/thumbnails128x128/69677.png inflating: /content/faces/thumbnails128x128/69678.png inflating: /content/faces/thumbnails128x128/69679.png inflating: /content/faces/thumbnails128x128/69680.png inflating: /content/faces/thumbnails128x128/69681.png inflating: /content/faces/thumbnails128x128/69682.png inflating: /content/faces/thumbnails128x128/69683.png inflating: /content/faces/thumbnails128x128/69684.png inflating: /content/faces/thumbnails128x128/69685.png inflating: /content/faces/thumbnails128x128/69686.png inflating: /content/faces/thumbnails128x128/69687.png inflating: /content/faces/thumbnails128x128/69688.png inflating: /content/faces/thumbnails128x128/69689.png inflating: /content/faces/thumbnails128x128/69690.png inflating: /content/faces/thumbnails128x128/69691.png inflating: /content/faces/thumbnails128x128/69692.png inflating: /content/faces/thumbnails128x128/69693.png inflating: /content/faces/thumbnails128x128/69694.png inflating: /content/faces/thumbnails128x128/69695.png inflating: /content/faces/thumbnails128x128/69696.png inflating: /content/faces/thumbnails128x128/69697.png inflating: /content/faces/thumbnails128x128/69698.png inflating: /content/faces/thumbnails128x128/69699.png inflating: /content/faces/thumbnails128x128/69700.png inflating: /content/faces/thumbnails128x128/69701.png inflating: /content/faces/thumbnails128x128/69702.png inflating: /content/faces/thumbnails128x128/69703.png inflating: /content/faces/thumbnails128x128/69704.png inflating: /content/faces/thumbnails128x128/69705.png inflating: /content/faces/thumbnails128x128/69706.png inflating: /content/faces/thumbnails128x128/69707.png inflating: /content/faces/thumbnails128x128/69708.png inflating: /content/faces/thumbnails128x128/69709.png inflating: /content/faces/thumbnails128x128/69710.png inflating: /content/faces/thumbnails128x128/69711.png inflating: /content/faces/thumbnails128x128/69712.png inflating: /content/faces/thumbnails128x128/69713.png inflating: /content/faces/thumbnails128x128/69714.png inflating: /content/faces/thumbnails128x128/69715.png inflating: /content/faces/thumbnails128x128/69716.png inflating: /content/faces/thumbnails128x128/69717.png inflating: /content/faces/thumbnails128x128/69718.png inflating: /content/faces/thumbnails128x128/69719.png inflating: /content/faces/thumbnails128x128/69720.png inflating: /content/faces/thumbnails128x128/69721.png inflating: /content/faces/thumbnails128x128/69722.png inflating: /content/faces/thumbnails128x128/69723.png inflating: /content/faces/thumbnails128x128/69724.png inflating: /content/faces/thumbnails128x128/69725.png inflating: /content/faces/thumbnails128x128/69726.png inflating: /content/faces/thumbnails128x128/69727.png inflating: /content/faces/thumbnails128x128/69728.png inflating: /content/faces/thumbnails128x128/69729.png inflating: /content/faces/thumbnails128x128/69730.png inflating: /content/faces/thumbnails128x128/69731.png inflating: /content/faces/thumbnails128x128/69732.png inflating: /content/faces/thumbnails128x128/69733.png inflating: /content/faces/thumbnails128x128/69734.png inflating: /content/faces/thumbnails128x128/69735.png inflating: /content/faces/thumbnails128x128/69736.png inflating: /content/faces/thumbnails128x128/69737.png inflating: /content/faces/thumbnails128x128/69738.png inflating: /content/faces/thumbnails128x128/69739.png inflating: /content/faces/thumbnails128x128/69740.png inflating: /content/faces/thumbnails128x128/69741.png inflating: /content/faces/thumbnails128x128/69742.png inflating: /content/faces/thumbnails128x128/69743.png inflating: /content/faces/thumbnails128x128/69744.png inflating: /content/faces/thumbnails128x128/69745.png inflating: /content/faces/thumbnails128x128/69746.png inflating: /content/faces/thumbnails128x128/69747.png inflating: /content/faces/thumbnails128x128/69748.png inflating: /content/faces/thumbnails128x128/69749.png inflating: /content/faces/thumbnails128x128/69750.png inflating: /content/faces/thumbnails128x128/69751.png inflating: /content/faces/thumbnails128x128/69752.png inflating: /content/faces/thumbnails128x128/69753.png inflating: /content/faces/thumbnails128x128/69754.png inflating: /content/faces/thumbnails128x128/69755.png inflating: /content/faces/thumbnails128x128/69756.png inflating: /content/faces/thumbnails128x128/69757.png inflating: /content/faces/thumbnails128x128/69758.png inflating: /content/faces/thumbnails128x128/69759.png inflating: /content/faces/thumbnails128x128/69760.png inflating: /content/faces/thumbnails128x128/69761.png inflating: /content/faces/thumbnails128x128/69762.png inflating: /content/faces/thumbnails128x128/69763.png inflating: /content/faces/thumbnails128x128/69764.png inflating: /content/faces/thumbnails128x128/69765.png inflating: /content/faces/thumbnails128x128/69766.png inflating: /content/faces/thumbnails128x128/69767.png inflating: /content/faces/thumbnails128x128/69768.png inflating: /content/faces/thumbnails128x128/69769.png inflating: /content/faces/thumbnails128x128/69770.png inflating: /content/faces/thumbnails128x128/69771.png inflating: /content/faces/thumbnails128x128/69772.png inflating: /content/faces/thumbnails128x128/69773.png inflating: /content/faces/thumbnails128x128/69774.png inflating: /content/faces/thumbnails128x128/69775.png inflating: /content/faces/thumbnails128x128/69776.png inflating: /content/faces/thumbnails128x128/69777.png inflating: /content/faces/thumbnails128x128/69778.png inflating: /content/faces/thumbnails128x128/69779.png inflating: /content/faces/thumbnails128x128/69780.png inflating: /content/faces/thumbnails128x128/69781.png inflating: /content/faces/thumbnails128x128/69782.png inflating: /content/faces/thumbnails128x128/69783.png inflating: /content/faces/thumbnails128x128/69784.png inflating: /content/faces/thumbnails128x128/69785.png inflating: /content/faces/thumbnails128x128/69786.png inflating: /content/faces/thumbnails128x128/69787.png inflating: /content/faces/thumbnails128x128/69788.png inflating: /content/faces/thumbnails128x128/69789.png inflating: /content/faces/thumbnails128x128/69790.png inflating: /content/faces/thumbnails128x128/69791.png inflating: /content/faces/thumbnails128x128/69792.png inflating: /content/faces/thumbnails128x128/69793.png inflating: /content/faces/thumbnails128x128/69794.png inflating: /content/faces/thumbnails128x128/69795.png inflating: /content/faces/thumbnails128x128/69796.png inflating: /content/faces/thumbnails128x128/69797.png inflating: /content/faces/thumbnails128x128/69798.png inflating: /content/faces/thumbnails128x128/69799.png inflating: /content/faces/thumbnails128x128/69800.png inflating: /content/faces/thumbnails128x128/69801.png inflating: /content/faces/thumbnails128x128/69802.png inflating: /content/faces/thumbnails128x128/69803.png inflating: /content/faces/thumbnails128x128/69804.png inflating: /content/faces/thumbnails128x128/69805.png inflating: /content/faces/thumbnails128x128/69806.png inflating: /content/faces/thumbnails128x128/69807.png inflating: /content/faces/thumbnails128x128/69808.png inflating: /content/faces/thumbnails128x128/69809.png inflating: /content/faces/thumbnails128x128/69810.png inflating: /content/faces/thumbnails128x128/69811.png inflating: /content/faces/thumbnails128x128/69812.png inflating: /content/faces/thumbnails128x128/69813.png inflating: /content/faces/thumbnails128x128/69814.png inflating: /content/faces/thumbnails128x128/69815.png inflating: /content/faces/thumbnails128x128/69816.png inflating: /content/faces/thumbnails128x128/69817.png inflating: /content/faces/thumbnails128x128/69818.png inflating: /content/faces/thumbnails128x128/69819.png inflating: /content/faces/thumbnails128x128/69820.png inflating: /content/faces/thumbnails128x128/69821.png inflating: /content/faces/thumbnails128x128/69822.png inflating: /content/faces/thumbnails128x128/69823.png inflating: /content/faces/thumbnails128x128/69824.png inflating: /content/faces/thumbnails128x128/69825.png inflating: /content/faces/thumbnails128x128/69826.png inflating: /content/faces/thumbnails128x128/69827.png inflating: /content/faces/thumbnails128x128/69828.png inflating: /content/faces/thumbnails128x128/69829.png inflating: /content/faces/thumbnails128x128/69830.png inflating: /content/faces/thumbnails128x128/69831.png inflating: /content/faces/thumbnails128x128/69832.png inflating: /content/faces/thumbnails128x128/69833.png inflating: /content/faces/thumbnails128x128/69834.png inflating: /content/faces/thumbnails128x128/69835.png inflating: /content/faces/thumbnails128x128/69836.png inflating: /content/faces/thumbnails128x128/69837.png inflating: /content/faces/thumbnails128x128/69838.png inflating: /content/faces/thumbnails128x128/69839.png inflating: /content/faces/thumbnails128x128/69840.png inflating: /content/faces/thumbnails128x128/69841.png inflating: /content/faces/thumbnails128x128/69842.png inflating: /content/faces/thumbnails128x128/69843.png inflating: /content/faces/thumbnails128x128/69844.png inflating: /content/faces/thumbnails128x128/69845.png inflating: /content/faces/thumbnails128x128/69846.png inflating: /content/faces/thumbnails128x128/69847.png inflating: /content/faces/thumbnails128x128/69848.png inflating: /content/faces/thumbnails128x128/69849.png inflating: /content/faces/thumbnails128x128/69850.png inflating: /content/faces/thumbnails128x128/69851.png inflating: /content/faces/thumbnails128x128/69852.png inflating: /content/faces/thumbnails128x128/69853.png inflating: /content/faces/thumbnails128x128/69854.png inflating: /content/faces/thumbnails128x128/69855.png inflating: /content/faces/thumbnails128x128/69856.png inflating: /content/faces/thumbnails128x128/69857.png inflating: /content/faces/thumbnails128x128/69858.png inflating: /content/faces/thumbnails128x128/69859.png inflating: /content/faces/thumbnails128x128/69860.png inflating: /content/faces/thumbnails128x128/69861.png inflating: /content/faces/thumbnails128x128/69862.png inflating: /content/faces/thumbnails128x128/69863.png inflating: /content/faces/thumbnails128x128/69864.png inflating: /content/faces/thumbnails128x128/69865.png inflating: /content/faces/thumbnails128x128/69866.png inflating: /content/faces/thumbnails128x128/69867.png inflating: /content/faces/thumbnails128x128/69868.png inflating: /content/faces/thumbnails128x128/69869.png inflating: /content/faces/thumbnails128x128/69870.png inflating: /content/faces/thumbnails128x128/69871.png inflating: /content/faces/thumbnails128x128/69872.png inflating: /content/faces/thumbnails128x128/69873.png inflating: /content/faces/thumbnails128x128/69874.png inflating: /content/faces/thumbnails128x128/69875.png inflating: /content/faces/thumbnails128x128/69876.png inflating: /content/faces/thumbnails128x128/69877.png inflating: /content/faces/thumbnails128x128/69878.png inflating: /content/faces/thumbnails128x128/69879.png inflating: /content/faces/thumbnails128x128/69880.png inflating: /content/faces/thumbnails128x128/69881.png inflating: /content/faces/thumbnails128x128/69882.png inflating: /content/faces/thumbnails128x128/69883.png inflating: /content/faces/thumbnails128x128/69884.png inflating: /content/faces/thumbnails128x128/69885.png inflating: /content/faces/thumbnails128x128/69886.png inflating: /content/faces/thumbnails128x128/69887.png inflating: /content/faces/thumbnails128x128/69888.png inflating: /content/faces/thumbnails128x128/69889.png inflating: /content/faces/thumbnails128x128/69890.png inflating: /content/faces/thumbnails128x128/69891.png inflating: /content/faces/thumbnails128x128/69892.png inflating: /content/faces/thumbnails128x128/69893.png inflating: /content/faces/thumbnails128x128/69894.png inflating: /content/faces/thumbnails128x128/69895.png inflating: /content/faces/thumbnails128x128/69896.png inflating: /content/faces/thumbnails128x128/69897.png inflating: /content/faces/thumbnails128x128/69898.png inflating: /content/faces/thumbnails128x128/69899.png inflating: /content/faces/thumbnails128x128/69900.png inflating: /content/faces/thumbnails128x128/69901.png inflating: /content/faces/thumbnails128x128/69902.png inflating: /content/faces/thumbnails128x128/69903.png inflating: /content/faces/thumbnails128x128/69904.png inflating: /content/faces/thumbnails128x128/69905.png inflating: /content/faces/thumbnails128x128/69906.png inflating: /content/faces/thumbnails128x128/69907.png inflating: /content/faces/thumbnails128x128/69908.png inflating: /content/faces/thumbnails128x128/69909.png inflating: /content/faces/thumbnails128x128/69910.png inflating: /content/faces/thumbnails128x128/69911.png inflating: /content/faces/thumbnails128x128/69912.png inflating: /content/faces/thumbnails128x128/69913.png inflating: /content/faces/thumbnails128x128/69914.png inflating: /content/faces/thumbnails128x128/69915.png inflating: /content/faces/thumbnails128x128/69916.png inflating: /content/faces/thumbnails128x128/69917.png inflating: /content/faces/thumbnails128x128/69918.png inflating: /content/faces/thumbnails128x128/69919.png inflating: /content/faces/thumbnails128x128/69920.png inflating: /content/faces/thumbnails128x128/69921.png inflating: /content/faces/thumbnails128x128/69922.png inflating: /content/faces/thumbnails128x128/69923.png inflating: /content/faces/thumbnails128x128/69924.png inflating: /content/faces/thumbnails128x128/69925.png inflating: /content/faces/thumbnails128x128/69926.png inflating: /content/faces/thumbnails128x128/69927.png inflating: /content/faces/thumbnails128x128/69928.png inflating: /content/faces/thumbnails128x128/69929.png inflating: /content/faces/thumbnails128x128/69930.png inflating: /content/faces/thumbnails128x128/69931.png inflating: /content/faces/thumbnails128x128/69932.png inflating: /content/faces/thumbnails128x128/69933.png inflating: /content/faces/thumbnails128x128/69934.png inflating: /content/faces/thumbnails128x128/69935.png inflating: /content/faces/thumbnails128x128/69936.png inflating: /content/faces/thumbnails128x128/69937.png inflating: /content/faces/thumbnails128x128/69938.png inflating: /content/faces/thumbnails128x128/69939.png inflating: /content/faces/thumbnails128x128/69940.png inflating: /content/faces/thumbnails128x128/69941.png inflating: /content/faces/thumbnails128x128/69942.png inflating: /content/faces/thumbnails128x128/69943.png inflating: /content/faces/thumbnails128x128/69944.png inflating: /content/faces/thumbnails128x128/69945.png inflating: /content/faces/thumbnails128x128/69946.png inflating: /content/faces/thumbnails128x128/69947.png inflating: /content/faces/thumbnails128x128/69948.png inflating: /content/faces/thumbnails128x128/69949.png inflating: /content/faces/thumbnails128x128/69950.png inflating: /content/faces/thumbnails128x128/69951.png inflating: /content/faces/thumbnails128x128/69952.png inflating: /content/faces/thumbnails128x128/69953.png inflating: /content/faces/thumbnails128x128/69954.png inflating: /content/faces/thumbnails128x128/69955.png inflating: /content/faces/thumbnails128x128/69956.png inflating: /content/faces/thumbnails128x128/69957.png inflating: /content/faces/thumbnails128x128/69958.png inflating: /content/faces/thumbnails128x128/69959.png inflating: /content/faces/thumbnails128x128/69960.png inflating: /content/faces/thumbnails128x128/69961.png inflating: /content/faces/thumbnails128x128/69962.png inflating: /content/faces/thumbnails128x128/69963.png inflating: /content/faces/thumbnails128x128/69964.png inflating: /content/faces/thumbnails128x128/69965.png inflating: /content/faces/thumbnails128x128/69966.png inflating: /content/faces/thumbnails128x128/69967.png inflating: /content/faces/thumbnails128x128/69968.png inflating: /content/faces/thumbnails128x128/69969.png inflating: /content/faces/thumbnails128x128/69970.png inflating: /content/faces/thumbnails128x128/69971.png inflating: /content/faces/thumbnails128x128/69972.png inflating: /content/faces/thumbnails128x128/69973.png inflating: /content/faces/thumbnails128x128/69974.png inflating: /content/faces/thumbnails128x128/69975.png inflating: /content/faces/thumbnails128x128/69976.png inflating: /content/faces/thumbnails128x128/69977.png inflating: /content/faces/thumbnails128x128/69978.png inflating: /content/faces/thumbnails128x128/69979.png inflating: /content/faces/thumbnails128x128/69980.png inflating: /content/faces/thumbnails128x128/69981.png inflating: /content/faces/thumbnails128x128/69982.png inflating: /content/faces/thumbnails128x128/69983.png inflating: /content/faces/thumbnails128x128/69984.png inflating: /content/faces/thumbnails128x128/69985.png inflating: /content/faces/thumbnails128x128/69986.png inflating: /content/faces/thumbnails128x128/69987.png inflating: /content/faces/thumbnails128x128/69988.png inflating: /content/faces/thumbnails128x128/69989.png inflating: /content/faces/thumbnails128x128/69990.png inflating: /content/faces/thumbnails128x128/69991.png inflating: /content/faces/thumbnails128x128/69992.png inflating: /content/faces/thumbnails128x128/69993.png inflating: /content/faces/thumbnails128x128/69994.png inflating: /content/faces/thumbnails128x128/69995.png inflating: /content/faces/thumbnails128x128/69996.png inflating: /content/faces/thumbnails128x128/69997.png inflating: /content/faces/thumbnails128x128/69998.png inflating: /content/faces/thumbnails128x128/69999.png ###Markdown Load Data utility functionAs there are around 70k images and cannot fit on RAM so specify it according to your RAM size ###Code PATH=os.path.join("faces/thumbnails128x128") def load_data(path,upper_limit=20000,lower_limit=0): data=[] files=os.listdir(PATH) for file in files[lower_limit:upper_limit]: image=Image.open(os.path.join(path,file)) image=np.array(image) image=image/255.0 data.append(image) return np.array(data) ###Output _____no_output_____ ###Markdown Loading train and test images ###Code colored_train=load_data(PATH,lower_limit=35000,upper_limit=40000) colored_test=load_data(PATH,lower_limit=15000,upper_limit=18000) print(colored_train.shape) print(colored_test.shape) ###Output (5000, 128, 128, 3) (3000, 128, 128, 3) ###Markdown Converting RGB color space to Lab color space ###Code lab_train=rgb2lab(colored_train) lab_test=rgb2lab(colored_test) ###Output _____no_output_____ ###Markdown First Channel for input image for network ###Code input_train=lab_train[:,:,:,0] input_test=lab_test[:,:,:,0] ###Output _____no_output_____ ###Markdown Reshaping grayscale image So that it have 4 dimensions to feed to convolutional neural network ###Code input_train=input_train.reshape(-1,128,128,1) input_test=input_test.reshape(-1,128,128,1) ###Output _____no_output_____ ###Markdown Remaining 2 Channel for output for network ###Code output_train=lab_train[:,:,:,1:] output_test=lab_test[:,:,:,1:] output_train=output_train/128.0 # feature scaling output_test=output_test/128.0 print(output_train.shape) print(input_train.shape) print(output_test.shape) print(input_test.shape) ###Output (5000, 128, 128, 2) (5000, 128, 128, 1) (3000, 128, 128, 2) (3000, 128, 128, 1) ###Markdown Utility Functions 1. merge_channels: merge the L channel (input) with ab channel (output)2. decoded_images: do merge_channels for all decoder output image3. plot_images: plot images using numpy ###Code def merge_channels(input_L,output_ab): image=np.zeros(shape=(input_L.shape[0],input_L.shape[1],3)) image[:,:,0]=np.squeeze(input_L) image[:,:,1:]=output_ab return image def decoded_images(encoded_input,decoded_output): assert len(encoded_input)==len(decoded_output) images=[] for i in range(len(encoded_input)): img=merge_channels(encoded_input[i],decoded_output[i]*128.0) img=lab2rgb(img) images.append(img) return np.array(images) def plot_images(images,w=3,h=3,cmap=None): images=np.squeeze(images) fig, axes = plt.subplots(w, h) fig.subplots_adjust(hspace=0.3, wspace=0.3) for i, ax in enumerate(axes.flat): if cmap is not None: ax.imshow(images[i],cmap=cmap) else: ax.imshow(images[i]) ax.set_xticks([]) ax.set_yticks([]) plt.show() images=input_train[0:9] plot_images(images,cmap="gray",w=3,h=3) images=output_train[0:9][:,:,:,0] plot_images(images,cmap="gray",w=3,h=3) images=output_train[0:9][:,:,:,1] plot_images(images,cmap="gray",w=3,h=3) ###Output _____no_output_____ ###Markdown Lets test how our merge_channels function works ###Code images=decoded_images(input_train[0:9],output_train[0:9]) plot_images(images,w=3,h=3) ###Output _____no_output_____ ###Markdown Encoder and Decoder creationour encoder and decoder consists of convolutional and convolution transpose layers ###Code bottleneck_unit=8 inputs=keras.layers.Input(shape=(128,128,1),name="input") enc=keras.layers.Conv2D(filters=512,kernel_size=(3,3),strides=2,padding="same",activation="relu",name="encoder1")(inputs) enc=keras.layers.Conv2D(filters=256,kernel_size=(3,3),strides=1,padding="same",activation="relu",name="encoder2")(enc) enc=keras.layers.Conv2D(filters=128,kernel_size=(3,3),strides=1,padding="same",activation="relu",name="encoder3")(enc) enc=keras.layers.Conv2D(filters=64,kernel_size=(3,3),strides=2,padding="same",activation="relu",name="encoder4")(enc) enc=keras.layers.Conv2D(filters=32,kernel_size=(3,3),strides=1,padding="same",activation="relu",name="encoder5")(enc) enc=keras.layers.Conv2D(filters=16,kernel_size=(3,3),strides=2,padding="same",activation="relu",name="encoder6")(enc) bottleneck=keras.layers.Conv2D(filters=bottleneck_unit,kernel_size=(3,3),strides=2,padding="same",activation="relu",name="bottleneck")(enc) ###Output _____no_output_____ ###Markdown check if we have trained model on our gdrive to further tuning ###Code if os.path.isfile(os.path.join("..","gdrive","My Drive","gray2color","encoder.h5")): encoder=keras.models.load_model(os.path.join("..","gdrive","My Drive","gray2color","encoder.h5"),compile=True) print("resuming older model") else: encoder=keras.Model(inputs=inputs,outputs=bottleneck) print("No precious model found") encoder.summary() encoded_input=keras.layers.Input(shape=(8,8,bottleneck_unit)) dec=keras.layers.Conv2DTranspose(filters=16,strides=2,kernel_size=(3,3),padding="same",activation="relu",name="decoder1")(encoded_input) dec=keras.layers.Conv2DTranspose(filters=32,strides=1,kernel_size=(3,3),padding="same",activation="relu",name="decoder2")(dec) dec=keras.layers.Conv2DTranspose(filters=64,strides=2,kernel_size=(3,3),padding="same",activation="relu",name="decoder3")(dec) dec=keras.layers.Conv2DTranspose(filters=128,strides=1,kernel_size=(3,3),padding="same",activation="relu",name="decoder4")(dec) dec=keras.layers.Conv2DTranspose(filters=256,strides=1,kernel_size=(3,3),padding="same",activation="relu",name="decoder5")(dec) dec=keras.layers.Conv2DTranspose(filters=512,strides=2,kernel_size=(3,3),padding="same",activation="relu",name="decoder6")(dec) outputs=keras.layers.Conv2DTranspose(filters=2,strides=2,kernel_size=(3,3),padding="same",activation="tanh",name="output")(dec) if os.path.isfile(os.path.join("..","gdrive","My Drive","gray2color","decoder.h5")): decoder=keras.models.load_model(os.path.join("..","gdrive","My Drive","gray2color","decoder.h5")) print("resuming older model") else: decoder=keras.Model(inputs=encoded_input,outputs=outputs) print("No precious model found") decoder.summary() ###Output Model: "model_1" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= input_1 (InputLayer) [(None, 8, 8, 8)] 0 _________________________________________________________________ decoder1 (Conv2DTranspose) (None, 16, 16, 16) 1168 _________________________________________________________________ decoder2 (Conv2DTranspose) (None, 16, 16, 32) 4640 _________________________________________________________________ decoder3 (Conv2DTranspose) (None, 32, 32, 64) 18496 _________________________________________________________________ decoder4 (Conv2DTranspose) (None, 32, 32, 128) 73856 _________________________________________________________________ decoder5 (Conv2DTranspose) (None, 32, 32, 256) 295168 _________________________________________________________________ decoder6 (Conv2DTranspose) (None, 64, 64, 512) 1180160 _________________________________________________________________ output (Conv2DTranspose) (None, 128, 128, 2) 9218 ================================================================= Total params: 1,582,706 Trainable params: 1,582,706 Non-trainable params: 0 _________________________________________________________________ ###Markdown Creating our AutoEncoder model ###Code class AutoEncoder(keras.Model): def __init__(self,encoder,decoder): super(AutoEncoder, self).__init__() self.encoder=encoder self.decoder=decoder def call(self,inputs): m=self.encoder.layers[0](inputs) for i in range(1,len(self.encoder.layers)): m=self.encoder.layers[i](m) for i in range(1,len(self.decoder.layers)): m=self.decoder.layers[i](m) return m autoencoder=AutoEncoder(encoder,decoder) autoencoder.compile(optimizer="adam",loss="mse") reduce_lr=tf.keras.callbacks.ReduceLROnPlateau( monitor="loss", factor=0.1, patience=10, verbose=0, mode="auto", min_delta=0.0001, cooldown=0, min_lr=0 ) ###Output _____no_output_____ ###Markdown Training the model ###Code epochs=10 history=autoencoder.fit(input_train,output_train,batch_size=32,epochs=epochs,shuffle=True,callbacks=[reduce_lr]) ###Output Epoch 1/10 157/157 [==============================] - 21s 134ms/step - loss: 0.0095 - lr: 0.0010 Epoch 2/10 157/157 [==============================] - 21s 131ms/step - loss: 0.0093 - lr: 0.0010 Epoch 3/10 157/157 [==============================] - 21s 131ms/step - loss: 0.0092 - lr: 0.0010 Epoch 4/10 157/157 [==============================] - 21s 131ms/step - loss: 0.0091 - lr: 0.0010 Epoch 5/10 157/157 [==============================] - 21s 131ms/step - loss: 0.0090 - lr: 0.0010 Epoch 6/10 157/157 [==============================] - 21s 131ms/step - loss: 0.0089 - lr: 0.0010 Epoch 7/10 157/157 [==============================] - 21s 131ms/step - loss: 0.0088 - lr: 0.0010 Epoch 8/10 157/157 [==============================] - 21s 131ms/step - loss: 0.0086 - lr: 0.0010 Epoch 9/10 157/157 [==============================] - 21s 131ms/step - loss: 0.0085 - lr: 0.0010 Epoch 10/10 157/157 [==============================] - 20s 130ms/step - loss: 0.0083 - lr: 0.0010 ###Markdown Testing the model1. latent vectors are spit out by encoder2. these vectors are passes to decoder to generate colored images ###Code latent_vectors=encoder.predict(input_test) decoded_test_images=decoder.predict(latent_vectors) val_loss=keras.losses.mse(output_test,decoded_test_images) print(tf.reduce_mean(val_loss).numpy()) ###Output 0.009874075 ###Markdown Plotting test output images and original test colored images ###Code images=decoded_images(input_test[0:20],decoded_test_images[0:20]) plot_images(images) images=colored_test[0:20] plot_images(images) ###Output _____no_output_____ ###Markdown Saving the model to local or gdrive ###Code encoder.save("encoder.h5") decoder.save("decoder.h5") encoder.save("/gdrive/My Drive/gray2color/encoder.h5") decoder.save("/gdrive/My Drive/gray2color/decoder.h5") !ls '/gdrive/My Drive/' ###Output 'Colab Notebooks' dogs_classifier_mobilenetv2.h5 Data gray2color datasets Images dogs_classifier_mobilenet1.h5 Keystores dogs_classifier_mobilenet.h5 tarun_bisht_coat11.jpg
Regular Expressions.ipynb
###Markdown Regular Expressions* The syntax for the regex library is to always to pass the pattern first, and then the string secondReferences* [Introduction to NLP in Python](https://campus.datacamp.com/courses/introduction-to-natural-language-processing-in-python/regular-expressions-word-tokenization) ###Code import re ###Output _____no_output_____ ###Markdown SearchSearch will go through the entire string ###Code match = re.search(r"\d+", "M123") print("Match:", match) print("Start:", match.start()) print("End:", match.end()) re.search(r"^([A-Z])", "M123").group(1) ###Output _____no_output_____ ###Markdown MatchWhereas match will start at the beginning ###Code re.match(r"\d+", "M123") # No match ###Output _____no_output_____ ###Markdown Split ###Code re.split(r"\s+", "Machine learning is fun") ###Output _____no_output_____ ###Markdown Find All ###Code re.findall(r"[A-Z]+", "THIS documentation ROCKS") ###Output _____no_output_____ ###Markdown Regular Expression Quick Guide^ Matches the beginning of the line$ Matches the end of the line. Matches any character\s Matches whitespace\S Matches any non-whitespace character* Repeats a character zero or more times*? Repeats a character zero or more times (non-greedy)+ Repeats a character one or more times+? Repeats a character one or more times (non-greedy)[aeiou] Matches a single character in the listed set[^XYZ] Matches a single character not in the listed set[a-z0-9] The set of characters can include a range( Indicates where string extraction is to start) Indicates where string extraction is to end{}[]\d any digit\D anything except for a digit\w equals to [a-zA-Z0-9_]\W equals to [^a-zA-Z0-9_] ###Code Class Recording 8 from 56:13. Edureka ###Output _____no_output_____ ###Markdown A character can be a-z, A-Z, 0-9, special chars(~!@$%^&*), spaces hare = open('C:\\Users\\HP\\Desktop\\Sudheer DS\\LearnDataScience-master\\Python_Handouts\\hare.txt','w+')hare.write("This notepad file is created using Anaconda")hare.write(" This is a great day")hare.seek(0)print(hare.read())hare.close() ###Code # find(), re.search(), startswith() ###Output _____no_output_____ ###Markdown Using re.search() like find()hand = open('mbox-short.txt')for line in hand: line = line.rstrip() if line.find('From:') >= 0: print(line) Using re.search() like find()import rehand = open('mbox-short.txt')for line in hand: line = line.rstrip() if re.search('From:',line): print(line) Using re.search() like startswith()hand = open('mbox-short.txt')for line in hand: line = line.rstrip() if line.startswith('From:'): print(line) Using re.search() like startswith() Ignoring import re and came to know that it is not necessary to call on the library everytime.hand = open('mbox-short.txt')for line in hand: line = line.rstrip() if re.search('^From:',line): Observe the previous codes ^ in not necessary also here. print(line) Wild-card charactersimport rehand = open('rexp.txt')for line in hand: line = line.rstrip()if re.search('^X.*:', line) : print(line) Fine tuning your Matchhand = open('rexp.txt')for line in hand: line = line.rstrip()if re.search('^X-\S+:', line) : print(line) Matching and Extracting Dataimport rex = 'My 2 favorite numbers are 19 and 42'y = re.findall('[0-9]+',x)print(y) import rex = 'My 2 favorite numbers are 19 and 42'y = re.findall('[0-9]+',x) re.findall always returns a listprint(y) Experimentimport rex = 'My 2 favorite numbers are 19 and 42'y = re.findall('[0-9]',x) re.findall always returns a listprint(y) Experimentimport rex = 'My 2 favorite numbers are 19 and 42'y = re.findall('[0-9]*',x) re.findall always returns a listprint(y) Experiment - It Workedimport rex = 'My 2 favorite numbers are 19 and 42'y = re.findall('[^A-z\s]+',x) re.findall always returns a listprint(y) y = re.findall('[AEIOU]+',x)print(y) Experimenty = re.findall('[AEIOUM]+',x)print(y) Experimentx = 'My 2 favorite numbers are 19 and 42'y = re.findall('[A-z]+',x)print(y) Warning: Greedy Matchingimport rex = 'From: Using the : character'y = re.findall('^F.+:', x) the last colon of the stringprint(y) Non-Greedy Matchingimport rex = 'From: Using the : character'y = re.findall('^F.+?:',x) the first colon of the stringprint(y) Fine-Tuning String Extractionx = 'From [email protected] Sat Jan 5 09:14:16 2008'y= re.findall('\S+@\S+',x)print(y) Fine-Tuning String Extractionx = 'From [email protected] Sat Jan 5 09:14:16 2008'y= re.findall('\S+@\S+',x)print(y)y = re.findall('^From:.*? (\S+@\S+)',x)print(y) data = 'From [email protected] Sat Jan 5 09:14:16 2008'atpos = data.find('.')print(atpos) data = 'From [email protected] Sat Jan 5 09:14:16 2008'atpos = data.find('o')print(atpos) data = 'From [email protected] Sat Jan 5 09:14:16 2008'atpos = data.find('@')print(atpos) Denotes space which comes after atpos - here it is "@"sppos = data.find(' ',atpos)print(sppos) host = data[atpos+1 : sppos]print(host) The Double Split Patternline = 'From [email protected] Sat Jan 5 09:14:16 2008'words = line.split()words email = words[1]email pieces = email.split('@')pieces print(pieces[1]) The Regex Version import relin = 'From [email protected] Sat Jan 5 09:14:16 2008'y = re.findall('@([^ ]*)',lin) [^ ] Match non blank character [^ ]* Match non blank character Match many of themprint(y) Experimentimport relin = 'From [email protected] Sat Jan 5 09:14:16 2008'y = re.findall('@([^ ]+)',lin) [^ ] Match non blank character [^ ]* Match non blank character Match many of themprint(y) Even Cooler Regex Versionimport relin = 'From [email protected] Sat Jan 5 09:14:16 2008'y = re.findall('^From .*@([^ ]*)',lin) ( Start extracting ) Stop extractingprint(y) Result will always be a extracted portion Spam Confidenceimport rehand = open('mbox-short.txt')numlist = list()for line in hand: line = line.rstrip() stuff = re.findall('^X-DSPAM-Confidence: ([0-9.]+)',line) if len(stuff) !=1 : continue num = float(stuff[0]) numlist.append(num)print('Maximum:', max(numlist)) Escape Character If you want a special regular expression character to just behave normally(most of the time) you prefix it with '\'import rex = 'We just received $10.00 for cookies.'y = re.findall('\$[0-9.]+',x)print(y) ###Code wazzzzzup = 'waz{3,5}up' wazzzup = 'waz{3,5}up' but does not match wazup as there is only one z aa+b*c+ a's atleast two or more b atleast 0 or more c atleast one or more aaaabcc aabbbbc aacc but does not match a \d+ files? found\? \d+ to match number of more than one digit files? to match file word and 's' zero or one time found\? to match found word and \ treats ? as character and not as regex shortcut 1 file found? 2 files found? 24 files found? but does not match No files found.[] ###Output _____no_output_____ ###Markdown [1-9]\.\s+abc[1-9] matches any digit from 1 to 9\. treats . as a character instead of regex shortcut\s+ whitespace character one or more timesabc letters abc1. abc2. abc3. abcbut does not match4.abc '^(file.+)\.pdf$'file_record_transcript.pdffile_0.7241999.pdf but does not matchtestfile_fake.pdf.tmp ([A-Za-z]+ ([0-9]+))[A-Za-z]+ Starting with capital Letters between A-Z followed by a-z one or manyfollowed by space[0-9]+ One or more digits\dJan 1987May 1969Aug 2011 I love (cats|dogs)I love catsI love dogsdoes not matchI love logsI love cogs ^-?\d+(,\d+)*(\.\d+(e\d+)?)?$^-? starts with - or not ^-?\d+ starts with - or not, followed by one or more digits(,\d+)* comma, then followed by one or more digits ------------ this pattern zero or more times(\.\d+(e\d+)?)? followed by a "." and then by one or more digits, then followed by letter "e" and then followed by 1 or more digits --------- this pattern present or notabove two lines present or not.3.14529-255.341281.9e10123,340.00but does not match720p ###Code import re p = re.compile(r'\sclass\s') print(p.search('no class at all')) # Takes class along with spaces p = re.compile(r'\bclass\b') print(p.search('no class at all')) # Takes class without spaces ###Output <re.Match object; span=(2, 9), match=' class '> <re.Match object; span=(3, 8), match='class'> ###Markdown Regular Expressionss = regex or regexp is essentially a search query for text that's expressed by string patternWhen you run a search against a particular piece of text, anything that matches a regular expression pattern you specified, is returned as a result of the search. Regular expressions let you answer the questions like what are all the four-letter words in a file? Or how many different error types are there in this error log? ###Code import re log = "July 31 07:51:48 mycomputer bad_process[12345]: ERROR Performing package upgrade" regex1 = r"\[(\d+)]" result =re.search(regex1, log) print(result[1]) ###Output 12345 ###Markdown grep() = command line regex tool ###Code import re result = re.search(r"aza","plaza") #always use raw string on regular python print(result) result = re.search(r"aza","bazaar") print(result) result = re.search(r"aza","market") print(result) print(re.search(r"p.ng","penguin")) print(re.search(r"^aag","aagtwg")) import re def check_aei (text): result = re.search(r"a.e.i", text) return result != None print(check_aei("academia")) # True print(check_aei("aerial")) # False print(check_aei("paramedic")) # True print(re.search(r"a.t","aagTwg", re.IGNORECASE)) ###Output <re.Match object; span=(1, 4), match='agT'> ###Markdown Wildcards and Char Classes ###Code print(re.search(r"[Pp]ython","Python")) print(re.search(r"[a-z]ython","Python")) print(re.search(r"[a-z]on","Prett cool qython")) print(re.search("cloud[a-zA-Z0-9]on","cloud2on")) import re def check_punctuation (text): result = re.search(r"[!-),-/<-?]", text) return result != None print(check_punctuation("This is a sentence that ends with a period.")) # True print(check_punctuation("This is a sentence fragment without a period")) # False print(check_punctuation("Aren't regular expressions awesome?")) # True print(check_punctuation("Wow! We're really picking up some steam now!")) # True print(check_punctuation("End of the line")) # False print(re.search(r"[^a-zA-Z]","This is a sentence with a space")) print(re.search(r"[^a-zA-Z ]","This is a sentence with a space.")) #because we add space, so it exclude the char inside[] print(re.search(r"[^a-zA-Z ]","This is a sentence with a space2")) #because we add space, so it exclude the char inside[] print(re.search(r"cat|dog","This is sentence with a space catttin and hotdog.")) #because we add space, so it exclude the char inside[] print(re.findall(r"cat|dog [a-z]","This is a sentence with a space dog cat dog cat cata cata.")) #because we add space, so it exclude the char inside[] ###Output ['dog c', 'dog c', 'cat', 'cat'] ###Markdown Repetition Qualifiers ###Code print(re.search(r"py.*y","pylahropy")) #the stars take the between leters print(re.search(r"o+l+","wool woool lllle lily nonon")) print(re.search(r"p?each","I like each")) print(re.search(r"p?each","I like peach")) ###Output <re.Match object; span=(7, 12), match='peach'>
015_Seaborn_Factor_Plot.ipynb
###Markdown All the IPython Notebooks in this lecture series by Dr. Milan Parmar are available @ **[GitHub](https://github.com/milaan9/12_Python_Seaborn_Module)** Seaborn: Factor Plot Welcome back to another lecture on *Data Visualization with Seaborn*! This lecture is kind of a continuation to **FacetGrid** that we had been discussing in previous lecture. Today our major emphasis is once again going to be on plotting *Categorical Data*. Well, you might think that we have already done enough of these in previous section, when we covered visualization methods like **Swarm Plot**, **Strip Plot**, **Box Plot**, **Violin Plot**, **Bar Plot** and **Point Plot**. News for you today is that all the above mentioned plots are generally considered *low-level methods* as they all plot onto a specific pair of *Matplotlib axes*.Today we shall discuss a higher level function, i.e. **Factor Plot**, which combines all the *low-level functions* with our **FacetGrid** to apply a *Categorical plot* across a grid of figure panels on **[Tidy DataFrame](http://vita.had.co.nz/papers/tidy-data.pdf)**. I shall attach a link in our notebook for you to better assess *Tidy Data* as defined by official page. Let us now dive little deeper to understand this *high-level* **Factor Plot**; not actually in terms of underlying code but in terms of the conceptual foundation, where it majorly holds relevance to **Factor Analysis**. To give you an overview, **[Factor Analysis](https://en.wikipedia.org/wiki/Factor_analysis)** is again a *statistical method* that describes *variability* among observed, correlated variables in terms of a potentially lower number of unobserved variables, which are referred to as **Factors**. **Factor Analysis** searches for similar Joint variations in response to an unobserved set of **[Latent variables](https://en.wikipedia.org/wiki/Latent_variable)**. The term **Latent** refers to the fact that even though these variables were not measured directly in a research design, still they are the ultimate goal of that project. Hence, the observed variables are modelled as *Linear combinations of potential factors*, plus *"error"* terms. Our Factor analysis aims to find such independent Latent variables and the theory behind these methods is that the *Information* gained about the inter-dependencies between observed variables can be used later to reduce the set of variables in a dataset. In short, we may say that Factor analysis is related to **Principal Component Analysis (PCA)**, though those two are not identical. During visualization, A *Factor plot* simply drafts the same plot generated for different response and factor variables and arranged on a single page. Here, the underlying plot generated can be any *Univariate* or *Bivariate* plot, and *Scatter Plot* serves this purpose quite frequently than others.Let us now get our package dependancies and plot a simple Factor Plot to understand the parameters offered by Seaborn to make our task easier: ###Code # Importing intrinsic libraries: import numpy as np import pandas as pd np.random.seed(44) import matplotlib.pyplot as plt import seaborn as sns %matplotlib inline sns.set(style="ticks", palette="hsv") import warnings warnings.filterwarnings("ignore") # Let us also get tableau colors we defined earlier: tableau_20 = [(31, 119, 180), (174, 199, 232), (255, 127, 14), (255, 187, 120), (44, 160, 44), (152, 223, 138), (214, 39, 40), (255, 152, 150), (148, 103, 189), (197, 176, 213), (140, 86, 75), (196, 156, 148), (227, 119, 194), (247, 182, 210), (127, 127, 127), (199, 199, 199), (188, 189, 34), (219, 219, 141), (23, 190, 207), (158, 218, 229)] # Scaling above RGB values to [0, 1] range, which is Matplotlib acceptable format: for i in range(len(tableau_20)): r, g, b = tableau_20[i] tableau_20[i] = (r / 255., g / 255., b / 255.) # Loading Built-in Dataset: exercise = sns.load_dataset("exercise") # Pre-viewing Dataset: exercise.head(10) #exercise.columns # Creating a basic Factor Plot: sns.factorplot(x="time", y="pulse", hue="kind", data=exercise, size=6) ###Output _____no_output_____ ###Markdown This looks quite informative and we already know how to interpret a **Point Plot** that we have on screen right now. If an individual is at **`rest`**, his/her **`pulse`** remains pretty constant at approximately 90, when measured for a **`time`** interval from 1 minute to half an hour. Even while **`walking`**, the **`pulse`** soars high for first **`15 minutes`**, but then stabilizes around 93, and then remains constant at that **`pulse`**. But the story is totally different when the individual is **`running`**, because the pulse then takes a major upwards leap in first **`time`** segment, and constantly keeps pounding with increase in **`time`**. Let us now look at the parameters offered by Seaborn to expand our horizon with **Factor Plot**:**`seaborn.factorplot(x=None, y=None, hue=None, data=None, row=None, col=None, col_wrap=None, estimator=, ci=95, n_boot=1000, units=None, order=None, hue_order=None, row_order=None, col_order=None, kind='point', size=4, aspect=1, orient=None, color=None, palette=None, legend=True, legend_out=True, sharex=True, sharey=True, margin_titles=False, facet_kws=None)`**The good news is that Seaborn seems to offer almost all the *optional parameters* that we've covered till now; and the other good news is that there isn't any extra parameter for us to fiddle with. So instead, let us play around with few more *Factor Plots* to visualize the difference. As of now, we just have a **Point Plot** on one facet, so we shall eventually even try to draw subplots to get further acquainted with the syntax and corresponding results. There isn't much we need to do in terms of *inference*, so let us run through few examples: ###Code # Let us begin by altering the type of plot to a "BarPlot" on our FactorPlot: sns.factorplot(x="time", y="pulse", hue="kind", data=exercise, size=7, kind="bar", palette="rocket") ###Output _____no_output_____ ###Markdown As you would have guessed by now, the **`kind`** parameter by default is for **Point Plot**, but just like this, we may modify it to **`box`**, **`violin`**, **`swarm`**, etc.Let us now *facet our plot along variable columns in the same row*: ###Code # For a change, here we shall use a "Box Plot", instead of a "Bar Plot" to visualize the difference: #sns.factorplot(x="time", y="pulse", hue="kind", col="diet", data=exercise, size=6, kind="box", palette="rocket") # Let us pull our Legend inside the plot: sns.factorplot(x="time", y="pulse", hue="kind", col="diet", data=exercise, size=7, kind="box", palette="rocket", legend_out=False) ###Output _____no_output_____ ###Markdown That comfortably divides **`exercise`** datapoints with respect to the *factor* whether they are on **`low fat`** *or* on **`no fat`** *diet*. Now let us add in few more *optional parameters* and tweak our presentation, and for this purpose we shall use our *Tips* dataset, so we shall commence by reloading this built-in dataset: ###Code # Loading Built-in Tips Dataset: tips = sns.load_dataset("tips") # Let us get all the facets of our grid vertically stacked this time: sns.factorplot(x="day", y="total_bill", hue="smoker", row="time", data=tips, orient="v", size=4, aspect=2.5, palette="bwr", kind="point", dodge=True, cut=0, bw=.2, margin_titles=True) ###Output _____no_output_____ ###Markdown Hmmm! So we have a competent plot here presenting the variations. Let us now try to extemporize few modifications in our **Barplot** that we plotted earlier using **`exercise`** dataset: ###Code # Let us also assign a variable "ax" to it: ax = sns.factorplot(x="time", y="pulse", hue="kind", col="diet", data=exercise, size=7, kind="bar", palette="rocket", ci=None) # Let us now customize it by using methods on our FacetGrid: ax.set_axis_labels("Time Taken", "Pulse Rate") ax.set_xticklabels(["1 minute", "15 minutes", "30 minutes"]) ax.set_titles("{col_name} {col_var}") ax.despine(left=True) ###Output _____no_output_____
doc/REPORT.ipynb
###Markdown Modesolver_helper 연구실 심화 실습 Project----- Project members+ Sanghoon Kim+ Kangseok Kim+ Seoungmin Park ---- Introduction> * Background of Si photonics> + The problem is that electronics reduces energy efficiency when processing and processing data in a data center. > + Data center capacity is increasing every year due to other Internet activities.> + An alternative to Si Electronics is the introduction of Si Photonics. >> * Waveguide and Mode> + Waveguide traps light in the Waveguide structure and transmits it to a specific location. > + Waveguide consists of Core and Cladding.Due to the geometric morphology of Waveguide and the refractive index of the components, there is a specific Efield profile in the Waveguide section, which is called Mode.> + If the structure of the Waveguide does not change in the direction of light travel, the light will proceed with the probability image in the form of an E-field profile in a particular mode. Waveguide's design parameters have different mode characteristics, so the process of optimizing Waveguide to achieve the desired mode is a must in Waveguide design. Requirements Getting Stared + Entered the Terminal, write down 'pip install -r requirements.txt' and download it. \```pip install -r requirements.txt```----- MotivationI've practice to solve modes in optical-waveguide by using modesolverpy in lab practice course in my university. However, It seems there are some problems when you run modesolverpy in Windows. So, I make a decision to help running modesolverpy in windows. And also, There are some example codes in this repository. modesolverpy> + photonic mode solver with a nice interface and output > + simple structure drawing.> + automated data saving and plotting via Gnuplotm> + some limited (at this stage) data processing (finding MFD of fundamental mode), and> + easily extensible library> > The documentation for this project can be found here.https://modesolverpy.readthedocs.io/en/latest/index.html Structure> example>> Parameter> + > + t_slab> + ----- BaseCode Import Library ###Code import modesolverpy.mode_solver as ms import modesolverpy.structure as st import numpy as mp import pandas as pd ###Output _____no_output_____ ###Markdown Make a list of parameters ###Code # Key Element : [w_wg,t_soi,t_slab,n_eff] # sweep : 0.01 um total_list = [] w_wg_list = [round(0.2 +0.01*i,2) for i in range(0,81) ] # 0.2 ~ 1 t_soi_list = [round(0.2 + 0.01*i,2) for i in range(0,31)] # 0.2 ~ 0.5 t_slab_list = [] # 0 ~ t_soi/2 ###Output _____no_output_____ ###Markdown Draw Structure ###Code """ Param : wg_width[um], film_thickness[um], wg_height[um] Base wavelength of light : 1350nm Return structure profile """ def draw_structure(wg_width=0.22,film_thickness=0.1,wg_height=0.05)->float: struct = st.RidgeWaveguide(x_step = 0.02, y_step = 0.02, wg_height = wg_height, wg_width = wg_width, sub_height = 0.5, sub_width = 2., n_sub = 1.4, n_wg = 3.4, n_clad = 1., wavelength = 1.350, angle = 90., clad_height = 0.5, film_thickness=film_thickness) return struct ###Output _____no_output_____ ###Markdown * Solve effective index of refraction ###Code """ ==== param : wg_width[um], film_thickness[um], wg_height[um] Return n_eff of waveguide with mode 0 and 1. ===== """ def find_n_eff_mode_zero(wg_width=0.22,film_thickness=0.1,wg_height=0.05)->float: structure = draw_structure(wg_width,film_thickness,wg_height) mode_solver = ms.ModeSolverSemiVectorial(2, semi_vectorial_method='Ex') a = mode_solver.solve(structure) return a["n_effs"][0].real # 구해야 하는 것 [mode : 0에서의 n_eff] def find_n_eff_mode_one(wg_width=0.22,film_thickness=0.1,wg_height=0.05)->float: structure = draw_structure(wg_width,film_thickness,wg_height) mode_solver = ms.ModeSolverSemiVectorial(4, semi_vectorial_method='Ex') a = mode_solver.solve(structure) return a["n_effs"][1].real # 구해야 하는 것 [mode : 0에서의 n_eff] ###Output _____no_output_____ ###Markdown Case1) param : Waveguide length> Draw structure & solve n_effs ###Code t_soi = 0.22 t_slab = 0.1 # t_slab = film_thickness - wg_height # mode_solver = ms.ModeSolverSemiVectorial(4,semi_vectorial_method='Ex') for i in w_wg_list: x = draw_structure(i,t_soi,t_slab) # x.write_to_file('.\\struct\\struct_w_wg={0}.dat'.format(i)) mode_solver.solve(x) # mode_solver.write_modes_to_file('mode_w_wg={0}.dat'.format(i)) ###Output _____no_output_____
Linear Regression/India_per_capita-2030.ipynb
###Markdown Machine Learning With Python: Linear Regression With One Variable **Problem Statement**: Predict India's per capita income in year 2023 using india_gdp.csv ###Code import pandas as pd import numpy as np from sklearn import linear_model import matplotlib.pyplot as plt df = pd.read_csv('../datasets/gdp/india_gdp.csv') df %matplotlib inline plt.xlabel('Year') plt.ylabel('GDP (US$)') plt.scatter(df.year,df.gdp,color='red',marker='+') year = df[['year']] year gdp = df.gdp gdp # Create linear regression object reg = linear_model.LinearRegression() reg.fit(year,gdp) reg.predict([[2023]]) reg.coef_ reg.intercept_ year_df = pd.read_csv("../datasets/year.csv") year_df.head(3) p = reg.predict(year_df) p year_df['income']=p year_df from sklearn.model_selection import train_test_split x_train, x_test, y_train, y_test = train_test_split(year,gdp,test_size=0.2,random_state=0) from sklearn.linear_model import LinearRegression regressor = LinearRegression() regressor.fit(x_train, y_train) lr = LinearRegression().fit(x_train, y_train) y_pred = regressor.predict([[2023]]) #y_pred = regressor.predict(x_test) y_pred year_df = pd.read_csv("../datasets/year.csv") year_df.head(3) q = regressor.predict(year_df) q year_df['income']=q year_df print("Training set score: {:.2f}".format(lr.score(x_train, y_train))) print("Test set score: {:.7f}".format(lr.score(x_test, y_test))) from sklearn.preprocessing import PolynomialFeatures from sklearn.pipeline import Pipeline steps = [ ('poly', PolynomialFeatures(degree=2)), ('model', LinearRegression()) ] pipeline = Pipeline(steps) pipeline.fit(x_train, y_train) print('Training score: {}'.format(pipeline.score(x_train, y_train))) print('Test score: {}'.format(pipeline.score(x_test, y_test))) pipeline.predict([[2023]]) # Now Read Years year_f = pd.read_csv("../datasets/year.csv") year_f.head(3) qr = pipeline.predict(year_f) qr year_f['gdp']=qr print('Forecast per capita GDP (US$) : ') year_f %matplotlib inline plt.xlabel('Year') plt.ylabel('GDP (US$)') plt.plot(df.year,df.gdp,color='g',marker='+') plt.plot(year_f.year,qr,color='red',marker='+') ###Output _____no_output_____
4_PyTorch_Example.ipynb
###Markdown Minimal PyTorch Example This notebooks shows a very minimal example on how to use PyTorch for training a neural network on the Iris data set. 0. Preamble ###Code import torch import torch.nn.functional as F import torch.nn as nn torch.manual_seed(1) ###Output _____no_output_____ ###Markdown The following lines checks for GPU availability on the machine and sets the GPU as processing device (if available).If you are on Colab you can enable GPU support in the menu via "Runtime > Change runtime type" and select "GPU" as hardware accelerator. ###Code if(torch.cuda.is_available()): processing_chip = "cuda:0" print(f"{torch.cuda.get_device_name(0)} available") else: processing_chip = "cpu" print("No GPU available") device = torch.device(processing_chip) device ###Output No GPU available ###Markdown 1. Data Preperation For this small example we use the [Iris flower data set](https://en.wikipedia.org/wiki/Iris_flower_data_set). The data set consists of 50 samples from each of three species of Iris (Iris setosa, Iris virginica and Iris versicolor). Four features were measured from each sample: the length and the width of the sepals and petals, in centimeters. Based on these four features, we want to train a model that can predict the species.In the first step we load the data into a Pandas. ###Code import pandas as pd url = 'data/iris.csv' dataset = pd.read_csv(url) dataset.head(5) ###Output _____no_output_____ ###Markdown To be able to train a model, we first need to transform the *species* column into a numeric: ###Code dataset.loc[dataset.species=='Iris-setosa', 'species'] = 0 dataset.loc[dataset.species=='Iris-versicolor', 'species'] = 1 dataset.loc[dataset.species=='Iris-virginica', 'species'] = 2 dataset.head() ###Output _____no_output_____ ###Markdown Next, we specify which columns we want to use as features and which as label: ###Code X = dataset[dataset.columns[0:4]].values y = dataset.species.values.astype(int) ###Output _____no_output_____ ###Markdown We then split our data into training and test data. ###Code from sklearn.model_selection import train_test_split train_X, test_X, train_y, test_y = train_test_split(X, y, test_size=0.2) print(train_X.shape, test_X.shape) ###Output (120, 4) (30, 4) ###Markdown To be able to use the data in PyTorch, we need to convert them into PyTorch tensors. Such a tensor can be thought of an efficient way to represent lists and matrices (similar to Numpy), with the additional benefit that they can be moved to the GPU (the `.to(device)` part in the code below) and that they support automatic backpropagation (more on this later): ###Code train_x = torch.Tensor(train_X).float().to(device) test_x = torch.Tensor(test_X).float().to(device) train_y =torch.Tensor(train_y).long().to(device) test_y = torch.Tensor(test_y).long().to(device) ###Output _____no_output_____ ###Markdown 2. Model definitionWe define now the strucutre of our neural network. For this we create a class that is a subclass from PyTorch's `nn.Module`.By convention we put in the `__init__` method the layers we want to use in the network and in the `forward` method how data flows through this network.Our network has 4 input features, 7 hidden layer nodes and 3 output neurons. The hidden layer uses a Relu activation function. Note that the output layer does not have a softmax activation (unlike we have seen it in the lecture). It rather gives out a raw score for each class (more on this later). ###Code class Net(nn.Module): def __init__(self): super(Net, self).__init__() self.hidden = nn.Linear(4, 7) self.output = nn.Linear(7, 3) def forward(self, x): z1 = self.hidden(x) z2 = F.relu(z1) z3 = self.output(z2) # no softmax. see CrossEntropyLoss() return z3 ###Output _____no_output_____ ###Markdown 3. Model TrainingWe can now start training our network. We run several epochs in which we first predict on the training data with our network and than backpropagate the loss. For this we use PyTorch's build-in optimizer that runs gradient descent on the weights of the network. Hence, in every episode we reduce the loss on the training data and improve our network.As loss function we use cross entropy, which consumes the raw scores from the prediction and internally applies a softmax (that is why we do not need the softmax as last layer in the network).Note that all training data is passed at once to our network (line `net(train_x)`), since PyTorch will predict on all data points in parallel. ###Code # create network, move it to device (either CPU or GPU) net = Net().to(device) # define the parameters for training no_epochs = 100 learning_rate = 0.04 loss_func = nn.CrossEntropyLoss() # applies softmax() internally optimizer = torch.optim.SGD(net.parameters(), lr=learning_rate) print("\nStarting training ") train_losses = [] for epoch in range(0, no_epochs): optimizer.zero_grad() # set gradients to zero predictions = net(train_x) # predict on the training data, this calls net.forward() loss = loss_func(predictions, train_y) # compute loss between prediction and true labels loss.backward() # calculate the gradients for every weight optimizer.step() # do one step of gradient descent train_losses.append(loss.item()) if epoch % 10 == 0: print(f"Loss in epoch {epoch} is {loss.item()}") print("Done training ") import matplotlib.pyplot as plt fig = plt.figure() plt.plot(range(0, no_epochs), train_losses, color='blue') plt.legend(['Train Loss'], loc='upper right') plt.xlabel('number of epochs') plt.ylabel('loss') ###Output _____no_output_____ ###Markdown 4. Model EvaluationFinally, we check the model accuracy on the test data. For this we predict on the test data, identify the class with the highest score and compare it to the true label. ###Code predictions = net(test_x) _, predicted = torch.max(predictions.data, 1) # get the class with highest score correct = (predicted == test_y).sum().item() # compare predicted class with real class print(f"Accuarcy is {100. * correct / len(test_x)}%") ###Output Accuarcy is 76.66666666666667%
Tutorials/Tutorial3/DS_Tutorial3.ipynb
###Markdown Name:- Parshwa Shah Roll No:- 34 UID:- 2019230071 Tutorial 3 ###Code import numpy as np from numpy . random import randn import matplotlib . pyplot as plt from mpl_toolkits . mplot3d import Axes3D n =1000 mu1 = np. array ([2 ,1 , -3]) mu2 = np. array ([1 , -4 ,0]) mu3 = np. array ([2 ,4 ,0]) X1 = randn (n ,3) + mu1 X2 = randn (n ,3) + mu2 X3 = randn (n ,3) + mu3 fig = plt. figure () ax = fig.gca( projection ='3d' ,) ax.plot(X1 [: ,0] , X1 [: ,1] , X1 [: ,2] , 'r.',alpha =0.5 , markersize =2) ax.plot(X2 [: ,0] , X2 [: ,1] , X2 [: ,2] , 'b.',alpha =0.5 , markersize =2) ax.plot(X3 [: ,0] , X3 [: ,1] , X3 [: ,2] , 'g.',alpha =0.5 , markersize =2) ax. set_xlim3d ( -4 ,6) ax. set_ylim3d ( -5 ,5) ax. set_zlim3d ( -5 ,2) plt.show () !pip install pandas_datareader from pandas_datareader import * from numpy . linalg import svd , pinv mu21 = (mu2 - mu1). reshape (3 ,1) mu31 = (mu3 - mu1). reshape (3 ,1) W = np. hstack (( mu21 , mu31)) U,_,_ = svd(W) # we only need U P = W @ pinv(W) R = U.T @ P RX1 = (R @ X1.T).T RX2 = (R @ X2.T).T RX3 = (R @ X3.T).T plt.plot(RX1 [: ,0] , RX1 [: ,1] ,'b.',alpha =0.5 , markersize =2) plt.plot(RX2 [: ,0] , RX2 [: ,1] ,'g.',alpha =0.5 , markersize =2) plt.plot(RX3 [: ,0] , RX3 [: ,1] ,'r.',alpha =0.5 , markersize =2) plt.show () ###Output _____no_output_____
vermeerkat/plugins/fleetingpol/diagnostics/Polcal solutions.ipynb
###Markdown Crosshand delays ###Code with tbl(KX) as t: delays = t.getcol("FPARAM") ants = t.getcol("ANTENNA1") field = t.getcol("FIELD_ID") flags = t.getcol("FLAG") time = t.getcol("TIME") with tbl("%s::ANTENNA" % KX) as t: antnames = t.getcol("NAME") delays[flags] = np.nan hrloc = mdates.HourLocator() minloc = mdates.MinuteLocator() dtFmt = mdates.DateFormatter('%hh%mm%ss') collections = [] collections_time = [] pcmax = -np.inf pcmin = np.inf for a in np.unique(ants): asel = ants == a unflagged = np.logical_not(flags[:, 0, 0][asel]) collections.append(delays[:, 0, 0][asel][unflagged]) pcmax = max(pcmax, np.nanpercentile(delays[:, 0, 0][asel][unflagged],98)) pcmin = min(pcmin, np.nanpercentile(delays[:, 0, 0][asel][unflagged],2)) collections_time.append(time[asel][unflagged]) labels=[antnames[ai] for ai in np.unique(ants)] plt.figure(figsize=(25, 6)) plt.title("Crosshand delays") plt.boxplot(collections, 0, '', labels=labels) plt.ylabel("Delay (ns)") plt.show() fig, ax = plt.subplots(figsize=(25, 6)) for t,a,aname in zip(collections_time, collections, labels): ax.plot(convertMJD2unix(t), a, label=aname) ax.set_ylabel("Delay (ns) [98%]") ax.set_xlabel("Time (start: %s)" % str(convertMJD2unix([np.min(time)])[0])) ax.legend(loc = (1.01,0)) ax.grid(True) hfmt = mdates.DateFormatter('%H:%M') ax.xaxis.set_major_formatter(hfmt) limmean = np.nanmean(delays) lim = np.nanstd(delays) ax.set_ylim(pcmin, pcmax) plt.show() ###Output _____no_output_____ ###Markdown Crosshand phase gain stability ###Code with tbl(Xref) as t: bpgain = t.getcol("CPARAM") ants = t.getcol("ANTENNA1") field = t.getcol("FIELD_ID") flags = t.getcol("FLAG") time = t.getcol("TIME") with tbl("%s::ANTENNA" % Xref) as t: antnames = t.getcol("NAME") bpgain[flags] = np.nan for corr in range(bpgain.shape[2]): collections = [] collections_std = [] collections_time = [] for a in np.unique(ants): asel = ants == a bpgain[flags] = np.nan ang = np.angle(bpgain[asel, :, corr]) collections.append(np.nanmedian(ang, axis=1)) collections_std.append((np.nanpercentile(ang, 75.0, axis=1) - np.nanpercentile(ang, 25.0, axis=1))*0.5) collections_time.append(time[asel]) labels=[antnames[ai] for ai in np.unique(ants)] fig, ax = plt.subplots(figsize=(25, 6)) for t,a,s,aname in zip(collections_time, collections, collections_std, labels): ax.errorbar(convertMJD2unix(t), np.rad2deg(a), yerr=np.rad2deg(s), label=aname) ax.set_title("Crosshand phase DC") ax.set_ylabel("Phase [deg]") ax.set_xlabel("Time (start: %s)" % str(convertMJD2unix([np.min(time)])[0])) ax.legend(loc = (1.01,0)) ax.grid(True) hfmt = mdates.DateFormatter('%H:%M') ax.xaxis.set_major_formatter(hfmt) plt.show() ###Output _____no_output_____ ###Markdown Crosshand phase ###Code with tbl(Xfreq) as t: xfsols = t.getcol("CPARAM") ants = t.getcol("ANTENNA1") field = t.getcol("FIELD_ID") flags = t.getcol("FLAG") time = t.getcol("TIME") with tbl("%s::ANTENNA" % Xfreq) as t: antnames = t.getcol("NAME") with tbl("%s::SPECTRAL_WINDOW" % Xfreq) as t: freqs = t.getcol("CHAN_FREQ")/1.0e6 xfsols[flags] = np.nan collections = [] for a in np.unique(ants): asel = ants == a collections.append(xfsols[:, :, corr][asel]) labels=[antnames[ai] for ai in np.unique(ants)] fig, ax = plt.subplots(figsize=(25, 6)) for a,aname in zip(collections, labels): ax.scatter(np.tile(freqs, (1, a.shape[0])), np.rad2deg(np.angle(a)), s=1.5, label=aname) ax.set_title("Crosshand phase") ax.set_ylabel("Phase [deg]") ax.set_xlabel("Frequency (MHz)") ax.legend(loc = (1.01,0)) ax.grid(True) plt.show() ###Output _____no_output_____ ###Markdown First order leakage gain stability ###Code with tbl(Dref) as t: dgain = t.getcol("CPARAM") ants = t.getcol("ANTENNA1") field = t.getcol("FIELD_ID") flags = t.getcol("FLAG") time = t.getcol("TIME") with tbl("%s::ANTENNA" % Dref) as t: antnames = t.getcol("NAME") dgain[flags] = np.nan collections = [] collections_time = [] for a in np.unique(ants): asel = ants == a unflagged = np.logical_not(flags[:, 0, 0][asel]) collections.append(dgain[:, 0, 0][asel][unflagged]) collections_time.append(time[asel][unflagged]) labels=[antnames[ai] for ai in np.unique(ants)] fig, ax = plt.subplots(figsize=(25, 6)) for t,a,aname in zip(collections_time, collections, labels): ax.plot(convertMJD2unix(t), 10*np.log10(np.abs(a)), label=aname) ax.set_title("Leakage gain") ax.set_ylabel("Amplitude [dB]") ax.set_xlabel("Time (start: %s)" % str(convertMJD2unix([np.min(time)])[0])) ax.legend(loc = (1.01,0)) ax.grid(True) hfmt = mdates.DateFormatter('%H:%M') ax.xaxis.set_major_formatter(hfmt) plt.show() plt.figure(figsize=(25, 6)) plt.title("DC leakage") plt.boxplot([10*np.log10(np.abs(c)) for c in collections], 0, '', labels=labels) plt.ylabel("DC leakage") plt.show() ###Output _____no_output_____ ###Markdown Leakage ###Code with tbl(Dfreq) as t: dfsols = t.getcol("CPARAM") ants = t.getcol("ANTENNA1") field = t.getcol("FIELD_ID") flags = t.getcol("FLAG") time = t.getcol("TIME") with tbl("%s::ANTENNA" % Dfreq) as t: antnames = t.getcol("NAME") with tbl("%s::SPECTRAL_WINDOW" % Dfreq) as t: freqs = t.getcol("CHAN_FREQ")/1.0e6 dfsols[flags] = np.nan collections = [] for a in np.unique(ants): asel = ants == a collections.append(dfsols[:, :, corr][asel]) labels=[antnames[ai] for ai in np.unique(ants)] fig, ax = plt.subplots(figsize=(25, 6)) for a,aname in zip(collections, labels): ax.scatter(np.tile(freqs, (1, a.shape[0])), 10*np.log10(np.abs(a)), s=1.5, label=aname) ax.set_title("Leakage") ax.set_ylabel("Leakage [dB]") ax.set_xlabel("Frequency (MHz)") ax.legend(loc = (1.01,0)) ax.grid(True) plt.show() ###Output _____no_output_____
documents/presentation-5/script7.ipynb
###Markdown Statistical Analysis of Data Environment SettingsAn statistical Analysis of the data captured will be performed.The environment configuration is the following:- A rectangle area is used whose dimension is 2 x 1.5 meters. - A custom robot similar to an epuck was used.- The robot starts in the middle of the arena.- The robot moves in a random fashion way around the environment avoiding obstacles.- The robot has 8 sensors that measure the distance between the robot and the walls.- Some noise was introduced in the sensors measurements of the robot using the concept of [lookup tables](https://cyberbotics.com/doc/reference/distancesensor) in the Webots simulator which according to Webots documentation "The first column of the table specifies the input distances, the second column specifies the corresponding desired response values, and the third column indicates the desired standard deviation of the noise. The noise on the return value is computed according to a gaussian random number distribution whose range is calculated as a percent of the response value (two times the standard deviation is often referred to as the signal quality)". The following values were taken: -First experiment: - (0, 0, 0.01) - (10, 10, 0.01) -Second experiment: - (0, 0, 0.2) - (10, 10, 0.2)- The simulator runs during 10 minutes in fast mode which is translated into 12 hours of collected data. ###Code # Install a pip package in the current Jupyter kernel import sys !{sys.executable} -m pip install scikit-learn !{sys.executable} -m pip install keras import pandas as pd import tensorflow as tf import numpy as np import math from sklearn.ensemble import RandomForestRegressor from keras import models from keras import layers from keras import regularizers import matplotlib.pyplot as plt from keras import optimizers ###Output Requirement already satisfied: scikit-learn in /usr/local/lib/python3.7/site-packages (0.22) Requirement already satisfied: scipy>=0.17.0 in /usr/local/lib/python3.7/site-packages (from scikit-learn) (1.1.0) Requirement already satisfied: numpy>=1.11.0 in /usr/local/lib/python3.7/site-packages (from scikit-learn) (1.16.1) Requirement already satisfied: joblib>=0.11 in /usr/local/lib/python3.7/site-packages (from scikit-learn) (0.14.1) Requirement already satisfied: keras in /usr/local/lib/python3.7/site-packages (2.3.1) Requirement already satisfied: scipy>=0.14 in /usr/local/lib/python3.7/site-packages (from keras) (1.1.0) Requirement already satisfied: keras-applications>=1.0.6 in /usr/local/lib/python3.7/site-packages (from keras) (1.0.7) Requirement already satisfied: numpy>=1.9.1 in /usr/local/lib/python3.7/site-packages (from keras) (1.16.1) Requirement already satisfied: six>=1.9.0 in /usr/local/lib/python3.7/site-packages (from keras) (1.12.0) Requirement already satisfied: keras-preprocessing>=1.0.5 in /usr/local/lib/python3.7/site-packages (from keras) (1.0.9) Requirement already satisfied: pyyaml in /usr/local/lib/python3.7/site-packages (from keras) (5.2) Requirement already satisfied: h5py in /usr/local/lib/python3.7/site-packages (from keras) (2.9.0) ###Markdown First Experiment ###Code csv_file = 'robot_info_dataset-jumped.csv' df = pd.read_csv(csv_file) df.head() ###Output _____no_output_____ ###Markdown Data pre-processing The data collected 1384848 samples. ###Code df.shape ###Output _____no_output_____ ###Markdown The data set contains some null values so they should be deleted from the samples. ###Code df = df.dropna() ###Output _____no_output_____ ###Markdown Now the data will be normalized. ###Code normalized_df=(df-df.min())/(df.max()-df.min()) normalized_df.describe() ###Output _____no_output_____ ###Markdown Input and output variables The data will be split into training, testing and validation sets. 60% of the data will be used for training, 20% for training and 20% of validation. ###Code # train size test_size_percentage = .2 train_size_percentage = .6 ds_size = normalized_df.shape[0] train_size = int(train_size_percentage * ds_size) test_size = int(test_size_percentage * ds_size) # shuffle dataset normalized_df = normalized_df.sample(frac=1) # separate inputs from outputs inputs = normalized_df[['x', 'y', 'theta']] targets = normalized_df[['sensor_1', 'sensor_2', 'sensor_3', 'sensor_4', 'sensor_5', 'sensor_6', 'sensor_7', 'sensor_8']] # train train_inputs = inputs[:train_size] train_targets = targets[:train_size] # test test_inputs = inputs[train_size:(train_size + test_size)] test_targets = targets[train_size:(train_size + test_size)] # validation validation_inputs = inputs[(train_size + test_size):] validation_targets = targets[(train_size + test_size):] ###Output _____no_output_____ ###Markdown Neural Network As input the neural network receives the x, y coordinates and rotation angle $\theta$. The output are the sensor measurements. One model per sensor will be created. ###Code def get_model(): # neural network with a 10-neuron hidden layer model = models.Sequential() model.add(layers.Dense(10, activation='relu', input_shape=(3,))) # model.add(layers.Dropout(0.5)) model.add(layers.Dense(6, activation='relu')) model.add(layers.Dense(3, activation='relu')) model.add(layers.Dense(1)) # rmsprop = optimizers.RMSprop(learning_rate=0.01) model.compile(optimizer='rmsprop', loss='mse', metrics=['mae']) return model model = get_model() history = model.fit(inputs, targets[['sensor_7']], epochs=75, batch_size=1, verbose=1) history.history['mae'] model.save("nn_sensor_7.h5") ###Output WARNING:tensorflow:From /usr/local/lib/python3.7/site-packages/tensorflow/python/ops/resource_variable_ops.py:435: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version. Instructions for updating: Colocations handled automatically by placer. WARNING:tensorflow:From /usr/local/lib/python3.7/site-packages/tensorflow/python/ops/math_ops.py:3066: to_int32 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version. Instructions for updating: Use tf.cast instead. Epoch 1/75 65341/65341 [==============================] - 211s 3ms/step - loss: 0.0155 - mae: 0.0960 Epoch 2/75 65341/65341 [==============================] - 212s 3ms/step - loss: 0.0096 - mae: 0.0740 Epoch 3/75 65341/65341 [==============================] - 222s 3ms/step - loss: 0.0067 - mae: 0.0608 Epoch 4/75 65341/65341 [==============================] - 211s 3ms/step - loss: 0.0064 - mae: 0.0591 Epoch 5/75 65341/65341 [==============================] - 233s 4ms/step - loss: 0.0063 - mae: 0.0583 Epoch 6/75 65341/65341 [==============================] - 239s 4ms/step - loss: 0.0062 - mae: 0.0572 Epoch 7/75 65341/65341 [==============================] - 220s 3ms/step - loss: 0.0060 - mae: 0.0566 Epoch 9/75 65341/65341 [==============================] - 227s 3ms/step - loss: 0.0059 - mae: 0.0563 Epoch 10/75 65341/65341 [==============================] - 231s 4ms/step - loss: 0.0058 - mae: 0.0561 Epoch 11/75 65341/65341 [==============================] - 178s 3ms/step - loss: 0.0058 - mae: 0.0558 Epoch 17/75 65341/65341 [==============================] - 176s 3ms/step - loss: 0.0058 - mae: 0.0557 Epoch 18/75 65341/65341 [==============================] - 176s 3ms/step - loss: 0.0059 - mae: 0.0559 Epoch 19/75 65341/65341 [==============================] - 178s 3ms/step - loss: 0.0058 - mae: 0.0560 Epoch 20/75 65341/65341 [==============================] - 175s 3ms/step - loss: 0.0058 - mae: 0.0557 0s - loss: 0.0058 - Epoch 22/75 65341/65341 [==============================] - 178s 3ms/step - loss: 0.0057 - mae: 0.0556 Epoch 23/75 65341/65341 [==============================] - 167s 3ms/step - loss: 0.0057 - mae: 0.0554 Epoch 24/75 65341/65341 [==============================] - 154s 2ms/step - loss: 0.0056 - mae: 0.0549 Epoch 25/75 65341/65341 [==============================] - 160s 2ms/step - loss: 0.0056 - mae: 0.0547 Epoch 26/75 65341/65341 [==============================] - 209s 3ms/step - loss: 0.0056 - mae: 0.0549 Epoch 28/75 65341/65341 [==============================] - 208s 3ms/step - loss: 0.0057 - mae: 0.0553 Epoch 29/75 65341/65341 [==============================] - 196s 3ms/step - loss: 0.0057 - mae: 0.0554 Epoch 30/75 65341/65341 [==============================] - 208s 3ms/step - loss: 0.0058 - mae: 0.0557 Epoch 31/75 65341/65341 [==============================] - 206s 3ms/step - loss: 0.0057 - mae: 0.0554 Epoch 32/75 65341/65341 [==============================] - 192s 3ms/step - loss: 0.0056 - mae: 0.0552 Epoch 33/75 65341/65341 [==============================] - 192s 3ms/step - loss: 0.0056 - mae: 0.0547 Epoch 34/75 65341/65341 [==============================] - 195s 3ms/step - loss: 0.0055 - mae: 0.0543 Epoch 35/75 65341/65341 [==============================] - 192s 3ms/step - loss: 0.0055 - mae: 0.0541 Epoch 36/75 65341/65341 [==============================] - 186s 3ms/step - loss: 0.0055 - mae: 0.0539 Epoch 37/75 65341/65341 [==============================] - 162s 2ms/step - loss: 0.0054 - mae: 0.0539 Epoch 38/75 65341/65341 [==============================] - 198s 3ms/step - loss: 0.0053 - mae: 0.0529 Epoch 40/75 65341/65341 [==============================] - 201s 3ms/step - loss: 0.0052 - mae: 0.0525 0s - loss: 0.0052 - ma Epoch 41/75 53410/65341 [=======================>......] - ETA: 35s - loss: 0.0052 - mae: 0.0521
HFI - A Brief Examination of Religious Freedom.ipynb
###Markdown A Brief Examination of World Religious Freedom The goal of this notebook is to perform a brief exploratory analysis of the human freedom index dataset, particularly with regards to religious freedom. We will begin by briefly looking at overall human freedom around the world, and then dive a little bit deeper into the trends in religious freedom. We will set out to answer some basic questions:1. How has world religious freedom varied over the 2008-2016 period?2. How is religious freedom connected to overall human freedom?3. How are government restrictions on religious freedom related to religious harassment?4. What countries represent the best and worst in terms of religious freedom, and how have those fluctuated in most recent times (2015-2016)? ###Code # Import Libraries import pandas as pd import numpy as np import matplotlib.pyplot as plt import seaborn as sns from IPython.display import Image from IPython.core.display import HTML sns.set() %matplotlib inline # Read in data and view the first 5 rows hfi_data = pd.read_csv('hfi_cc_2018.csv') hfi_data = hfi_data[hfi_data['ISO_code'] != 'BRN'] # We drop this country as it does not contain pf_religion data hfi_data.head() # Define function for plotting yearly median for given metrics with std_dev error bars def plot_med_year(df, metric, title=None): df_year = df.groupby('year') df_med = pd.merge(df_year[metric].median().to_frame().reset_index(), df_year[metric].std().to_frame(name='std').reset_index()) plt.figure() plt.errorbar(x=df_med['year'], y=df_med[metric], yerr=df_med['std'], linestyle='None', marker='s') if title: plt.title(title) else: plt.title('Median {} by Year'.format(metric)) ###Output _____no_output_____ ###Markdown Overall FreedomHere, we will view the overall changes in average world freedom for the time period contained in the datset (2008-2016). In particular, we will view the trends in overall human freedom and in personal freedom. While the human freedom score is made up of an average of personal freedom and economic freedom, we will not be analyzing economic freedom as our interest is primairly in religious freedom (a sub-category of personal freedom). ###Code # Plot World Average hfi by year plot_med_year(hfi_data, 'hf_score', 'World Median HF Score by Year') # Plot World Average personal freedom by year plot_med_year(hfi_data, 'pf_score', 'World Median PF Score by Year') ###Output _____no_output_____ ###Markdown The above graphs show a downward trend in both personal and overall human freedom from 2008 to 2016. Let's break these scores down further by world region. ###Code # Explore yearly regional metrics in human freedom, personal freedom, and religious freedom # Other religion related metrics are excluded as many have missing data region_yr = hfi_data.groupby(['region', 'year']).median().reset_index() # Plot human freedom by region plt.figure() sns.relplot(x='year', y='hf_score', hue='region', data=region_yr) plt.title('Median Yearly hf_score by Region') #Plot personal freeodm by region plt.figure() sns.relplot(x='year', y='pf_score', hue='region', data=region_yr) plt.title('Median Yearly pf_score by Region') ###Output _____no_output_____ ###Markdown Not surprisingly, North America and Western Europe consistently exhibit the highest scores for personal and overall human freedom. What is more surprising from these visualizations is the large spread in human and personal freedom between the five highest reigons and the five lowest regions in the world. These groups are spread by nearly a full point or more in both the personal and human freedom scores. Let's dive deeper into the sub-category of interest here - religious freedom, and see if similar trends exist in religious freedom. Religious Freedom ###Code # Plot World Median religious freedom score by year plot_med_year(hfi_data, 'pf_religion', 'World Med Religious Freedom Score by Year') ###Output _____no_output_____ ###Markdown Religious freedom exhibits an overall downard trend over the 2008 to 2016 period, similar to those of personal and human freedom, albeit with more fluctuation. Next, we will explore the regional breakdown of religious freedom over this period. ###Code #Plot religious freedom by region plt.figure() sns.relplot(x='year', y='pf_religion', hue='region', data=region_yr) plt.title('Median Yearly pf_religion Score by Region') ###Output _____no_output_____ ###Markdown From this regional breakdown, it is clear that there has been more significant variance in religious freedom for each region over 2008-2016 than there has been for overall personal or human freedom. Additionally, while Western Europe has generally ranked very highly in personal and human freedom, it ranks in the middle for religious freedom. Let's look further into the overall and regional medians for government religious restrictions and religious harassment. ###Code #Plot religious freeodm - restrictions plot_med_year(hfi_data, 'pf_religion_restrictions', 'World Med Religious Freedom Restrictions Score by Year') #Plot religious freeodm - harassment plot_med_year(hfi_data, 'pf_religion_harassment', 'World Med Religious Freedom Harassment Score by Year') #Plot religious freeodm - restrictions by region plt.figure() sns.relplot(x='year', y='pf_religion_restrictions', hue='region', data=region_yr) plt.title('Median Yearly pf_religion_restrictions Score by Region') #Plot religious freeodm - harassment by region plt.figure() sns.relplot(x='year', y='pf_religion_harassment', hue='region', data=region_yr) plt.title('Median Yearly pf_religion_harassment Score by Region') ###Output _____no_output_____ ###Markdown Western Europe is again low in both sub-categories of religious freedom. There is also an interesting and somewhat surprising downward trend for North America, particularly from 2014-2016. To get a better understanding of how these metrics relate to each other and to overall religious, personal, and human freedom, we will view the correlation matrix. ###Code # First, we will create a data frame containing only the metrics of interest religion = hfi_data[['countries', 'region', 'year', 'pf_religion_harassment','pf_religion_restrictions', 'pf_religion', 'pf_score', 'hf_score']] # Next, we create a heat map of the correlation matrix plt.figure() sns.heatmap(religion.drop(columns='year').corr(), annot=True) ###Output _____no_output_____ ###Markdown Not surprisingly given the trends observed above, around the world there is a high degree of correlation between government restrictions on religious freedom and the presence of religious harassment. Additionally, there is a somewhat strong correlation (.49) between pf_religion and overall pf_score. Somewhat surprisingly however, the correlation between pf_religion and overall hf_score is only .39. Thus, it exhbits a somewhat strong but not overwhelmingly strong correlation. Let's view these metrics another way to get a better sense for how they relate. ###Code # Create a scatter matrix to view correlation at a more granular level plt.figure() sns.pairplot(religion.drop(columns='year'), hue='region') ###Output /anaconda3/lib/python3.6/site-packages/statsmodels/nonparametric/kde.py:448: RuntimeWarning: invalid value encountered in greater X = X[np.logical_and(X > clip[0], X < clip[1])] # won't work for two columns. /anaconda3/lib/python3.6/site-packages/statsmodels/nonparametric/kde.py:448: RuntimeWarning: invalid value encountered in less X = X[np.logical_and(X > clip[0], X < clip[1])] # won't work for two columns. /anaconda3/lib/python3.6/site-packages/scipy/stats/stats.py:1713: FutureWarning: Using a non-tuple sequence for multidimensional indexing is deprecated; use `arr[tuple(seq)]` instead of `arr[seq]`. In the future this will be interpreted as an array index, `arr[np.array(seq)]`, which will result either in an error or a different result. return np.add.reduce(sorted[indexer] * weights, axis=axis) / sumval ###Markdown The above scatter matrix allows us to visualize the correlation between each of the key metrics in a more granular way. This allows us to see what the correlaiton matrix already told us, that pf_religion is only mildly correlated with hf_score, but that govrnment restrictions of religion and religious harassment are more strongly correlated. Next, we will examine religious freedom at an even more granular level, and look at the world's 5 best and 5 worst countries in terms of religious freedom for 2015 and 2016. ###Code # Create new df for these years only years = [2015, 2016] religion_15_16 = religion[religion['year'].isin(years)] # Bottom 5 - 2015 religion_15_16[religion_15_16['year'] == 2015].nsmallest(n=5, columns='pf_religion') ###Output _____no_output_____ ###Markdown From here, we see that the five lowest countries for pf_religion also have very low scores for religious restrictions, but moderate scores for religious harassment. Let's see if anything changed from 2015 to 2016. ###Code # Bottom 5 - 2016 religion_15_16[religion_15_16['year'] == 2016].nsmallest(n=5, columns='pf_religion') ###Output _____no_output_____ ###Markdown Interestingly, Iran is no longer in the bottom 5 for religious freedom in 2016, but Malaysia now is. The other four countries remain the same, albeit in a different order. Additionally, the scores in government restrictions in the bottom 5 all decreased from 2015-2016. This is not surprising given the overall world downward trend in restrictions scores from 2015 to 2016. Let's see if there is a similar trend amongst the 5 best countries in the world with regards to religious freedom. ###Code # Top 5 - 2015 religion_15_16[religion_15_16['year'] == 2015].nlargest(n=5, columns='pf_religion') # Top 5 - 2016 religion_15_16[religion_15_16['year'] == 2016].nlargest(n=5, columns='pf_religion') ###Output _____no_output_____ ###Markdown Unlike the lowest countries for religious freedom, there is a higher degree of variance amongst the countries included in the top 5. Additionally, the countries in the top 5 do not seem to exhibit the same downward trend in government restrictions that the world in general, and the bottom 5 in particular, exhibit. Of course, this is not a perfect comparison given the changes in the countries included in the top 5, but for our brief analysis, it is still helpful. Lastly, it is interesting to note that there is slightly more regional variation amongst the top 5 countries than amongst the bottom 5. ConclusionThus, we have briefly answered the basic questions we set out to answer at the beginning of this analysis. While we have not gone too in depth, the analysis has shown us that there is a correlation between high levels of government restrictions on religion and religious harassment. Additionally, one somewhat surprising finding was the downward trend in government religious restrictions scores in Western Europe. We will end this analysis with three geographic visualizations which show how religious freedom scores are distributed around the globe. ###Code %%html <div class='tableauPlaceholder' id='viz1548565321797' style='position: relative'><noscript><a href='#'><img alt=' ' src='https:&#47;&#47;public.tableau.com&#47;static&#47;images&#47;hf&#47;hfi_analysis_v1&#47;RelFreedom&#47;1_rss.png' style='border: none' /></a></noscript><object class='tableauViz' style='display:none;'><param name='host_url' value='https%3A%2F%2Fpublic.tableau.com%2F' /> <param name='embed_code_version' value='3' /> <param name='site_root' value='' /><param name='name' value='hfi_analysis_v1&#47;RelFreedom' /><param name='tabs' value='no' /><param name='toolbar' value='yes' /><param name='static_image' value='https:&#47;&#47;public.tableau.com&#47;static&#47;images&#47;hf&#47;hfi_analysis_v1&#47;RelFreedom&#47;1.png' /> <param name='animate_transition' value='yes' /><param name='display_static_image' value='yes' /><param name='display_spinner' value='yes' /><param name='display_overlay' value='yes' /><param name='display_count' value='yes' /><param name='filter' value='publish=yes' /></object></div> <script type='text/javascript'> var divElement = document.getElementById('viz1548565321797'); var vizElement = divElement.getElementsByTagName('object')[0]; vizElement.style.width='100%';vizElement.style.height=(divElement.offsetWidth*0.75)+'px'; var scriptElement = document.createElement('script'); scriptElement.src = 'https://public.tableau.com/javascripts/api/viz_v1.js'; vizElement.parentNode.insertBefore(scriptElement, vizElement); </script> %%html <div class='tableauPlaceholder' id='viz1548565707738' style='position: relative'><noscript><a href='#'><img alt=' ' src='https:&#47;&#47;public.tableau.com&#47;static&#47;images&#47;hf&#47;hfi_analysis_v1&#47;RelRestrictions&#47;1_rss.png' style='border: none' /></a></noscript><object class='tableauViz' style='display:none;'><param name='host_url' value='https%3A%2F%2Fpublic.tableau.com%2F' /> <param name='embed_code_version' value='3' /> <param name='site_root' value='' /><param name='name' value='hfi_analysis_v1&#47;RelRestrictions' /><param name='tabs' value='no' /><param name='toolbar' value='yes' /><param name='static_image' value='https:&#47;&#47;public.tableau.com&#47;static&#47;images&#47;hf&#47;hfi_analysis_v1&#47;RelRestrictions&#47;1.png' /> <param name='animate_transition' value='yes' /><param name='display_static_image' value='yes' /><param name='display_spinner' value='yes' /><param name='display_overlay' value='yes' /><param name='display_count' value='yes' /></object></div> <script type='text/javascript'> var divElement = document.getElementById('viz1548565707738'); var vizElement = divElement.getElementsByTagName('object')[0]; vizElement.style.width='100%';vizElement.style.height=(divElement.offsetWidth*0.75)+'px'; var scriptElement = document.createElement('script'); scriptElement.src = 'https://public.tableau.com/javascripts/api/viz_v1.js'; vizElement.parentNode.insertBefore(scriptElement, vizElement); </script> %%html <div class='tableauPlaceholder' id='viz1548565668582' style='position: relative'><noscript><a href='#'><img alt=' ' src='https:&#47;&#47;public.tableau.com&#47;static&#47;images&#47;hf&#47;hfi_analysis_v1&#47;RelHarassment&#47;1_rss.png' style='border: none' /></a></noscript><object class='tableauViz' style='display:none;'><param name='host_url' value='https%3A%2F%2Fpublic.tableau.com%2F' /> <param name='embed_code_version' value='3' /> <param name='site_root' value='' /><param name='name' value='hfi_analysis_v1&#47;RelHarassment' /><param name='tabs' value='no' /><param name='toolbar' value='yes' /><param name='static_image' value='https:&#47;&#47;public.tableau.com&#47;static&#47;images&#47;hf&#47;hfi_analysis_v1&#47;RelHarassment&#47;1.png' /> <param name='animate_transition' value='yes' /><param name='display_static_image' value='yes' /><param name='display_spinner' value='yes' /><param name='display_overlay' value='yes' /><param name='display_count' value='yes' /></object></div> <script type='text/javascript'> var divElement = document.getElementById('viz1548565668582'); var vizElement = divElement.getElementsByTagName('object')[0]; vizElement.style.width='100%';vizElement.style.height=(divElement.offsetWidth*0.75)+'px'; var scriptElement = document.createElement('script'); scriptElement.src = 'https://public.tableau.com/javascripts/api/viz_v1.js'; vizElement.parentNode.insertBefore(scriptElement, vizElement); </script> ###Output _____no_output_____
GuidedTour/GuidedTour.ipynb
###Markdown Data ScrapingFor analyzing wallstreetbets data, we recommend downloading full.csv from [url] and putting it in ../Data/subreddit_wallstreetbets.If you want to scrape a different subreddit, you can use the following file. You will need API.env with appropriate credentials in /Automated/ ###Code start = dt.datetime(2020, 1, 1) end = dt.datetime(2020, 1, 30) if not os.path.exists(f"../Data/subreddit_{subreddit}/full.csv"): print("Did not find scraped data, scraping.") RedditScraper.scrape_data(subreddits = [subreddit], start = start, end = end) ###Output _____no_output_____ ###Markdown Change Point AnalysisThe next cell will open full.csv , compute the words that are among the top daily_words most popular words on any day, and then run the change point analysis model on each of them.The first time this is a run, a cleaned up version of the dataframe will be created for ease of processing. ###Code up_to = None # Only calculate change points for up_to of the popular words. Set to None to do all of them. daily_words = 1 # Get the daily_words most popular posts on each day. # Compute the changepoints # ChangePointAnalysis.changepointanalysis([subreddit], up_to = up_to, daily_words = daily_words) # The output has been commented out because it is very long. ###Output _____no_output_____ ###Markdown After running, these files will in ../Data/subreddit_subreddit/Changepoints/Metropolis_30000Draws_5000TuneThe final folder name corresponds to the parameters of the Markov chain used by pymc3 for the inference. Organizing the changepointsA table of the keywords considered with parameters estimated by the model is stored in : ../Data/subreddit_{subreddit}/Changepoints/results.csvYou can then sort through these to find the keywords the model detected a change in.There are two key parameters:change_point_confidence (also denoted p): the models belief that there was a changepoint. p = 1 indicates yes, p = 0 indicates no.mu_diff : a measurement of the size of the changepoint. ###Code results = pd.read_csv(f"../Data/subreddit_{subreddit}/Changepoints/results.csv") results = results.rename( columns = {"Unnamed: 0" : "keyword" }) filtered = results[(results.change_point_confidence == 1) & (results.mu_diff.apply(lambda x : np.abs(x)) > .03)] filtered = filtered.sort_values(by = "mu_diff", ascending = False) for row in filtered.iterrows(): word = row[1]["keyword"] print(word, " Change point confidence: ", row[1]["change_point_confidence"], " Change point magnitude: ", row[1]["mu_diff"]) img = mpimg.imread(f'../Data/subreddit_WallStreetBets/Changepoints/Metropolis_30000Draws_5000Tune/ChangePoint_{word}.png') imgplot = plt.imshow(img) plt.axis('off') plt.show() ###Output gme Change point confidence: 1.0 Change point magnitude: 0.0870740331085252 ###Markdown Warning: Sometimes change point confidence alone is not enough. For instance, it was very confident (p = 1) that there was a changepoint in the following, although it wouldn't be reported as a change point because mu_2 - mu_1 was very small (~.003):![title](../Data/subreddit_WallStreetBets/Changepoints/Metropolis_30000Draws_5000Tune/ChangePoint_new.png) Brief explanation of how this model works:The Bayesian model is as follows:1. A coin is flipped with probability p.2. If the coin comes up heads, then there is a change point. Otherwise, there is no change point.3. It is assumed that the frequency random variable consists of independent draws from a beta distribution. If the coin decided there would be no change point, it is the same beta distribution at all times. Otherwise, it is a different beta on the different sides of the change points.The posterior distribution of p is the models confidence that there is a change point, and the posterior distribution of tau represents its guess about when it occured.The variable mu_1 represents the mean of the beta distribution before the change point, and mu_2 represents the mean of the beta distribution after the changepoint.Of course, this is not a realistic picture of the process; the independence of the different draws from the betas is especially unlike the data. However, it appears to be good enough to discover change points, especially when p and mu_2 - mu_1 are used together.As currently written, it only handles one change point, however this can be improved.(This model was inspired by the changepoint example from Chapter 1 of Bayesian Methods for Hackers: Probabilistic Programming and Bayesian Inference.) Neural NetsThe following code will train a neural net that predicts, given a submission's title text and time of posting, whether that submission's score will be above the median. We use pre-trained GloVe word embeddings in order to convert the title text into a vector that can be used in the neural net. These word embeddings are tuned along with the model parameters as the model is being trained. This technique and the neural net's architecture are taken from a blog post of Max Woolf, https://minimaxir.com/2017/06/reddit-deep-learning/. ###Code model, accuracies, word_tokenizer, df = CreateNeuralNets.buildnets(['wallstreetbets'])[0] ###Output Starting Post Classification Model. ###Markdown Predicted popularity as a time seriesWe now show how the predicted popularity of a post depends on the day on which it was posted. We plot the prediction for the same title, "GME GME GME GME GME GME", as if it were posted at noon each day. It is interesting to note that the variance seems to decrease after the GameStop short squeeze of early 2021. ###Code text = "GME GME GME GME GME GME" CreateNeuralNets.timeseries(df, text, model, word_tokenizer) ###Output _____no_output_____ ###Markdown This will produce a picture like the following:![title](../Data/subreddit_WallStreetBets/6_GME.png) Workshopping exampleHere we start with a potential title (to be posted at noon on April 1, 2021) and attempt to improve it based on the model's prediction. ###Code #this is the date information for April 1, 2021. #Note we normalize so the earliest year in our data set (2020) #and the earliest day of the year correspond to the number 0 input_hour = np.array([12]) input_dayofweek = np.array([3]) input_minute = np.array([0]) input_dayofyear = np.array([91]) input_year = np.array([0]) input_info=[input_hour,input_dayofweek, input_minute, input_dayofyear, input_year] #given a list of potential titles, predict the success of each one def CheckPopularity(potential_titles): for title in potential_titles: print(model.predict([CreateNeuralNets.encode_text(title,word_tokenizer)] + input_info)[0][0][0]) potential_titles = ["Buy TSLA", "Buy TSLA! I like the stock", "Buy TSLA! Elon likes the stock", "TSLA is the next GME. Elon likes the stock", "TSLA is the next GME. To the moon! Elon likes the stock"] CheckPopularity(potential_titles) ###Output 0.9536921 0.957647 0.9620316 0.98298347 0.983858
evaluations/sars-cov-2/4-query.case-5000.ipynb
###Markdown 1. Parameters ###Code cases_dir = 'cases/unset' metadata_file = 'input/metadata-subsample-pangolin.tsv' build_tree = False # Parameters cases_dir = "cases/case-5000" iterations = 3 number_samples = 5000 build_tree = False from pathlib import Path import imp fp, pathname, description = imp.find_module('gdi_benchmark', ['../../lib']) gdi_benchmark = imp.load_module('gdi_benchmark', fp, pathname, description) cases_dir_path = Path(cases_dir) case_name = str(cases_dir_path.name) index_path = cases_dir_path / 'index' output_api_path = cases_dir_path / 'query-api.tsv' output_cli_path = cases_dir_path / 'query-cli.tsv' ###Output _____no_output_____ ###Markdown 2. Benchmark command-line ###Code import pandas as pd import genomics_data_index.api as gdi def benchmark_cli_index(name: str, index_path: Path, build_tree: bool) -> pd.DataFrame: benchmark_commands = { 'query hasa': f'gdi --project-dir {index_path} --ncores 1 query "hasa:hgvs_gn:NC_045512.2:S:p.D614G"', 'query isa': f'gdi --project-dir {index_path} --ncores 1 query "isa:Switzerland/100108/2020"', 'query --summary': f'gdi --project-dir {index_path} --ncores 1 query "hasa:hgvs_gn:NC_045512.2:S:p.D614G" --summary', 'query --features-summary': f'gdi --project-dir {index_path} --ncores 1 query --features-summary mutations', 'list samples': f'gdi --project-dir {index_path} --ncores 1 list samples', } if build_tree: benchmark_commands['query isin'] = f'gdi --project-dir {index_path} --ncores 1 query --reference-name NC_045512 "isin_5_substitutions:Switzerland/100108/2020"' db = gdi.GenomicsDataIndex.connect(index_path) number_samples = db.count_samples() number_features_no_unknown = db.count_mutations(reference_genome='NC_045512', include_unknown=False) number_features_all = db.count_mutations(reference_genome='NC_045512', include_unknown=True) iterations = 10 benchmarker = gdi_benchmark.QueryBenchmarkHandler() return benchmarker.benchmark_cli(name=name, kind_commands=benchmark_commands, number_samples=number_samples, number_features_no_unknown=number_features_no_unknown, number_features_all=number_features_all, iterations=iterations) cli_df = benchmark_cli_index(name=case_name, index_path=index_path, build_tree=build_tree) cli_df.head(3) cli_df.to_csv(output_cli_path, sep='\t', index=False) ###Output _____no_output_____ ###Markdown 3. Test query API 3.1. Load (example) metadataThe simulated data is based off of real sample names and a real tree. So I can load up real metadata and attach it to a query (though the mutations and reference genome are all simulated). ###Code import pandas as pd metadata_df = pd.read_csv(metadata_file, sep='\t') metadata_df.head(2) ###Output _____no_output_____ ###Markdown 3.2. Define benchmark cases ###Code from typing import List import genomics_data_index.api as gdi def benchmark_api_index(name: str, index_path: Path, build_tree: bool) -> pd.DataFrame: db = gdi.GenomicsDataIndex.connect(index_path) q_no_join = db.samples_query(reference_name='NC_045512', universe='mutations') q_join = db.samples_query(reference_name='NC_045512', universe='mutations').join(metadata_df, sample_names_column='strain') q = q_join.hasa('hgvs_gn:NC_045512.2:S:p.D614G') r = q_join.hasa('hgvs_gn:NC_045512.2:N:p.R203K') number_samples = db.count_samples() number_features_no_unknown = db.count_mutations(reference_genome='NC_045512', include_unknown=False) number_features_all = db.count_mutations(reference_genome='NC_045512', include_unknown=True) repeat = 10 benchmark_cases = { 'db.samples_query': lambda: db.samples_query(reference_name='NC_045512', universe='mutations'), 'q.join': lambda: q_no_join.join(metadata_df, sample_names_column='strain'), 'q.features_summary': lambda: q_join.features_summary(), 'q.features_comparison': lambda: q_join.features_comparison(sample_categories='lineage', categories_kind='dataframe', kind='mutations', unit='proportion'), 'q.hasa': lambda: q_join.hasa("hgvs_gn:NC_045512.2:N:p.R203K"), 'q.isa': lambda: q_join.isa("Switzerland/100112/2020"), 'q AND r': lambda: q & r, 'q.toframe': lambda: q_join.toframe(), 'q.summary': lambda: q_join.summary(), } if build_tree: benchmark_cases['q.isin (distance)'] = lambda: q_join.isin("Switzerland/100108/2020", kind='distance', distance=5, units='substitutions') benchmark_cases['q.isin (mrca)'] = lambda: q_join.isin(["Switzerland/100108/2020", "FR993751"], kind='mrca') benchmarker = gdi_benchmark.QueryBenchmarkHandler() return benchmarker.benchmark_api(name=name, kind_functions=benchmark_cases, number_samples=number_samples, number_features_no_unknown=number_features_no_unknown, number_features_all=number_features_all, repeat=repeat) ###Output _____no_output_____ ###Markdown 3.3. Benchmark reads index ###Code api_df = benchmark_api_index(name=case_name, index_path=index_path, build_tree=build_tree) api_df.head(5) api_df.to_csv(output_api_path, sep='\t', index=False) ###Output _____no_output_____
src/modeling/Apply Models to Figshare Talk Corpus.ipynb
###Markdown Load Models ###Code tasks = ['attack', 'toxicity', 'aggression'] model_dict = {} for task in tasks: os.system("python get_prod_models.py --task %s" % task) model_dict[task] = joblib.load("/tmp/%s_linear_char_oh_pipeline.pkl" % task) def apply_models(df): comments = df['comment'] for task, model in model_dict.items(): scores = model.predict_proba(comments)[:,1] df['pred_%s_score' % task] = scores return df def pred_helper(df): if len(df) == 0: return None return df.assign(timestamp = lambda x: pd.to_datetime(x.timestamp), comment = lambda x: x['comment'].astype(str))\ .pipe(apply_models) def prep_in_parallel(path, k = 8): df = pd.read_csv(path, sep = '\t', encoding = 'utf-8') m = df.shape[0] if m < 15000: n_groups = 1 else: n_groups = int(m / 10000.0) df['key'] = np.random.randint(0, high=n_groups, size=m) dfs = [e[1] for e in df.groupby('key')] #dfs = [pred_helper(d) for d in dfs] p = mp.Pool(k) dfs = p.map(pred_helper, dfs) p.close() p.join() return pd.concat(dfs) base = '../../data/figshare/' nss = ['user', 'article'] years = range(2001, 2016) for ns in nss: for year in years: dirname = "comments_%s_%d" % (ns, year) print(dirname) indir = os.path.join(base, dirname + ".tar.gz") os.system("mkdir ", os.path.join(base, "scored")) outf = os.path.join(base, "scored", dirname + ".tsv.gz") os.system("cp %s ." % indir) os.system("tar -zxvf %s.tar.gz" % dirname) dfs = [] for inf in os.listdir(dirname): print(inf) if inf.endswith(".tsv"): df = prep_in_parallel(os.path.join(dirname, inf), k = 8) dfs.append(df) os.system("rm -rf %s" % dirname) os.system("rm -rf %s.tar.gz" % dirname) pd.concat(dfs).to_csv(outf, sep = '\t', index = False, compression = "gzip") df.sort_values("pred_toxicity_score").tail() ###Output _____no_output_____
docs/feature-store/end-to-end-demo/01-ingest-datasources.ipynb
###Markdown Part 1: Data Ingestion This demo showcases financial fraud prevention. It uses the MLRun feature store to define complex features that help identify fraud. Fraud prevention is a special challenge since it requires processing raw transaction and events in real-time and being able to quickly respond and block transactions before they occur.To address this, you'll create a development pipeline and a production pipeline. Both pipelines share the same feature engineering and model code, but serve data very differently. Furthermore, MLRun automates the data and model monitoring process, drift identification, and trigger retraining in a CI/CD pipeline. This process is described in the diagram below:![Feature store demo diagram - fraud prevention](../../_static/images/feature_store_demo_diagram.png) The raw data is described as follows:| TRANSACTIONS || &x2551; |USER EVENTS || |-----------------|----------------------------------------------------------------|----------|-----------------|----------------------------------------------------------------|| **age** | age group value 0-6. Some values are marked as U for unknown | &x2551; | **source** | The party/entity related to the event || **gender** | A character to define the age | &x2551; | **event** | event, such as login or password change || **zipcodeOri** | ZIP code of the person originating the transaction | &x2551; | **timestamp** | The date and time of the event || **zipMerchant** | ZIP code of the merchant receiving the transaction | &x2551; | | || **category** | category of the transaction (e.g., transportation, food, etc.) | &x2551; | | || **amount** | the total amount of the transaction | &x2551; | | || **fraud** | whether the transaction is fraudulent | &x2551; | | || **timestamp** | the date and time in which the transaction took place | &x2551; | | || **source** | the ID of the party/entity performing the transaction | &x2551; | | || **target** | the ID of the party/entity receiving the transaction | &x2551; | | || **device** | the device ID used to perform the transaction | &x2551; | | | This notebook introduces how to **Ingest** different data sources to the **Feature Store**.The following FeatureSets are created:- **Transactions**: Monetary transactions between a source and a target.- **Events**: Account events such as account login or a password change.- **Label**: Fraud label for the data.By the end of this tutorial you’ll know how to:- Create an ingestion pipeline for each data source.- Define preprocessing, aggregation, and validation of the pipeline.- Run the pipeline locally within the notebook.- Launch a real-time function to ingest live data.- Schedule a cron to run the task when needed. ###Code project_name = 'fraud-demo' import mlrun # Initialize the MLRun project object project = mlrun.get_or_create_project(project_name, context="./", user_project=True) ###Output > 2022-03-16 05:45:07,703 [info] loaded project fraud-demo from MLRun DB ###Markdown Step 1 - Fetch, process and ingest the datasets 1.1 - Transactions Transactions ###Code # Helper functions to adjust the timestamps of our data # while keeping the order of the selected events and # the relative distance from one event to the other def date_adjustment(sample, data_max, new_max, old_data_period, new_data_period): ''' Adjust a specific sample's date according to the original and new time periods ''' sample_dates_scale = ((data_max - sample) / old_data_period) sample_delta = new_data_period * sample_dates_scale new_sample_ts = new_max - sample_delta return new_sample_ts def adjust_data_timespan(dataframe, timestamp_col='timestamp', new_period='2d', new_max_date_str='now'): ''' Adjust the dataframe timestamps to the new time period ''' # Calculate old time period data_min = dataframe.timestamp.min() data_max = dataframe.timestamp.max() old_data_period = data_max-data_min # Set new time period new_time_period = pd.Timedelta(new_period) new_max = pd.Timestamp(new_max_date_str) new_min = new_max-new_time_period new_data_period = new_max-new_min # Apply the timestamp change df = dataframe.copy() df[timestamp_col] = df[timestamp_col].apply(lambda x: date_adjustment(x, data_max, new_max, old_data_period, new_data_period)) return df import pandas as pd # Fetch the transactions dataset from the server transactions_data = pd.read_csv('https://s3.wasabisys.com/iguazio/data/fraud-demo-mlrun-fs-docs/data.csv', parse_dates=['timestamp'], nrows=500) # Adjust the samples timestamp for the past 2 days transactions_data = adjust_data_timespan(transactions_data, new_period='2d') # Preview transactions_data.head(3) ###Output _____no_output_____ ###Markdown Transactions - create a feature set and preprocessing pipelineCreate the feature set (data pipeline) definition for the **credit transaction processing** that describes the offline/online data transformations and aggregations.The feature store automatically adds an offline `parquet` target and an online `NoSQL` target by using `set_targets()`.The data pipeline consists of:* **Extracting** the data components (hour, day of week)* **Mapping** the age values* **One hot encoding** for the transaction category and the gender* **Aggregating** the amount (avg, sum, count, max over 2/12/24 hour time windows)* **Aggregating** the transactions per category (over 14 days time windows)* **Writing** the results to **offline** (Parquet) and **online** (NoSQL) targets ###Code # Import MLRun's Feature Store import mlrun.feature_store as fstore from mlrun.feature_store.steps import OneHotEncoder, MapValues, DateExtractor # Define the transactions FeatureSet transaction_set = fstore.FeatureSet("transactions", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="transactions feature set") # Define and add value mapping main_categories = ["es_transportation", "es_health", "es_otherservices", "es_food", "es_hotelservices", "es_barsandrestaurants", "es_tech", "es_sportsandtoys", "es_wellnessandbeauty", "es_hyper", "es_fashion", "es_home", "es_contents", "es_travel", "es_leisure"] # One Hot Encode the newly defined mappings one_hot_encoder_mapping = {'category': main_categories, 'gender': list(transactions_data.gender.unique())} # Define the graph steps transaction_set.graph\ .to(DateExtractor(parts = ['hour', 'day_of_week'], timestamp_col = 'timestamp'))\ .to(MapValues(mapping={'age': {'U': '0'}}, with_original_features=True))\ .to(OneHotEncoder(mapping=one_hot_encoder_mapping)) # Add aggregations for 2, 12, and 24 hour time windows transaction_set.add_aggregation(name='amount', column='amount', operations=['avg','sum', 'count','max'], windows=['2h', '12h', '24h'], period='1h') # Add the category aggregations over a 14 day window for category in main_categories: transaction_set.add_aggregation(name=category,column=f'category_{category}', operations=['count'], windows=['14d'], period='1d') # Add default (offline-parquet & online-nosql) targets transaction_set.set_targets() # Plot the pipeline so we can see the different steps transaction_set.plot(rankdir="LR", with_targets=True) ###Output _____no_output_____ ###Markdown Transactions - ingestion ###Code # Ingest the transactions dataset through the defined pipeline transactions_df = fstore.ingest(transaction_set, transactions_data, infer_options=fstore.InferOptions.default()) transactions_df.head(3) ###Output persist count = 0 persist count = 100 persist count = 200 persist count = 300 persist count = 400 persist count = 500 persist count = 600 persist count = 700 persist count = 800 persist count = 900 persist count = 1000 ###Markdown 1.2 - User events User events - fetching ###Code # Fetch the user_events dataset from the server user_events_data = pd.read_csv('https://s3.wasabisys.com/iguazio/data/fraud-demo-mlrun-fs-docs/events.csv', index_col=0, quotechar="\'", parse_dates=['timestamp'], nrows=500) # Adjust to the last 2 days to see the latest aggregations in our online feature vectors user_events_data = adjust_data_timespan(user_events_data, new_period='2d') # Preview user_events_data.head(3) ###Output _____no_output_____ ###Markdown User events - create a feature set and preprocessing pipelineDefine the events feature set.This is a fairly straightforward pipeline in which you only "one hot encode" the event categories and save the data to the default targets. ###Code user_events_set = fstore.FeatureSet("events", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="user events feature set") # Define and add value mapping events_mapping = {'event': list(user_events_data.event.unique())} # One Hot Encode user_events_set.graph.to(OneHotEncoder(mapping=events_mapping)) # Add default (offline-parquet & online-nosql) targets user_events_set.set_targets() # Plot the pipeline so we can see the different steps user_events_set.plot(rankdir="LR", with_targets=True) ###Output _____no_output_____ ###Markdown User events - ingestion ###Code # Ingestion of the newly created events feature set events_df = fstore.ingest(user_events_set, user_events_data) events_df.head(3) ###Output persist count = 0 persist count = 100 persist count = 200 persist count = 300 persist count = 400 persist count = 500 ###Markdown Step 2 - Create a labels dataset for model training Label set - create a feature setThis feature set contains the label for the fraud demo, it is ingested directly to the default targets without any changes ###Code def create_labels(df): labels = df[['fraud','source','timestamp']].copy() labels = labels.rename(columns={"fraud": "label"}) labels['timestamp'] = labels['timestamp'].astype("datetime64[ms]") labels['label'] = labels['label'].astype(int) labels.set_index('source', inplace=True) return labels # Define the "labels" feature set labels_set = fstore.FeatureSet("labels", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="training labels", engine="pandas") labels_set.graph.to(name="create_labels", handler=create_labels) # specify only Parquet (offline) target since its not used for real-time labels_set.set_targets(['parquet'], with_defaults=False) labels_set.plot(with_targets=True) ###Output _____no_output_____ ###Markdown Label set - ingestion ###Code # Ingest the labels feature set labels_df = fstore.ingest(labels_set, transactions_data) labels_df.head(3) ###Output _____no_output_____ ###Markdown Step 3 - Deploy a real-time pipelineWhen dealing with real-time aggregation, it's important to be able to update these aggregations in real-time.For this purpose, you'll create live serving functions that update the online feature store of the `transactions` FeatureSet and `Events` FeatureSet.Using MLRun's `serving` runtime, create a nuclio function loaded with the feature set's computational graph definitionand an `HttpSource` to define the HTTP trigger.Notice that the implementation below does not require any rewrite of the pipeline logic. 3.1 - Transactions Transactions - deploy the feature set live endpoint ###Code # Create iguazio v3io stream and transactions push API endpoint transaction_stream = f'v3io:///projects/{project.name}/streams/transaction' transaction_pusher = mlrun.datastore.get_stream_pusher(transaction_stream) # Define the source stream trigger (use v3io streams) # Define the `key` and `time` fields (extracted from the Json message). source = mlrun.datastore.sources.StreamSource(path=transaction_stream , key_field='source', time_field='timestamp') # Deploy the transactions feature set's ingestion service over a real-time (Nuclio) serverless function # you can use the run_config parameter to pass function/service specific configuration transaction_set_endpoint = fstore.deploy_ingestion_service(featureset=transaction_set, source=source) ###Output > 2022-03-16 05:45:43,035 [info] Starting remote function deploy 2022-03-16 05:45:43 (info) Deploying function 2022-03-16 05:45:43 (info) Building 2022-03-16 05:45:43 (info) Staging files and preparing base images 2022-03-16 05:45:43 (warn) Python 3.6 runtime is deprecated and will soon not be supported. Please migrate your code and use Python 3.7 runtime (`python:3.7`) or higher 2022-03-16 05:45:43 (info) Building processor image 2022-03-16 05:47:03 (info) Build complete 2022-03-16 05:47:08 (info) Function deploy complete > 2022-03-16 05:47:08,835 [info] successfully deployed function: {'internal_invocation_urls': ['nuclio-fraud-demo-admin-transactions-ingest.default-tenant.svc.cluster.local:8080'], 'external_invocation_urls': ['fraud-demo-admin-transactions-ingest-fraud-demo-admin.default-tenant.app.xtvtjecfcssi.iguazio-cd1.com/']} ###Markdown Transactions - test the feature set HTTP endpoint By defining the `transactions` feature set you can now use MLRun and Storey to deploy it as a live endpoint, ready to ingest new data!Using MLRun's `serving` runtime, create a nuclio function loaded with the feature set's computational graph definition and an `HttpSource` to define the HTTP trigger. ###Code import requests import json # Select a sample from the dataset and serialize it to JSON transaction_sample = json.loads(transactions_data.sample(1).to_json(orient='records'))[0] transaction_sample['timestamp'] = str(pd.Timestamp.now()) transaction_sample # Post the sample to the ingestion endpoint requests.post(transaction_set_endpoint, json=transaction_sample).text ###Output _____no_output_____ ###Markdown 3.2 - User events User events - deploy the feature set live endpointDeploy the events feature set's ingestion service using the feature set and all the previously defined resources. ###Code # Create iguazio v3io stream and transactions push API endpoint events_stream = f'v3io:///projects/{project.name}/streams/events' events_pusher = mlrun.datastore.get_stream_pusher(events_stream) # Define the source stream trigger (use v3io streams) # Define the `key` and `time` fields (extracted from the Json message). source = mlrun.datastore.sources.StreamSource(path=events_stream , key_field='source', time_field='timestamp') # Deploy the transactions feature set's ingestion service over a real-time (Nuclio) serverless function # you can use the run_config parameter to pass function/service specific configuration events_set_endpoint = fstore.deploy_ingestion_service(featureset=user_events_set, source=source) ###Output > 2022-03-16 05:47:09,035 [info] Starting remote function deploy 2022-03-16 05:47:09 (info) Deploying function 2022-03-16 05:47:09 (info) Building 2022-03-16 05:47:09 (info) Staging files and preparing base images 2022-03-16 05:47:09 (warn) Python 3.6 runtime is deprecated and will soon not be supported. Please migrate your code and use Python 3.7 runtime (`python:3.7`) or higher 2022-03-16 05:47:09 (info) Building processor image ###Markdown User events - test the feature set HTTP endpoint ###Code # Select a sample from the events dataset and serialize it to JSON user_events_sample = json.loads(user_events_data.sample(1).to_json(orient='records'))[0] user_events_sample['timestamp'] = str(pd.Timestamp.now()) user_events_sample # Post the sample to the ingestion endpoint requests.post(events_set_endpoint, json=user_events_sample).text ###Output _____no_output_____ ###Markdown Part 1: Data Ingestion In this notebook we will learn how to **Ingest** different data sources to our **Feature Store**. Specifically, this patient data has been successfully used to treat hospitalized COVID-19 patients prior to their condition becoming severe or critical. To do this we will use a medical dataset which includes three types of data: - **Healthcare systems**: Batch updated dataset, containing different lab test results (Blood test results for ex.).- **Patient Records**: Static dataset containing general patient details.- **Real-time sensors**: Real-Time patient metric monitoring sensor. We will walk through creation of ingestion pipeline for each datasource with all the needed preprocessing and validation. We will run the pipeline locally within the notebook and then launch a real-time function to **ingest live data** or schedule a cron to run the task when needed. Environment SetupSince our work is done in a this project scope, first define the project itself for all our MLRun work in this notebook. ###Code import mlrun from os import getenv mlrun.set_environment(project='fsdemo', user_project=True) # location of the output data files data_path = f"{getenv('V3IO_HOME_URL')}/demos/feature-store/data/" def move_timestamps(df, shift='0s'): ''' Update timetsamps to current time so we can see live aggregations ''' now = pd.to_datetime('now') max_time = df['timestamp'].max() time_shift = now-max_time tmp_df = df.copy() tmp_df['timestamp'] = tmp_df['timestamp'].apply(lambda t: t + time_shift + pd.to_timedelta(shift)) return tmp_df ###Output _____no_output_____ ###Markdown Create Ingestion Pipeline With MLRunIn this section we will ingest the lab measurements data using MLRun and Storey. Storey is the underlying implementation of the feature store which is used by MLRun. It is the engine that allows you to define and execute complex graphs that create the feature engineering pipeline. With storey, you can define source, transformations and targets, many actions are available as part of the Storey library, but you can define additional actions easily. We will see these custom actions in later sections.For the execution, it is possible to also use Spark. The main difference between Storey and Spark pipelines is that Storey blocks are built for Real-Time workloads while Spark is more Batch oriented. We will now do the following:- Create the `measurements` FeatureSet- Define Preprocessing graph including aggregations- Ingest the data using the defined pipeline ###Code # Import MLRun's Feature Store import mlrun.feature_store as fs # Import MLRun's Data Sources to set the wanted ingestion pipeline from mlrun.datastore.sources import CSVSource, ParquetSource, HttpSource # Import storey so it will be available on our scope # when testing the pipeline import storey # Define the Lab Measurements FeatureSet measurements_set = fs.FeatureSet("measurements", entities=[fs.Entity("patient_id")], timestamp_key='timestamp', description="various patient health measurements") # Get FeatureSet computation graph measurements_graph = measurements_set.graph ###Output _____no_output_____ ###Markdown Define the processing pipeline- Transformation function- Sliding window aggregation- Set targets (NoSQL and Parquet) ###Code # Import pandas and load the sample CSV and load it as a datasource # for our ingestion import pandas as pd measurements_df = pd.read_csv('https://s3.wasabisys.com/iguazio/data/patients/measurements.csv', index_col=0) measurements_df['timestamp'] = pd.to_datetime(measurements_df['timestamp']) measurements_df['timestamp'] = measurements_df['timestamp'].astype("datetime64[ms]") measurements_df = pd.concat([move_timestamps(measurements_df, '-1h'), move_timestamps(measurements_df)]) # update timestamps ###Output _____no_output_____ ###Markdown Take a look at the measurements dataset. This dataset includes a a single measurement per row. The measurement type is defined by the `source` and `parameter` column. We would like to transform this data, so each patient has multiple measurement columns. To do that, we will need to create a new column for each `source` and `parameter` combination. For example, if `source` is 3 and `parameter` is 0, then our transformed dataset will have the measurement value in a new feature named `sp_3_0`.Following that, we will create a sliding window aggregation that averages the values across that time window. ###Code measurements_df.head() ###Output _____no_output_____ ###Markdown The following code performs the transformation, adds the aggregation and sets the target to store the values to a NoSQL database for online retrieval and parquet files for batch processing. ###Code # Define transform to create sparse dataset for aggregation # adding an extra column for the specific source-parameter pair's measurement # ex: source=3, parameter=4, measurement=100 -> add extra column sp_3_4=100 def transform(event): event["_".join(['sp', str(event["source"]), str(event["parameter"])])] = event["measurement"] return event # Define Measurement FeatureSet pipeline measurements_graph.to( "storey.Map", _fn="transform" ) # Get the available source, parameter pairs for our aggregation sps = list(measurements_df.apply(lambda x: '_'.join(['sp', str(x['source']), str(x['parameter'])]), axis=1).unique()) # Add aggregations on top of the created sparse # features by the transorm function for col in sps: measurements_set.add_aggregation(name=f'agg_{col}', column=col, operations=['avg'], windows='1h', period='30m') # Add default (NoSQL via KV and Parquet) targets to save # the ingestion results to measurements_set.set_targets() ###Output _____no_output_____ ###Markdown You can plot the graph to visalize the pipeline: ###Code # Plot the ingestion pipeline we defined measurements_set.plot(rankdir='LR', with_targets=True) ###Output _____no_output_____ ###Markdown Run ingestion task using MLRun & StoreyIn order to ingest the dataframe to the featureset, use the `ingest` function. ###Code # User our loaded DF as the datasource and ingest it through # the define pipeline resp = fs.ingest(measurements_set, measurements_df, infer_options=fs.InferOptions.default()) resp.head() # Save the FeatureSet and pipeline definition measurements_set.save() ###Output _____no_output_____ ###Markdown Ingest Patient Details Features In this section we will use MLRun to create our patient details datasource. We will do the following:- Create a `patient_details` FeatureSet- Add preprocessing transformations to the pipeline - Map ages to buckets and One Hot Encode them - Impute missing values- Test the processing pipeline with sample data- Run ingestion pipeline on top of the cluster Create the FeatureSet ###Code # add feature set without time column (stock ticker metadata) patients_set = fs.FeatureSet("patient_details", entities=[fs.Entity("patient_id")], description="personal and medical patient details") # Get FeatureSet computation graph graph = patients_set.spec.graph ###Output _____no_output_____ ###Markdown Define the computation pipeline ###Code # Define age buckets for our age value mapping personal_details = {'age': {'ranges': [{'range': [0, 3], "value": "toddler"}, {'range': [3, 18], "value": "child"}, {'range': [18, 65], "value": "adult"}, {'range': [65, 120], "value": "elder"}]}} # Define one hot encoding values map one_hot_encoder_mapping = {'age_mapped': ['toddler', 'child', 'adult', 'elder']} # Import MLRun's FeatureStore steps for easy # use in our pipeline from mlrun.feature_store.steps import * # Define the pipeline for our FeatureSet graph.to(MapValues(mapping=personal_details, with_original_features=True))\ .to(OneHotEncoder(mapping=one_hot_encoder_mapping))\ .to(Imputer(method='values', default_value=1, mapping={})) # Add default NoSQL & Parquet ingestion targets patients_set.set_targets() # Plot the FeatureSet pipeline patients_set.plot(rankdir='LR', with_targets=True) ###Output _____no_output_____ ###Markdown Test the Feature transformation pipelineCreating a transformation pipeline requires some trial and error. Therefore, it is useful to run the pipeline in memory without storing the resultant data. For this purpose, `infer` is used. This function receives as input any sample DataFrame, performs all the graph steps and outputs the transformed DataFrame. ###Code # Load the sample patient details data patients_df = pd.read_parquet('https://s3.wasabisys.com/iguazio/data/patients/patient_details.parquet') # Run local ingestion test fs.infer(patients_set, patients_df.head()) ###Output _____no_output_____ ###Markdown Save the FeatureSet and run full ingestion taskOnce you are satisfied with the transformation pipeline, ingest that full DataFrame and store the data. ###Code # Save the FeatureSet patients_set.save() # Run Ingestion task resp = fs.ingest(patients_set, patients_df, infer_options=fs.InferOptions.default()) ###Output _____no_output_____ ###Markdown Start Immediate or Scheduled Ingestion Job (over Kubernetes)Another useful method to ingest data, is by creating a Kubernetes job. This may be necessary to process large amounts of data as well as to process any recurring data. With MLRun it is easy to take the pipeline and run it as a job. This is done by:1. Define a source, specifically here we define a parquet file source2. Define a configuration where `local` is set to `False`3. Mount to the provisioned storage by calling `auto_mount`4. Run `ingest` with the source and run configuration ###Code source = ParquetSource('pq', 'https://s3.wasabisys.com/iguazio/data/patients/patient_details.parquet') config = fs.RunConfig(local=False).apply(mlrun.platforms.auto_mount()) fs.ingest(patients_set, source, run_config=config) ###Output > 2021-05-06 15:27:08,769 [info] starting run patient_details_ingest uid=76f197b8ab3347d1b995a5ea55d0a98a DB=http://mlrun-api:8080 > 2021-05-06 15:27:09,022 [info] Job is running in the background, pod: patient-details-ingest-g9hgn > 2021-05-06 15:27:15,073 [info] starting ingestion task to store://feature-sets/fsdemo-admin/patient_details:latest > 2021-05-06 15:27:15,745 [info] ingestion task completed, targets: > 2021-05-06 15:27:15,746 [info] [{'name': 'parquet', 'kind': 'parquet', 'path': 'v3io:///projects/fsdemo-admin/fs/parquet/sets/patient_details-latest.parquet', 'status': 'created', 'updated': '2021-05-06T15:27:15.432576+00:00'}, {'name': 'nosql', 'kind': 'nosql', 'path': 'v3io:///projects/fsdemo-admin/fs/nosql/sets/patient_details-latest', 'status': 'created', 'updated': '2021-05-06T15:27:15.432947+00:00'}] > 2021-05-06 15:27:15,936 [info] run executed, status=completed final state: completed ###Markdown Real-time Early-Sense Sensor Ingestion (HTTP or Stream Processing With Nuclio) In this section we will use MLRun to create our Early Sense Sensor datasource. We will do the following:- Create early sense FeatureSet- Add Preprocessing transformations to the Pipeline using custom functions - Drop and Rename columns - Aggregations- Add Feature Validator to detect bad sensor readings- Test the processing pipeline with sample data- Deploy the FeatureSet ingestion service as a live rest endpoint ###Code early_sense_set = fs.FeatureSet("early_sense", entities=[fs.Entity("patient_id")], timestamp_key='timestamp', description="real time patient bed sensor data") ###Output _____no_output_____ ###Markdown Define data validation & quality policyWe can define validations on the feature level. For example, define here validation to check if the heart-rate value is between 0 and 220 and respitory rate is between 0 and 25. ###Code from mlrun.features import MinMaxValidator early_sense_set["hr"] = fs.Feature(validator = MinMaxValidator(min=0, max=220, severity="info")) early_sense_set["rr"] = fs.Feature(validator = MinMaxValidator(min=0, max=25, severity="info")) ###Output _____no_output_____ ###Markdown Define custom processing classesIn the previous sections we used transformation steps that are available as part of Storey. Here we show how to create custom transformation classes. We will later run these functions as part of a Nuclio serverless real-time function, therefore, we also use the nuclio `start-code` and `end-code` comments. ###Code # nuclio: start-code # We will import storey here too so it will # be included in our function code (within the nuclio comment block) import json import storey from typing import List, Dict # The custom functions are based on `storey.MapClass` # when they are called in the graph the `do(self, event)` # function will be activated. # A to_dict(self) function is also required by MLRun # to allow the class creation on remote functions class DropColumns(storey.MapClass): def __init__(self, columns: List[str], **kwargs): super().__init__(**kwargs) self.columns = columns def do(self, event): for col in self.columns: if col in event: del event[col] return event def to_dict(self): return { "class_name": "DropColumns", "name": self.name or "DropColumns", "class_args": { "columns": self.columns }, } class RenameColumns(storey.MapClass): def __init__(self, mapping: Dict[str, str], **kwargs): super().__init__(**kwargs) self.mapping = mapping def do(self, event): for old_col, new_col in self.mapping.items(): try: event[new_col] = event.pop(old_col) except Exception as e: print(f'{old_col} doesnt exist') return event def to_dict(self): return { "class_name": "RenameColumns", "name": self.name or "RenameColumns", "class_args": {"mapping": self.mapping}, } # nuclio: end-code ###Output _____no_output_____ ###Markdown Define the Real-Time PipelineDefine the transoformation pipeline below. This is done just like the previous sections. ###Code # Configure the list of columns to drop from # the raw data drop_columns = ['hr_is_error', 'rr_is_error', 'spo2_is_error', 'movements_is_error', 'turn_count_is_error', 'is_in_bed_is_error'] # Define the computationala graph including our custom functions early_sense_set.graph.to(DropColumns(drop_columns), after='start')\ .to(RenameColumns(mapping={'bad': 'bed'})) # Add real-time aggreagations on top of our sensor readings for col in ['hr', 'rr', 'spo2', 'movements', 'turn_count']: early_sense_set.add_aggregation(col + "_h", col, ['avg', 'max', 'min'], "1h") early_sense_set.add_aggregation(col + "_d", col, ['avg', 'max', 'min'], "1d") early_sense_set.add_aggregation('in_bed_h', 'is_in_bed', ['avg'], "1h") early_sense_set.add_aggregation('in_bed_d', 'is_in_bed', ['avg'], "1d") # Set NoSQL and Parquet default targets early_sense_set.set_targets() # Plot the pipeline early_sense_set.plot(rankdir='LR', with_targets=True) ###Output _____no_output_____ ###Markdown Test/debug the real-time pipeline locally in the notebook ###Code # infer schema + stats, show the final feature set (after the data pipeline) early_sense_df = pd.read_parquet('https://s3.wasabisys.com/iguazio/data/patients/early_sense.parquet') early_sense_df['timestamp'] = pd.to_datetime(early_sense_df['timestamp']) early_sense_df = move_timestamps(early_sense_df) # update timestamps fs.infer(early_sense_set, early_sense_df.head()) # Run ingest pipeline df=fs.ingest(early_sense_set, early_sense_df) # Save the early-sense Featureset early_sense_set.save() # print the FeatureSet spec print(early_sense_set.status.targets.to_dict()) ###Output [{'name': 'parquet', 'kind': 'parquet', 'path': 'v3io:///projects/fsdemo-admin/fs/parquet/sets/early_sense-latest.parquet', 'status': 'created', 'updated': '2021-05-06T15:27:46.222973+00:00'}, {'name': 'nosql', 'kind': 'nosql', 'path': 'v3io:///projects/fsdemo-admin/fs/nosql/sets/early_sense-latest', 'status': 'created', 'updated': '2021-05-06T15:27:46.223349+00:00'}] ###Markdown Deploy as Real-Time Stream Processing Function (Nuclio Serverless)Features are not static. For example, it is common that features include different aggregations that need to be updated as data continues to flow. A real-time pipeline requires this data to be up date. Therefore, we need a convenient way to ingest data, not just as batch, but per specific input.MLRun can convert any code to a real-time serverless function, including the pipeline. This is done by performing the following steps:1. Define a source, in this case it's an HTTP source2. Convert the previously defined code to a serving function3. Create a configuration to run the function4. Deploy an ingestion service with the Featureset, source and the configuration ###Code # Set a new HTTPSource, this will tell our ingestion service # to setup a Nuclio function to act as the rest endpoint # to which we would receive the data source = HttpSource(key_field='patient_id', time_field='timestamp') # Take the relevant code parts from this notebook and create # an MLRun function from them so we can run the pipeline # as a Nuclio function func = mlrun.code_to_function("ingest", kind="serving") nuclio_config = fs.RunConfig(function=func, local=False).apply(mlrun.platforms.auto_mount()) # Deploy the Online ingestion service using the pipeline definition from before # with our new HTTP Source and our define Function server = fs.deploy_ingestion_service(early_sense_set, source, run_config=nuclio_config) ###Output > 2021-05-06 15:29:52,032 [info] Starting remote function deploy 2021-05-06 15:29:52 (info) Deploying function {'level': 'info', 'message': 'Deploying function', 'name': 'fsdemo-admin-ingest', 'time': 1620314992169.7139} 2021-05-06 15:29:52 (info) Building {'level': 'info', 'message': 'Building', 'name': 'fsdemo-admin-ingest', 'time': 1620314992169.7478, 'versionInfo': 'Label: 1.5.16, Git commit: ae43a6a560c2bec42d7ccfdf6e8e11a1e3cc3774, OS: linux, Arch: amd64, Go version: go1.14.3'} 2021-05-06 15:29:52 (info) Staging files and preparing base images {'level': 'info', 'message': 'Staging files and preparing base images', 'name': 'deployer', 'time': 1620314992237.7905} 2021-05-06 15:29:52 (info) Building processor image {'imageName': 'fsdemo-admin-fsdemo-admin-ingest-processor:latest', 'level': 'info', 'message': 'Building processor image', 'name': 'deployer', 'time': 1620314992238.347} 2021-05-06 15:29:55 (info) Build complete {'level': 'info', 'message': 'Build complete', 'name': 'deployer', 'result': {'Image': 'fsdemo-admin-fsdemo-admin-ingest-processor:latest', 'UpdatedFunctionConfig': {'metadata': {'annotations': {'nuclio.io/generated_by': 'function generated from https://github.com/mlrun/mlrun#004d7b6797e3292525d220bb4389470342ebe752:ingest.ipynb'}, 'labels': {'mlrun/class': 'serving', 'nuclio.io/project-name': 'fsdemo-admin'}, 'name': 'fsdemo-admin-ingest', 'namespace': 'default-tenant'}, 'spec': {'build': {'baseImage': 'mlrun/mlrun:0.6.3-rc9', 'codeEntryType': 'sourceCode', 'functionSourceCode': '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', 'noBaseImagesPull': True, 'offline': True, 'registry': 'docker-registry.default-tenant.app.yh30.iguazio-c0.com'}, 'env': [{'name': 'V3IO_API', 'value': 'v3io-webapi.default-tenant.svc:8081'}, {'name': 'V3IO_USERNAME', 'value': 'admin'}, {'name': 'V3IO_ACCESS_KEY', 'value': '142a98fa-bef9-4095-b2d0-cab733f53238'}, {'name': 'MLRUN_LOG_LEVEL', 'value': 'DEBUG'}, {'name': 'MLRUN_DEFAULT_PROJECT', 'value': 'fsdemo-admin'}, {'name': 'MLRUN_DBPATH', 'value': 'http://mlrun-api:8080'}, {'name': 'MLRUN_NAMESPACE', 'value': 'default-tenant'}, {'name': 'SERVING_SPEC_ENV', 'value': '{"function_uri": "fsdemo-admin/ingest", "version": "v2", "parameters": {"infer_options": 0, "featureset": "store://feature-sets/fsdemo-admin/early_sense", "source": {"kind": "http", "path": "None", "key_field": "patient_id", "time_field": "timestamp", "online": true}}, "graph": {"states": {"DropColumns": {"kind": "task", "class_name": "DropColumns", "class_args": {"columns": ["hr_is_error", "rr_is_error", "spo2_is_error", "movements_is_error", "turn_count_is_error", "is_in_bed_is_error"]}}, "RenameColumns": {"kind": "task", "class_name": "RenameColumns", "class_args": {"mapping": {"bad": "bed"}}, "after": ["DropColumns"]}, "Aggregates": {"kind": "task", "class_name": "storey.AggregateByKey", "class_args": {"aggregates": [{"name": "hr", "column": "hr", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "rr", "column": "rr", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "spo2", "column": "spo2", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "movements", "column": "movements", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "turn_count", "column": "turn_count", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "in_bed", "column": "is_in_bed", "operations": ["avg"], "windows": ["1h", "1d"]}], "table": "."}, "after": ["RenameColumns"]}}}, "load_mode": null, "functions": {}, "graph_initializer": "mlrun.feature_store.ingestion.featureset_initializer", "error_stream": null, "track_models": null}'}], 'eventTimeout': '', 'handler': '01-ingest-datasources:handler', 'maxReplicas': 4, 'minReplicas': 1, 'platform': {}, 'resources': {}, 'runtime': 'python:3.6', 'securityContext': {}, 'serviceType': 'NodePort', 'triggers': {'default-http': {'attributes': {'serviceType': 'NodePort'}, 'class': '', 'kind': 'http', 'maxWorkers': 1, 'name': 'default-http'}}, 'volumes': [{'volume': {'flexVolume': {'driver': 'v3io/fuse', 'options': {'accessKey': '142a98fa-bef9-4095-b2d0-cab733f53238'}}, 'name': 'v3io'}, 'volumeMount': {'mountPath': '/v3io', 'name': 'v3io'}}, {'volume': {'flexVolume': {'driver': 'v3io/fuse', 'options': {'accessKey': '142a98fa-bef9-4095-b2d0-cab733f53238'}}, 'name': 'v3io'}, 'volumeMount': {'mountPath': '/User', 'name': 'v3io', 'subPath': 'users/admin'}}]}}}, 'time': 1620314995613.7964} > 2021-05-06 15:30:03,749 [info] function deployed, address=default-tenant.app.yh30.iguazio-c0.com:31610 ###Markdown Test the function by sending data to the HTTP endpoint ###Code test_data = {'patient_id': '838-21-8151', 'bad': 38, 'department': '01e9fe31-76de-45f0-9aed-0f94cc97bca0', 'room': 1, 'hr': 220.0, 'hr_is_error': True, 'rr': 5, 'rr_is_error': True, 'spo2': 85, 'spo2_is_error': True, 'movements': 0.0, 'movements_is_error': True, 'turn_count': 0.0, 'turn_count_is_error': True, 'is_in_bed': 1, 'is_in_bed_is_error': False, 'timestamp': 1606843455.906352 } import requests import json response = requests.post(server, json=test_data) response.text ###Output _____no_output_____ ###Markdown Ingest labelsFinally, we define label data, this will be useful in the next notebook where we train a model Create Labels Set ###Code # Define labels metric from the early sense error data error_columns = [c for c in early_sense_df.columns if 'error' in c] labels = early_sense_df.loc[:, ['patient_id', 'timestamp'] + error_columns] labels['label'] = labels.apply(lambda x: sum([x[c] for c in error_columns])>(len(error_columns)*0.7), axis=1) labels.to_parquet(data_path + 'labels.parquet') #labels_df = pd.read_parquet('labels.parquet') labels_set = fs.FeatureSet("labels", entities=[fs.Entity("patient_id")], timestamp_key='timestamp', description="training labels") labels_set.set_targets() df = fs.infer(labels_set, data_path + 'labels.parquet') df.head() df = fs.ingest(labels_set, data_path + 'labels.parquet') labels_set.save() ###Output _____no_output_____ ###Markdown Part 1: Data Ingestion This demo showcases financial fraud prevention and using the MLRun feature store to define complex features that help identify fraud. Fraud prevention specifically is a challenge as it requires processing raw transaction and events in real-time and being able to quickly respond and block transactions before they occur.To address this, we create a development pipeline and a production pipeline. Both pipelines share the same feature engineering and model code, but serve data very differently. Furthermore, we automate the data and model monitoring process, identify drift and trigger retraining in a CI/CD pipeline. This process is described in the diagram below:![Feature store demo diagram - fraud prevention](../../_static/images/feature_store_demo_diagram.png) The raw data is described as follows:| TRANSACTIONS || &x2551; |USER EVENTS || |-----------------|----------------------------------------------------------------|----------|-----------------|----------------------------------------------------------------|| **age** | age group value 0-6. Some values are marked as U for unknown | &x2551; | **source** | The party/entity related to the event || **gender** | A character to define the age | &x2551; | **event** | event, such as login or password change || **zipcodeOri** | ZIP code of the person originating the transaction | &x2551; | **timestamp** | The date and time of the event || **zipMerchant** | ZIP code of the merchant receiving the transaction | &x2551; | | || **category** | category of the transaction (e.g., transportation, food, etc.) | &x2551; | | || **amount** | the total amount of the transaction | &x2551; | | || **fraud** | whether the transaction is fraudulent | &x2551; | | || **timestamp** | the date and time in which the transaction took place | &x2551; | | || **source** | the ID of the party/entity performing the transaction | &x2551; | | || **target** | the ID of the party/entity receiving the transaction | &x2551; | | || **device** | the device ID used to perform the transaction | &x2551; | | | This notebook introduces how to **Ingest** different data sources to the **Feature Store**.The following FeatureSets will be created:- **Transactions**: Monetary transactions between a source and a target.- **Events**: Account events such as account login or a password change.- **Label**: Fraud label for the data.By the end of this tutorial you’ll learn how to:- Create an ingestion pipeline for each data source.- Define preprocessing, aggregation and validation of the pipeline.- Run the pipeline locally within the notebook.- Launch a real-time function to ingest live data.- Schedule a cron to run the task when needed. ###Code project_name = 'fraud-demo' import mlrun # Initialize the MLRun project object project = mlrun.get_or_create_project(project_name, context="./", user_project=True) ###Output > 2021-10-28 11:25:47,346 [info] loaded project fraud-demo from MLRun DB ###Markdown Step 1 - Fetch, Process and Ingest our datasets 1.1 - Transactions Transactions ###Code # Helper functions to adjust the timestamps of our data # while keeping the order of the selected events and # the relative distance from one event to the other def date_adjustment(sample, data_max, new_max, old_data_period, new_data_period): ''' Adjust a specific sample's date according to the original and new time periods ''' sample_dates_scale = ((data_max - sample) / old_data_period) sample_delta = new_data_period * sample_dates_scale new_sample_ts = new_max - sample_delta return new_sample_ts def adjust_data_timespan(dataframe, timestamp_col='timestamp', new_period='2d', new_max_date_str='now'): ''' Adjust the dataframe timestamps to the new time period ''' # Calculate old time period data_min = dataframe.timestamp.min() data_max = dataframe.timestamp.max() old_data_period = data_max-data_min # Set new time period new_time_period = pd.Timedelta(new_period) new_max = pd.Timestamp(new_max_date_str) new_min = new_max-new_time_period new_data_period = new_max-new_min # Apply the timestamp change df = dataframe.copy() df[timestamp_col] = df[timestamp_col].apply(lambda x: date_adjustment(x, data_max, new_max, old_data_period, new_data_period)) return df import pandas as pd # Fetch the transactions dataset from the server transactions_data = pd.read_csv('https://s3.wasabisys.com/iguazio/data/fraud-demo-mlrun-fs-docs/data.csv', parse_dates=['timestamp']) # Adjust the samples timestamp for the past 2 days transactions_data = adjust_data_timespan(transactions_data.sample(50000), new_period='2d') # Preview transactions_data.head(3) ###Output _____no_output_____ ###Markdown Transactions - Create a FeatureSet and Preprocessing PipelineCreate the FeatureSet (data pipeline) definition for the **credit transaction processing** which describes the offline/online data transformations and aggregations.The feature store will automatically add an offline `parquet` target and an online `NoSQL` target by using `set_targets()`.The data pipeline consists of:* **Extracting** the data components (hour, day of week)* **Mapping** the age values* **One hot encoding** for the transaction category and the gender* **Aggregating** the amount (avg, sum, count, max over 2/12/24 hour time windows)* **Aggregating** the transactions per category (over 14 days time windows)* **Writing** the results to **offline** (Parquet) and **online** (NoSQL) targets ###Code # Import MLRun's Feature Store import mlrun.feature_store as fstore from mlrun.feature_store.steps import OneHotEncoder, MapValues, DateExtractor # Define the transactions FeatureSet transaction_set = fstore.FeatureSet("transactions", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="transactions feature set") # Define and add value mapping main_categories = ["es_transportation", "es_health", "es_otherservices", "es_food", "es_hotelservices", "es_barsandrestaurants", "es_tech", "es_sportsandtoys", "es_wellnessandbeauty", "es_hyper", "es_fashion", "es_home", "es_contents", "es_travel", "es_leisure"] # One Hot Encode the newly defined mappings one_hot_encoder_mapping = {'category': main_categories, 'gender': list(transactions_data.gender.unique())} # Define the graph steps transaction_set.graph\ .to(DateExtractor(parts = ['hour', 'day_of_week'], timestamp_col = 'timestamp'))\ .to(MapValues(mapping={'age': {'U': '0'}}, with_original_features=True))\ .to(OneHotEncoder(mapping=one_hot_encoder_mapping)) # Add aggregations for 2, 12, and 24 hour time windows transaction_set.add_aggregation(name='amount', column='amount', operations=['avg','sum', 'count','max'], windows=['2h', '12h', '24h'], period='1h') # Add the category aggregations over a 14 day window for category in main_categories: transaction_set.add_aggregation(name=category,column=f'category_{category}', operations=['count'], windows=['14d'], period='1d') # Add default (offline-parquet & online-nosql) targets transaction_set.set_targets() # Plot the pipeline so we can see the different steps transaction_set.plot(rankdir="LR", with_targets=True) ###Output _____no_output_____ ###Markdown Transactions - Ingestion ###Code # Ingest our transactions dataset through our defined pipeline transactions_df = fstore.ingest(transaction_set, transactions_data, infer_options=fstore.InferOptions.default()) transactions_df.head(3) ###Output _____no_output_____ ###Markdown 1.2 - User Events User Events - Fetching ###Code # Fetch our user_events dataset from the server user_events_data = pd.read_csv('https://s3.wasabisys.com/iguazio/data/fraud-demo-mlrun-fs-docs/events.csv', index_col=0, quotechar="\'", parse_dates=['timestamp']) # Adjust to the last 2 days to see the latest aggregations in our online feature vectors user_events_data = adjust_data_timespan(user_events_data, new_period='2d') # Preview user_events_data.head(3) ###Output _____no_output_____ ###Markdown User Events - Create a FeatureSet and Preprocessing PipelineNow we will define the events feature set.This is a pretty straight forward pipeline in which we only one hot encode the event categories and save the data to the default targets. ###Code user_events_set = fstore.FeatureSet("events", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="user events feature set") # Define and add value mapping events_mapping = {'event': list(user_events_data.event.unique())} # One Hot Encode user_events_set.graph.to(OneHotEncoder(mapping=events_mapping)) # Add default (offline-parquet & online-nosql) targets user_events_set.set_targets() # Plot the pipeline so we can see the different steps user_events_set.plot(rankdir="LR", with_targets=True) ###Output _____no_output_____ ###Markdown User Events - Ingestion ###Code # Ingestion of our newly created events feature set events_df = fstore.ingest(user_events_set, user_events_data) events_df.head(3) ###Output _____no_output_____ ###Markdown Step 2 - Create a labels dataset for model training Label Set - Create a FeatureSetThis feature set contains the label for the fraud demo, it will be ingested directly to the default targets without any changes ###Code def create_labels(df): labels = df[['fraud','source','timestamp']].copy() labels = labels.rename(columns={"fraud": "label"}) labels['timestamp'] = labels['timestamp'].astype("datetime64[ms]") labels['label'] = labels['label'].astype(int) labels.set_index('source', inplace=True) return labels # Define the "labels" feature set labels_set = fstore.FeatureSet("labels", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="training labels", engine="pandas") labels_set.graph.to(name="create_labels", handler=create_labels) # specify only Parquet (offline) target since its not used for real-time labels_set.set_targets(['parquet'], with_defaults=False) labels_set.plot(with_targets=True) ###Output _____no_output_____ ###Markdown Label Set - Ingestion ###Code # Ingest the labels feature set labels_df = fstore.ingest(labels_set, transactions_data) labels_df.head(3) ###Output 1000 None ###Markdown Step 3 - Deploy a real-time pipelineWhen dealing with real-time aggregation, it's important to be able to update these aggregations in real-time.For this purpose, we will create live serving functions that will update the online feature store of the `transactions` FeatureSet and `Events` FeatureSet.Using MLRun's `serving` runtime, craetes a nuclio function loaded with our feature set's computational graph definitionand an `HttpSource` to define the HTTP trigger.Notice that the implementation below does not require any rewrite of the pipeline logic. 3.1 - Transactions Transactions - Deploy our FeatureSet live endpoint ###Code # Create iguazio v3io stream and transactions push API endpoint transaction_stream = f'v3io:///projects/{project.name}/streams/transaction' transaction_pusher = mlrun.datastore.get_stream_pusher(transaction_stream) # Define the HTTP Source to_dictable the HTTP trigger on our function and expose the endpoint. # as any other datasource, we will define the `key` and `time` fields here too. http_source = mlrun.datastore.sources.HttpSource(key_field='source', time_field='timestamp') transaction_set.spec.source = http_source # Create a real-time serverless function definition to deploy the ingestion pipeline on. # the serving runtimes enables the deployment of our feature set's computational graph function = (mlrun.new_function('ingest-transactions', kind='serving', image='mlrun/mlrun')).with_code(body=" ") # Add stream trigger (must first create the stream) function.add_v3io_stream_trigger(transaction_stream) # run_config = fstore.RunConfig(function=function, local=False).apply(mlrun.mount_v3io()) # Deploy the transactions feature set's ingestion service using the feature set # and all the defined resources above. transaction_set_endpoint = fstore.deploy_ingestion_service(featureset=transaction_set, run_config=run_config) ###Output > 2021-09-19 17:58:50,402 [info] Starting remote function deploy 2021-09-19 17:58:50 (info) Deploying function 2021-09-19 17:58:50 (info) Building 2021-09-19 17:58:50 (info) Staging files and preparing base images 2021-09-19 17:58:50 (info) Building processor image 2021-09-19 17:58:52 (info) Build complete 2021-09-19 17:59:01 (info) Function deploy complete > 2021-09-19 17:59:01,461 [info] successfully deployed function: {'internal_invocation_urls': ['nuclio-fraud-demo-admin-ingest-transactions.default-tenant.svc.cluster.local:8080'], 'external_invocation_urls': ['fraud-demo-admin-ingest-transactions-fraud-demo-admin.default-tenant.app.jnewriujxdig.iguazio-cd1.com/']} ###Markdown Transactions - Test the feature set HTTP endpoint By defining our `transactions` feature set we can now use MLRun and Storey to deploy it as a live endpoint, ready to ingest new data!Using MLRun's `serving` runtime, we will create a nuclio function loaded with our feature set's computational graph definition and an `HttpSource` to define the HTTP trigger. ###Code import requests import json # Select a sample from the dataset and serialize it to JSON transaction_sample = json.loads(transactions_data.sample(1).to_json(orient='records'))[0] transaction_sample['timestamp'] = str(pd.Timestamp.now()) transaction_sample # Post the sample to the ingestion endpoint requests.post(transaction_set_endpoint, json=transaction_sample).text ###Output _____no_output_____ ###Markdown 3.2 - User Events User Events - Deploy our FeatureSet live endpointDeploy the events feature set's ingestion service using the feature set and all the previously defined resources. ###Code # Create iguazio v3io stream and transactions push API endpoint events_stream = f'v3io:///projects/{project.name}/streams/events' events_pusher = mlrun.datastore.get_stream_pusher(events_stream) # Create a `serving` "base function" to deploy the ingestion function on # the serving runtimes enables the deployment of our feature set's computational graph function = (mlrun.new_function('ingest-events', kind='serving', image='mlrun/mlrun')).with_code(body=" ") # Add stream trigger function.add_v3io_stream_trigger(events_stream) # run_config = fstore.RunConfig(function=function, local=False).apply(mlrun.mount_v3io()) # Deploy the transactions feature set's ingestion service using the feature set # and all the defined resources above. events_set_endpoint = fstore.deploy_ingestion_service(name="ingest-events", featureset=user_events_set, source=http_source, run_config=run_config) ###Output > 2021-09-19 17:59:01,795 [info] Starting remote function deploy 2021-09-19 17:59:02 (info) Deploying function 2021-09-19 17:59:02 (info) Building 2021-09-19 17:59:02 (info) Staging files and preparing base images 2021-09-19 17:59:02 (info) Building processor image 2021-09-19 17:59:03 (info) Build complete > 2021-09-19 17:59:12,356 [info] successfully deployed function: {'internal_invocation_urls': ['nuclio-fraud-demo-admin-ingest-events.default-tenant.svc.cluster.local:8080'], 'external_invocation_urls': ['fraud-demo-admin-ingest-events-fraud-demo-admin.default-tenant.app.jnewriujxdig.iguazio-cd1.com/']} ###Markdown User Events - Test the feature set HTTP endpoint ###Code # Select a sample from the events dataset and serialize it to JSON user_events_sample = json.loads(user_events_data.sample(1).to_json(orient='records'))[0] user_events_sample['timestamp'] = str(pd.Timestamp.now()) user_events_sample # Post the sample to the ingestion endpoint requests.post(events_set_endpoint, json=user_events_sample).text ###Output _____no_output_____ ###Markdown Part 1: Data Ingestion This demo showcases financial fraud prevention and using the MLRun feature store to define complex features that help identify fraud. Fraud prevention specifically is a challenge as it requires processing raw transaction and events in real-time and being able to quickly respond and block transactions before they occur.To address this, we create a development pipeline and a production pipeline. Both pipelines share the same feature engineering and model code, but serve data very differently. Furthermore, we automate the data and model monitoring process, identify drift and trigger retraining in a CI/CD pipeline. This process is described in the diagram below:![Feature store demo diagram - fraud prevention](../../_static/images/feature_store_demo_diagram.png) The raw data is described as follows:| TRANSACTIONS || &x2551; |USER EVENTS || |-----------------|----------------------------------------------------------------|----------|-----------------|----------------------------------------------------------------|| **age** | age group value 0-6. Some values are marked as U for unknown | &x2551; | **source** | The party/entity related to the event || **gender** | A character to define the age | &x2551; | **event** | event, such as login or password change || **zipcodeOri** | ZIP code of the person originating the transaction | &x2551; | **timestamp** | The date and time of the event || **zipMerchant** | ZIP code of the merchant receiving the transaction | &x2551; | | || **category** | category of the transaction (e.g., transportation, food, etc.) | &x2551; | | || **amount** | the total amount of the transaction | &x2551; | | || **fraud** | whether the transaction is fraudulent | &x2551; | | || **timestamp** | the date and time in which the transaction took place | &x2551; | | || **source** | the ID of the party/entity performing the transaction | &x2551; | | || **target** | the ID of the party/entity receiving the transaction | &x2551; | | || **device** | the device ID used to perform the transaction | &x2551; | | | This notebook introduces how to **Ingest** different data sources to the **Feature Store**.The following FeatureSets will be created:- **Transactions**: Monetary transactions between a source and a target.- **Events**: Account events such as account login or a password change.- **Label**: Fraud label for the data.By the end of this tutorial you’ll learn how to:- Create an ingestion pipeline for each data source.- Define preprocessing, aggregation and validation of the pipeline.- Run the pipeline locally within the notebook.- Launch a real-time function to ingest live data.- Schedule a cron to run the task when needed. ###Code project_name = 'fraud-demo' import mlrun # Initialize the MLRun project object project = mlrun.get_or_create_project(project_name, context="./", user_project=True) ###Output > 2021-09-19 17:55:08,313 [info] created and saved project fraud-demo ###Markdown Step 1 - Fetch, Process and Ingest our datasets 1.1 - Transactions Transactions ###Code # Helper functions to adjust the timestamps of our data # while keeping the order of the selected events and # the relative distance from one event to the other def date_adjustment(sample, data_max, new_max, old_data_period, new_data_period): ''' Adjust a specific sample's date according to the original and new time periods ''' sample_dates_scale = ((data_max - sample) / old_data_period) sample_delta = new_data_period * sample_dates_scale new_sample_ts = new_max - sample_delta return new_sample_ts def adjust_data_timespan(dataframe, timestamp_col='timestamp', new_period='2d', new_max_date_str='now'): ''' Adjust the dataframe timestamps to the new time period ''' # Calculate old time period data_min = dataframe.timestamp.min() data_max = dataframe.timestamp.max() old_data_period = data_max-data_min # Set new time period new_time_period = pd.Timedelta(new_period) new_max = pd.Timestamp(new_max_date_str) new_min = new_max-new_time_period new_data_period = new_max-new_min # Apply the timestamp change df = dataframe.copy() df[timestamp_col] = df[timestamp_col].apply(lambda x: date_adjustment(x, data_max, new_max, old_data_period, new_data_period)) return df import pandas as pd # Fetch the transactions dataset from the server transactions_data = pd.read_csv('https://s3.wasabisys.com/iguazio/data/fraud-demo-mlrun-fs-docs/data.csv', parse_dates=['timestamp']) # Adjust the samples timestamp for the past 2 days transactions_data = adjust_data_timespan(transactions_data.sample(50000), new_period='2d') # Preview transactions_data.head(3) ###Output _____no_output_____ ###Markdown Transactions - Create a FeatureSet and Preprocessing PipelineCreate the FeatureSet (data pipeline) definition for the **credit transaction processing** which describes the offline/online data transformations and aggregations.The feature store will automatically add an offline `parquet` target and an online `NoSQL` target by using `set_targets()`.The data pipeline consists of:* **Extracting** the data components (hour, day of week)* **Mapping** the age values* **One hot encoding** for the transaction category and the gender* **Aggregating** the amount (avg, sum, count, max over 2/12/24 hour time windows)* **Aggregating** the transactions per category (over 14 days time windows)* **Writing** the results to **offline** (Parquet) and **online** (NoSQL) targets ###Code # Import MLRun's Feature Store import mlrun.feature_store as fstore from mlrun.feature_store.steps import OneHotEncoder, MapValues, DateExtractor # Define the transactions FeatureSet transaction_set = fstore.FeatureSet("transactions", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="transactions feature set") # Define and add value mapping main_categories = ["es_transportation", "es_health", "es_otherservices", "es_food", "es_hotelservices", "es_barsandrestaurants", "es_tech", "es_sportsandtoys", "es_wellnessandbeauty", "es_hyper", "es_fashion", "es_home", "es_contents", "es_travel", "es_leisure"] # One Hot Encode the newly defined mappings one_hot_encoder_mapping = {'category': main_categories, 'gender': list(transactions_data.gender.unique())} # Define the graph steps transaction_set.graph\ .to(DateExtractor(parts = ['hour', 'day_of_week'], timestamp_col = 'timestamp'))\ .to(MapValues(mapping={'age': {'U': '0'}}, with_original_features=True))\ .to(OneHotEncoder(mapping=one_hot_encoder_mapping)) # Add aggregations for 2, 12, and 24 hour time windows transaction_set.add_aggregation(name='amount', column='amount', operations=['avg','sum', 'count','max'], windows=['2h', '12h', '24h'], period='1h') # Add the category aggregations over a 14 day window for category in main_categories: transaction_set.add_aggregation(name=category,column=f'category_{category}', operations=['count'], windows=['14d'], period='1d') # Add default (offline-parquet & online-nosql) targets transaction_set.set_targets() # Plot the pipeline so we can see the different steps transaction_set.plot(rankdir="LR", with_targets=True) ###Output _____no_output_____ ###Markdown Transactions - Ingestion ###Code # Ingest our transactions dataset through our defined pipeline transactions_df = fstore.ingest(transaction_set, transactions_data, infer_options=fstore.InferOptions.default()) transactions_df.head(3) ###Output _____no_output_____ ###Markdown 1.2 - User Events User Events - Fetching ###Code # Fetch our user_events dataset from the server user_events_data = pd.read_csv('https://s3.wasabisys.com/iguazio/data/fraud-demo-mlrun-fs-docs/events.csv', index_col=0, quotechar="\'", parse_dates=['timestamp']) # Adjust to the last 2 days to see the latest aggregations in our online feature vectors user_events_data = adjust_data_timespan(user_events_data, new_period='2d') # Preview user_events_data.head(3) ###Output _____no_output_____ ###Markdown User Events - Create a FeatureSet and Preprocessing PipelineNow we will define the events feature set.This is a pretty straight forward pipeline in which we only one hot encode the event categories and save the data to the default targets. ###Code user_events_set = fstore.FeatureSet("events", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="user events feature set") # Define and add value mapping events_mapping = {'event': list(user_events_data.event.unique())} # One Hot Encode user_events_set.graph.to(OneHotEncoder(mapping=events_mapping)) # Add default (offline-parquet & online-nosql) targets user_events_set.set_targets() # Plot the pipeline so we can see the different steps user_events_set.plot(rankdir="LR", with_targets=True) ###Output _____no_output_____ ###Markdown User Events - Ingestion ###Code # Ingestion of our newly created events feature set events_df = fstore.ingest(user_events_set, user_events_data) events_df.head(3) ###Output _____no_output_____ ###Markdown Step 2 - Create a labels dataset for model training Label Set - Create a FeatureSetThis feature set contains the label for the fraud demo, it will be ingested directly to the default targets without any changes ###Code def create_labels(df): labels = df[['fraud','source','timestamp']].copy() labels = labels.rename(columns={"fraud": "label"}) labels['timestamp'] = labels['timestamp'].astype("datetime64[ms]") labels['label'] = labels['label'].astype(int) labels.set_index('source', inplace=True) return labels # Define the "labels" feature set labels_set = fstore.FeatureSet("labels", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="training labels", engine="pandas") labels_set.graph.to(name="create_labels", handler=create_labels) # specify only Parquet (offline) target since its not used for real-time labels_set.set_targets(['parquet'], with_defaults=False) labels_set.plot(with_targets=True) ###Output _____no_output_____ ###Markdown Label Set - Ingestion ###Code # Ingest the labels feature set labels_df = fstore.ingest(labels_set, transactions_data) labels_df.head(3) ###Output _____no_output_____ ###Markdown Step 3 - Deploy a real-time pipelineWhen dealing with real-time aggregation, it's important to be able to update these aggregations in real-time.For this purpose, we will create live serving functions that will update the online feature store of the `transactions` FeatureSet and `Events` FeatureSet.Using MLRun's `serving` runtime, craetes a nuclio function loaded with our feature set's computational graph definitionand an `HttpSource` to define the HTTP trigger.Notice that the implementation below does not require any rewrite of the pipeline logic. 3.1 - Transactions Transactions - Deploy our FeatureSet live endpoint ###Code # Create iguazio v3io stream and transactions push API endpoint transaction_stream = f'v3io:///projects/{project.name}/streams/transaction' transaction_pusher = mlrun.datastore.get_stream_pusher(transaction_stream) # Define the HTTP Source to_dictable the HTTP trigger on our function and expose the endpoint. # as any other datasource, we will define the `key` and `time` fields here too. http_source = mlrun.datastore.sources.HttpSource(key_field='source', time_field='timestamp') transaction_set.spec.source = http_source # Create a real-time serverless function definition to deploy the ingestion pipeline on. # the serving runtimes enables the deployment of our feature set's computational graph function = (mlrun.new_function('ingest-transactions', kind='serving', image='mlrun/mlrun')).with_code(body=" ") # Add stream trigger (must first create the stream) function.add_v3io_stream_trigger(transaction_stream) # run_config = fstore.RunConfig(function=function, local=False).apply(mlrun.mount_v3io()) # Deploy the transactions feature set's ingestion service using the feature set # and all the defined resources above. transaction_set_endpoint = fstore.deploy_ingestion_service(featureset=transaction_set, run_config=run_config) ###Output > 2021-09-19 17:58:50,402 [info] Starting remote function deploy 2021-09-19 17:58:50 (info) Deploying function 2021-09-19 17:58:50 (info) Building 2021-09-19 17:58:50 (info) Staging files and preparing base images 2021-09-19 17:58:50 (info) Building processor image 2021-09-19 17:58:52 (info) Build complete 2021-09-19 17:59:01 (info) Function deploy complete > 2021-09-19 17:59:01,461 [info] successfully deployed function: {'internal_invocation_urls': ['nuclio-fraud-demo-admin-ingest-transactions.default-tenant.svc.cluster.local:8080'], 'external_invocation_urls': ['fraud-demo-admin-ingest-transactions-fraud-demo-admin.default-tenant.app.jnewriujxdig.iguazio-cd1.com/']} ###Markdown Transactions - Test the feature set HTTP endpoint By defining our `transactions` feature set we can now use MLRun and Storey to deploy it as a live endpoint, ready to ingest new data!Using MLRun's `serving` runtime, we will create a nuclio function loaded with our feature set's computational graph definition and an `HttpSource` to define the HTTP trigger. ###Code import requests import json # Select a sample from the dataset and serialize it to JSON transaction_sample = json.loads(transactions_data.sample(1).to_json(orient='records'))[0] transaction_sample['timestamp'] = str(pd.Timestamp.now()) transaction_sample # Post the sample to the ingestion endpoint requests.post(transaction_set_endpoint, json=transaction_sample).text ###Output _____no_output_____ ###Markdown 3.2 - User Events User Events - Deploy our FeatureSet live endpointDeploy the events feature set's ingestion service using the feature set and all the previously defined resources. ###Code # Create iguazio v3io stream and transactions push API endpoint events_stream = f'v3io:///projects/{project.name}/streams/events' events_pusher = mlrun.datastore.get_stream_pusher(events_stream) # Create a `serving` "base function" to deploy the ingestion function on # the serving runtimes enables the deployment of our feature set's computational graph function = (mlrun.new_function('ingest-events', kind='serving', image='mlrun/mlrun')).with_code(body=" ") # Add stream trigger function.add_v3io_stream_trigger(events_stream) # run_config = fstore.RunConfig(function=function, local=False).apply(mlrun.mount_v3io()) # Deploy the transactions feature set's ingestion service using the feature set # and all the defined resources above. events_set_endpoint = fstore.deploy_ingestion_service(name="ingest-events", featureset=user_events_set, source=http_source, run_config=run_config) ###Output > 2021-09-19 17:59:01,795 [info] Starting remote function deploy 2021-09-19 17:59:02 (info) Deploying function 2021-09-19 17:59:02 (info) Building 2021-09-19 17:59:02 (info) Staging files and preparing base images 2021-09-19 17:59:02 (info) Building processor image 2021-09-19 17:59:03 (info) Build complete > 2021-09-19 17:59:12,356 [info] successfully deployed function: {'internal_invocation_urls': ['nuclio-fraud-demo-admin-ingest-events.default-tenant.svc.cluster.local:8080'], 'external_invocation_urls': ['fraud-demo-admin-ingest-events-fraud-demo-admin.default-tenant.app.jnewriujxdig.iguazio-cd1.com/']} ###Markdown User Events - Test the feature set HTTP endpoint ###Code # Select a sample from the events dataset and serialize it to JSON user_events_sample = json.loads(user_events_data.sample(1).to_json(orient='records'))[0] user_events_sample['timestamp'] = str(pd.Timestamp.now()) user_events_sample # Post the sample to the ingestion endpoint requests.post(events_set_endpoint, json=user_events_sample).text ###Output _____no_output_____ ###Markdown Part 1: Data Ingestion This demo showcases financial fraud prevention and using the MLRun feature store to define complex features that help identify fraud. Fraud prevention specifically is a challenge as it requires processing raw transaction and events in real-time and being able to quickly respond and block transactions before they occur.To address this, we create a development pipeline and a production pipeline. Both pipelines share the same feature engineering and model code, but serve data very differently. Furthermore, we automate the data and model monitoring process, identify drift and trigger retraining in a CI/CD pipeline. This process is described in the diagram below:![Feature store demo diagram - fraud prevention](../../_static/images/feature_store_demo_diagram.png) The raw data is described as follows:| TRANSACTIONS || &x2551; |USER EVENTS || |-----------------|----------------------------------------------------------------|----------|-----------------|----------------------------------------------------------------|| **age** | age group value 0-6. Some values are marked as U for unknown | &x2551; | **source** | The party/entity related to the event || **gender** | A character to define the age | &x2551; | **event** | event, such as login or password change || **zipcodeOri** | ZIP code of the person originating the transaction | &x2551; | **timestamp** | The date and time of the event || **zipMerchant** | ZIP code of the merchant receiving the transaction | &x2551; | | || **category** | category of the transaction (e.g., transportation, food, etc.) | &x2551; | | || **amount** | the total amount of the transaction | &x2551; | | || **fraud** | whether the transaction is fraudulent | &x2551; | | || **timestamp** | the date and time in which the transaction took place | &x2551; | | || **source** | the ID of the party/entity performing the transaction | &x2551; | | || **target** | the ID of the party/entity receiving the transaction | &x2551; | | || **device** | the device ID used to perform the transaction | &x2551; | | | This notebook introduces how to **Ingest** different data sources to the **Feature Store**.The following FeatureSets will be created:- **Transactions**: Monetary transactions between a source and a target.- **Events**: Account events such as account login or a password change.- **Label**: Fraud label for the data.By the end of this tutorial you’ll learn how to:- Create an ingestion pipeline for each data source.- Define preprocessing, aggregation and validation of the pipeline.- Run the pipeline locally within the notebook.- Launch a real-time function to ingest live data.- Schedule a cron to run the task when needed. ###Code project_name = 'fraud-demo' import mlrun # Initialize the MLRun project object project = mlrun.get_or_create_project(project_name, context="./", user_project=True) ###Output > 2022-03-16 05:45:07,703 [info] loaded project fraud-demo from MLRun DB ###Markdown Step 1 - Fetch, Process and Ingest our datasets 1.1 - Transactions Transactions ###Code # Helper functions to adjust the timestamps of our data # while keeping the order of the selected events and # the relative distance from one event to the other def date_adjustment(sample, data_max, new_max, old_data_period, new_data_period): ''' Adjust a specific sample's date according to the original and new time periods ''' sample_dates_scale = ((data_max - sample) / old_data_period) sample_delta = new_data_period * sample_dates_scale new_sample_ts = new_max - sample_delta return new_sample_ts def adjust_data_timespan(dataframe, timestamp_col='timestamp', new_period='2d', new_max_date_str='now'): ''' Adjust the dataframe timestamps to the new time period ''' # Calculate old time period data_min = dataframe.timestamp.min() data_max = dataframe.timestamp.max() old_data_period = data_max-data_min # Set new time period new_time_period = pd.Timedelta(new_period) new_max = pd.Timestamp(new_max_date_str) new_min = new_max-new_time_period new_data_period = new_max-new_min # Apply the timestamp change df = dataframe.copy() df[timestamp_col] = df[timestamp_col].apply(lambda x: date_adjustment(x, data_max, new_max, old_data_period, new_data_period)) return df import pandas as pd # Fetch the transactions dataset from the server transactions_data = pd.read_csv('https://s3.wasabisys.com/iguazio/data/fraud-demo-mlrun-fs-docs/data.csv', parse_dates=['timestamp'], nrows=500) # Adjust the samples timestamp for the past 2 days transactions_data = adjust_data_timespan(transactions_data, new_period='2d') # Preview transactions_data.head(3) ###Output _____no_output_____ ###Markdown Transactions - Create a FeatureSet and Preprocessing PipelineCreate the FeatureSet (data pipeline) definition for the **credit transaction processing** which describes the offline/online data transformations and aggregations.The feature store will automatically add an offline `parquet` target and an online `NoSQL` target by using `set_targets()`.The data pipeline consists of:* **Extracting** the data components (hour, day of week)* **Mapping** the age values* **One hot encoding** for the transaction category and the gender* **Aggregating** the amount (avg, sum, count, max over 2/12/24 hour time windows)* **Aggregating** the transactions per category (over 14 days time windows)* **Writing** the results to **offline** (Parquet) and **online** (NoSQL) targets ###Code # Import MLRun's Feature Store import mlrun.feature_store as fstore from mlrun.feature_store.steps import OneHotEncoder, MapValues, DateExtractor # Define the transactions FeatureSet transaction_set = fstore.FeatureSet("transactions", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="transactions feature set") # Define and add value mapping main_categories = ["es_transportation", "es_health", "es_otherservices", "es_food", "es_hotelservices", "es_barsandrestaurants", "es_tech", "es_sportsandtoys", "es_wellnessandbeauty", "es_hyper", "es_fashion", "es_home", "es_contents", "es_travel", "es_leisure"] # One Hot Encode the newly defined mappings one_hot_encoder_mapping = {'category': main_categories, 'gender': list(transactions_data.gender.unique())} # Define the graph steps transaction_set.graph\ .to(DateExtractor(parts = ['hour', 'day_of_week'], timestamp_col = 'timestamp'))\ .to(MapValues(mapping={'age': {'U': '0'}}, with_original_features=True))\ .to(OneHotEncoder(mapping=one_hot_encoder_mapping)) # Add aggregations for 2, 12, and 24 hour time windows transaction_set.add_aggregation(name='amount', column='amount', operations=['avg','sum', 'count','max'], windows=['2h', '12h', '24h'], period='1h') # Add the category aggregations over a 14 day window for category in main_categories: transaction_set.add_aggregation(name=category,column=f'category_{category}', operations=['count'], windows=['14d'], period='1d') # Add default (offline-parquet & online-nosql) targets transaction_set.set_targets() # Plot the pipeline so we can see the different steps transaction_set.plot(rankdir="LR", with_targets=True) ###Output _____no_output_____ ###Markdown Transactions - Ingestion ###Code # Ingest our transactions dataset through our defined pipeline transactions_df = fstore.ingest(transaction_set, transactions_data, infer_options=fstore.InferOptions.default()) transactions_df.head(3) ###Output persist count = 0 persist count = 100 persist count = 200 persist count = 300 persist count = 400 persist count = 500 persist count = 600 persist count = 700 persist count = 800 persist count = 900 persist count = 1000 ###Markdown 1.2 - User Events User Events - Fetching ###Code # Fetch our user_events dataset from the server user_events_data = pd.read_csv('https://s3.wasabisys.com/iguazio/data/fraud-demo-mlrun-fs-docs/events.csv', index_col=0, quotechar="\'", parse_dates=['timestamp'], nrows=500) # Adjust to the last 2 days to see the latest aggregations in our online feature vectors user_events_data = adjust_data_timespan(user_events_data, new_period='2d') # Preview user_events_data.head(3) ###Output _____no_output_____ ###Markdown User Events - Create a FeatureSet and Preprocessing PipelineNow we will define the events feature set.This is a pretty straight forward pipeline in which we only one hot encode the event categories and save the data to the default targets. ###Code user_events_set = fstore.FeatureSet("events", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="user events feature set") # Define and add value mapping events_mapping = {'event': list(user_events_data.event.unique())} # One Hot Encode user_events_set.graph.to(OneHotEncoder(mapping=events_mapping)) # Add default (offline-parquet & online-nosql) targets user_events_set.set_targets() # Plot the pipeline so we can see the different steps user_events_set.plot(rankdir="LR", with_targets=True) ###Output _____no_output_____ ###Markdown User Events - Ingestion ###Code # Ingestion of our newly created events feature set events_df = fstore.ingest(user_events_set, user_events_data) events_df.head(3) ###Output persist count = 0 persist count = 100 persist count = 200 persist count = 300 persist count = 400 persist count = 500 ###Markdown Step 2 - Create a labels dataset for model training Label Set - Create a FeatureSetThis feature set contains the label for the fraud demo, it will be ingested directly to the default targets without any changes ###Code def create_labels(df): labels = df[['fraud','source','timestamp']].copy() labels = labels.rename(columns={"fraud": "label"}) labels['timestamp'] = labels['timestamp'].astype("datetime64[ms]") labels['label'] = labels['label'].astype(int) labels.set_index('source', inplace=True) return labels # Define the "labels" feature set labels_set = fstore.FeatureSet("labels", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="training labels", engine="pandas") labels_set.graph.to(name="create_labels", handler=create_labels) # specify only Parquet (offline) target since its not used for real-time labels_set.set_targets(['parquet'], with_defaults=False) labels_set.plot(with_targets=True) ###Output _____no_output_____ ###Markdown Label Set - Ingestion ###Code # Ingest the labels feature set labels_df = fstore.ingest(labels_set, transactions_data) labels_df.head(3) ###Output _____no_output_____ ###Markdown Step 3 - Deploy a real-time pipelineWhen dealing with real-time aggregation, it's important to be able to update these aggregations in real-time.For this purpose, we will create live serving functions that will update the online feature store of the `transactions` FeatureSet and `Events` FeatureSet.Using MLRun's `serving` runtime, craetes a nuclio function loaded with our feature set's computational graph definitionand an `HttpSource` to define the HTTP trigger.Notice that the implementation below does not require any rewrite of the pipeline logic. 3.1 - Transactions Transactions - Deploy our FeatureSet live endpoint ###Code # Create iguazio v3io stream and transactions push API endpoint transaction_stream = f'v3io:///projects/{project.name}/streams/transaction' transaction_pusher = mlrun.datastore.get_stream_pusher(transaction_stream) # Define the source stream trigger (use v3io streams) # we will define the `key` and `time` fields (extracted from the Json message). source = mlrun.datastore.sources.StreamSource(path=transaction_stream , key_field='source', time_field='timestamp') # Deploy the transactions feature set's ingestion service over a real-time (Nuclio) serverless function # you can use the run_config parameter to pass function/service specific configuration transaction_set_endpoint = fstore.deploy_ingestion_service(featureset=transaction_set, source=source) ###Output > 2022-03-16 05:45:43,035 [info] Starting remote function deploy 2022-03-16 05:45:43 (info) Deploying function 2022-03-16 05:45:43 (info) Building 2022-03-16 05:45:43 (info) Staging files and preparing base images 2022-03-16 05:45:43 (warn) Python 3.6 runtime is deprecated and will soon not be supported. Please migrate your code and use Python 3.7 runtime (`python:3.7`) or higher 2022-03-16 05:45:43 (info) Building processor image 2022-03-16 05:47:03 (info) Build complete 2022-03-16 05:47:08 (info) Function deploy complete > 2022-03-16 05:47:08,835 [info] successfully deployed function: {'internal_invocation_urls': ['nuclio-fraud-demo-admin-transactions-ingest.default-tenant.svc.cluster.local:8080'], 'external_invocation_urls': ['fraud-demo-admin-transactions-ingest-fraud-demo-admin.default-tenant.app.xtvtjecfcssi.iguazio-cd1.com/']} ###Markdown Transactions - Test the feature set HTTP endpoint By defining our `transactions` feature set we can now use MLRun and Storey to deploy it as a live endpoint, ready to ingest new data!Using MLRun's `serving` runtime, we will create a nuclio function loaded with our feature set's computational graph definition and an `HttpSource` to define the HTTP trigger. ###Code import requests import json # Select a sample from the dataset and serialize it to JSON transaction_sample = json.loads(transactions_data.sample(1).to_json(orient='records'))[0] transaction_sample['timestamp'] = str(pd.Timestamp.now()) transaction_sample # Post the sample to the ingestion endpoint requests.post(transaction_set_endpoint, json=transaction_sample).text ###Output _____no_output_____ ###Markdown 3.2 - User Events User Events - Deploy our FeatureSet live endpointDeploy the events feature set's ingestion service using the feature set and all the previously defined resources. ###Code # Create iguazio v3io stream and transactions push API endpoint events_stream = f'v3io:///projects/{project.name}/streams/events' events_pusher = mlrun.datastore.get_stream_pusher(events_stream) # Define the source stream trigger (use v3io streams) # we will define the `key` and `time` fields (extracted from the Json message). source = mlrun.datastore.sources.StreamSource(path=events_stream , key_field='source', time_field='timestamp') # Deploy the transactions feature set's ingestion service over a real-time (Nuclio) serverless function # you can use the run_config parameter to pass function/service specific configuration events_set_endpoint = fstore.deploy_ingestion_service(featureset=user_events_set, source=source) ###Output > 2022-03-16 05:47:09,035 [info] Starting remote function deploy 2022-03-16 05:47:09 (info) Deploying function 2022-03-16 05:47:09 (info) Building 2022-03-16 05:47:09 (info) Staging files and preparing base images 2022-03-16 05:47:09 (warn) Python 3.6 runtime is deprecated and will soon not be supported. Please migrate your code and use Python 3.7 runtime (`python:3.7`) or higher 2022-03-16 05:47:09 (info) Building processor image ###Markdown User Events - Test the feature set HTTP endpoint ###Code # Select a sample from the events dataset and serialize it to JSON user_events_sample = json.loads(user_events_data.sample(1).to_json(orient='records'))[0] user_events_sample['timestamp'] = str(pd.Timestamp.now()) user_events_sample # Post the sample to the ingestion endpoint requests.post(events_set_endpoint, json=user_events_sample).text ###Output _____no_output_____ ###Markdown Part 1: Data Ingestion In this notebook we will learn how to **Ingest** different data sources to our **Feature Store**. Specifically, this patient data has been successfully used to treat hospitalized COVID-19 patients prior to their condition becoming severe or critical. To do this we will use a medical dataset which includes three types of data: - **Healthcare systems**: Batch updated dataset, containing different lab test results (Blood test results for ex.).- **Patient Records**: Static dataset containing general patient details.- **Real-time sensors**: Real-Time patient metric monitoring sensor. We will walk through creation of ingestion pipeline for each datasource with all the needed preprocessing and validation. We will run the pipeline locally within the notebook and then launch a real-time function to **ingest live data** or schedule a cron to run the task when needed. Environment SetupSince our work is done in a this project scope, first define the project itself for all our MLRun work in this notebook. ###Code import mlrun from os import getenv mlrun.set_environment(project='fsdemo', user_project=True) # location of the output data files data_path = f"{getenv('V3IO_HOME_URL')}/demos/feature-store/data/" def move_timestamps(df, shift='0s'): ''' Update timetsamps to current time so we can see live aggregations ''' now = pd.to_datetime('now') max_time = df['timestamp'].max() time_shift = now-max_time tmp_df = df.copy() tmp_df['timestamp'] = tmp_df['timestamp'].apply(lambda t: t + time_shift + pd.to_timedelta(shift)) return tmp_df ###Output _____no_output_____ ###Markdown Create Ingestion Pipeline With MLRunIn this section we will ingest the lab measurements data using MLRun and Storey. Storey is the underlying implementation of the feature store which is used by MLRun. It is the engine that allows you to define and execute complex graphs that create the feature engineering pipeline. With storey, you can define source, transformations and targets, many actions are available as part of the Storey library, but you can define additional actions easily. We will see these custom actions in later sections.For the execution, it is possible to also use Spark. The main difference between Storey and Spark pipelines is that Storey blocks are built for Real-Time workloads while Spark is more Batch oriented. We will now do the following:- Create the `measurements` FeatureSet- Define Preprocessing graph including aggregations- Ingest the data using the defined pipeline ###Code # Import MLRun's Feature Store import mlrun.feature_store as fs # Import MLRun's Data Sources to set the wanted ingestion pipeline from mlrun.datastore.sources import CSVSource, ParquetSource, HttpSource # Import storey so it will be available on our scope # when testing the pipeline import storey # Define the Lab Measurements FeatureSet measurements_set = fs.FeatureSet("measurements", entities=[fs.Entity("patient_id")], timestamp_key='timestamp', description="various patient health measurements") # Get FeatureSet computation graph measurements_graph = measurements_set.graph ###Output _____no_output_____ ###Markdown Define the processing pipeline- Transformation function- Sliding window aggregation- Set targets (NoSQL and Parquet) ###Code # Import pandas and load the sample CSV and load it as a datasource # for our ingestion import pandas as pd measurements_df = pd.read_csv('https://s3.wasabisys.com/iguazio/data/patients/measurements.csv', index_col=0) measurements_df['timestamp'] = pd.to_datetime(measurements_df['timestamp']) measurements_df['timestamp'] = measurements_df['timestamp'].astype("datetime64[ms]") measurements_df = pd.concat([move_timestamps(measurements_df, '-1h'), move_timestamps(measurements_df)]) # update timestamps ###Output _____no_output_____ ###Markdown Take a look at the measurements dataset. This dataset includes a a single measurement per row. The measurement type is defined by the `source` and `parameter` column. We would like to transform this data, so each patient has multiple measurement columns. To do that, we will need to create a new column for each `source` and `parameter` combination. For example, if `source` is 3 and `parameter` is 0, then our transformed dataset will have the measurement value in a new feature named `sp_3_0`.Following that, we will create a sliding window aggregation that averages the values across that time window. ###Code measurements_df.head() ###Output _____no_output_____ ###Markdown The following code performs the transformation, adds the aggregation and sets the target to store the values to a NoSQL database for online retrieval and parquet files for batch processing. ###Code # Define transform to create sparse dataset for aggregation # adding an extra column for the specific source-parameter pair's measurement # ex: source=3, parameter=4, measurement=100 -> add extra column sp_3_4=100 def transform(event): event["_".join(['sp', str(event["source"]), str(event["parameter"])])] = event["measurement"] return event # Define Measurement FeatureSet pipeline measurements_graph.to( "storey.Map", _fn="transform" ) # Get the available source, parameter pairs for our aggregation sps = list(measurements_df.apply(lambda x: '_'.join(['sp', str(x['source']), str(x['parameter'])]), axis=1).unique()) # Add aggregations on top of the created sparse # features by the transorm function for col in sps: measurements_set.add_aggregation(name=f'agg_{col}', column=col, operations=['avg'], window='1h', period='30m') # Add default (NoSQL via KV and Parquet) targets to save # the ingestion results to measurements_set.set_targets() ###Output _____no_output_____ ###Markdown You can plot the graph to visalize the pipeline: ###Code # Plot the ingestion pipeline we defined measurements_set.plot(rankdir='LR', with_targets=True) ###Output _____no_output_____ ###Markdown Run ingestion task using MLRun & StoreyIn order to ingest the dataframe to the featureset, use the `ingest` function. ###Code # User our loaded DF as the datasource and ingest it through # the define pipeline resp = fs.ingest(measurements_set, measurements_df, infer_options=fs.InferOptions.default()) resp.head() # Save the FeatureSet and pipeline definition measurements_set.save() ###Output _____no_output_____ ###Markdown Ingest Patient Details Features In this section we will use MLRun to create our patient details datasource. We will do the following:- Create a `patient_details` FeatureSet- Add preprocessing transformations to the pipeline - Map ages to buckets and One Hot Encode them - Impute missing values- Test the processing pipeline with sample data- Run ingestion pipeline on top of the cluster Create the FeatureSet ###Code # add feature set without time column (stock ticker metadata) patients_set = fs.FeatureSet("patient_details", entities=[fs.Entity("patient_id")], description="personal and medical patient details") # Get FeatureSet computation graph graph = patients_set.spec.graph ###Output _____no_output_____ ###Markdown Define the computation pipeline ###Code # Define age buckets for our age value mapping personal_details = {'age': {'ranges': [{'range': [0, 3], "value": "toddler"}, {'range': [3, 18], "value": "child"}, {'range': [18, 65], "value": "adult"}, {'range': [65, 120], "value": "elder"}]}} # Define one hot encoding values map one_hot_encoder_mapping = {'age_mapped': ['toddler', 'child', 'adult', 'elder']} # Import MLRun's FeatureStore steps for easy # use in our pipeline from mlrun.feature_store.steps import * # Define the pipeline for our FeatureSet graph.to(MapValues(mapping=personal_details, with_original_features=True))\ .to(OneHotEncoder(mapping=one_hot_encoder_mapping))\ .to(Imputer(method='values', default_value=1, mapping={})) # Add default NoSQL & Parquet ingestion targets patients_set.set_targets() # Plot the FeatureSet pipeline patients_set.plot(rankdir='LR', with_targets=True) ###Output _____no_output_____ ###Markdown Test the Feature transformation pipelineCreating a transformation pipeline requires some trial and error. Therefore, it is useful to run the pipeline in memory without storing the resultant data. For this purpose, `infer` is used. This function receives as input any sample DataFrame, performs all the graph steps and outputs the transformed DataFrame. ###Code # Load the sample patient details data patients_df = pd.read_parquet('https://s3.wasabisys.com/iguazio/data/patients/patient_details.parquet') # Run local ingestion test fs.infer(patients_set, patients_df.head()) ###Output _____no_output_____ ###Markdown Save the FeatureSet and run full ingestion taskOnce you are satisfied with the transformation pipeline, ingest that full DataFrame and store the data. ###Code # Save the FeatureSet patients_set.save() # Run Ingestion task resp = fs.ingest(patients_set, patients_df, infer_options=fs.InferOptions.default()) ###Output _____no_output_____ ###Markdown Start Immediate or Scheduled Ingestion Job (over Kubernetes)Another useful method to ingest data, is by creating a Kubernetes job. This may be necessary to process large amounts of data as well as to process any recurring data. With MLRun it is easy to take the pipeline and run it as a job. This is done by:1. Define a source, specifically here we define a parquet file source2. Define a configuration where `local` is set to `False`3. Mount to the provisioned storage by calling `auto_mount`4. Run `ingest` with the source and run configuration ###Code source = ParquetSource('pq', 'https://s3.wasabisys.com/iguazio/data/patients/patient_details.parquet') config = fs.RunConfig(local=False).apply(mlrun.platforms.auto_mount()) fs.ingest(patients_set, source, run_config=config) ###Output > 2021-05-06 15:27:08,769 [info] starting run patient_details_ingest uid=76f197b8ab3347d1b995a5ea55d0a98a DB=http://mlrun-api:8080 > 2021-05-06 15:27:09,022 [info] Job is running in the background, pod: patient-details-ingest-g9hgn > 2021-05-06 15:27:15,073 [info] starting ingestion task to store://feature-sets/fsdemo-admin/patient_details:latest > 2021-05-06 15:27:15,745 [info] ingestion task completed, targets: > 2021-05-06 15:27:15,746 [info] [{'name': 'parquet', 'kind': 'parquet', 'path': 'v3io:///projects/fsdemo-admin/fs/parquet/sets/patient_details-latest.parquet', 'status': 'created', 'updated': '2021-05-06T15:27:15.432576+00:00'}, {'name': 'nosql', 'kind': 'nosql', 'path': 'v3io:///projects/fsdemo-admin/fs/nosql/sets/patient_details-latest', 'status': 'created', 'updated': '2021-05-06T15:27:15.432947+00:00'}] > 2021-05-06 15:27:15,936 [info] run executed, status=completed final state: completed ###Markdown Real-time Early-Sense Sensor Ingestion (HTTP or Stream Processing With Nuclio) In this section we will use MLRun to create our Early Sense Sensor datasource. We will do the following:- Create early sense FeatureSet- Add Preprocessing transformations to the Pipeline using custom functions - Drop and Rename columns - Aggregations- Add Feature Validator to detect bad sensor readings- Test the processing pipeline with sample data- Deploy the FeatureSet ingestion service as a live rest endpoint ###Code early_sense_set = fs.FeatureSet("early_sense", entities=[fs.Entity("patient_id")], timestamp_key='timestamp', description="real time patient bed sensor data") ###Output _____no_output_____ ###Markdown Define data validation & quality policyWe can define validations on the feature level. For example, define here validation to check if the heart-rate value is between 0 and 220 and respitory rate is between 0 and 25. ###Code from mlrun.features import MinMaxValidator early_sense_set["hr"] = fs.Feature(validator = MinMaxValidator(min=0, max=220, severity="info")) early_sense_set["rr"] = fs.Feature(validator = MinMaxValidator(min=0, max=25, severity="info")) ###Output _____no_output_____ ###Markdown Define custom processing classesIn the previous sections we used transformation steps that are available as part of Storey. Here we show how to create custom transformation classes. We will later run these functions as part of a Nuclio serverless real-time function, therefore, we also use the nuclio `start-code` and `end-code` comments. ###Code # nuclio: start-code # We will import storey here too so it will # be included in our function code (within the nuclio comment block) import json import storey from typing import List, Dict # The custom functions are based on `storey.MapClass` # when they are called in the graph the `do(self, event)` # function will be activated. # A to_dict(self) function is also required by MLRun # to allow the class creation on remote functions class DropColumns(storey.MapClass): def __init__(self, columns: List[str], **kwargs): super().__init__(**kwargs) self.columns = columns def do(self, event): for col in self.columns: if col in event: del event[col] return event def to_dict(self): return { "class_name": "DropColumns", "name": self.name or "DropColumns", "class_args": { "columns": self.columns }, } class RenameColumns(storey.MapClass): def __init__(self, mapping: Dict[str, str], **kwargs): super().__init__(**kwargs) self.mapping = mapping def do(self, event): for old_col, new_col in self.mapping.items(): try: event[new_col] = event.pop(old_col) except Exception as e: print(f'{old_col} doesnt exist') return event def to_dict(self): return { "class_name": "RenameColumns", "name": self.name or "RenameColumns", "class_args": {"mapping": self.mapping}, } # nuclio: end-code ###Output _____no_output_____ ###Markdown Define the Real-Time PipelineDefine the transoformation pipeline below. This is done just like the previous sections. ###Code # Configure the list of columns to drop from # the raw data drop_columns = ['hr_is_error', 'rr_is_error', 'spo2_is_error', 'movements_is_error', 'turn_count_is_error', 'is_in_bed_is_error'] # Define the computationala graph including our custom functions early_sense_set.graph.to(DropColumns(drop_columns), after='start')\ .to(RenameColumns(mapping={'bad': 'bed'})) # Add real-time aggreagations on top of our sensor readings for col in ['hr', 'rr', 'spo2', 'movements', 'turn_count']: early_sense_set.add_aggregation(col + "_h", col, ['avg', 'max', 'min'], "1h") early_sense_set.add_aggregation(col + "_d", col, ['avg', 'max', 'min'], "1d") early_sense_set.add_aggregation('in_bed_h', 'is_in_bed', ['avg'], "1h") early_sense_set.add_aggregation('in_bed_d', 'is_in_bed', ['avg'], "1d") # Set NoSQL and Parquet default targets early_sense_set.set_targets() # Plot the pipeline early_sense_set.plot(rankdir='LR', with_targets=True) ###Output _____no_output_____ ###Markdown Test/debug the real-time pipeline locally in the notebook ###Code # infer schema + stats, show the final feature set (after the data pipeline) early_sense_df = pd.read_parquet('https://s3.wasabisys.com/iguazio/data/patients/early_sense.parquet') early_sense_df['timestamp'] = pd.to_datetime(early_sense_df['timestamp']) early_sense_df = move_timestamps(early_sense_df) # update timestamps fs.infer(early_sense_set, early_sense_df.head()) # Run ingest pipeline df=fs.ingest(early_sense_set, early_sense_df) # Save the early-sense Featureset early_sense_set.save() # print the FeatureSet spec print(early_sense_set.status.targets.to_dict()) ###Output [{'name': 'parquet', 'kind': 'parquet', 'path': 'v3io:///projects/fsdemo-admin/fs/parquet/sets/early_sense-latest.parquet', 'status': 'created', 'updated': '2021-05-06T15:27:46.222973+00:00'}, {'name': 'nosql', 'kind': 'nosql', 'path': 'v3io:///projects/fsdemo-admin/fs/nosql/sets/early_sense-latest', 'status': 'created', 'updated': '2021-05-06T15:27:46.223349+00:00'}] ###Markdown Deploy as Real-Time Stream Processing Function (Nuclio Serverless)Features are not static. For example, it is common that features include different aggregations that need to be updated as data continues to flow. A real-time pipeline requires this data to be up date. Therefore, we need a convenient way to ingest data, not just as batch, but per specific input.MLRun can convert any code to a real-time serverless function, including the pipeline. This is done by performing the following steps:1. Define a source, in this case it's an HTTP source2. Convert the previously defined code to a serving function3. Create a configuration to run the function4. Deploy an ingestion service with the Featureset, source and the configuration ###Code # Set a new HTTPSource, this will tell our ingestion service # to setup a Nuclio function to act as the rest endpoint # to which we would receive the data source = HttpSource(key_field='patient_id', time_field='timestamp') # Take the relevant code parts from this notebook and create # an MLRun function from them so we can run the pipeline # as a Nuclio function func = mlrun.code_to_function("ingest", kind="serving") nuclio_config = fs.RunConfig(function=func, local=False).apply(mlrun.platforms.auto_mount()) # Deploy the Online ingestion service using the pipeline definition from before # with our new HTTP Source and our define Function server = fs.deploy_ingestion_service(early_sense_set, source, run_config=nuclio_config) ###Output > 2021-05-06 15:29:52,032 [info] Starting remote function deploy 2021-05-06 15:29:52 (info) Deploying function {'level': 'info', 'message': 'Deploying function', 'name': 'fsdemo-admin-ingest', 'time': 1620314992169.7139} 2021-05-06 15:29:52 (info) Building {'level': 'info', 'message': 'Building', 'name': 'fsdemo-admin-ingest', 'time': 1620314992169.7478, 'versionInfo': 'Label: 1.5.16, Git commit: ae43a6a560c2bec42d7ccfdf6e8e11a1e3cc3774, OS: linux, Arch: amd64, Go version: go1.14.3'} 2021-05-06 15:29:52 (info) Staging files and preparing base images {'level': 'info', 'message': 'Staging files and preparing base images', 'name': 'deployer', 'time': 1620314992237.7905} 2021-05-06 15:29:52 (info) Building processor image {'imageName': 'fsdemo-admin-fsdemo-admin-ingest-processor:latest', 'level': 'info', 'message': 'Building processor image', 'name': 'deployer', 'time': 1620314992238.347} 2021-05-06 15:29:55 (info) Build complete {'level': 'info', 'message': 'Build complete', 'name': 'deployer', 'result': {'Image': 'fsdemo-admin-fsdemo-admin-ingest-processor:latest', 'UpdatedFunctionConfig': {'metadata': {'annotations': {'nuclio.io/generated_by': 'function generated from https://github.com/mlrun/mlrun#004d7b6797e3292525d220bb4389470342ebe752:ingest.ipynb'}, 'labels': {'mlrun/class': 'serving', 'nuclio.io/project-name': 'fsdemo-admin'}, 'name': 'fsdemo-admin-ingest', 'namespace': 'default-tenant'}, 'spec': {'build': {'baseImage': 'mlrun/mlrun:0.6.3-rc9', 'codeEntryType': 'sourceCode', 'functionSourceCode': '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', 'noBaseImagesPull': True, 'offline': True, 'registry': 'docker-registry.default-tenant.app.yh30.iguazio-c0.com'}, 'env': [{'name': 'V3IO_API', 'value': 'v3io-webapi.default-tenant.svc:8081'}, {'name': 'V3IO_USERNAME', 'value': 'admin'}, {'name': 'V3IO_ACCESS_KEY', 'value': '142a98fa-bef9-4095-b2d0-cab733f53238'}, {'name': 'MLRUN_LOG_LEVEL', 'value': 'DEBUG'}, {'name': 'MLRUN_DEFAULT_PROJECT', 'value': 'fsdemo-admin'}, {'name': 'MLRUN_DBPATH', 'value': 'http://mlrun-api:8080'}, {'name': 'MLRUN_NAMESPACE', 'value': 'default-tenant'}, {'name': 'SERVING_SPEC_ENV', 'value': '{"function_uri": "fsdemo-admin/ingest", "version": "v2", "parameters": {"infer_options": 0, "featureset": "store://feature-sets/fsdemo-admin/early_sense", "source": {"kind": "http", "path": "None", "key_field": "patient_id", "time_field": "timestamp", "online": true}}, "graph": {"states": {"DropColumns": {"kind": "task", "class_name": "DropColumns", "class_args": {"columns": ["hr_is_error", "rr_is_error", "spo2_is_error", "movements_is_error", "turn_count_is_error", "is_in_bed_is_error"]}}, "RenameColumns": {"kind": "task", "class_name": "RenameColumns", "class_args": {"mapping": {"bad": "bed"}}, "after": ["DropColumns"]}, "Aggregates": {"kind": "task", "class_name": "storey.AggregateByKey", "class_args": {"aggregates": [{"name": "hr", "column": "hr", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "rr", "column": "rr", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "spo2", "column": "spo2", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "movements", "column": "movements", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "turn_count", "column": "turn_count", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "in_bed", "column": "is_in_bed", "operations": ["avg"], "windows": ["1h", "1d"]}], "table": "."}, "after": ["RenameColumns"]}}}, "load_mode": null, "functions": {}, "graph_initializer": "mlrun.feature_store.ingestion.featureset_initializer", "error_stream": null, "track_models": null}'}], 'eventTimeout': '', 'handler': '01-ingest-datasources:handler', 'maxReplicas': 4, 'minReplicas': 1, 'platform': {}, 'resources': {}, 'runtime': 'python:3.6', 'securityContext': {}, 'serviceType': 'NodePort', 'triggers': {'default-http': {'attributes': {'serviceType': 'NodePort'}, 'class': '', 'kind': 'http', 'maxWorkers': 1, 'name': 'default-http'}}, 'volumes': [{'volume': {'flexVolume': {'driver': 'v3io/fuse', 'options': {'accessKey': '142a98fa-bef9-4095-b2d0-cab733f53238'}}, 'name': 'v3io'}, 'volumeMount': {'mountPath': '/v3io', 'name': 'v3io'}}, {'volume': {'flexVolume': {'driver': 'v3io/fuse', 'options': {'accessKey': '142a98fa-bef9-4095-b2d0-cab733f53238'}}, 'name': 'v3io'}, 'volumeMount': {'mountPath': '/User', 'name': 'v3io', 'subPath': 'users/admin'}}]}}}, 'time': 1620314995613.7964} > 2021-05-06 15:30:03,749 [info] function deployed, address=default-tenant.app.yh30.iguazio-c0.com:31610 ###Markdown Test the function by sending data to the HTTP endpoint ###Code test_data = {'patient_id': '838-21-8151', 'bad': 38, 'department': '01e9fe31-76de-45f0-9aed-0f94cc97bca0', 'room': 1, 'hr': 220.0, 'hr_is_error': True, 'rr': 5, 'rr_is_error': True, 'spo2': 85, 'spo2_is_error': True, 'movements': 0.0, 'movements_is_error': True, 'turn_count': 0.0, 'turn_count_is_error': True, 'is_in_bed': 1, 'is_in_bed_is_error': False, 'timestamp': 1606843455.906352 } import requests import json response = requests.post(server, json=test_data) response.text ###Output _____no_output_____ ###Markdown Ingest labelsFinally, we define label data, this will be useful in the next notebook where we train a model Create Labels Set ###Code # Define labels metric from the early sense error data error_columns = [c for c in early_sense_df.columns if 'error' in c] labels = early_sense_df.loc[:, ['patient_id', 'timestamp'] + error_columns] labels['label'] = labels.apply(lambda x: sum([x[c] for c in error_columns])>(len(error_columns)*0.7), axis=1) labels.to_parquet(data_path + 'labels.parquet') #labels_df = pd.read_parquet('labels.parquet') labels_set = fs.FeatureSet("labels", entities=[fs.Entity("patient_id")], timestamp_key='timestamp', description="training labels") labels_set.set_targets() df = fs.infer(labels_set, data_path + 'labels.parquet') df.head() df = fs.ingest(labels_set, data_path + 'labels.parquet') labels_set.save() ###Output _____no_output_____ ###Markdown Part 1: Data Ingestion In this notebook we will learn how to **Ingest** different data sources to our **Feature Store**. Specifically, this patient data has been successfully used to treat hospitalized COVID-19 patients prior to their condition becoming severe or critical. To do this we will use a medical dataset which includes three types of data: - **Healthcare systems**: Batch updated dataset, containing different lab test results (Blood test results for ex.).- **Patient Records**: Static dataset containing general patient details.- **Real-time sensors**: Real-Time patient metric monitoring sensor. We will walk through creation of ingestion pipeline for each datasource with all the needed preprocessing and validation. We will run the pipeline locally within the notebook and then launch a real-time function to **ingest live data** or schedule a cron to run the task when needed. Environment SetupSince our work is done in a this project scope, first define the project itself for all our MLRun work in this notebook. ###Code import mlrun project, artifact_path = mlrun.set_environment(project='fsdemo', user_project=True) # location of the output data files data_path = f"{artifact_path}/data/" def move_timestamps(df, shift='0s'): ''' Update timetsamps to current time so we can see live aggregations ''' now = pd.to_datetime('now') max_time = df['timestamp'].max() time_shift = now-max_time tmp_df = df.copy() tmp_df['timestamp'] = tmp_df['timestamp'].apply(lambda t: t + time_shift + pd.to_timedelta(shift)) return tmp_df ###Output _____no_output_____ ###Markdown Create Ingestion Pipeline With MLRunIn this section we will ingest the lab measurements data using MLRun and Storey. Storey is the underlying implementation of the feature store which is used by MLRun. It is the engine that allows you to define and execute complex graphs that create the feature engineering pipeline. With storey, you can define source, transformations and targets, many actions are available as part of the Storey library, but you can define additional actions easily. We will see these custom actions in later sections.For the execution, it is possible to also use Spark. The main difference between Storey and Spark pipelines is that Storey blocks are built for Real-Time workloads while Spark is more Batch oriented. We will now do the following:- Create the `measurements` FeatureSet- Define Preprocessing graph including aggregations- Ingest the data using the defined pipeline ###Code # Import MLRun's Feature Store import mlrun.feature_store as fstore # Import MLRun's Data Sources to set the wanted ingestion pipeline from mlrun.datastore.sources import CSVSource, ParquetSource, HttpSource # Import storey so it will be available on our scope # when testing the pipeline import storey # Define the Lab Measurements FeatureSet measurements_set = fstore.FeatureSet("measurements", entities=[fstore.Entity("patient_id")], timestamp_key='timestamp', description="various patient health measurements") # Get FeatureSet computation graph measurements_graph = measurements_set.graph ###Output _____no_output_____ ###Markdown Define the processing pipeline- Transformation function- Sliding window aggregation- Set targets (NoSQL and Parquet) ###Code # Import pandas and load the sample CSV and load it as a datasource # for our ingestion import pandas as pd measurements_df = pd.read_csv('https://s3.wasabisys.com/iguazio/data/patients/measurements.csv', index_col=0) measurements_df['timestamp'] = pd.to_datetime(measurements_df['timestamp']) measurements_df['timestamp'] = measurements_df['timestamp'].astype("datetime64[ms]") measurements_df = pd.concat([move_timestamps(measurements_df, '-1h'), move_timestamps(measurements_df)]) # update timestamps ###Output _____no_output_____ ###Markdown Take a look at the measurements dataset. This dataset includes a a single measurement per row. The measurement type is defined by the `source` and `parameter` column. We would like to transform this data, so each patient has multiple measurement columns. To do that, we will need to create a new column for each `source` and `parameter` combination. For example, if `source` is 3 and `parameter` is 0, then our transformed dataset will have the measurement value in a new feature named `sp_3_0`.Following that, we will create a sliding window aggregation that averages the values across that time window. ###Code measurements_df.head() ###Output _____no_output_____ ###Markdown The following code performs the transformation, adds the aggregation and sets the target to store the values to a NoSQL database for online retrieval and parquet files for batch processing. ###Code # Define transform to create sparse dataset for aggregation # adding an extra column for the specific source-parameter pair's measurement # ex: source=3, parameter=4, measurement=100 -> add extra column sp_3_4=100 def transform(event): event["_".join(['sp', str(event["source"]), str(event["parameter"])])] = event["measurement"] return event # Define Measurement FeatureSet pipeline measurements_graph.to( "storey.Map", _fn="transform" ) # Get the available source, parameter pairs for our aggregation sps = list(measurements_df.apply(lambda x: '_'.join(['sp', str(x['source']), str(x['parameter'])]), axis=1).unique()) # Add aggregations on top of the created sparse # features by the transorm function for col in sps: measurements_set.add_aggregation(name=f'agg_{col}', column=col, operations=['avg'], windows='1h', period='30m') # Add default (NoSQL via KV and Parquet) targets to save # the ingestion results to measurements_set.set_targets() ###Output _____no_output_____ ###Markdown You can plot the graph to visualize the pipeline: ###Code # Plot the ingestion pipeline we defined measurements_set.plot(rankdir='LR', with_targets=True) ###Output _____no_output_____ ###Markdown Run ingestion task using MLRun & StoreyIn order to ingest the dataframe to the featureset, use the `ingest` function. ###Code # User our loaded DF as the datasource and ingest it through # the define pipeline resp = fstore.ingest(measurements_set, measurements_df, infer_options=fstore.InferOptions.default()) resp.head() # Save the FeatureSet and pipeline definition measurements_set.save() ###Output _____no_output_____ ###Markdown Ingest Patient Details Features In this section we will use MLRun to create our patient details datasource. We will do the following:- Create a `patient_details` FeatureSet- Add preprocessing transformations to the pipeline - Map ages to buckets and One Hot Encode them - Impute missing values- Test the processing pipeline with sample data- Run ingestion pipeline on top of the cluster Create the FeatureSet ###Code # add feature set without time column (stock ticker metadata) patients_set = fstore.FeatureSet("patient_details", entities=[fstore.Entity("patient_id")], description="personal and medical patient details") # Get FeatureSet computation graph graph = patients_set.spec.graph ###Output _____no_output_____ ###Markdown Define the computation pipeline ###Code # Define age buckets for our age value mapping personal_details = {'age': {'ranges': [{'range': [0, 3], "value": "toddler"}, {'range': [3, 18], "value": "child"}, {'range': [18, 65], "value": "adult"}, {'range': [65, 120], "value": "elder"}]}} # Define one hot encoding values map one_hot_encoder_mapping = {'age_mapped': ['toddler', 'child', 'adult', 'elder']} # Import MLRun's FeatureStore steps for easy # use in our pipeline from mlrun.feature_store.steps import * # Define the pipeline for our FeatureSet graph.to(MapValues(mapping=personal_details, with_original_features=True))\ .to(OneHotEncoder(mapping=one_hot_encoder_mapping))\ .to(Imputer(method='values', default_value=1, mapping={})) # Add default NoSQL & Parquet ingestion targets patients_set.set_targets() # Plot the FeatureSet pipeline patients_set.plot(rankdir='LR', with_targets=True) ###Output _____no_output_____ ###Markdown Test the Feature transformation pipelineCreating a transformation pipeline requires some trial and error. Therefore, it is useful to run the pipeline in memory without storing the resultant data. For this purpose, `infer` is used. This function receives as input any sample DataFrame, performs all the graph steps and outputs the transformed DataFrame. ###Code # Load the sample patient details data patients_df = pd.read_parquet('https://s3.wasabisys.com/iguazio/data/patients/patient_details.parquet') # Run local ingestion test fstore.preview(patients_set, patients_df.head()) ###Output _____no_output_____ ###Markdown Save the FeatureSet and run full ingestion taskOnce you are satisfied with the transformation pipeline, ingest that full DataFrame and store the data. ###Code # Save the FeatureSet patients_set.save() # Run Ingestion task resp = fstore.ingest(patients_set, patients_df, infer_options=fstore.InferOptions.default()) ###Output _____no_output_____ ###Markdown Start Immediate or Scheduled Ingestion Job (over Kubernetes)Another useful method to ingest data, is by creating a Kubernetes job. This may be necessary to process large amounts of data as well as to process any recurring data. With MLRun it is easy to take the pipeline and run it as a job. This is done by:1. Define a source, specifically here we define a parquet file source2. Define a configuration where `local` is set to `False`3. Mount to the provisioned storage by calling `auto_mount`4. Run `ingest` with the source and run configuration ###Code source = ParquetSource('pq', 'https://s3.wasabisys.com/iguazio/data/patients/patient_details.parquet') config = fstore.RunConfig(local=False).apply(mlrun.platforms.auto_mount()) fstore.ingest(patients_set, source, run_config=config) ###Output > 2021-05-06 15:27:08,769 [info] starting run patient_details_ingest uid=76f197b8ab3347d1b995a5ea55d0a98a DB=http://mlrun-api:8080 > 2021-05-06 15:27:09,022 [info] Job is running in the background, pod: patient-details-ingest-g9hgn > 2021-05-06 15:27:15,073 [info] starting ingestion task to store://feature-sets/fsdemo-admin/patient_details:latest > 2021-05-06 15:27:15,745 [info] ingestion task completed, targets: > 2021-05-06 15:27:15,746 [info] [{'name': 'parquet', 'kind': 'parquet', 'path': 'v3io:///projects/fsdemo-admin/fs/parquet/sets/patient_details-latest.parquet', 'status': 'created', 'updated': '2021-05-06T15:27:15.432576+00:00'}, {'name': 'nosql', 'kind': 'nosql', 'path': 'v3io:///projects/fsdemo-admin/fs/nosql/sets/patient_details-latest', 'status': 'created', 'updated': '2021-05-06T15:27:15.432947+00:00'}] > 2021-05-06 15:27:15,936 [info] run executed, status=completed final state: completed ###Markdown Real-time Early-Sense Sensor Ingestion (HTTP or Stream Processing With Nuclio) In this section we will use MLRun to create our Early Sense Sensor datasource. We will do the following:- Create early sense FeatureSet- Add Preprocessing transformations to the Pipeline using custom functions - Drop and Rename columns - Aggregations- Add Feature Validator to detect bad sensor readings- Test the processing pipeline with sample data- Deploy the FeatureSet ingestion service as a live rest endpoint ###Code early_sense_set = fstore.FeatureSet("early_sense", entities=[fstore.Entity("patient_id")], timestamp_key='timestamp', description="real time patient bed sensor data") ###Output _____no_output_____ ###Markdown Define data validation & quality policyWe can define validations on the feature level. For example, define here validation to check if the heart-rate value is between 0 and 220 and respiratory rate is between 0 and 25. ###Code from mlrun.features import MinMaxValidator early_sense_set["hr"] = fstore.Feature(validator = MinMaxValidator(min=0, max=220, severity="info")) early_sense_set["rr"] = fstore.Feature(validator = MinMaxValidator(min=0, max=25, severity="info")) ###Output _____no_output_____ ###Markdown Define custom processing classesIn the previous sections we used transformation steps that are available as part of Storey. Here we show how to create custom transformation classes. We will later run these functions as part of a Nuclio serverless real-time function, therefore, we also use the nuclio `start-code` and `end-code` comments. ###Code # nuclio: start-code # We will import storey here too so it will # be included in our function code (within the nuclio comment block) import json import storey from typing import List, Dict # The custom functions are based on `storey.MapClass` # when they are called in the graph the `do(self, event)` # function will be activated. # A to_dict(self) function is also required by MLRun # to allow the class creation on remote functions class DropColumns(storey.MapClass): def __init__(self, columns: List[str], **kwargs): super().__init__(**kwargs) self.columns = columns def do(self, event): for col in self.columns: if col in event: del event[col] return event def to_dict(self): return { "class_name": "DropColumns", "name": self.name or "DropColumns", "class_args": { "columns": self.columns }, } class RenameColumns(storey.MapClass): def __init__(self, mapping: Dict[str, str], **kwargs): super().__init__(**kwargs) self.mapping = mapping def do(self, event): for old_col, new_col in self.mapping.items(): try: event[new_col] = event.pop(old_col) except Exception as e: print(f'{old_col} doesnt exist') return event def to_dict(self): return { "class_name": "RenameColumns", "name": self.name or "RenameColumns", "class_args": {"mapping": self.mapping}, } # nuclio: end-code ###Output _____no_output_____ ###Markdown Define the Real-Time PipelineDefine the transformation pipeline below. This is done just like the previous sections. ###Code # Configure the list of columns to drop from # the raw data drop_columns = ['hr_is_error', 'rr_is_error', 'spo2_is_error', 'movements_is_error', 'turn_count_is_error', 'is_in_bed_is_error'] # Define the computationala graph including our custom functions early_sense_set.graph.to(DropColumns(drop_columns), after='start')\ .to(RenameColumns(mapping={'bad': 'bed'})) # Add real-time aggreagations on top of our sensor readings for col in ['hr', 'rr', 'spo2', 'movements', 'turn_count']: early_sense_set.add_aggregation(col + "_h", col, ['avg', 'max', 'min'], "1h") early_sense_set.add_aggregation(col + "_d", col, ['avg', 'max', 'min'], "1d") early_sense_set.add_aggregation('in_bed_h', 'is_in_bed', ['avg'], "1h") early_sense_set.add_aggregation('in_bed_d', 'is_in_bed', ['avg'], "1d") # Set NoSQL and Parquet default targets early_sense_set.set_targets() # Plot the pipeline early_sense_set.plot(rankdir='LR', with_targets=True) ###Output _____no_output_____ ###Markdown Test/debug the real-time pipeline locally in the notebook ###Code # infer schema + stats, show the final feature set (after the data pipeline) early_sense_df = pd.read_parquet('https://s3.wasabisys.com/iguazio/data/patients/early_sense.parquet') early_sense_df['timestamp'] = pd.to_datetime(early_sense_df['timestamp']) early_sense_df = move_timestamps(early_sense_df) # update timestamps fstore.preview(early_sense_set, early_sense_df.head()) # Run ingest pipeline df=fstore.ingest(early_sense_set, early_sense_df) # Save the early-sense Featureset early_sense_set.save() # print the FeatureSet spec print(early_sense_set.status.targets.to_dict()) ###Output [{'name': 'parquet', 'kind': 'parquet', 'path': 'v3io:///projects/fsdemo-admin/fs/parquet/sets/early_sense-latest.parquet', 'status': 'created', 'updated': '2021-05-06T15:27:46.222973+00:00'}, {'name': 'nosql', 'kind': 'nosql', 'path': 'v3io:///projects/fsdemo-admin/fs/nosql/sets/early_sense-latest', 'status': 'created', 'updated': '2021-05-06T15:27:46.223349+00:00'}] ###Markdown Deploy as Real-Time Stream Processing Function (Nuclio Serverless)Features are not static. For example, it is common that features include different aggregations that need to be updated as data continues to flow. A real-time pipeline requires this data to be up date. Therefore, we need a convenient way to ingest data, not just as batch, but per specific input.MLRun can convert any code to a real-time serverless function, including the pipeline. This is done by performing the following steps:1. Define a source, in this case it's an HTTP source2. Convert the previously defined code to a serving function3. Create a configuration to run the function4. Deploy an ingestion service with the Featureset, source and the configuration ###Code # Set a new HTTPSource, this will tell our ingestion service # to setup a Nuclio function to act as the rest endpoint # to which we would receive the data source = HttpSource(key_field='patient_id', time_field='timestamp') # Take the relevant code parts from this notebook and create # an MLRun function from them so we can run the pipeline # as a Nuclio function func = mlrun.code_to_function("ingest", kind="serving") nuclio_config = fstore.RunConfig(function=func, local=False).apply(mlrun.platforms.auto_mount()) # Deploy the Online ingestion service using the pipeline definition from before # with our new HTTP Source and our define Function server = fstore.deploy_ingestion_service(early_sense_set, source, run_config=nuclio_config) ###Output > 2021-05-06 15:29:52,032 [info] Starting remote function deploy 2021-05-06 15:29:52 (info) Deploying function {'level': 'info', 'message': 'Deploying function', 'name': 'fsdemo-admin-ingest', 'time': 1620314992169.7139} 2021-05-06 15:29:52 (info) Building {'level': 'info', 'message': 'Building', 'name': 'fsdemo-admin-ingest', 'time': 1620314992169.7478, 'versionInfo': 'Label: 1.5.16, Git commit: ae43a6a560c2bec42d7ccfdf6e8e11a1e3cc3774, OS: linux, Arch: amd64, Go version: go1.14.3'} 2021-05-06 15:29:52 (info) Staging files and preparing base images {'level': 'info', 'message': 'Staging files and preparing base images', 'name': 'deployer', 'time': 1620314992237.7905} 2021-05-06 15:29:52 (info) Building processor image {'imageName': 'fsdemo-admin-fsdemo-admin-ingest-processor:latest', 'level': 'info', 'message': 'Building processor image', 'name': 'deployer', 'time': 1620314992238.347} 2021-05-06 15:29:55 (info) Build complete {'level': 'info', 'message': 'Build complete', 'name': 'deployer', 'result': {'Image': 'fsdemo-admin-fsdemo-admin-ingest-processor:latest', 'UpdatedFunctionConfig': {'metadata': {'annotations': {'nuclio.io/generated_by': 'function generated from https://github.com/mlrun/mlrun#004d7b6797e3292525d220bb4389470342ebe752:ingest.ipynb'}, 'labels': {'mlrun/class': 'serving', 'nuclio.io/project-name': 'fsdemo-admin'}, 'name': 'fsdemo-admin-ingest', 'namespace': 'default-tenant'}, 'spec': {'build': {'baseImage': 'mlrun/mlrun:0.6.3-rc9', 'codeEntryType': 'sourceCode', 'functionSourceCode': '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', 'noBaseImagesPull': True, 'offline': True, 'registry': 'docker-registry.default-tenant.app.yh30.iguazio-c0.com'}, 'env': [{'name': 'V3IO_API', 'value': 'v3io-webapi.default-tenant.svc:8081'}, {'name': 'V3IO_USERNAME', 'value': 'admin'}, {'name': 'V3IO_ACCESS_KEY', 'value': '142a98fa-bef9-4095-b2d0-cab733f53238'}, {'name': 'MLRUN_LOG_LEVEL', 'value': 'DEBUG'}, {'name': 'MLRUN_DEFAULT_PROJECT', 'value': 'fsdemo-admin'}, {'name': 'MLRUN_DBPATH', 'value': 'http://mlrun-api:8080'}, {'name': 'MLRUN_NAMESPACE', 'value': 'default-tenant'}, {'name': 'SERVING_SPEC_ENV', 'value': '{"function_uri": "fsdemo-admin/ingest", "version": "v2", "parameters": {"infer_options": 0, "featureset": "store://feature-sets/fsdemo-admin/early_sense", "source": {"kind": "http", "path": "None", "key_field": "patient_id", "time_field": "timestamp", "online": true}}, "graph": {"states": {"DropColumns": {"kind": "task", "class_name": "DropColumns", "class_args": {"columns": ["hr_is_error", "rr_is_error", "spo2_is_error", "movements_is_error", "turn_count_is_error", "is_in_bed_is_error"]}}, "RenameColumns": {"kind": "task", "class_name": "RenameColumns", "class_args": {"mapping": {"bad": "bed"}}, "after": ["DropColumns"]}, "Aggregates": {"kind": "task", "class_name": "storey.AggregateByKey", "class_args": {"aggregates": [{"name": "hr", "column": "hr", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "rr", "column": "rr", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "spo2", "column": "spo2", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "movements", "column": "movements", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "turn_count", "column": "turn_count", "operations": ["avg", "max", "min"], "windows": ["1h", "1d"]}, {"name": "in_bed", "column": "is_in_bed", "operations": ["avg"], "windows": ["1h", "1d"]}], "table": "."}, "after": ["RenameColumns"]}}}, "load_mode": null, "functions": {}, "graph_initializer": "mlrun.feature_store.ingestion.featureset_initializer", "error_stream": null, "track_models": null}'}], 'eventTimeout': '', 'handler': '01-ingest-datasources:handler', 'maxReplicas': 4, 'minReplicas': 1, 'platform': {}, 'resources': {}, 'runtime': 'python:3.6', 'securityContext': {}, 'serviceType': 'NodePort', 'triggers': {'default-http': {'attributes': {'serviceType': 'NodePort'}, 'class': '', 'kind': 'http', 'maxWorkers': 1, 'name': 'default-http'}}, 'volumes': [{'volume': {'flexVolume': {'driver': 'v3io/fuse', 'options': {'accessKey': '142a98fa-bef9-4095-b2d0-cab733f53238'}}, 'name': 'v3io'}, 'volumeMount': {'mountPath': '/v3io', 'name': 'v3io'}}, {'volume': {'flexVolume': {'driver': 'v3io/fuse', 'options': {'accessKey': '142a98fa-bef9-4095-b2d0-cab733f53238'}}, 'name': 'v3io'}, 'volumeMount': {'mountPath': '/User', 'name': 'v3io', 'subPath': 'users/admin'}}]}}}, 'time': 1620314995613.7964} > 2021-05-06 15:30:03,749 [info] function deployed, address=default-tenant.app.yh30.iguazio-c0.com:31610 ###Markdown Test the function by sending data to the HTTP endpoint ###Code test_data = {'patient_id': '838-21-8151', 'bad': 38, 'department': '01e9fe31-76de-45f0-9aed-0f94cc97bca0', 'room': 1, 'hr': 220.0, 'hr_is_error': True, 'rr': 5, 'rr_is_error': True, 'spo2': 85, 'spo2_is_error': True, 'movements': 0.0, 'movements_is_error': True, 'turn_count': 0.0, 'turn_count_is_error': True, 'is_in_bed': 1, 'is_in_bed_is_error': False, 'timestamp': 1606843455.906352 } import requests import json response = requests.post(server, json=test_data) response.text ###Output _____no_output_____ ###Markdown Ingest labelsFinally, we define label data, this will be useful in the next notebook where we train a model Create Labels Set ###Code # Define labels metric from the early sense error data error_columns = [c for c in early_sense_df.columns if 'error' in c] labels = early_sense_df.loc[:, ['patient_id', 'timestamp'] + error_columns] labels['label'] = labels.apply(lambda x: sum([x[c] for c in error_columns])>(len(error_columns)*0.7), axis=1) labels.to_parquet(data_path + 'labels.parquet') #labels_df = pd.read_parquet('labels.parquet') labels_set = fstore.FeatureSet("labels", entities=[fstore.Entity("patient_id")], timestamp_key='timestamp', description="training labels") labels_set.set_targets() df = fstore.preview(labels_set, data_path + 'labels.parquet') df.head() df = fstore.ingest(labels_set, data_path + 'labels.parquet') labels_set.save() ###Output _____no_output_____ ###Markdown Part 1: Data Ingestion In this notebook we will learn how to **Ingest** different data sources to our **Feature Store**. Specifically, this patient data has been successfully used to treat hospitalized COVID-19 patients prior to their condition becoming severe or critical. To do this we will use a medical dataset which includes three types of data: - **Healthcare systems**: Batch updated dataset, containing different lab test results (Blood test results for ex.).- **Patient Records**: Static dataset containing general patient details.- **Real-time sensors**: Real-Time patient metric monitoring sensor. We will walk through creation of ingestion pipeline for each datasource with all the needed preprocessing and validation. We will run the pipeline locally within the notebook and then launch a real-time function to **ingest live data** or schedule a cron to run the task when needed. Environment SetupSince our work is done in a this project scope, first define the project itself for all our MLRun work in this notebook. ###Code import mlrun project, _ = mlrun.set_environment(project='fsdemo', user_project=True) def move_timestamps(df, shift='0s'): ''' Update timetsamps to current time so we can see live aggregations ''' now = pd.to_datetime('now') max_time = df['timestamp'].max() time_shift = now-max_time tmp_df = df.copy() tmp_df['timestamp'] = tmp_df['timestamp'].apply(lambda t: t + time_shift + pd.to_timedelta(shift)) return tmp_df ###Output _____no_output_____ ###Markdown Create Ingestion Pipeline With MLRunIn this section we will ingest the lab measurements data using MLRun and Storey. Storey is the underlying implementation of the feature store which is used by MLRun. It is the engine that allows you to define and execute complex graphs that create the feature engineering pipeline. With storey, you can define source, transformations and targets, many actions are available as part of the Storey library, but you can define additional actions easily. We will see these custom actions in later sections.For the execution, it is possible to also use Spark. The main difference between Storey and Spark pipelines is that Storey blocks are built for Real-Time workloads while Spark is more Batch oriented. We will now do the following:- Create the `measurements` FeatureSet- Define Preprocessing graph including aggregations- Ingest the data using the defined pipeline ###Code # Import MLRun's Feature Store import mlrun.feature_store as fstore # Import MLRun's Data Sources to set the wanted ingestion pipeline from mlrun.datastore.sources import CSVSource, ParquetSource, HttpSource # Import storey so it will be available on our scope # when testing the pipeline import storey # Define the Lab Measurements FeatureSet measurements_set = fstore.FeatureSet("measurements", entities=[fstore.Entity("patient_id")], timestamp_key='timestamp', description="various patient health measurements") # Get FeatureSet computation graph measurements_graph = measurements_set.graph ###Output _____no_output_____ ###Markdown Define the processing pipeline- Transformation function- Sliding window aggregation- Set targets (NoSQL and Parquet) ###Code # Import pandas and load the sample CSV and load it as a datasource # for our ingestion import pandas as pd measurements_df = pd.read_csv('https://s3.wasabisys.com/iguazio/data/patients/measurements.csv', index_col=0) measurements_df['timestamp'] = pd.to_datetime(measurements_df['timestamp']) measurements_df['timestamp'] = measurements_df['timestamp'].astype("datetime64[ms]") measurements_df = pd.concat([move_timestamps(measurements_df, '-1h'), move_timestamps(measurements_df)]) # update timestamps ###Output _____no_output_____ ###Markdown Take a look at the measurements dataset. This dataset includes a a single measurement per row. The measurement type is defined by the `source` and `parameter` column. We would like to transform this data, so each patient has multiple measurement columns. To do that, we will need to create a new column for each `source` and `parameter` combination. For example, if `source` is 3 and `parameter` is 0, then our transformed dataset will have the measurement value in a new feature named `sp_3_0`.Following that, we will create a sliding window aggregation that averages the values across that time window. ###Code measurements_df.head() ###Output _____no_output_____ ###Markdown The following code performs the transformation, adds the aggregation and sets the target to store the values to a NoSQL database for online retrieval and parquet files for batch processing. ###Code # Define transform to create sparse dataset for aggregation # adding an extra column for the specific source-parameter pair's measurement # ex: source=3, parameter=4, measurement=100 -> add extra column sp_3_4=100 def transform(event): event["_".join(['sp', str(event["source"]), str(event["parameter"])])] = event["measurement"] return event # Define Measurement FeatureSet pipeline measurements_graph.to( "storey.Map", _fn="transform" ) # Get the available source, parameter pairs for our aggregation sps = list(measurements_df.apply(lambda x: '_'.join(['sp', str(x['source']), str(x['parameter'])]), axis=1).unique()) # Add aggregations on top of the created sparse # features by the transorm function for col in sps: measurements_set.add_aggregation(name=f'agg_{col}', column=col, operations=['avg'], windows='1h', period='30m') # Add default (NoSQL via KV and Parquet) targets to save # the ingestion results to measurements_set.set_targets() ###Output _____no_output_____ ###Markdown You can plot the graph to visualize the pipeline: ###Code # Plot the ingestion pipeline we defined measurements_set.plot(rankdir='LR', with_targets=True) ###Output _____no_output_____ ###Markdown Run ingestion task using MLRun & StoreyIn order to ingest the dataframe to the featureset, use the `ingest` function. ###Code # User our loaded DF as the datasource and ingest it through # the define pipeline resp = fstore.ingest(measurements_set, measurements_df, infer_options=fstore.InferOptions.default()) resp.head() # Save the FeatureSet and pipeline definition measurements_set.save() ###Output _____no_output_____ ###Markdown Ingest Patient Details Features In this section we will use MLRun to create our patient details datasource. We will do the following:- Create a `patient_details` FeatureSet- Add preprocessing transformations to the pipeline - Map ages to buckets and One Hot Encode them - Impute missing values- Test the processing pipeline with sample data- Run ingestion pipeline on top of the cluster Create the FeatureSet ###Code # add feature set without time column (stock ticker metadata) patients_set = fstore.FeatureSet("patient_details", entities=[fstore.Entity("patient_id")], description="personal and medical patient details") # Get FeatureSet computation graph graph = patients_set.spec.graph ###Output _____no_output_____ ###Markdown Define the computation pipeline ###Code # Define age buckets for our age value mapping personal_details = {'age': {'ranges': [{'range': [0, 3], "value": "toddler"}, {'range': [3, 18], "value": "child"}, {'range': [18, 65], "value": "adult"}, {'range': [65, 120], "value": "elder"}]}} # Define one hot encoding values map one_hot_encoder_mapping = {'age_mapped': ['toddler', 'child', 'adult', 'elder']} # Import MLRun's FeatureStore steps for easy # use in our pipeline from mlrun.feature_store.steps import * # Define the pipeline for our FeatureSet graph.to(MapValues(mapping=personal_details, with_original_features=True))\ .to(OneHotEncoder(mapping=one_hot_encoder_mapping))\ .to(Imputer(method='values', default_value=1, mapping={})) # Add default NoSQL & Parquet ingestion targets patients_set.set_targets() # Plot the FeatureSet pipeline patients_set.plot(rankdir='LR', with_targets=True) ###Output _____no_output_____ ###Markdown Test the Feature transformation pipelineCreating a transformation pipeline requires some trial and error. Therefore, it is useful to run the pipeline in memory without storing the resultant data. For this purpose, `infer` is used. This function receives as input any sample DataFrame, performs all the graph steps and outputs the transformed DataFrame. ###Code # Load the sample patient details data patients_df = pd.read_parquet('https://s3.wasabisys.com/iguazio/data/patients/patient_details.parquet') # Run local ingestion test fstore.preview(patients_set, patients_df.head()) ###Output _____no_output_____ ###Markdown Save the FeatureSet and run full ingestion taskOnce you are satisfied with the transformation pipeline, ingest that full DataFrame and store the data. ###Code # Save the FeatureSet patients_set.save() # Run Ingestion task resp = fstore.ingest(patients_set, patients_df, infer_options=fstore.InferOptions.default()) ###Output _____no_output_____ ###Markdown Start Immediate or Scheduled Ingestion Job (over Kubernetes)Another useful method to ingest data, is by creating a Kubernetes job. This may be necessary to process large amounts of data as well as to process any recurring data. With MLRun it is easy to take the pipeline and run it as a job. This is done by:1. Define a source, specifically here we define a parquet file source2. Define a configuration where `local` is set to `False`3. Mount to the provisioned storage by calling `auto_mount`4. Run `ingest` with the source and run configuration ###Code source = ParquetSource('pq', 'https://s3.wasabisys.com/iguazio/data/patients/patient_details.parquet') config = fstore.RunConfig(local=False).apply(mlrun.platforms.auto_mount()) fstore.ingest(patients_set, source, run_config=config) ###Output > 2021-07-12 13:51:19,037 [info] starting run patient_details_ingest uid=5da83655a87c492eaa1065eb2b5ca501 DB=http://mlrun-api:8080 > 2021-07-12 13:51:19,134 [info] Job is running in the background, pod: patient-details-ingest-btft8 > 2021-07-12 13:51:24,554 [info] starting ingestion task to store://feature-sets/fsdemo-iguazio/patient_details:latest > 2021-07-12 13:51:27,785 [info] ingestion task completed, targets: > 2021-07-12 13:51:27,785 [info] [{'name': 'parquet', 'kind': 'parquet', 'path': 'v3io:///projects/fsdemo-iguazio/FeatureStore/patient_details/parquet/sets/patient_details-latest', 'status': 'created', 'updated': '2021-07-12T13:51:26.477162+00:00'}, {'name': 'nosql', 'kind': 'nosql', 'path': 'v3io:///projects/fsdemo-iguazio/FeatureStore/patient_details/nosql/sets/patient_details-latest', 'status': 'created', 'updated': '2021-07-12T13:51:26.477760+00:00'}] > 2021-07-12 13:51:27,811 [info] run executed, status=completed final state: completed ###Markdown Real-time Early-Sense Sensor Ingestion (HTTP or Stream Processing With Nuclio) In this section we will use MLRun to create our Early Sense Sensor datasource. We will do the following:- Create early sense FeatureSet- Add Preprocessing transformations to the Pipeline using custom functions - Drop and Rename columns - Aggregations- Add Feature Validator to detect bad sensor readings- Test the processing pipeline with sample data- Deploy the FeatureSet ingestion service as a live rest endpoint ###Code early_sense_set = fstore.FeatureSet("early_sense", entities=[fstore.Entity("patient_id")], timestamp_key='timestamp', description="real time patient bed sensor data") ###Output _____no_output_____ ###Markdown Define data validation & quality policyWe can define validations on the feature level. For example, define here validation to check if the heart-rate value is between 0 and 220 and respiratory rate is between 0 and 25. ###Code from mlrun.features import MinMaxValidator early_sense_set["hr"] = fstore.Feature(validator = MinMaxValidator(min=0, max=220, severity="info")) early_sense_set["rr"] = fstore.Feature(validator = MinMaxValidator(min=0, max=25, severity="info")) ###Output _____no_output_____ ###Markdown Define custom processing classesIn the previous sections we used transformation steps that are available as part of Storey. Here we show how to create custom transformation classes. We will later run these functions as part of a Nuclio serverless real-time function, therefore, we also use the nuclio `start-code` and `end-code` comments. ###Code # mlrun: start-code # We will import storey here too so it will # be included in our function code (within the nuclio comment block) import json import storey from typing import List, Dict # The custom functions are based on `storey.MapClass` # when they are called in the graph the `do(self, event)` # function will be activated. # A to_dict(self) function is also required by MLRun # to allow the class creation on remote functions class DropColumns(storey.MapClass): def __init__(self, columns: List[str], **kwargs): super().__init__(**kwargs) self.columns = columns def do(self, event): for col in self.columns: if col in event: del event[col] return event def to_dict(self): return { "class_name": "DropColumns", "name": self.name or "DropColumns", "class_args": { "columns": self.columns }, } class RenameColumns(storey.MapClass): def __init__(self, mapping: Dict[str, str], **kwargs): super().__init__(**kwargs) self.mapping = mapping def do(self, event): for old_col, new_col in self.mapping.items(): try: event[new_col] = event.pop(old_col) except Exception as e: print(f'{old_col} doesnt exist') return event def to_dict(self): return { "class_name": "RenameColumns", "name": self.name or "RenameColumns", "class_args": {"mapping": self.mapping}, } # mlrun: end-code ###Output _____no_output_____ ###Markdown Define the Real-Time PipelineDefine the transformation pipeline below. This is done just like the previous sections. ###Code # Configure the list of columns to drop from # the raw data drop_columns = ['hr_is_error', 'rr_is_error', 'spo2_is_error', 'movements_is_error', 'turn_count_is_error', 'is_in_bed_is_error'] # Define the computationala graph including our custom functions early_sense_set.graph.to(DropColumns(drop_columns), after='start')\ .to(RenameColumns(mapping={'bad': 'bed'})) # Add real-time aggreagations on top of our sensor readings for col in ['hr', 'rr', 'spo2', 'movements', 'turn_count']: early_sense_set.add_aggregation(col + "_h", col, ['avg', 'max', 'min'], "1h") early_sense_set.add_aggregation(col + "_d", col, ['avg', 'max', 'min'], "1d") early_sense_set.add_aggregation('in_bed_h', 'is_in_bed', ['avg'], "1h") early_sense_set.add_aggregation('in_bed_d', 'is_in_bed', ['avg'], "1d") # Set NoSQL and Parquet default targets early_sense_set.set_targets() # Plot the pipeline early_sense_set.plot(rankdir='LR', with_targets=True) ###Output _____no_output_____ ###Markdown Test/debug the real-time pipeline locally in the notebook ###Code # infer schema + stats, show the final feature set (after the data pipeline) early_sense_df = pd.read_parquet('https://s3.wasabisys.com/iguazio/data/patients/early_sense_v2.parquet') early_sense_df['timestamp'] = pd.to_datetime(early_sense_df['timestamp']) early_sense_df = move_timestamps(early_sense_df) # update timestamps fstore.preview(early_sense_set, early_sense_df.head()) # Run ingest pipeline df=fstore.ingest(early_sense_set, early_sense_df) # Save the early-sense Featureset early_sense_set.save() # print the FeatureSet spec print(early_sense_set.status.targets.to_dict()) ###Output [{'name': 'parquet', 'kind': 'parquet', 'path': 'v3io:///projects/fsdemo-iguazio/FeatureStore/early_sense/parquet/sets/early_sense-latest', 'status': 'created', 'updated': '2021-07-12T13:52:09.109041+00:00'}, {'name': 'nosql', 'kind': 'nosql', 'path': 'v3io:///projects/fsdemo-iguazio/FeatureStore/early_sense/nosql/sets/early_sense-latest', 'status': 'created', 'updated': '2021-07-12T13:52:09.109597+00:00'}] ###Markdown Deploy as Real-Time Stream Processing Function (Nuclio Serverless)Features are not static. For example, it is common that features include different aggregations that need to be updated as data continues to flow. A real-time pipeline requires this data to be up date. Therefore, we need a convenient way to ingest data, not just as batch, but per specific input.MLRun can convert any code to a real-time serverless function, including the pipeline. This is done by performing the following steps:1. Define a source, in this case it's an HTTP source2. Convert the previously defined code to a serving function3. Create a configuration to run the function4. Deploy an ingestion service with the Featureset, source and the configuration ###Code # Set a new HTTPSource, this will tell our ingestion service # to setup a Nuclio function to act as the rest endpoint # to which we would receive the data source = HttpSource(key_field='patient_id', time_field='timestamp') # Take the relevant code parts from this notebook and create # an MLRun function from them so we can run the pipeline # as a Nuclio function func = mlrun.code_to_function("ingest", kind="serving") nuclio_config = fstore.RunConfig(function=func, local=False).apply(mlrun.platforms.auto_mount()) # Deploy the Online ingestion service using the pipeline definition from before # with our new HTTP Source and our define Function server = fstore.deploy_ingestion_service(early_sense_set, source, run_config=nuclio_config) ###Output > 2021-07-12 13:53:06,632 [info] Starting remote function deploy 2021-07-12 13:53:06 (info) Deploying function 2021-07-12 13:53:06 (info) Building 2021-07-12 13:53:06 (info) Staging files and preparing base images 2021-07-12 13:53:07 (info) Building processor image 2021-07-12 13:53:08 (info) Build complete > 2021-07-12 13:53:16,889 [info] function deployed, address=default-tenant.app.dev65.lab.iguazeng.com:31969 ###Markdown Test the function by sending data to the HTTP endpoint ###Code test_data = {'patient_id': '838-21-8151', 'bad': 38, 'department': '01e9fe31-76de-45f0-9aed-0f94cc97bca0', 'room': 1, 'hr': 220.0, 'hr_is_error': True, 'rr': 5, 'rr_is_error': True, 'spo2': 85, 'spo2_is_error': True, 'movements': 0.0, 'movements_is_error': True, 'turn_count': 0.0, 'turn_count_is_error': True, 'is_in_bed': 1, 'is_in_bed_is_error': False, 'timestamp': 1606843455.906352 } import requests import json response = requests.post(server, json=test_data) response.text ###Output _____no_output_____ ###Markdown Ingest labelsFinally, we define label data, this will be useful in the next notebook where we train a model Create Labels Set ###Code labels_df = pd.read_parquet('https://s3.wasabisys.com/iguazio/data/patients/labels.parquet') labels_df['timestamp'] = pd.to_datetime(labels_df['timestamp']) labels_df = move_timestamps(labels_df) # update timestamps labels_set = fstore.FeatureSet("labels", entities=[fstore.Entity("patient_id")], timestamp_key='timestamp', description="training labels") labels_set.set_targets() df = fstore.preview(labels_set, labels_df) df.head() df = fstore.ingest(labels_set, labels_df) labels_set.save() ###Output _____no_output_____ ###Markdown Part 1: Data Ingestion This demo showcases financial fraud prevention and using the MLRun feature store to define complex features that help identify fraud. Fraud prevention specifically is a challenge as it requires processing raw transaction and events in real-time and being able to quickly respond and block transactions before they occur.To address this, we create a development pipeline and a production pipeline. Both pipelines share the same feature engineering and model code, but serve data very differently. Furthermore, we automate the data and model monitoring process, identify drift and trigger retraining in a CI/CD pipeline. This process is described in the diagram below:![Feature store demo diagram - fraud prevention](../../_static/images/feature_store_demo_diagram.png) The raw data is described as follows:| TRANSACTIONS || &x2551; |USER EVENTS || |-----------------|----------------------------------------------------------------|----------|-----------------|----------------------------------------------------------------|| **age** | age group value 0-6. Some values are marked as U for unknown | &x2551; | **source** | The party/entity related to the event || **gender** | A character to define the age | &x2551; | **event** | event, such as login or password change || **zipcodeOri** | ZIP code of the person originating the transaction | &x2551; | **timestamp** | The date and time of the event || **zipMerchant** | ZIP code of the merchant receiving the transaction | &x2551; | | || **category** | category of the transaction (e.g., transportation, food, etc.) | &x2551; | | || **amount** | the total amount of the transaction | &x2551; | | || **fraud** | whether the transaction is fraudulent | &x2551; | | || **timestamp** | the date and time in which the transaction took place | &x2551; | | || **source** | the ID of the party/entity performing the transaction | &x2551; | | || **target** | the ID of the party/entity receiving the transaction | &x2551; | | || **device** | the device ID used to perform the transaction | &x2551; | | | This notebook introduces how to **Ingest** different data sources to the **Feature Store**.The following FeatureSets will be created:- **Transactions**: Monetary transactions between a source and a target.- **Events**: Account events such as account login or a password change.- **Label**: Fraud label for the data.By the end of this tutorial you’ll learn how to:- Create an ingestion pipeline for each data source.- Define preprocessing, aggregation and validation of the pipeline.- Run the pipeline locally within the notebook.- Launch a real-time function to ingest live data.- Schedule a cron to run the task when needed. ###Code project_name = 'fraud-demo' import mlrun # Initialize the MLRun project object project = mlrun.get_or_create_project(project_name, context="./", user_project=True) ###Output > 2022-01-10 21:27:08,455 [warning] Failed resolving version info. Ignoring and using defaults > 2022-01-10 21:27:11,778 [warning] Server or client version is unstable. Assuming compatible: {'server_version': '0.9.1', 'client_version': '0.0.0+unstable'} > 2022-01-10 21:27:11,813 [info] created and saved project fraud-demo ###Markdown Step 1 - Fetch, Process and Ingest our datasets 1.1 - Transactions Transactions ###Code # Helper functions to adjust the timestamps of our data # while keeping the order of the selected events and # the relative distance from one event to the other def date_adjustment(sample, data_max, new_max, old_data_period, new_data_period): ''' Adjust a specific sample's date according to the original and new time periods ''' sample_dates_scale = ((data_max - sample) / old_data_period) sample_delta = new_data_period * sample_dates_scale new_sample_ts = new_max - sample_delta return new_sample_ts def adjust_data_timespan(dataframe, timestamp_col='timestamp', new_period='2d', new_max_date_str='now'): ''' Adjust the dataframe timestamps to the new time period ''' # Calculate old time period data_min = dataframe.timestamp.min() data_max = dataframe.timestamp.max() old_data_period = data_max-data_min # Set new time period new_time_period = pd.Timedelta(new_period) new_max = pd.Timestamp(new_max_date_str) new_min = new_max-new_time_period new_data_period = new_max-new_min # Apply the timestamp change df = dataframe.copy() df[timestamp_col] = df[timestamp_col].apply(lambda x: date_adjustment(x, data_max, new_max, old_data_period, new_data_period)) return df import pandas as pd # Fetch the transactions dataset from the server transactions_data = pd.read_csv('https://s3.wasabisys.com/iguazio/data/fraud-demo-mlrun-fs-docs/data.csv', parse_dates=['timestamp']) # Adjust the samples timestamp for the past 2 days transactions_data = adjust_data_timespan(transactions_data.sample(50000), new_period='2d') # Preview transactions_data.head(3) ###Output _____no_output_____ ###Markdown Transactions - Create a FeatureSet and Preprocessing PipelineCreate the FeatureSet (data pipeline) definition for the **credit transaction processing** which describes the offline/online data transformations and aggregations.The feature store will automatically add an offline `parquet` target and an online `NoSQL` target by using `set_targets()`.The data pipeline consists of:* **Extracting** the data components (hour, day of week)* **Mapping** the age values* **One hot encoding** for the transaction category and the gender* **Aggregating** the amount (avg, sum, count, max over 2/12/24 hour time windows)* **Aggregating** the transactions per category (over 14 days time windows)* **Writing** the results to **offline** (Parquet) and **online** (NoSQL) targets ###Code # Import MLRun's Feature Store import mlrun.feature_store as fstore from mlrun.feature_store.steps import OneHotEncoder, MapValues, DateExtractor # Define the transactions FeatureSet transaction_set = fstore.FeatureSet("transactions", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="transactions feature set") # Define and add value mapping main_categories = ["es_transportation", "es_health", "es_otherservices", "es_food", "es_hotelservices", "es_barsandrestaurants", "es_tech", "es_sportsandtoys", "es_wellnessandbeauty", "es_hyper", "es_fashion", "es_home", "es_contents", "es_travel", "es_leisure"] # One Hot Encode the newly defined mappings one_hot_encoder_mapping = {'category': main_categories, 'gender': list(transactions_data.gender.unique())} # Define the graph steps transaction_set.graph\ .to(DateExtractor(parts = ['hour', 'day_of_week'], timestamp_col = 'timestamp'))\ .to(MapValues(mapping={'age': {'U': '0'}}, with_original_features=True))\ .to(OneHotEncoder(mapping=one_hot_encoder_mapping)) # Add aggregations for 2, 12, and 24 hour time windows transaction_set.add_aggregation(name='amount', column='amount', operations=['avg','sum', 'count','max'], windows=['2h', '12h', '24h'], period='1h') # Add the category aggregations over a 14 day window for category in main_categories: transaction_set.add_aggregation(name=category,column=f'category_{category}', operations=['count'], windows=['14d'], period='1d') # Add default (offline-parquet & online-nosql) targets transaction_set.set_targets() # Plot the pipeline so we can see the different steps transaction_set.plot(rankdir="LR", with_targets=True) ###Output _____no_output_____ ###Markdown Transactions - Ingestion ###Code # Ingest our transactions dataset through our defined pipeline transactions_df = fstore.ingest(transaction_set, transactions_data, infer_options=fstore.InferOptions.default()) transactions_df.head(3) ###Output _____no_output_____ ###Markdown 1.2 - User Events User Events - Fetching ###Code # Fetch our user_events dataset from the server user_events_data = pd.read_csv('https://s3.wasabisys.com/iguazio/data/fraud-demo-mlrun-fs-docs/events.csv', index_col=0, quotechar="\'", parse_dates=['timestamp']) # Adjust to the last 2 days to see the latest aggregations in our online feature vectors user_events_data = adjust_data_timespan(user_events_data, new_period='2d') # Preview user_events_data.head(3) ###Output _____no_output_____ ###Markdown User Events - Create a FeatureSet and Preprocessing PipelineNow we will define the events feature set.This is a pretty straight forward pipeline in which we only one hot encode the event categories and save the data to the default targets. ###Code user_events_set = fstore.FeatureSet("events", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="user events feature set") # Define and add value mapping events_mapping = {'event': list(user_events_data.event.unique())} # One Hot Encode user_events_set.graph.to(OneHotEncoder(mapping=events_mapping)) # Add default (offline-parquet & online-nosql) targets user_events_set.set_targets() # Plot the pipeline so we can see the different steps user_events_set.plot(rankdir="LR", with_targets=True) ###Output _____no_output_____ ###Markdown User Events - Ingestion ###Code # Ingestion of our newly created events feature set events_df = fstore.ingest(user_events_set, user_events_data) events_df.head(3) ###Output _____no_output_____ ###Markdown Step 2 - Create a labels dataset for model training Label Set - Create a FeatureSetThis feature set contains the label for the fraud demo, it will be ingested directly to the default targets without any changes ###Code def create_labels(df): labels = df[['fraud','source','timestamp']].copy() labels = labels.rename(columns={"fraud": "label"}) labels['timestamp'] = labels['timestamp'].astype("datetime64[ms]") labels['label'] = labels['label'].astype(int) labels.set_index('source', inplace=True) return labels # Define the "labels" feature set labels_set = fstore.FeatureSet("labels", entities=[fstore.Entity("source")], timestamp_key='timestamp', description="training labels", engine="pandas") labels_set.graph.to(name="create_labels", handler=create_labels) # specify only Parquet (offline) target since its not used for real-time labels_set.set_targets(['parquet'], with_defaults=False) labels_set.plot(with_targets=True) ###Output _____no_output_____ ###Markdown Label Set - Ingestion ###Code # Ingest the labels feature set labels_df = fstore.ingest(labels_set, transactions_data) labels_df.head(3) ###Output _____no_output_____ ###Markdown Step 3 - Deploy a real-time pipelineWhen dealing with real-time aggregation, it's important to be able to update these aggregations in real-time.For this purpose, we will create live serving functions that will update the online feature store of the `transactions` FeatureSet and `Events` FeatureSet.Using MLRun's `serving` runtime, craetes a nuclio function loaded with our feature set's computational graph definitionand an `HttpSource` to define the HTTP trigger.Notice that the implementation below does not require any rewrite of the pipeline logic. 3.1 - Transactions Transactions - Deploy our FeatureSet live endpoint ###Code # Create iguazio v3io stream and transactions push API endpoint transaction_stream = f'v3io:///projects/{project.name}/streams/transaction' transaction_pusher = mlrun.datastore.get_stream_pusher(transaction_stream) # Define the source stream trigger (use v3io streams) # we will define the `key` and `time` fields (extracted from the Json message). source = mlrun.datastore.sources.StreamSource(path=transaction_stream , key_field='source', time_field='timestamp') # Deploy the transactions feature set's ingestion service over a real-time (Nuclio) serverless function # you can use the run_config parameter to pass function/service specific configuration transaction_set_endpoint = fstore.deploy_ingestion_service(featureset=transaction_set, source=source) ###Output > 2022-01-10 22:08:17,147 [info] Starting remote function deploy 2022-01-10 22:08:17 (info) Deploying function 2022-01-10 22:08:17 (info) Building 2022-01-10 22:08:18 (info) Staging files and preparing base images 2022-01-10 22:08:18 (info) Building processor image 2022-01-10 22:08:20 (info) Build complete 2022-01-10 22:08:25 (info) Function deploy complete > 2022-01-10 22:08:26,697 [info] successfully deployed function: {'internal_invocation_urls': ['nuclio-fraud-demo-admin-transactions-ingest.default-tenant.svc.cluster.local:8080'], 'external_invocation_urls': ['fraud-demo-admin-transactions-ingest-fraud-demo-admin.default-tenant.app.yh41.iguazio-cd1.com/']} ###Markdown Transactions - Test the feature set HTTP endpoint By defining our `transactions` feature set we can now use MLRun and Storey to deploy it as a live endpoint, ready to ingest new data!Using MLRun's `serving` runtime, we will create a nuclio function loaded with our feature set's computational graph definition and an `HttpSource` to define the HTTP trigger. ###Code import requests import json # Select a sample from the dataset and serialize it to JSON transaction_sample = json.loads(transactions_data.sample(1).to_json(orient='records'))[0] transaction_sample['timestamp'] = str(pd.Timestamp.now()) transaction_sample # Post the sample to the ingestion endpoint requests.post(transaction_set_endpoint, json=transaction_sample).text ###Output _____no_output_____ ###Markdown 3.2 - User Events User Events - Deploy our FeatureSet live endpointDeploy the events feature set's ingestion service using the feature set and all the previously defined resources. ###Code # Create iguazio v3io stream and transactions push API endpoint events_stream = f'v3io:///projects/{project.name}/streams/events' events_pusher = mlrun.datastore.get_stream_pusher(events_stream) # Define the source stream trigger (use v3io streams) # we will define the `key` and `time` fields (extracted from the Json message). source = mlrun.datastore.sources.StreamSource(path=events_stream , key_field='source', time_field='timestamp') # Deploy the transactions feature set's ingestion service over a real-time (Nuclio) serverless function # you can use the run_config parameter to pass function/service specific configuration events_set_endpoint = fstore.deploy_ingestion_service(featureset=user_events_set, source=source) ###Output > 2022-01-10 22:10:02,576 [info] Starting remote function deploy 2022-01-10 22:10:03 (info) Deploying function 2022-01-10 22:10:03 (info) Building 2022-01-10 22:10:04 (info) Staging files and preparing base images 2022-01-10 22:10:04 (info) Building processor image 2022-01-10 22:10:06 (info) Build complete 2022-01-10 22:10:11 (info) Function deploy complete > 2022-01-10 22:10:12,856 [info] successfully deployed function: {'internal_invocation_urls': ['nuclio-fraud-demo-admin-events-ingest.default-tenant.svc.cluster.local:8080'], 'external_invocation_urls': ['fraud-demo-admin-events-ingest-fraud-demo-admin.default-tenant.app.yh41.iguazio-cd1.com/']} ###Markdown User Events - Test the feature set HTTP endpoint ###Code # Select a sample from the events dataset and serialize it to JSON user_events_sample = json.loads(user_events_data.sample(1).to_json(orient='records'))[0] user_events_sample['timestamp'] = str(pd.Timestamp.now()) user_events_sample # Post the sample to the ingestion endpoint requests.post(events_set_endpoint, json=user_events_sample).text ###Output _____no_output_____
Ensemble.ipynb
###Markdown Grab external data ###Code img_ids = [] labels = [] for img in os.listdir('data/Abnormal/'): img_ids.append(os.path.join('Abnormal', img)) labels.append([0, 1, 0, 0, 0]) for img in os.listdir('data/ETIS-LaribPolypDB/'): img_ids.append(os.path.join('ETIS-LaribPolypDB/', img)) labels.append([0, 0, 0, 0, 1]) for img in os.listdir('data/Kvasir-SEG/images/'): img_ids.append(os.path.join('Kvasir-SEG/images/', img)) labels.append([0, 0, 0, 0, 1]) ext_df = pd.concat([pd.Series(img_ids), pd.DataFrame(labels)], 1) ext_df.columns = ['img', 'BE', 'suspicious', 'HGD', 'cancer', 'polyp'] ext_df.to_csv('./data/external_data.csv') ###Output _____no_output_____ ###Markdown Split original data ###Code imgs_dir = "data/originalImages/" masks_dir = "data/masks/" classes = [ "BE", "suspicious", "HGD", "cancer", "polyp", ] img_labels = [] img_ids = [] for img in os.listdir(imgs_dir): img_ids.append(img) img_path = os.path.join(imgs_dir, img) img_label = [] for cls in classes: mask_path = os.path.join(masks_dir, img.replace(".jpg", f"_{cls}.tif")) if os.path.exists(mask_path): img_label.append(1) else: img_label.append(0) img_labels.append(img_label) df = pd.concat([pd.Series(img_ids), pd.DataFrame(img_labels)], axis=1) df.columns = ["img"] + classes for cls in classes: print(f"Class {cls} - num. samples {df[cls].value_counts()[0]}") NUM_FOLDS = 5 SEED = 2709 iterkfold = IterativeStratification(n_splits=5, random_state=SEED) x, y = df.iloc[:, 0].values, df.iloc[:, 1:].values for i, (train, test) in enumerate(iterkfold.split(x, y)): print(x[train].shape, x[test].shape) df.loc[train].to_csv(f"data/train_fold{i}.csv", index=False) df.loc[test].to_csv(f"data/valid_fold{i}.csv", index=False) ###Output _____no_output_____ ###Markdown Search thresholds ###Code seg_threshold = [0.5] * 5 _grid_thresholds = np.linspace(0.1, 0.6, 100) classes = [ "BE", "suspicious", "HGD", "cancer", "polyp" ] models = ['b4_unet', 'b3_unet', 'resnet50_fpn'] def search_threshold(inputs, targets, grid_thresholds=np.linspace(0.1, 0.6, 100), metric_func=f1_score): num_classes = inputs.shape[1] best_cls_thresholds = [] for i in range(num_classes): class_inp = inputs[:, i] class_tar = targets[:, i] grid_scores = [] for thresh in _grid_thresholds: grid_scores.append(metric_func(class_tar, class_inp > thresh)) best_t = grid_thresholds[np.argmax(grid_scores)] best_score = np.max(grid_scores) best_cls_thresholds.append(best_t) return best_cls_thresholds ens_mask_output_list = [] valid_mask_list = [] for seg_weights in [ [1., 0., 0.], [0., 1., 0.], [0., 0., 1.], # [0.5, 0.4, 0.1], # [0.6, 0.3, 0.1], # [0.7, 0.2, 0.1], [1./3, 1./3, 1./3] ]: for i in range(5): valid_mask = torch.load(f'thresholds_tuning/mask_{i}.pth') ens_mask_output = 0 for model, w in zip(models, seg_weights): if model == 'b4_unet': model_output = torch.load(f'thresholds_tuning/{model}_fold{i}.pth') model_output = F.interpolate(model_output, 384, mode='bilinear', align_corners=False) else: model_output = torch.load(f'thresholds_tuning/{model}_fold{i}.pth') ens_mask_output += model_output * w ens_mask_output_list.append(ens_mask_output) valid_mask_list.append(valid_mask) ens_mask_output = torch.cat(ens_mask_output_list, 0) valid_mask = torch.cat(valid_mask_list, 0) dice_score = binary_dice_metric(ens_mask_output, valid_mask, seg_threshold).mean().item() iou = binary_iou_metric(ens_mask_output, valid_mask, seg_threshold).mean().item() print(f'\nWeights: {seg_weights} Dice score: {dice_score} - IoU: {iou}') seg_weights = [0.5, 0.4, 0.1] num_folds = 5 ens_mask_pred = 0 for m, seg_w in zip(models, seg_weights): single_mask_pred = 0 for f in range(num_folds): # folds if m == "b4_unet": single_mask_pred += F.interpolate( torch.load(f'./thresholds_tuning/{m}_fold{f}_test.pth'), 384, align_corners=False, mode='bilinear' ) / num_folds else: single_mask_pred += torch.load(f'./thresholds_tuning/{m}_fold{f}_test.pth') / num_folds ens_mask_pred += single_mask_pred * seg_w ens_mask_pred = torch.where(ens_mask_pred!=0, torch.sigmoid(ens_mask_pred), ens_mask_pred) ens_mask_pred = torch.stack([ ens_mask_pred[:, i, ...] > th for i, th in enumerate(best_seg_thresholds)], 1) ens_mask_pred = ens_mask_pred.float() for out, i, o_sz in zip(ens_mask_pred, img_id, orig_size): out = F.interpolate(out.unsqueeze(0), o_sz, mode="bilinear", align_corners=False) out = out.squeeze(0) out = out.cpu().numpy().astype(np.uint8) * 255 save_path = os.path.join(mask_pred_dir, i.replace(".jpg", ".tif")) tiff.imwrite(save_path, out) ###Output _____no_output_____ ###Markdown TRANQUANGDAT ATOMIC BOMB ###Code best_seg_thresholds = [0.5] * 5 # models = ['rx101-x448', 'rx50-x384-iter-focal'] # seg_weights = [.7, .3] # models = ['rx101-x448', 'rx50-x384-iter-focal', 'rx101-fpn'] # seg_weights = [.5, .3, .2] models = ['rx101-x448', 'rx50-x384-iter-focal', 'rx101-fpn', 'b4-fpn'] seg_weights = [.45, .3, .2, .05] out_dir = 'dattran2346_kfold/' img_id = torch.load(f'{out_dir}test_img_ids.pth') orig_size = torch.load(f'{out_dir}test_sizes.pth') mask_pred_dir = out_dir num_folds = 3 ens_mask_pred = 0 for m, seg_w in zip(models, seg_weights): single_mask_pred = 0 for f in range(num_folds): # folds if m == "rx101-x448" or m == "rx101-fpn" or m == "b4-fpn": single_mask_pred += F.interpolate( torch.load(f'{out_dir}{m}_test_{f}.pth'), 384, align_corners=False, mode='bilinear' ) / num_folds else: single_mask_pred += torch.load(f'{out_dir}{m}_test_{f}.pth') / num_folds ens_mask_pred += single_mask_pred * seg_w ens_mask_pred = torch.where(ens_mask_pred!=0, torch.sigmoid(ens_mask_pred), ens_mask_pred) ens_mask_pred = torch.stack([ ens_mask_pred[:, i, ...] > th for i, th in enumerate(best_seg_thresholds)], 1) ens_mask_pred = ens_mask_pred.float() min_ins_ratio = 0.000927 min_art_ratio = 0.000293 min_sat_ratio = 0.000380 for out, i, o_sz in zip(ens_mask_pred, img_id, orig_size): out = F.interpolate(out.unsqueeze(0), o_sz, mode="bilinear", align_corners=False) out = out.squeeze(0) area = np.prod(out.shape[1:]) instrument_area = out[0].sum() artefact_area = out[2].sum() saturation_area = out[-1].sum() if instrument_area > 0: ins_ratio = instrument_area / area if ins_ratio < min_ins_ratio: # less than min area in training set print('Instrument ', i) out[0] = 0 if artefact_area > 0: art_ratio = artefact_area / area if art_ratio < min_art_ratio: print('Artefact ', i) out[2] = 0 if saturation_area > 0: sat_ratio = saturation_area / area if sat_ratio < min_sat_ratio: print('Saturation ', i) out[-1] = 0 out = out.cpu().numpy().astype(np.uint8) * 255 save_path = os.path.join(mask_pred_dir, i+'.tif') tiff.imwrite(save_path, out) min_ins_ratio = 0.000927 min_art_ratio = 0.000293 min_sat_ratio = 0.000380 for out, i, o_sz in zip(ens_mask_pred, img_id, orig_size): out = F.interpolate(out.unsqueeze(0), o_sz, mode="bilinear", align_corners=False) out = out.squeeze(0) area = np.prod(out.shape[1:]) instrument_area = out[0].sum() artefact_area = out[2].sum() saturation_area = out[-1].sum() if instrument_area > 0: ins_ratio = instrument_area / area if ins_ratio < min_ins_ratio: # less than min area in training set print('Instrument ', i) out[0] = 0 if artefact_area > 0: art_ratio = artefact_area / area if art_ratio < min_art_ratio: print('Artefact ', i) out[2] = 0 if saturation_area > 0: sat_ratio = saturation_area / area if sat_ratio < min_sat_ratio: print('Saturation ', i) out[-1] = 0 out = out.cpu().numpy().astype(np.uint8) * 255 save_path = os.path.join(mask_pred_dir, i+'.tif') tiff.imwrite(save_path, out) ###Output _____no_output_____ ###Markdown Search Segmentation threshold ###Code # best_seg_thresholds = [] # for i in range(5): # 5 classes # cls_out = ens_mask_output[:, i, ...].unsqueeze(1) # cls_mask = valid_mask[:, i, ...].unsqueeze(1) # _grid_dice_scores = [] # _grid_ious = [] # for thresh in _grid_thresholds: # _grid_dice_scores.append(binary_dice_metric(cls_out, cls_mask, thresh).mean().item()) # _grid_ious.append(binary_iou_metric(cls_out, cls_mask, thresh).mean().item()) # best_t = _grid_thresholds[np.argmax(_grid_dice_scores)] # # best_t = _grid_thresholds[np.argmax(_grid_ious)] # best_dice = np.max(_grid_dice_scores) # best_iou = np.max(_grid_ious) # best_seg_thresholds.append(best_t) # # for i in range(5): # for i in range(1): # valid_dice = binary_dice_metric( # ens_mask_output, valid_mask, best_seg_thresholds) # valid_iou = binary_iou_metric( # ens_mask_output, valid_mask, best_seg_thresholds) # print(f'Dice Score - Fold {i}: ', valid_dice.mean(0).mean(0).item()) # print(f'IoU - Fold {i}: ', valid_iou.mean(0).mean(0).item()) ###Output _____no_output_____ ###Markdown Seed 20+300+321 0.83193277 (FL) Seed 50+300+321 0.834033 (FL) Seed 20+300+321 0.731092 (SL) Seed 50+300+321 0.733193 (SL) Seed 20+50+300 0.73949 (SL) ###Code def get_wrong_records(pred, levelFlag): test = pd.read_csv("data/feedback/test_set.csv") test['predLabel']=pred test = test[test[levelFlag] != test['predLabel']] return test get_wrong_records(pred_ensemble_fl, fl) get_wrong_records(pred_ensemble_sl, sl) ###Output _____no_output_____ ###Markdown EnsembleThis notebook completed to ensemble 3 submissions based on 3 xgboost models, and 3 neural networks. ###Code import numpy as np # linear algebra import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv) # Input data files are available in the "../input/" directory. # For example, running this (by clicking run or pressing Shift+Enter) will list the files in the input directory import os # Any results you write to the current directory are saved as output. print("Reading the data...\n") df1 = pd.read_csv('./submission/y_pred1.csv') df2 = pd.read_csv('./submission/y_pred2.csv') df3 = pd.read_csv('./submission/y_pred3.csv') df4 = pd.read_csv('./submission/model_0.csv') df1.head() df3.head() # df3.iloc[:, 1:2] # df4.iloc[:,1:2] # models = { 'df1' :{ 'name':'dnn1', 'score':81.9738, 'df':df1 }, # 'df2' :{ 'name':'dnn2', # 'score':81.9694, # 'df':df2 }, # 'df3' :{ 'name':'dnn3', # 'score':82.0703, # 'df':df3 }, 'df4' :{ 'name':'xgboost', 'score':93.5107, 'df':df4 } } df1.head() isa_lg = 0 isa_hm = 0 isa_am = 0 isa_gm=0 print("Blending...\n") for df in models.keys() : if df == 'df4': isa_lg += np.log(models[df]['df'].pred_prob_0) isa_hm += 1/(models[df]['df'].pred_prob_0) isa_am +=isa_am isa_gm *= isa_gm else: isa_lg += np.log(models[df]['df'][u'0']) isa_hm += 1/(models[df]['df'][u'0']) isa_am +=isa_am isa_gm *= isa_gm isa_lg = np.exp(isa_lg/5) isa_hm = 5/isa_hm isa_am = isa_am/5 isa_gm = (isa_gm)**(1/5) print("Isa log\n") print(isa_lg[:5]) print() print("Isa harmo\n") print(isa_hm[:5]) # sub_log = pd.DataFrame() # sub_log['click_id'] = df1['click_id'] # sub_log['is_attributed'] = isa_lg # sub_log.head() # sub_hm = pd.DataFrame() # sub_hm['click_id'] = df1['click_id'] # sub_hm['is_attributed'] = isa_hm # sub_hm.head() sub_fin=pd.DataFrame() #sub_fin['click_id']=df1['click_id'] sub_fin['is_attributed']= (5*isa_lg+3*isa_hm+2*isa_am)/10 print("Writing...") # sub_log.to_csv('submission_log2.csv', index=False, float_format='%.9f') # sub_hm.to_csv('submission_hm2.csv', index=False, float_format='%.9f') sub_fin.to_csv('submission_esb_1x_1n.csv', index=False, float_format='%.9f') #sub_fin.to_csv('submission_esb_3n.csv', index=False, float_format='%.9f') print("DONE!") sub_fin.head() ###Output _____no_output_____ ###Markdown Ensemble ###Code from __future__ import division from IPython.display import display from matplotlib import pyplot as plt %matplotlib inline import numpy as np import pandas as pd import random, sys, os, re ###Output _____no_output_____ ###Markdown The test set has duplicates so we get the list of IDs in the sample file in order ###Code id_list = [] with open('../submissions/Submission_Format.csv', 'r') as f: lines = f.read().splitlines() for line in lines: ID,prob = line.split(',') if ID == '': continue id_list.append(ID) def get_filepaths(directory): """ This function will generate the file names in a directory tree by walking the tree either top-down or bottom-up. For each directory in the tree rooted at directory top (including top itself), it yields a 3-tuple list (dirpath, dirnames, filenames). """ import os file_paths = [] # List which will store all of the full filepaths. # Walk the tree. for root, directories, files in os.walk(directory): for filename in files: # Join the two strings in order to form the full filepath. filepath = os.path.join(root, filename) file_paths.append(filepath) # Add it to the list. return file_paths ###Output _____no_output_____ ###Markdown Get the list of submission files * remove the example file * and all ensembles BEFORE ###Code file_list = get_filepaths('../submissions') file_list ###Output _____no_output_____ ###Markdown AFTER ###Code # why do it more than once? For some reason it doesn't work if only run once. Who knows? # ====================================================================================== for i in range(3): for file_name in file_list: if 'Format' in file_name: file_list.remove(file_name) if 'Ensemble' in file_name: file_list.remove(file_name) if 'ensemble' in file_name: file_list.remove(file_name) file_list.sort(key=lambda x: x[26:32]) from copy import copy file_list_all = copy(file_list) file_list ###Output _____no_output_____ ###Markdown --------------------------------------------- Ensemble ALL the submissions --------------------------------------------- Find the average probability for all IDs ###Code from collections import defaultdict aggregates = defaultdict(list) averages = defaultdict(list) # 1. collect the probabilities for each ID from all the submission files # ====================================================================== for file_name in file_list: with open(file_name, 'r') as f: lines = f.read().splitlines() for line in lines: ID,prob = line.split(',') if ID == '': continue aggregates[ID].append(prob) # 2. find the average of all the probabilities for each ID # ======================================================== averages.update((ID, np.mean(map(float, probs))) for ID, probs in aggregates.items()) aggregates['1'],averages['1'] len(aggregates),len(averages) ###Output _____no_output_____ ###Markdown Create a submission file of the ensemble of averages ###Code # f = open("../submissions/submission_EnsembleOfAveragesALL.csv", "w") # f.write(",Made Donation in March 2007\n") # for ID in id_list: # f.write("{},{}\n".format(ID, averages[ID])) # f.close() ###Output _____no_output_____ ###Markdown --------------------------------------------------------------- Ensemble the submissions with high scores --------------------------------------------------------------- BEFORE ###Code file_list ###Output _____no_output_____ ###Markdown AFTER ###Code # why do it more than once? For some reason it doesn't work if only run once. Who knows? # ====================================================================================== for _ in range(2): for _ in range(4): for file_name in file_list: if 'Format' in file_name: file_list.remove(file_name) if 'Ensemble' in file_name: file_list.remove(file_name) # scores of 0.4... or 0.3... are good # files with SEED... are good-scoring models that were re-run with different random seeds if ('bagged_nolearn' not in file_name): file_list.remove(file_name) file_list from collections import defaultdict aggregates = defaultdict(list) averages = defaultdict(list) # 1. collect the probabilities for each ID from all the submission files # ====================================================================== for file_name in file_list: with open(file_name, 'r') as f: lines = f.read().splitlines() for line in lines: ID,prob = line.split(',') if ID == '': continue aggregates[ID].append(prob) # 2. find the average of all the probabilities for each ID # ======================================================== averages.update((ID, np.mean(map(float, probs))) for ID, probs in aggregates.items()) aggregates['1'],averages['1'] len(aggregates),len(averages) f = open("../submissions/submission_EnsembleOfAveragesBEST_SEED.csv", "w") f.write(",Made Donation in March 2007\n") for ID in id_list: f.write("{},{}\n".format(ID, averages[ID])) f.close() ###Output _____no_output_____ ###Markdown --------------------------------------------------------------- Ensemble the least-correlated submissions --------------------------------------------------------------- Create a dataframe with one column per submission ###Code from os.path import split corr_table = pd.read_csv(file_list_all[0],names=['id',split(file_list_all[0])[1][11:-4]],header=0,index_col=0) corr_table.head() for file_path in file_list_all[1:]: temp = pd.read_csv(file_path,names=['id',split(file_path)[1][11:-4]],header=0,index_col=0) corr_table[temp.columns[0]] = temp[[temp.columns[0]]] corr_table.head() ###Output _____no_output_____ ###Markdown Display the correlations among the submissions ###Code import seaborn as sns # Compute the correlation matrix corr_matrix = corr_table.corr() # Generate a mask for the upper triangle mask = np.zeros_like(corr_matrix, dtype=np.bool) mask[np.triu_indices_from(mask)] = True # Set up the matplotlib figure f, ax = plt.subplots(figsize=(11, 9)) # Generate a custom diverging colormap cmap = sns.diverging_palette(220, 10, as_cmap=True) # Draw the heatmap with the mask and correct aspect ratio sns.heatmap(corr_matrix, mask=mask, cmap=cmap, vmax=.9, square=True, xticklabels=4, yticklabels=3, linewidths=.5, cbar_kws={"shrink": .5}, ax=ax) plt.show() ###Output _____no_output_____ ###Markdown Find the least-correlated pairs of submissions ###Code corr_threshold = 0.20 indices = np.where(corr_matrix < corr_threshold) indices = [(corr_matrix.index[x], corr_matrix.columns[y], corr_matrix.ix[x,y]) for x, y in zip(*indices) if x != y and x < y] from operator import itemgetter indices.sort(key=itemgetter(2)) len(indices),indices least_corr = set(set(['../submissions/submission_'+a+'.csv' for a,b,c in indices]).\ union(set(['../submissions/submission_'+b+'.csv' for a,b,c in indices]))) len(least_corr), least_corr from collections import defaultdict aggregates = defaultdict(list) averages = defaultdict(list) # 1. collect the probabilities for each ID from all the submission files # ====================================================================== for file_name in least_corr: with open(file_name, 'r') as f: lines = f.read().splitlines() for line in lines: ID,prob = line.split(',') if ID == '': continue aggregates[ID].append(prob) # 2. find the average of all the probabilities for each ID # ======================================================== averages.update((ID, np.mean(map(float, probs))) for ID, probs in aggregates.items()) aggregates['1'],averages['1'] # f = open("../submissions/submission_EnsembleOfAveragesLeastCorr.csv", "w") # f.write(",Made Donation in March 2007\n") # for ID in id_list: # f.write("{},{}\n".format(ID, averages[ID])) # f.close() ###Output _____no_output_____ ###Markdown Machine Learning Models and Ensemble Method---1. Split X-Features and y-labels2. 80-20 train-validation split3. fit models (on train), evaluate (on validation): 1. DNN 2. SVM 3. RF 4. XGB 5. LogReg4. __Manual__ Ensemble Method: 1. Evaluate Ensemble on Validation data5. Evaluate Ensemble on TEST data ###Code import os os.environ["CUDA_VISIBLE_DEVICES"] = "-1" # disable GPU from tqdm import tqdm # progress bar import pandas as pd import numpy as np import matplotlib.pyplot as plt import seaborn as sns %matplotlib inline # processing / validation from sklearn.model_selection import train_test_split from sklearn.preprocessing import MinMaxScaler # keras/tf import tensorflow as tf from tensorflow.keras.models import Sequential from tensorflow.keras.layers import Dense, Activation,Dropout # models from sklearn.svm import SVC from sklearn.ensemble import RandomForestRegressor from xgboost import XGBClassifier from sklearn.linear_model import LogisticRegression #from sklearn.model_selection import GridSearchCV # hp-tuning # metrics from sklearn import metrics from sklearn.metrics import classification_report from sklearn.metrics import roc_auc_score # constant seed for reproducibility SEED = 111 os.environ['PYTHONHASHSEED'] = str(SEED) np.random.seed(SEED) tf.random.set_seed(SEED) ###Output /Users/rezanaghshineh/opt/anaconda3/lib/python3.7/site-packages/statsmodels/tools/_testing.py:19: FutureWarning: pandas.util.testing is deprecated. Use the functions in the public API at pandas.testing instead. import pandas.util.testing as tm ###Markdown 1 & 2 - X/y + Train/Test Split: ###Code df = pd.read_csv("data/UFC_TRAIN.csv") # tackling imbalance issue #theMin = df["Winner"].value_counts().min() #minority = df[df["Winner"]==1].iloc[0:theMin] #undersampleMaj = df[df["Winner"]==0].iloc[0:theMin] #df = pd.concat([minority, undersampleMaj], axis=0) #df["Winner"].value_counts() # feature/label and train/test split X = df.drop(["date","Winner","B_fighter","R_fighter"], axis=1).values y = df["Winner"].values X_TRAIN, X_VAL, y_TRAIN, y_VAL = train_test_split(X,y, test_size=0.20, random_state=SEED) ###Output _____no_output_____ ###Markdown Baseline: Always predict red (i.e: 0) ###Code metrics.accuracy_score(np.zeros(len(df.index)),df["Winner"]) ###Output _____no_output_____ ###Markdown Baseline accuracy is 67.96 % in unbalanced dataset 3- ML Models A: DNN - Using a deep neural network with early stopping functionality to prevent divergence of loss & val_loss: ###Code # scaling scaler = MinMaxScaler() scaler.fit(X_TRAIN) X_train_scaled = scaler.transform(X_TRAIN) X_val_scaled = scaler.transform(X_VAL) print(f"X_train_scaled shape: {X_train_scaled.shape} | X_val_scaled shape: {X_val_scaled.shape} | y_train shape: {y_TRAIN.shape} | y_val shape: {y_VAL.shape}") # model dnnClf = Sequential() # first hiden layer dnnClf.add(Dense(units=20, input_dim=42,activation='relu')) #dnnClf.add(Dropout(0.5)) # deactivates 50% of nodes dnnClf.add(Dense(units=10, activation='relu')) dnnClf.add(Dropout(0.5)) # deactivates 50% of nodes # output layer dnnClf.add(Dense(units=1, activation='sigmoid')) dnnClf.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy']) from tensorflow.keras.callbacks import EarlyStopping # prevent divergence of loss & val_loss early_stop = EarlyStopping(monitor='val_loss', mode='min', verbose=1, patience=16) dnnClf.fit(x=X_train_scaled, y=y_TRAIN, epochs=400, validation_data=(X_val_scaled, y_VAL), verbose=1, callbacks=[early_stop] ) model_loss = pd.DataFrame(dnnClf.history.history) model_loss.plot() dnnPreds = dnnClf.predict(scaler.transform(X_VAL)) dnnPreds = [round(i[0]) for i in dnnPreds] target_names = ['class 0', 'class 1'] print("DNN Validation Performance on UNBALANCED(!):\n------------------\n",classification_report(y_VAL, dnnPreds , target_names=target_names)) print("AUC: ",roc_auc_score(y_VAL, dnnPreds)) ###Output DNN Validation Performance on UNBALANCED(!): ------------------ precision recall f1-score support class 0 0.71 0.91 0.80 590 class 1 0.58 0.25 0.35 291 accuracy 0.69 881 macro avg 0.64 0.58 0.57 881 weighted avg 0.67 0.69 0.65 881 AUC: 0.5787960859688974 ###Markdown B: SVM - Support Vector Machine: ###Code svmClf = SVC(kernel="linear") svmClf.fit(X_TRAIN,y_TRAIN) svmPreds = svmClf.predict(X_VAL) print("SVM Validation Performance on UNBALANCED(!):\n------------------\n",classification_report(y_VAL, svmPreds , target_names=target_names)) print("AUC: ",roc_auc_score(y_VAL, svmPreds)) ###Output SVM Validation Performance on UNBALANCED(!): ------------------ precision recall f1-score support class 0 0.68 0.99 0.81 590 class 1 0.71 0.05 0.10 291 accuracy 0.68 881 macro avg 0.70 0.52 0.45 881 weighted avg 0.69 0.68 0.57 881 AUC: 0.5206884501135768 ###Markdown C: RF - RandomForest: ###Code rfClf = RandomForestRegressor(n_estimators = 1000) rfClf.fit(X_TRAIN, y_TRAIN) rfPreds = rfClf.predict(X_VAL) rfPreds = [round(i) for i in rfPreds] print("RF Validation Performance on UNBALANCED(!):\n------------------\n",classification_report(y_VAL, rfPreds , target_names=target_names)) print("AUC: ",roc_auc_score(y_VAL, rfPreds)) ###Output RF Validation Performance on UNBALANCED(!): ------------------ precision recall f1-score support class 0 0.71 0.89 0.79 590 class 1 0.55 0.26 0.36 291 accuracy 0.68 881 macro avg 0.63 0.58 0.57 881 weighted avg 0.66 0.68 0.65 881 AUC: 0.5780651173626886 ###Markdown D: XGB - Gradient Boost: ###Code xgbClf = XGBClassifier(n_estimators=200) xgbClf.fit(X_TRAIN, y_TRAIN) xgbPreds = xgbClf.predict(X_VAL) print("XGB Validation Performance on UNBALANCED(!):\n------------------\n",classification_report(y_VAL, xgbPreds , target_names=target_names)) print("AUC: ",roc_auc_score(y_VAL, xgbPreds)) ###Output XGB Validation Performance on UNBALANCED(!): ------------------ precision recall f1-score support class 0 0.71 0.91 0.79 590 class 1 0.56 0.23 0.33 291 accuracy 0.69 881 macro avg 0.63 0.57 0.56 881 weighted avg 0.66 0.69 0.64 881 AUC: 0.5702050206768011 ###Markdown E: LR - Logistic Regression: ###Code lrClf = LogisticRegression(solver="newton-cg") lrClf.fit(X_TRAIN, y_TRAIN) lrPreds = lrClf.predict(X_VAL) print("LogReg Validation Performance on UNBALANCED(!):\n------------------\n",classification_report(y_VAL, lrPreds , target_names=target_names)) print("AUC: ",roc_auc_score(y_VAL, lrPreds)) ###Output LogReg Validation Performance on UNBALANCED(!): ------------------ precision recall f1-score support class 0 0.70 0.92 0.80 590 class 1 0.56 0.20 0.29 291 accuracy 0.68 881 macro avg 0.63 0.56 0.54 881 weighted avg 0.65 0.68 0.63 881 AUC: 0.5606733065408586 ###Markdown 4- Ensemble Method with Validation Performance:Ensemble method aggregates the votes of each model and gives the most frequent vote as output ###Code def predictEnsemble(sample, models=0): """predicts the label of a given sample by aggregating votes of number of models. by default, models = 0, takes into account all models. Otherwise, for a given list of codes, it involves the corresponsing model. codes: 1: dnn | 2: svm | 3: rf | 4: xgb | 5: lr """ modelsDict = { # models predictions dictionary 1:dnnClf.predict(scaler.transform(sample.reshape(1,-1))).tolist()[0][0], 2:svmClf.predict(sample.reshape(1,-1)).tolist()[0], 3:rfClf.predict(sample.reshape(1,-1)).tolist()[0], 4:xgbClf.predict(sample.reshape(1,-1)).tolist()[0], 5:lrClf.predict(sample.reshape(1,-1)).tolist()[0] } preds = [] if models == 0: # use all models [preds.append(model) for model in modelsDict.values()] else: # use only specified models for model_code in models: preds.append(modelsDict[model_code]) #print(preds) preds = [round(i) for i in preds] # transform probability to label (threshold 0.5) #print(preds) #print(max(set(preds), key=preds.count)) return(max(set(preds), key=preds.count)) ensPreds = [] [ensPreds.append(predictEnsemble(sample, models=[3,4,5])) for sample in tqdm(X_VAL)] print("Ensemble Validation Performance on UNBALANCED(!):\n------------------\n",classification_report(y_VAL, ensPreds , target_names=target_names)) print("AUC: ",roc_auc_score(y_VAL, ensPreds)) ###Output 100%|██████████| 881/881 [02:23<00:00, 6.13it/s] ###Markdown 5- Performance Evaluation on TEST (unseen data) ###Code TEST = pd.read_csv("data/UFC_TEST.csv") X_TEST = TEST.drop(["date","B_fighter","R_fighter","Winner"],axis=1).values y_TEST = TEST["Winner"].values ensPreds_TEST = [] [ensPreds_TEST.append(predictEnsemble(test_sample, models=[3,4,5])) for test_sample in tqdm(X_TEST)] print("Ensemble TEST Performance on UNBALANCED(!):\n------------------\n",classification_report(y_TEST, ensPreds_TEST , target_names=target_names)) print("AUC: ",roc_auc_score(y_TEST, ensPreds_TEST)) # save models to disk #import pickle #dnnClf.save('resources/dnn_model.h5') #pickle.dump(svmClf, open('resources/svm_model.sav', 'wb')) #pickle.dump(rfClf, open('resources/rf_model.sav', 'wb')) #pickle.dump(xgbClf, open('resources/xgb_model.sav', 'wb')) #pickle.dump(lrClf, open('resources/lr_model.sav', 'wb')) #pickle.dump(scaler, open('resources/scaler.pkl', 'wb')) # notes: ''' dnnPreds2 = dnnClf.predict(scaler.transform(X_TEST)) dnnPreds2 = [round(i[0]) for i in dnnPreds2] dnnAcc2 = metrics.accuracy_score(dnnPreds2, y_TEST) print("DNN Accuracy:",round(dnnAcc2,3)) svmPreds2 = svmClf.predict(X_TEST) svmAcc2 = metrics.accuracy_score(svmPreds2, y_TEST) print("SVM Accuracy:",round(svmAcc2,3)) rfPreds2 = rfClf.predict(X_TEST) rfPreds2 = [round(i) for i in rfPreds2] rfAcc2 = metrics.accuracy_score(rfPreds2, y_TEST) print("RF Accuracy:",round(rfAcc2,3)) xgbPreds2 = xgbClf.predict(X_TEST) xgbAcc2 = metrics.accuracy_score(xgbPreds2, y_TEST) print("XGB Accuracy:",round(xgbAcc2,3)) lrPreds2 = lrClf.predict(X_TEST) lrAcc2 = metrics.accuracy_score(lrPreds2, y_TEST) print("LogReg Accuracy:",round(lrAcc2,3)) accTable = pd.DataFrame({"Model":["DNN", "SVM", "RF", "XGB", "LogReg", "Ensemble"], "Val_Accuracy":[dnnAcc, svmAcc, rfAcc, xgbAcc, lrAcc, ensAcc], "Test_Accuracy":[dnnAcc2, svmAcc2, rfAcc2, xgbAcc2, lrAcc2, ensAcc2]}) accTable.plot(kind="bar",ylim=(0.5,0.8),x="Model",title="Models Performance on Validation and Test Data") # grid-search hyper-parameter tuning # svm hp-tuning with gridSearch #svm_param = {"kernel":("linear","poly","rbf", "sigmoid"), # "C":[1,52,10], # "degree":[3,8], # "gamma":("auto","scale"), # "coef0":[0.001,10,0.5]} #svmClf = SVC() #svmGrid = GridSearchCV(svmClf, svm_param,cv=2) #svmGrid.fit(X_TRAIN, y_TRAIN) ''' ###Output _____no_output_____ ###Markdown > This notebook aims to push the public LB under 0.50. Certainly, the competition is not yet at its peak and there clearly remains room for improvement. Credits and comments on changesThis notebook is based on [m5-first-public-notebook-under-0-50](https://www.kaggle.com/kneroma/m5-first-public-notebook-under-0-50) v.6 by @kkiller Presently it's sole purpose is to test accelerated prediction stage (vs original notebook) where I generate lag features only for the days that need sales forecasts. Everything else is unchanged vs the original _kkiller's_ notebook (as in version 6). ###Code CAL_DTYPES={"event_name_1": "category", "event_name_2": "category", "event_type_1": "category", "event_type_2": "category", "weekday": "category", 'wm_yr_wk': 'int16', "wday": "int16", "month": "int16", "year": "int16", "snap_CA": "float32", 'snap_TX': 'float32', 'snap_WI': 'float32' } PRICE_DTYPES = {"store_id": "category", "item_id": "category", "wm_yr_wk": "int16","sell_price":"float32" } pd.options.display.max_columns = 50 h = 28 max_lags = 57 tr_last = 1913 fday = datetime(2016,4, 25) fday def create_dt(is_train = True, nrows = None, first_day = 1200): prices = pd.read_csv("../input/m5-forecasting-accuracy/sell_prices.csv", dtype = PRICE_DTYPES) for col, col_dtype in PRICE_DTYPES.items(): if col_dtype == "category": prices[col] = prices[col].cat.codes.astype("int16") prices[col] -= prices[col].min() cal = pd.read_csv("../input/m5-forecasting-accuracy/calendar.csv", dtype = CAL_DTYPES) cal["date"] = pd.to_datetime(cal["date"]) for col, col_dtype in CAL_DTYPES.items(): if col_dtype == "category": cal[col] = cal[col].cat.codes.astype("int16") cal[col] -= cal[col].min() start_day = max(1 if is_train else tr_last-max_lags, first_day) numcols = [f"d_{day}" for day in range(start_day,tr_last+1)] catcols = ['id', 'item_id', 'dept_id','store_id', 'cat_id', 'state_id'] dtype = {numcol:"float32" for numcol in numcols} dtype.update({col: "category" for col in catcols if col != "id"}) dt = pd.read_csv("../input/m5-forecasting-accuracy/sales_train_validation.csv", nrows = nrows, usecols = catcols + numcols, dtype = dtype) for col in catcols: if col != "id": dt[col] = dt[col].cat.codes.astype("int16") dt[col] -= dt[col].min() if not is_train: for day in range(tr_last+1, tr_last+ 28 +1): dt[f"d_{day}"] = np.nan dt = pd.melt(dt, id_vars = catcols, value_vars = [col for col in dt.columns if col.startswith("d_")], var_name = "d", value_name = "sales") dt = dt.merge(cal, on= "d", copy = False) dt = dt.merge(prices, on = ["store_id", "item_id", "wm_yr_wk"], copy = False) return dt def create_fea(dt): lags = [7, 28] lag_cols = [f"lag_{lag}" for lag in lags ] for lag, lag_col in zip(lags, lag_cols): dt[lag_col] = dt[["id","sales"]].groupby("id")["sales"].shift(lag) wins = [7, 28] for win in wins : for lag,lag_col in zip(lags, lag_cols): dt[f"rmean_{lag}_{win}"] = dt[["id", lag_col]].groupby("id")[lag_col].transform(lambda x : x.rolling(win).mean()) date_features = { "wday": "weekday", "week": "weekofyear", "month": "month", "quarter": "quarter", "year": "year", "mday": "day", # "ime": "is_month_end", # "ims": "is_month_start", } # dt.drop(["d", "wm_yr_wk", "weekday"], axis=1, inplace = True) for date_feat_name, date_feat_func in date_features.items(): if date_feat_name in dt.columns: dt[date_feat_name] = dt[date_feat_name].astype("int16") else: dt[date_feat_name] = getattr(dt["date"].dt, date_feat_func).astype("int16") FIRST_DAY = 350 # If you want to load all the data set it to '1' --> Great memory overflow risk ! %%time df = create_dt(is_train=True, first_day= FIRST_DAY) df.shape df.head() df.info() %%time create_fea(df) df.shape df.info() df.head() df.dropna(inplace = True) df.shape cat_feats = ['item_id', 'dept_id','store_id', 'cat_id', 'state_id'] + ["event_name_1", "event_name_2", "event_type_1", "event_type_2"] useless_cols = ["id", "date", "sales","d", "wm_yr_wk", "weekday"] train_cols = df.columns[~df.columns.isin(useless_cols)] X_train = df[train_cols] y_train = df["sales"] %%time np.random.seed(777) fake_valid_inds = np.random.choice(X_train.index.values, 2_000_000, replace = False) train_inds = np.setdiff1d(X_train.index.values, fake_valid_inds) X_train=X_train.loc[fake_valid_inds] y_train = y_train.loc[fake_valid_inds] #del df, X_train, y_train, fake_valid_inds,train_inds ; gc.collect() from sklearn.pipeline import make_pipeline from sklearn.preprocessing import RobustScaler from sklearn.model_selection import KFold, cross_val_score from sklearn.metrics import mean_squared_error from sklearn.preprocessing import LabelEncoder #ML Algoirthm from sklearn.linear_model import ElasticNetCV, LassoCV, RidgeCV import sklearn.linear_model as linear_model from sklearn.svm import SVR from lightgbm import LGBMRegressor from sklearn.ensemble import GradientBoostingRegressor,RandomForestRegressor from xgboost import XGBRegressor from mlxtend.regressor import StackingCVRegressor kf = KFold(n_splits=12, random_state=42, shuffle=True) # Define error metrics def cv_rmse(model, X=X_train): rmse = np.sqrt(-cross_val_score(model, X, y_train, scoring="neg_mean_squared_error", cv=kf)) return (rmse) from xgboost import XGBRegressor from lightgbm import LGBMRegressor from mlxtend.regressor import StackingCVRegressor ridge_alphas = [1e-15, 1e-10, 1e-8, 9e-4, 7e-4, 5e-4, 3e-4, 1e-4, 1e-3, 5e-2, 1e-2, 0.1, 0.3, 1, 3, 5, 10, 15, 18, 20, 30, 50, 75, 100] ridge = make_pipeline(RobustScaler(), RidgeCV(alphas=ridge_alphas, cv=kf)) # Support Vector Regressor #svr = make_pipeline(RobustScaler(), SVR(C= 5, epsilon= 0.008, gamma=0.0003)) # Gradient Boosting Regressor gbr = GradientBoostingRegressor(n_estimators=100, learning_rate=0.075) rf=RandomForestRegressor(n_estimators=10) lightgbm1 = LGBMRegressor(objective='poisson', metric ='rmse', learning_rate = 0.075, sub_row = 0.75, bagging_freq = 1, lambda_l2 = 0.1, verbosity= 1, n_estimators = 200, num_leaves= 128, min_data_in_leaf= 100) lightgbm2 = LGBMRegressor(objective='tweedie', metric ='rmse', learning_rate = 0.075, sub_row = 0.75, bagging_freq = 1, lambda_l2 = 0.1, verbosity= 1, n_estimators = 200, num_leaves= 128, min_data_in_leaf= 100) xgboost = XGBRegressor(objective='count:poisson', learning_rate=0.075, n_estimators=100, min_child_weight=50) stackReg = StackingCVRegressor(regressors=(lightgbm1,lightgbm2), meta_regressor=(xgboost), use_features_in_secondary=True, random_state=42) model_score = {} score = cv_rmse(lightgbm1) lgb_model1_full_data = lightgbm1.fit(X_train, y_train) print("lightgbm1: {:.4f}".format(score.mean())) model_score['lgb1'] = score.mean() score = cv_rmse(lightgbm2) lgb_model2_full_data = lightgbm2.fit(X_train, y_train) print("lightgbm2: {:.4f}".format(score.mean())) model_score['lgb2'] = score.mean() score = cv_rmse(xgboost) xgboost_full_data = xgboost.fit(X_train, y_train) print("xgboost: {:.4f}".format(score.mean())) model_score['xgb'] = score.mean() score = cv_rmse(ridge) ridge_full_data = ridge.fit(X_train, y_train) print("ridge: {:.4f}".format(score.mean())) model_score['ridge'] = score.mean() # score = cv_rmse(svr) # svr_full_data = svr.fit(X_train, y_train) # print("svr: {:.4f}".format(score.mean())) # model_score['svr'] = score.mean() score = cv_rmse(gbr) gbr_full_data = gbr.fit(X_train, y_train) print("gbr: {:.4f}".format(score.mean())) model_score['gbr'] = score.mean() score = cv_rmse(rf) rf_full_data = rf.fit(X_train, y_train) print("rf: {:.4f}".format(score.mean())) model_score['rf'] = score.mean() score = cv_rmse(stackReg) stackReg_full_data = stackReg.fit(X_train, y_train) print("stackReg: {:.4f}".format(score.mean())) model_score['stackReg'] = score.mean() def rmsle(y, y_pred): return np.sqrt(mean_squared_error(y, y_pred)) def blended_predictions(X_train,weight): return ((weight[0] * ridge_full_data.predict(X_train)) + \ (weight[1] * rf_full_data.predict(X_train)) + \ (weight[2] * gbr_full_data.predict(X_train)) + \ (weight[3] * xgboost_full_data.predict(X_train)) + \ (weight[4] * lgb_model1_full_data.predict(X_train)) + \ (weight[5] * stackReg_full_data.predict(np.array(X_train)))) # Blended model predictions blended_score = rmsle(y_train, blended_predictions(X_train,[0.15,0.2,0.18,0.1,0.27,0.1])) print("blended score: {:.4f}".format(blended_score)) model_score['blended_model'] = blended_score model_score #my_model = stacked_ensemble(X_train,y_train) import warnings warnings.filterwarnings("default") # %%time # blend= blended_predictions(X_train,[0.15,0.2,0.1,0.18,0.1,0.27]) ###Output _____no_output_____ ###Markdown Prediction stage(updated vs original) ###Code def create_lag_features_for_test(dt, day): # create lag feaures just for single day (faster) lags = [7, 28] lag_cols = [f"lag_{lag}" for lag in lags] for lag, lag_col in zip(lags, lag_cols): dt.loc[dt.date == day, lag_col] = \ dt.loc[dt.date ==day-timedelta(days=lag), 'sales'].values # !!! main windows = [7, 28] for window in windows: for lag in lags: df_window = dt[(dt.date <= day-timedelta(days=lag)) & (dt.date > day-timedelta(days=lag+window))] df_window_grouped = df_window.groupby("id").agg({'sales':'mean'}).reindex(dt.loc[dt.date==day,'id']) dt.loc[dt.date == day,f"rmean_{lag}_{window}"] = \ df_window_grouped.sales.values def create_date_features_for_test(dt): # copy of the code from `create_dt()` above date_features = { "wday": "weekday", "week": "weekofyear", "month": "month", "quarter": "quarter", "year": "year", "mday": "day", } for date_feat_name, date_feat_func in date_features.items(): if date_feat_name in dt.columns: dt[date_feat_name] = dt[date_feat_name].astype("int16") else: dt[date_feat_name] = getattr( dt["date"].dt, date_feat_func).astype("int16") %%time alphas = [1.028, 1.023, 1.018] weights = [1/len(alphas)]*len(alphas) # equal weights te0 = create_dt(False) # create master copy of `te` create_date_features_for_test (te0) for icount, (alpha, weight) in enumerate(zip(alphas, weights)): te = te0.copy() # just copy # te1 = te0.copy() cols = [f"F{i}" for i in range(1, 29)] for tdelta in range(0, 28): day = fday + timedelta(days=tdelta) print(tdelta, day.date()) tst = te[(te.date >= day - timedelta(days=max_lags)) & (te.date <= day)].copy() # tst1 = te1[(te1.date >= day - timedelta(days=max_lags)) # & (te1.date <= day)].copy() # create_fea(tst) # correct, but takes much time create_lag_features_for_test(tst, day) # faster tst = tst.loc[tst.date == day, train_cols] te.loc[te.date == day, "sales"] = \ alpha * blended_predictions(tst,[0.15,0.2,0.18,0.1,0.27,0.1]) # magic multiplier by kyakovlev # create_lag_features_for_test(tst1, day) # faster # tst1 = tst1.loc[tst1.date == day, train_cols] # te1.loc[te1.date == day, "sales"] = \ # alpha * m_lgb1.predict(tst1) # magic multiplier by kyakovlev te_sub = te.loc[te.date >= fday, ["id", "sales"]].copy() # te_sub1 = te1.loc[te1.date >= fday, ["id", "sales"]].copy() te_sub["F"] = [f"F{rank}" for rank in te_sub.groupby("id")[ "id"].cumcount()+1] # te_sub1["F"] = [f"F{rank}" for rank in te_sub1.groupby("id")[ # "id"].cumcount()+1] te_sub = te_sub.set_index(["id", "F"]).unstack()[ "sales"][cols].reset_index() # te_sub1 = te_sub1.set_index(["id", "F"]).unstack()[ # "sales"][cols].reset_index() te_sub.fillna(0., inplace=True) # te_sub1.fillna(0., inplace=True) te_sub.sort_values("id", inplace=True) # te_sub1.sort_values("id", inplace=True) te_sub.reset_index(drop=True, inplace=True) # te_sub1.reset_index(drop=True, inplace=True) te_sub.to_csv(f"submission_{icount}.csv", index=False) # te_sub1.to_csv(f"submission1_{icount}.csv", index=False) if icount == 0: sub = te_sub sub[cols] *= weight # sub1 = te_sub1 # sub1[cols] *= weight else: sub[cols] += te_sub[cols]*weight # sub1[cols] += te_sub1[cols]*weight print(icount, alpha, weight) sub.head(10) sub.id.nunique(), sub["id"].str.contains("validation$").sum() # sub1.id.nunique(), sub1["id"].str.contains("validation$").sum() sub.shape # sub1.shape sub2 = sub.copy() sub2["id"] = sub2["id"].str.replace("validation$", "evaluation") sub = pd.concat([sub, sub2], axis=0, sort=False) sub.to_csv("submissionp.csv",index=False) # sub3 = sub1.copy() # sub3["id"] = sub3["id"].str.replace("validation$", "evaluation") # sub1 = pd.concat([sub1, sub3], axis=0, sort=False) # sub.to_csv("submissiont.csv",index=False) # poisson = sub.sort_values(by = 'id').reset_index(drop = True) # tweedie = sub1.sort_values(by = 'id').reset_index(drop = True) # sub5 = poisson.copy() # for i in sub5.columns : # if i != 'id' : # sub5[i] = 0.5*poisson[i] + 0.5*tweedie[i] # sub5.to_csv('submissionavg.csv', index = False) ###Output _____no_output_____ ###Markdown Weighting ###Code w_soft = 0.275 #0.58 w_sig = 0.5 #0.609 w_crop = 0.05 #0.55 w_pred2 = 0.1 #0.562 w_pred3 = 0.075 #0.5 softmax['0'] = w_soft*softmax['0'] + w_crop*soft_crop['0'] + w_sig*sigmoid['0'] + w_pred2*pred_2['0'] + w_pred3*pred_3['0'] softmax['1'] = w_soft*softmax['1'] + w_crop*soft_crop['1'] + w_sig*sigmoid['1'] + w_pred2*pred_2['1'] + w_pred3*pred_3['1'] softmax.head() softmax['Predicted'] = np.where(softmax['1']>0.95, 1, 0) softmax.head(10) pred = softmax[['Id', 'Predicted']] pred.head(10) pred.to_csv('Ensemble.csv', index=False) ###Output _____no_output_____ ###Markdown Classificação Utilizando Ensemble Dataset: Telco Customer ChurnContext"Predict behavior to retain customers. You can analyze all relevant customer data and develop focused customer retention programs." [IBM Sample Data Sets]ContentEach row represents a customer, each column contains customer’s attributes described on the column Metadata.The data set includes information about:Customers who left within the last month – the column is called ChurnServices that each customer has signed up for – phone, multiple lines, internet, online security, online backup, device protection, tech support, and streaming TV and moviesCustomer account information – how long they’ve been a customer, contract, payment method, paperless billing, monthly charges, and total chargesDemographic info about customers – gender, age range, and if they have partners and dependentsInspirationTo explore this type of models and learn more about the subject.New version from IBM:https://community.ibm.com/community/user/businessanalytics/blogs/steven-macko/2019/07/11/telco-customer-churn-1113https://www.kaggle.com/blastchar/telco-customer-churn Obtendo Dados do Dataset ###Code import pandas as pd from sklearn.model_selection import train_test_split data = pd.read_csv("/content/sample_data/Telco-Churn.csv") data.head() data = data.replace({'Yes': 1, 'No': 0, "No internet service": 0}) data.head() data.info() def preprocessamento(x): colunasParaRemover = ["customerID", "OnlineSecurity", "MultipleLines", "InternetService", "gender", "Contract", "PaymentMethod", "TotalCharges","Churn"] onehot = pd.get_dummies(x["gender"], prefix='gender',prefix_sep='_') x = pd.concat([x, onehot],axis=1) onehot = pd.get_dummies(x["Contract"], prefix='Contract',prefix_sep='_') x = pd.concat([x, onehot],axis=1) onehot = pd.get_dummies(x["PaymentMethod"], prefix='PaymentMethod',prefix_sep='_') x = pd.concat([x, onehot],axis=1) y = x["Churn"] x = x.drop(colunasParaRemover, axis=1) return x, y print(data["TotalCharges"]) X, y = preprocessamento(data) X_tr, X_te, y_tr, y_te = train_test_split(X, y, random_state=42, train_size=0.7) X_tr.shape, X_te.shape, y_tr.shape, y_te.shape X_tr.info() ###Output <class 'pandas.core.frame.DataFrame'> Int64Index: 4930 entries, 1695 to 860 Data columns (total 21 columns): # Column Non-Null Count Dtype --- ------ -------------- ----- 0 SeniorCitizen 4930 non-null int64 1 Partner 4930 non-null int64 2 Dependents 4930 non-null int64 3 tenure 4930 non-null int64 4 PhoneService 4930 non-null int64 5 OnlineBackup 4930 non-null int64 6 DeviceProtection 4930 non-null int64 7 TechSupport 4930 non-null int64 8 StreamingTV 4930 non-null int64 9 StreamingMovies 4930 non-null int64 10 PaperlessBilling 4930 non-null int64 11 MonthlyCharges 4930 non-null float64 12 gender_Female 4930 non-null uint8 13 gender_Male 4930 non-null uint8 14 Contract_Month-to-month 4930 non-null uint8 15 Contract_One year 4930 non-null uint8 16 Contract_Two year 4930 non-null uint8 17 PaymentMethod_Bank transfer (automatic) 4930 non-null uint8 18 PaymentMethod_Credit card (automatic) 4930 non-null uint8 19 PaymentMethod_Electronic check 4930 non-null uint8 20 PaymentMethod_Mailed check 4930 non-null uint8 dtypes: float64(1), int64(11), uint8(9) memory usage: 544.0 KB ###Markdown Classificadores Simples KNeighborsClassifier ###Code from sklearn.neighbors import KNeighborsClassifier model = KNeighborsClassifier(n_neighbors=3) model.fit(X_tr,y_tr) ypred = model.predict(X_te) score = sum(ypred == y_te)/len(y_te) score ###Output _____no_output_____ ###Markdown DecisionTreeClassifier ###Code from sklearn.tree import DecisionTreeClassifier model = DecisionTreeClassifier() model.fit(X_tr,y_tr) ypred = model.predict(X_te) score = sum(ypred == y_te)/len(y_te) score ###Output _____no_output_____ ###Markdown Perceptron ###Code from sklearn.linear_model import Perceptron model = Perceptron() model.fit(X_tr,y_tr) ypred = model.predict(X_te) score = sum(ypred == y_te)/len(y_te) score ###Output _____no_output_____ ###Markdown Bagging BaggingClassifier ###Code # Diversificação por amostragem: pega amostras e substui por outras mantendo o mesmo tamanho do dataset # Bootstrap (bagging) utiliza aleatóriedade da amostragem para criar diversidade. As árvores geradas variam (esse algoritmo gera 10 árvores por padrão) # Por padrão BaggingClassifier utiliza arvore de decisao # A implementação abaixo é do RandomForest! (Random Forest é um Bagging de arvores de decisão) from sklearn.ensemble import BaggingClassifier model = BaggingClassifier(DecisionTreeClassifier(splitter='random'), n_estimators=100, max_features=0.5, random_state=42) model.fit(X_tr, y_tr) ypred = model.predict(X_te) score = sum(ypred == y_te)/len(y_te) score ###Output _____no_output_____ ###Markdown RandomForestClassifier ###Code from sklearn.ensemble import RandomForestClassifier model = RandomForestClassifier(random_state=42) model.fit(X_tr, y_tr) ypred = model.predict(X_te) score = sum(ypred == y_te)/len(y_te) score ###Output _____no_output_____ ###Markdown ExtraTreesClassifier ###Code # Florestas extremamente aleatórias from sklearn.ensemble import ExtraTreesClassifier model = ExtraTreesClassifier(random_state=42) model.fit(X_tr, y_tr) ypred = model.predict(X_te) score = sum(ypred == y_te)/len(y_te) score ###Output _____no_output_____ ###Markdown VotingClassifier ###Code from sklearn.ensemble import VotingClassifier model = VotingClassifier([ ('knn', KNeighborsClassifier(n_neighbors=1)), ('arvore', DecisionTreeClassifier()), ('perceptron', Perceptron())]) model.fit(X_tr,y_tr) vo_pred = model.predict(X_te) score = sum(vo_pred == y_te)/len(y_te) score ###Output _____no_output_____ ###Markdown Boosting AdaBoostClassifier ###Code from sklearn.ensemble import AdaBoostClassifier model = AdaBoostClassifier(random_state=42) model.fit(X_tr, y_tr) ypred = model.predict(X_te) score = sum(ypred == y_te)/len(y_te) score ###Output _____no_output_____ ###Markdown XGBClassifier ###Code from xgboost import XGBClassifier model = XGBClassifier(use_label_encoder=False,random_state=42) model.fit(X_tr, y_tr) ypred = model.predict(X_te) score = sum(ypred == y_te)/len(y_te) score ###Output _____no_output_____ ###Markdown Stacking ###Code from sklearn.ensemble import StackingClassifier from sklearn.model_selection import GridSearchCV voting = VotingClassifier([ ('knn', KNeighborsClassifier()), ('arvore', DecisionTreeClassifier()), ('perceptron', Perceptron())]) model = StackingClassifier([ ('voting', voting), ('xgbclassifier', XGBClassifier(use_label_encoder=False,random_state=42)), ('randomforest', RandomForestClassifier(random_state=42))]) params = [ { "voting__knn__n_neighbors" : [1,3,5], "voting__arvore__criterion" : ['gini', 'entropy'], "voting__arvore__max_depth" : [15,20,50], "randomforest__bootstrap" : [True, False], "randomforest__min_samples_leaf": [1,2,3] } ] grid_search_model = GridSearchCV(model, params, cv=3, verbose=3, return_train_score=True, n_jobs=1) grid_search_model.fit(X_tr, y_tr) grid_pred = grid_search_model.predict(X_te) score = sum(grid_pred == y_te)/len(y_te) score grid_search_model.best_estimator_ model.fit(X_tr,y_tr) stk_pred = model.predict(X_te) score = sum(stk_pred == y_te)/len(y_te) score from sklearn.metrics import confusion_matrix confusion_matrix(stk_pred, y_te) from sklearn import metrics import matplotlib.pyplot as plt import numpy as np fig, ax = plt.subplots() x = np.linspace(*ax.get_xlim()) ax.plot(x, x, color='green', linestyle='dashed', linewidth=1, markersize=1) metrics.plot_roc_curve(grid_search_model, X_te, y_te, ax=ax) plt.show() ###Output _____no_output_____ ###Markdown Logistic Regression ###Code #split dataset into train and test data X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, shuffle = True, random_state=1, stratify=y) #create knn model log_reg = LogisticRegression(penalty = 'l2',dual = False,tol= 0.0001, C = 1.0, fit_intercept = True,intercept_scaling = 1, class_weight = None, random_state = 666, solver = 'liblinear', max_iter = 100, multi_class = 'ovr', verbose = 0, warm_start = False, n_jobs = None) log_reg.fit(X_train, y_train) ###Output _____no_output_____ ###Markdown Random Forest ###Code # train the classifier rf = RandomForestClassifier() # dictionary of parameters to test params_rf = { 'bootstrap': [True], 'max_depth': [80, 90, 100, 110], 'max_features': [2, 3], 'min_samples_leaf': [3, 4, 5], 'min_samples_split': [8, 10, 12], 'n_estimators': [100, 200, 300, 1000] } #use gridsearch to test all values for n_estimators rf_gs = GridSearchCV(rf, params_rf, cv=5) #fit model to training data rf_gs.fit(X_train, y_train) rf_best = rf_gs.best_estimator_ #check best n_estimators value print(rf_gs.best_params_) ###Output {'bootstrap': True, 'max_depth': 100, 'max_features': 3, 'min_samples_leaf': 3, 'min_samples_split': 10, 'n_estimators': 100} ###Markdown XGB ###Code xgb = XGBClassifier() xgb.fit(X_train, y_train) ###Output _____no_output_____ ###Markdown SVM ###Code svclassifier = SVC(kernel='linear') Cs = [0.001, 0.01, 0.1, 1, 10] gammas = [0.001, 0.01, 0.1, 1] param_grid = {'C': Cs, 'gamma' : gammas} sv_gs = GridSearchCV(svclassifier, param_grid, cv=5) sv_gs.fit(X_train, y_train) sv_best = sv_gs.best_estimator_ print(sv_gs.best_params_) print('rf: {}'.format(rf_best.score(X_test, y_test))) print('log_reg: {}'.format(log_reg.score(X_test, y_test))) print('xgb: {}'.format(xgb.score(X_test, y_test))) print('SVM: {}'.format(sv_best.score(X_test, y_test))) ###Output rf: 0.63 log_reg: 0.67 xgb: 0.725 SVM: 0.68 ###Markdown Essemble Classifier ###Code from sklearn.ensemble import VotingClassifier #create a dictionary of our models estimators=[('xgb', xgb), ('rf', rf_best), ('log_reg', log_reg), ('svclassifier', svclassifier)] #create our voting classifier, inputting our models ensemble = VotingClassifier(estimators, voting='hard') #fit model to training data ensemble.fit(X_train, y_train) # make predictions for test data y_pred = ensemble.predict(X_test) #test our model on the test data ensemble.score(X_test, y_test) # Confusion Matrix import seaborn as sn genres = df["genre"].unique() cmx = confusion_matrix(y_test,y_pred) df_cm = pd.DataFrame(cmx,genres,genres) sn.set(font_scale = 1.4) sn.heatmap(df_cm, annot = True, annot_kws = {"size": 16}) plt.ylabel('True Label') plt.xlabel('Predicted Label') ###Output _____no_output_____
Chapter09/a_b_deployment_with_production_variants.ipynb
###Markdown Chapter 10 : SageMaker Endpoint Production Variants and Deployment StrategiesThis notebook demonstrates how to update a deployed model using SageMaker Endpoint Production variants. Specifically it demonstrates the A/B deployment strategy. You can use this notebook as a starting point to implement other strategies discussed in Chapter 10, since the APIs used to either deploy a new endpoint or update an existing endpoint remain the same. Overview1. Set up2. Prepare (Reuse or Train) models to deploy and update3. Create an endpoint (with single production variant)4. Invoke the endpoint5. Update endpoint (with two production variants)6. CloudWatch Analysis7. Update endpoint8. Clean up 1. Set up 1.1 Imports ###Code ##Imports import sagemaker import boto3 import time from datetime import datetime, timedelta from sagemaker import image_uris from sagemaker.session import Session from sagemaker.inputs import TrainingInput from sagemaker.session import production_variant from botocore.response import StreamingBody ###Output _____no_output_____ ###Markdown 1.2 Setup variables ###Code s3_bucket = 'datascience-environment-notebookinstance--06dc7a0224df' s3_prefix = 'prepared' m_prefix = 'xgboost-sample' sagemaker_session = sagemaker.Session() region = sagemaker_session.boto_region_name ###Output _____no_output_____ ###Markdown 1.3 Setup service clients ###Code sm = boto3.Session().client("sagemaker") smrt = boto3.Session().client("sagemaker-runtime") s3 = boto3.client("s3") ### Define variable to toggle between using trained models from previous chapters and training the models in this notebook ### Set use_trained_models to True, if you have XGBoost models trained in previous chapters, use those models to save training time and costs. ### To train models in this notebook set use_trained_model to False. #use_trained_models = 'False' use_trained_models = 'True' if use_trained_models == 'True': print("Using models trained before") else: print("Train the model") ###Output _____no_output_____ ###Markdown Section 2 - Prepare (Reuse or Train) models to deploy and update ###Code ### Use the XGBoost models previously trained ### Note: Update to use the models available in your datascience account if use_trained_models == 'True': model_name_1='sagemaker-xgboost-2021-06-24-02-34-20-510' model_name_2='sagemaker-xgboost-2021-06-24-02-47-08-912' if use_trained_models == 'False': # set an output path where the trained model will be saved output_path = 's3://{}/{}/{}/output'.format(s3_bucket, m_prefix, 'xgboost') # this line automatically looks for the XGBoost image URI and builds an XGBoost container. # specify the repo_version depending on your preference. xgboost_container = sagemaker.image_uris.retrieve("xgboost", region, "1.2-1") # define the data type and paths to the training and validation datasets content_type = "csv" train_input = TrainingInput("s3://{}/{}/{}/".format(s3_bucket, s3_prefix, 'train'), content_type=content_type) validation_input = TrainingInput("s3://{}/{}/{}/".format(s3_bucket, s3_prefix, 'validation'), content_type=content_type) #### Train and get the name of the first model # initialize hyperparameters hyperparameters_1 = { "max_depth":"5", "eta":"0.2", "gamma":"4", "min_child_weight":"6", "subsample":"0.7", "objective":"reg:squarederror", "num_round":"5"} # construct a SageMaker estimator that calls the xgboost-container estimator_1 = sagemaker.estimator.Estimator(image_uri=xgboost_container, hyperparameters=hyperparameters_1, role=sagemaker.get_execution_role(), instance_count=1, instance_type='ml.m5.12xlarge', volume_size=200, # 5 GB output_path=output_path) # execute the XGBoost training job estimator_1.fit({'train': train_input, 'validation': validation_input}) training_job_name_1 = estimator_1.latest_training_job.name model_name_1 = sagemaker_session.create_model_from_job(training_job_name_1) #### Train and get the name of the second model # initialize hyperparameters hyperparameters_2 = { "max_depth":"10", ##Different value of the hyperparameter "eta":"0.2", "gamma":"4", "min_child_weight":"6", "subsample":"0.7", "objective":"reg:squarederror", "num_round":"5"} # construct a SageMaker estimator that calls the xgboost-container estimator_2 = sagemaker.estimator.Estimator(image_uri=xgboost_container, hyperparameters=hyperparameters_2, role=sagemaker.get_execution_role(), instance_count=1, instance_type='ml.m5.12xlarge', volume_size=200, # 5 GB output_path=output_path) # execute the XGBoost training job estimator_2.fit({'train': train_input, 'validation': validation_input}) training_job_name_2 = estimator_2.latest_training_job.name model_name_2 = sagemaker_session.create_model_from_job(training_job_name_2) print("Model 1 : " , model_name_1) print("Model 2 : " , model_name_2) ###Output _____no_output_____ ###Markdown 3 Create an endpoint (with single production variant) ###Code #Create production variant A variantA = production_variant(model_name=model_name_1, instance_type="ml.m5.xlarge", initial_instance_count=1, variant_name='VariantA', initial_weight=1) #Variable for endpoint name endpoint_name=f"abtest-{datetime.now():%Y-%m-%d-%H-%M-%S}" ##First create an endpoint with single variant ##Note this step automatically creates an endpointconfig with same name as the endpoint, that you can update later #Create an endpoint with a single production variant sagemaker_session.endpoint_from_production_variants( name=endpoint_name, production_variants=[variantA] ) ###Output _____no_output_____ ###Markdown 4. Invoke the endpoint ###Code ##Get the file name at index from the 'prefix' folder def get_file_in_bucket(prefix,index): response = s3.list_objects( Bucket=s3_bucket, Prefix=s3_prefix + "/" + prefix ) ## At '0' index you will find the SUCCESS/FAILURE of file uploades to S3. First data file is at index 1 file_name = response['Contents'][index]['Key'] print("Returing file name : " + file_name) return file_name ##Download the test files to execute inferences s3.download_file(s3_bucket, get_file_in_bucket('test',1), 't_file.csv') with open('t_file.csv', 'r') as TF: t_lines = TF.readlines() ### Define a method to run inferences against the endpoint def get_predictions(): #Skip the first line since it has column headers for tl in t_lines[1:50]: #Remove the first column since it is the label test_list = tl.split(",") test_list.pop(0) test_string = ','.join([str(elem) for elem in test_list]) result = smrt.invoke_endpoint(EndpointName=endpoint_name, ContentType="text/csv", Body=test_string) #print(result) rbody = StreamingBody(raw_stream=result['Body'],content_length=int(result['ResponseMetadata']['HTTPHeaders']['content-length'])) print(f"Result from {result['InvokedProductionVariant']} = {rbody.read().decode('utf-8')}") #Get predictions get_predictions() ###Output _____no_output_____ ###Markdown 5. Update endpoint with two production variants ###Code #Create production variant B variantB = production_variant(model_name=model_name_2, instance_type="ml.m5.xlarge", initial_instance_count=1, variant_name='VariantB', initial_weight=1) ##Next update the endpoint to include both production variants endpoint_config_new =f"abtest-new-config-{datetime.now():%Y-%m-%d-%H-%M-%S}" sagemaker_session.create_endpoint_config_from_existing ( existing_config_name=endpoint_name, new_config_name=endpoint_config_new, new_production_variants=[variantA,variantB] ## Two production variants ) ##Update the endpoint sagemaker_session.update_endpoint(endpoint_name=endpoint_name, endpoint_config_name=endpoint_config_new, wait=False) #Show that you can still get inferences while the endpoint is being updated #Get predictions get_predictions() ###Output _____no_output_____ ###Markdown 6. CloudWatch Analysis Observe the CloudWatch metrics generated for the two variants to understand the endpoint behavior. Here we are plotting the number of invocations of each variant.You can use the same pattern to plot other metrics. ###Code ##Define utility methods to retrieve and plot cloudwatch metrics import pandas as pd cw = boto3.Session().client("cloudwatch") def get_invocation_metrics_for_endpoint_variant(endpoint_name, variant_name, start_time, end_time): metrics = cw.get_metric_statistics( Namespace="AWS/SageMaker", MetricName="Invocations", StartTime=start_time, EndTime=end_time, Period=60, Statistics=["Sum"], Dimensions=[ {"Name": "EndpointName", "Value": endpoint_name}, {"Name": "VariantName", "Value": variant_name}, ], ) return ( pd.DataFrame(metrics["Datapoints"]) .sort_values("Timestamp") .set_index("Timestamp") .drop("Unit", axis=1) .rename(columns={"Sum": variant_name}) ) def plot_endpoint_metrics(start_time=None): start_time = start_time or datetime.now() - timedelta(minutes=60) end_time = datetime.now() metrics_variant1 = get_invocation_metrics_for_endpoint_variant( endpoint_name, variantA["VariantName"], start_time, end_time ) metrics_variant2 = get_invocation_metrics_for_endpoint_variant( endpoint_name, variantB["VariantName"], start_time, end_time ) metrics_variants = metrics_variant1.join(metrics_variant2, how="outer") metrics_variants.plot() return metrics_variants ##Send traffic to endpoint for about 2 minutes. ##You should see both the variants serving traffic, after the endpoint is updated. print(f"Sending test traffic to the endpoint {endpoint_name}. \nPlease wait...") #Skip the first line since it has column headers for tl in t_lines[1:200]: #print(".", end="", flush=True) #Remove the first column since it is the label test_list = tl.split(",") test_list.pop(0) test_string = ','.join([str(elem) for elem in test_list]) result = smrt.invoke_endpoint(EndpointName=endpoint_name, ContentType="text/csv", Body=test_string) #print(result) rbody = StreamingBody(raw_stream=result['Body'],content_length=int(result['ResponseMetadata']['HTTPHeaders']['content-length'])) print(f"Result from {result['InvokedProductionVariant']} = {rbody.read().decode('utf-8')}") time.sleep(0.5) print("Done!") print("Waiting a minute for initial metric creation...") time.sleep(60) plot_endpoint_metrics() ###Output _____no_output_____ ###Markdown 7. Update endpoint to contain just the VariantB 7.1 - Gradually update the weights of each production variants ###Code #Update the product variant weight to route 60% of traffic to VariantB sm.update_endpoint_weights_and_capacities( EndpointName=endpoint_name, DesiredWeightsAndCapacities=[ {"DesiredWeight": 4, "VariantName": variantA["VariantName"]}, {"DesiredWeight": 6, "VariantName": variantB["VariantName"]}, ], ) ###Output _____no_output_____ ###Markdown 7.2 - Alternatively, update the endpoint to route all live traffic to VariantB in a single step ###Code ##Update the endpoint to point to VariantB endpoint_config_new =f"abtest-b-config-{datetime.now():%Y-%m-%d-%H-%M-%S}" sagemaker_session.create_endpoint_config_from_existing ( existing_config_name=endpoint_name, new_config_name=endpoint_config_new, new_production_variants=[variantB] ) ##Update the endpoint ##Note : This step will fail if the endpoint is still updating sagemaker_session.update_endpoint(endpoint_name=endpoint_name, endpoint_config_name=endpoint_config_new, wait=False) ###Output _____no_output_____ ###Markdown 8. Cleanup ###Code # If you do not plan to use this endpoint further, you should delete the endpoint to avoid incurring additional charges. sagemaker_session.delete_endpoint(endpoint_name) ###Output _____no_output_____
Homework_03.ipynb
###Markdown CPSC 4300/6300: Applied Data Science Homework 2: k-NN Regression**Clemson University****Fall 2021****Instructor(s):** Nina Hubig --- ###Code """ RUN THIS CELL TO GET THE RIGHT FORMATTING """ import requests from IPython.core.display import HTML css_file = 'https://raw.githubusercontent.com/bsethwalker/clemson-cs4300/main/css/cpsc6300.css' styles = requests.get(css_file).text HTML(styles) ###Output _____no_output_____ ###Markdown INSTRUCTIONS- To submit your assignment follow the instructions given in Canvas.- Restart the kernel and run the whole notebook again before you submit. - If you submit individually and you have worked with someone, please include the name of your [one] partner below. - As much as possible, try and stick to the hints and functions we import at the top of the homework, as those are the ideas and tools the class supports and is aiming to teach. And if a problem specifies a particular library you're required to use that library, and possibly others from the import list.- Please use .head() when viewing data. Do not submit a notebook that is excessively long because output was not suppressed or otherwise limited. ---In this homework, we will explore regression methods for predicting a quantitative variable. Specifically, we will build regression models that can predict the number of taxi pickups in New York City at any given time of the day. These prediction models will be useful, for example, in monitoring traffic in the city.The data set for this problem is given in the file `nyc_taxi.csv`. You will need to separate it into training and test sets. The first column contains the time of a day in minutes, and the second column contains the number of pickups observed at that time. The data set covers taxi pickups recorded in NYC during Jan 2015.We will fit models that use the time of the day (in minutes) as a predictor and predict the average number of taxi pickups at that time. The models will be fitted to the training set and evaluated on the test set. The performance of the models will be evaluated using the $R^2$ metric. ###Code import numpy as np import pandas as pd from sklearn.metrics import r2_score from sklearn.neighbors import KNeighborsRegressor from sklearn.linear_model import LinearRegression from sklearn.model_selection import train_test_split import statsmodels.api as sm from statsmodels.api import OLS import matplotlib.pyplot as plt import seaborn as sns sns.set() %matplotlib inline ###Output _____no_output_____ ###Markdown Question 1 We next consider simple linear regression, which we know from lecture is a parametric approach for regression that assumes that the response variable has a linear relationship with the predictor. Use the `statsmodels` module for Linear Regression. This module has built-in functions to summarize the results of regression and to compute confidence intervals for estimated regression parameters. Question 1.1 Use pandas to load the dataset from the csv file `nyc_taxi.csv` into a pandas data frame. Use the `train_test_split` method from `sklearn` with a `random_state` of 42 and a `test_size` of 0.2 to split the dataset into training and test sets. Store your train set data frame as `train_data` and your test set data frame as `test_data`. ###Code # Your code here nyc_taxi = pd.read_csv("nyc_taxi.csv") train_data, test_data = train_test_split(nyc_taxi, test_size=0.2, random_state=42) ###Output _____no_output_____ ###Markdown Question 1.2 Again choose `TimeMin` as your predictor and `PickupCount` as your response variable. Create an `OLS` class instance and use it to fit a Linear Regression model on the training set (`train_data`). Store your fitted model in the variable `OLSModel`. ###Code # Your code here x_train = train_data['TimeMin'] y_train = train_data['PickupCount'] x_test = test_data['TimeMin'] y_test = test_data['PickupCount'] X_train = sm.add_constant(x_train) X_test = sm.add_constant(x_test) OLS = sm.OLS(y_train, X_train) OLSModel = OLS.fit() print(OLSModel.summary()) ###Output OLS Regression Results ============================================================================== Dep. Variable: PickupCount R-squared: 0.243 Model: OLS Adj. R-squared: 0.242 Method: Least Squares F-statistic: 320.4 Date: Fri, 17 Sep 2021 Prob (F-statistic): 2.34e-62 Time: 02:27:17 Log-Likelihood: -4232.9 No. Observations: 1000 AIC: 8470. Df Residuals: 998 BIC: 8480. Df Model: 1 Covariance Type: nonrobust ============================================================================== coef std err t P>|t| [0.025 0.975] ------------------------------------------------------------------------------ const 16.7506 1.058 15.838 0.000 14.675 18.826 TimeMin 0.0233 0.001 17.900 0.000 0.021 0.026 ============================================================================== Omnibus: 203.688 Durbin-Watson: 1.910 Prob(Omnibus): 0.000 Jarque-Bera (JB): 462.910 Skew: 1.111 Prob(JB): 3.02e-101 Kurtosis: 5.485 Cond. No. 1.63e+03 ============================================================================== Notes: [1] Standard Errors assume that the covariance matrix of the errors is correctly specified. [2] The condition number is large, 1.63e+03. This might indicate that there are strong multicollinearity or other numerical problems. ###Markdown Question 1.3 Create a plot just like you did in question 2.2 from Homework 2 (but with fewer subplots): plot both the observed values and the predictions from `OLSModel` on the training and test set. You should have one figure with two subplots, one subplot for the training set and one for the test set.**Hints**:1. Each subplot should use different color and/or markers to distinguish Linear Regression prediction values from that of the actual data values.2. Each subplot must have appropriate axis labels, title, and legend.3. The overall figure should have a title. ###Code # Your code here ytrain_pred = OLSModel.predict(X_train) ytest_pred = OLSModel.predict(X_test) fig, (ax1, ax2) = plt.subplots(1,2, figsize = (15, 5)) fig.suptitle('Predictions vs Actuals', fontsize=14) ax1.scatter(x_train, y_train, color='b',label='Actual') ax1.scatter(x_train, ytrain_pred, color='tab:orange',label ='Predicted') ax1.set_title('Training Set') ax1.set_xlabel('Time of Day in Minutes') ax1.set_ylabel('Pickup Count') ax1.legend() ax2.scatter(x_test, y_test, label='Actual', color='b') ax2.scatter(x_test, ytest_pred, color='tab:orange',label ='Predicted') ax2.set_title('Test Set') ax2.set_xlabel('Time of Day in Minutes') ax2.set_ylabel('Pickup Count') ax2.legend() ###Output _____no_output_____ ###Markdown Question 1.4 Report the $R^2$ score for the fitted model on both the training and test sets. ###Code # Your code here r2_train = r2_score(y_train, ytrain_pred) r2_test = r2_score(y_test, ytest_pred) print('r2_training is', r2_train) print('r2_test is', r2_test) ###Output r2_training is 0.24302603531893352 r2_test is 0.240661535615741 ###Markdown Question 1.5 Report the estimates for the slope and intercept for the fitted linear model. ###Code # Your code here beta0_sm = OLSModel.params[0] beta1_sm = OLSModel.params[1] print(f'The regression coef from statsmodels are: beta_0 = {beta0_sm:8.6f} and beta_1 = {beta1_sm:8.6f}') ###Output The regression coef from statsmodels are: beta_0 = 16.750601 and beta_1 = 0.023335 ###Markdown Question 1.6 Report the $95\%$ confidence intervals (CIs) for the slope and intercept. ###Code # Your code here OLSModel.conf_int(alpha=0.05, cols =None) ###Output _____no_output_____ ###Markdown Question 1.7 Discuss the results:1. How does the test $R^2$ score compare with the best test $R^2$ value obtained with k-NN regression? Describe why this is not surprising for these data.2. What does the sign of the slope of the fitted linear model convey about the data? 3. Interpret the $95\%$ confidence intervals from 3.5. Based on these CIs is there evidence to suggest that the number of taxi pickups has a significant linear relationship with time of day? How do you know? 4. How would $99\%$ confidence intervals for the slope and intercept compare to the $95\%$ confidence intervals (in terms of midpoint and width)? Briefly explain your answer. 5. Based on the data structure, what restriction on the model would you put at the endpoints (at $x\approx0$ and $x\approx1440$)? What does this say about the appropriateness of a linear model? *your answer here* 1. The test R^2 score is smaller than the best test R^2 value (at K=75) with k-NN regression. Because these data have two trends, upward and downward, but linear regression has only one trend depending on the slope, thus the resulting model will provide a poor fit to the data than k-NN.2. The slope of the fitted linear model convey about the ratio between the predictor and the response variable. In other words, we can understand that the slope is the rate of change in the response variable when the predictor increases by 1.3. The 95% confidence intervals is a range of values that you can be 95% certain contains the slope and intercept of the fitted linear model. We are able to say that the number of taxi as a signigicant linear relationship with time of day. It is easy to know as the 95% confidence interval of the slope is very small.4. The 99% confidence intervals for slope and intercept are larger in width, and the same with the midpoint compare to the 95% confidence intervals. Because the fitted model has a certain slope and intercept, so the midpoint is unchanged. The 99% confidence intervals have a more extensive range of values, thus larger in width than 95% confidence intervals.5. At the endpoints 𝑥≈0 and 𝑥≈1440, it should be two values close to the same based on the data structure, but the fitted model calculates two very far different values. A linear model is only appropriate with the data has one clear tendency upward or downward. Question 2 You may recall from lectures that OLS Linear Regression can be susceptible to outliers in the data. We're going to look at a dataset that includes some outliers and get a sense for how that affects modeling data with Linear Regression. **Note, this is an open-ended question, there is not one correct solution (or one correct definition of an outlier).** Question 2.1 We've provided you with two files `outliers_train.csv` and `outliers_test.csv` corresponding to training set and test set data. What does a visual inspection of training set tell you about the existence of outliers in the data? ###Code # Your code here trainingset = pd.read_csv("outliers_train.csv") x_otrain = trainingset['X'] y_otrain = trainingset['Y'] fig, ax = plt.subplots(1, 1, figsize=(15,6)) ax.scatter(x_otrain, y_otrain, label='Training set', alpha=0.5) ax.set_title('Training set') ax.set_xlabel('x_otrain') ax.set_ylabel('y_otrain') ax.legend(); ###Output _____no_output_____ ###Markdown *Your answer here*As very clearly visible in the graph, we can see three points are out of the trend of the training set. Two points are at the left upper corner; one point is at the right lower corner. Question 2.2 Choose `X` as your feature variable and `Y` as your response variable. Use `statsmodel` to create a Linear Regression model on the training set data. Store your model in the variable `OutlierOLSModel` and display the model summary. ###Code # Your code here X_otrain = sm.add_constant(x_otrain) OutlierOLS = sm.OLS(y_otrain, X_otrain) OutlierOLSModel = OutlierOLS.fit() print(OutlierOLSModel.summary()) ###Output OLS Regression Results ============================================================================== Dep. Variable: Y R-squared: 0.084 Model: OLS Adj. R-squared: 0.066 Method: Least Squares F-statistic: 4.689 Date: Fri, 17 Sep 2021 Prob (F-statistic): 0.0351 Time: 02:27:44 Log-Likelihood: -343.59 No. Observations: 53 AIC: 691.2 Df Residuals: 51 BIC: 695.1 Df Model: 1 Covariance Type: nonrobust ============================================================================== coef std err t P>|t| [0.025 0.975] ------------------------------------------------------------------------------ const -9.5063 22.192 -0.428 0.670 -54.059 35.046 X 47.3554 21.869 2.165 0.035 3.452 91.259 ============================================================================== Omnibus: 2.102 Durbin-Watson: 1.758 Prob(Omnibus): 0.350 Jarque-Bera (JB): 1.251 Skew: 0.215 Prob(JB): 0.535 Kurtosis: 3.617 Cond. No. 1.06 ============================================================================== Notes: [1] Standard Errors assume that the covariance matrix of the errors is correctly specified. ###Markdown Question 2.3 You're given the knowledge ahead of time that there are 3 outliers in the training set data. The test set data doesn't have any outliers. You want to remove the 3 outliers in order to get the optimal intercept and slope. In the case that you're sure of the existence and number (3) of outliers ahead of time, one potential brute force method to outlier detection might be to find the best Linear Regression model on all possible subsets of the training set data with 3 points removed. Using this method, how many times will you have to calculate the Linear Regression coefficients on the training data? *Your answer here.*We have to choose 50 out of 53 points. It means the number of times we will have to calculate the Linear Regression coefficients on the training data are 'Combinations without repetition' C = 53!/((3!)((53-3)!)) = 23426. Question 2.4 In CPSC 4300 we're strong believers that creating heuristic models is a great way to build intuition. In that spirit, construct an approximate algorithm to find the 3 outlier candidates in the training data by taking advantage of the Linear Regression residuals. Place your algorithm in the function `find_outliers_simple`. It should take the parameters `dataset_x` and `dataset_y`, and `num_outliers` representing your features, response variable values (make sure your response variable is stored as a numpy column vector), and the number of outliers to remove. The return value should be a list `outlier_indices` representing the indices of the `num_outliers` outliers in the original datasets you passed in. Run your algorithm and remove the outliers that your algorithm identified, use `statsmodels` to create a Linear Regression model on the remaining training set data, and store your model in the variable `OutlierFreeSimpleModel` display the summary of this model. ###Code def find_outliers_simple(dataset_x, dataset_y, num_outliers): df = pd.concat([dataset_x, dataset_y], axis=1) # get predictions x = sm.add_constant(dataset_x) df['y_pred'] = OutlierOLSModel.predict(x) # get residuals df['y_residual'] = abs(df['Y'] - df['y_pred']) # sort by residual df = df.sort_values(by=['y_residual'], ascending=False) # identify top n residuals as outliers outliers = df[0:num_outliers] return outliers.index.tolist(); find_outliers_simple(x_otrain, y_otrain, 3) XTrain_out = x_otrain.drop([50, 51, 52]) YTrain_out = y_otrain.drop([50, 51, 52]) # Your code here OutlierFreeSimpleModel = sm.OLS(YTrain_out, XTrain_out) OutlierFreeSimpleModel = OutlierFreeSimpleModel.fit() # printing the summary table print(OutlierFreeSimpleModel.summary()) # Your code here X_otrain = sm.add_constant(x_otrain) OutlierOLS = sm.OLS(y_otrain, X_otrain) OutlierOLSModel = OutlierOLS.fit() OutlierFreeSimpleModel = sm.OLS(YTrain_out, XTrain_out) OutlierFreeSimpleModel = OutlierFreeSimpleModel.fit() y_otrainpred = OutlierOLSModel.predict(X_otrain) y_oftrainpred = OutlierFreeSimpleModel.predict(XTrain_out) ###Output _____no_output_____ ###Markdown Question 2.5 Create a figure with two subplots. The first is a scatterplot where the color of the points denotes the outliers from the non-outliers in the training set, and include two regression lines on this scatterplot: one fitted with the outliers included and one fitted with the outlier removed (all on the training set). The second plot should include a scatterplot of points from the test set with the same two regression lines fitted on the training set: with and without outliers. Visually which model fits the test set data more closely? ###Code # Your code here testset = pd.read_csv("outliers_test.csv") x_otest = testset['X'] y_otest = testset['Y'] fig, (ax1, ax2) = plt.subplots(1,2, figsize = (15, 5)) ax1.scatter(x_otrain, y_otrain, color='k',label='Outliers') ax1.scatter(XTrain_out, YTrain_out, color='tab:orange',label ='Non-outliers') ax1.scatter(x_otrain, y_otrainpred, color='g',label='Regression line with outliers') ax1.scatter(XTrain_out, y_oftrainpred, color='m',label='Regression line without outliers') ax1.set_title('Training Set') ax1.set_xlabel('X') ax1.set_ylabel('Y') ax1.legend(loc=3) ax2.scatter(x_otest, y_otest, label='Test set', color='r') ax2.scatter(x_otrain, y_otrainpred, color='g',label='Regression line with outliers') ax2.scatter(XTrain_out, y_oftrainpred, color='m',label='Regression line without outliers') ax2.set_title('Test Set') ax2.set_xlabel('X') ax2.set_ylabel('Y') ax2.legend(loc=3) ###Output _____no_output_____ ###Markdown *Your answer here*It can be seen that the model without outliers fits more closely than the model with outliers. Question 2.6 Calculate the $R^2$ score for the `OutlierOLSModel` and the `OutlierFreeSimpleModel` on the test set data. Which model produces a better $R^2$ score? ###Code # Your code here X_otest = sm.add_constant(x_otest) y_otestpred = OutlierOLSModel.predict(X_otest) y_oftestpred = OutlierFreeSimpleModel.predict(x_otest) r2_outliers = r2_score(y_otest, y_otestpred) r2_woutliers = r2_score(y_otest, y_oftestpred) print('r2_outliers is', r2_outliers) print('r2_outlierfree is', r2_woutliers) print('The model without outliers produces a better R^2 score') ###Output r2_outliers is 0.34085656043405654 r2_outlierfree is 0.4579491642913984 The model without outliers produces a better R^2 score ###Markdown 1. Попробуйте обучить нейронную сеть на TensorFlow 2 на датасете imdb_reviews. Опишите в комментарии к уроку - какого результата вы добились от нейросети? Что помогло вам улучшить ее точность?2. Поработайте с документацией TensorFlow 2. Найдите полезные команды не разобранные на уроке. ###Code from __future__ import absolute_import, division, print_function, unicode_literals # TensorFlow and tf.keras import tensorflow as tf from tensorflow import keras from keras.preprocessing import sequence # Helper libraries import numpy as np import matplotlib.pyplot as plt print(tf.__version__) from keras.datasets import imdb NUM_WORDS = 10000 (train_x, train_y), (test_x, test_y) = imdb.load_data() train_x.shape print(train_x[0]) train_y.shape train_y[0] print("Количество классов: ") print(np.unique(train_y)) print("Количество слов: ") print(len(np.unique(np.hstack(train_x)))) from matplotlib import pyplot print("Длина обзора: ") result = [len(x) for x in train_x] print("Средняя длина %.2f слов со стандартным отклонением (%.2f)" % (np.mean(result), np.std(result))) print("95 персентиль длины обзора: ", np.percentile(result, 95)) pyplot.boxplot(result) pyplot.show() test_x.shape test_y.shape index = imdb.get_word_index() reverse_index = dict([(value, key) for (key, value) in index.items()]) decoded = " ".join( [reverse_index.get(i - 3, "#") for i in train_x[0]] ) print(decoded) (train_x, train_y), (test_x, test_y) = imdb.load_data(num_words=NUM_WORDS) MAX_WORDS = 610 train_x = sequence.pad_sequences(train_x, maxlen=MAX_WORDS) test_x = sequence.pad_sequences(test_x, maxlen=MAX_WORDS) # train_y = np.array(train_y).astype("float32") # test_y = np.array(test_y).astype("float32") model = keras.Sequential([ keras.layers.Embedding(NUM_WORDS, 32, input_length=MAX_WORDS), keras.layers.Flatten(), keras.layers.Dense(512, activation='sigmoid'), keras.layers.Dropout(0.5), keras.layers.Dense(256, activation='sigmoid'), keras.layers.Dropout(0.5), keras.layers.Dense(256, activation='sigmoid'), keras.layers.Dropout(0.5), keras.layers.Dense(128, activation='sigmoid'), keras.layers.Dropout(0.5), keras.layers.Dense(1, activation='sigmoid') ]) print(model.summary()) model.compile(optimizer='adam', loss=tf.keras.losses.BinaryCrossentropy(), metrics=['accuracy']) model.fit(train_x, train_y, validation_data=(test_x, test_y), epochs=3, batch_size=32) test_loss, test_acc = model.evaluate(test_x, test_y, verbose=1) print('\nTest accuracy:', test_acc) ###Output 25000/25000 [==============================] - 3s 124us/sample - loss: 0.4030 - accuracy: 0.8748 Test accuracy: 0.87476
notebooks/.ipynb_checkpoints/2.0_clean_data-checkpoint.ipynb
###Markdown Clean DataThis notebook intends to clean the Raw DataFrame having as outcome one or more interim DataFrame, ready to has the features engineered (next step).The Clean Data steps that might be followed in order to clean the Raw DataFrame are:1. NaN2. Features that has same value in all rows3. Duplicated Features(identical to other existent column)4. High Correlation Features5. Window Selection (0-2) Sumário * [Importe das Bibliotecas](import) * [Leitura dos dados](leitura) * [Dados faltantes (NaN)](nan) * [Valores repetidos](repet) * [Features duplicadas](dup_feat) * [Nome features exame de sangue](rename) * [Seleção da Janela de tempo](win_select) * [Salva dados limpos](clean) * [Conclusão](conclusion) Importe das Bibliotecas Bibliotecas Externas ###Code import pandas as pd import numpy as np ###Output _____no_output_____ ###Markdown Bibliotecas Internas ###Code import sys sys.path.insert(1, "../src/") from clean import print_nan_count_by_feature, neighborhood_missing_data from clean import drop_features_with_same_value_for_all_observations, plot_features_with_same_value_for_all_observations from clean import drop_duplicated_features, plot_duplicated_features from clean import rename_portion_of_columns from clean import drop_patient_moved_to_icu_on_first_window, plot_patient_moved_to_icu_on_first_window, prepare_window ###Output _____no_output_____ ###Markdown -----------------Retornar ao [Sumário](sumario) Leitura dos dadosFaz a leitura dos dados que serão usados para a limpeza e futura modelagem.**```df : pd.DataFrame```** é o DataFrame que receberá os valores *raw* baixados do Kaggle. ###Code # Leitura do raw data desse projeto no Github df = pd.read_excel('https://github.com/fdrigui/covid19_icu_admission_prediction/raw/main/data/raw/Kaggle_Sirio_Libanes_ICU_Prediction.xlsx' ) # Imprime na tela todas as colunas em vez de as 10 primeiras e 10 ultimas pd.set_option('max_columns', df.shape[1]) df.head() ###Output _____no_output_____ ###Markdown -----------------Retornar ao [Sumário](sumario) Dados faltantes (NaN)Para saber mais detalhes sobre a estratégia de eliminação dos dados faltantes (NaN) foram tratados, veja o documento:[0.0_understanding_the_data.md](0.0_understanding_the_data.md), no tópico: **Dados faltantes (NaN)**. Existem dados faltantes no DataFrame?Essa pergunta é importante uma vez que muitos dos modelos de predição não conseguem trabalhar com dados NaN.A linha conta todas as ocorrencias de 'valores' ```NaN``` no **df**, e vemos que a quantidade é de **223863** NaN. ###Code print(f'Existem no DataFrame df {df.isna().sum().sum()} "valores" NaN') ###Output Existem no DataFrame df 223863 "valores" NaN ###Markdown Uma vez que sabemos que existem valores faltantes (NaN) vamos começar a trata-los. Limpando os NaN com a função 'neighborhood_missing_data'Usando a função ```neighborhood_missing_data``` para eliminar os NaN.**```df_1_without_nan : pd.DataFrame```** é o DataFrame após tirar os NaN com a função ```neighborhood_missing_data```.Após isso, uma contagem é ralizada para saber se ainda existem dados faltantes. ###Code # Usando a função neighborhood_missing_data para eliminar os NaN df_1_without_nan = neighborhood_missing_data(df, 'PATIENT_VISIT_IDENTIFIER') # Avaliando quantos dados ainda sobraram como NaN print(f'O total de NaN existentes no DataFrame df_1_without_nan é:{df_1_without_nan.isna().sum().sum()}') ###Output O total de NaN existentes no DataFrame df_1_without_nan é:2025 ###Markdown É possível observar uma considerável redução na quantidade de ```NaN```, que passou de: **223863** para: **2025**, mesmo assim, é necessário entender e eliminar esses dados remanescentes.A função ```print_nan_count_by_feature``` imprime todas as colunas, e mostra a quantidade de dados ```NaN``` das de cada uma.É possível observar que temos algumas colunas com 5 e outras com 10 NaN. Como a a janela possui 5 perídos ```(0-2, 2-4, 4-6, 6-12 e mais que 12)```, então temos 1 paciente para aquelas que apresentam 5 valores faltantes e 2 pacientes para aquelas que apresentam 10 valores faltantes. ###Code print_nan_count_by_feature(df_1_without_nan) ###Output Count - Feature Name -------------------- 0000 - PATIENT_VISIT_IDENTIFIER 0000 - AGE_ABOVE65 0000 - AGE_PERCENTIL 0000 - GENDER 0005 - DISEASE GROUPING 1 0005 - DISEASE GROUPING 2 0005 - DISEASE GROUPING 3 0005 - DISEASE GROUPING 4 0005 - DISEASE GROUPING 5 0005 - DISEASE GROUPING 6 0005 - HTN 0005 - IMMUNOCOMPROMISED 0005 - OTHER 0010 - ALBUMIN_MEDIAN 0010 - ALBUMIN_MEAN 0010 - ALBUMIN_MIN 0010 - ALBUMIN_MAX 0010 - ALBUMIN_DIFF 0010 - BE_ARTERIAL_MEDIAN 0010 - BE_ARTERIAL_MEAN 0010 - BE_ARTERIAL_MIN 0010 - BE_ARTERIAL_MAX 0010 - BE_ARTERIAL_DIFF 0010 - BE_VENOUS_MEDIAN 0010 - BE_VENOUS_MEAN 0010 - BE_VENOUS_MIN 0010 - BE_VENOUS_MAX 0010 - BE_VENOUS_DIFF 0010 - BIC_ARTERIAL_MEDIAN 0010 - BIC_ARTERIAL_MEAN 0010 - BIC_ARTERIAL_MIN 0010 - BIC_ARTERIAL_MAX 0010 - BIC_ARTERIAL_DIFF 0010 - BIC_VENOUS_MEDIAN 0010 - BIC_VENOUS_MEAN 0010 - BIC_VENOUS_MIN 0010 - BIC_VENOUS_MAX 0010 - BIC_VENOUS_DIFF 0010 - BILLIRUBIN_MEDIAN 0010 - BILLIRUBIN_MEAN 0010 - BILLIRUBIN_MIN 0010 - BILLIRUBIN_MAX 0010 - BILLIRUBIN_DIFF 0010 - BLAST_MEDIAN 0010 - BLAST_MEAN 0010 - BLAST_MIN 0010 - BLAST_MAX 0010 - BLAST_DIFF 0010 - CALCIUM_MEDIAN 0010 - CALCIUM_MEAN 0010 - CALCIUM_MIN 0010 - CALCIUM_MAX 0010 - CALCIUM_DIFF 0010 - CREATININ_MEDIAN 0010 - CREATININ_MEAN 0010 - CREATININ_MIN 0010 - CREATININ_MAX 0010 - CREATININ_DIFF 0010 - FFA_MEDIAN 0010 - FFA_MEAN 0010 - FFA_MIN 0010 - FFA_MAX 0010 - FFA_DIFF 0010 - GGT_MEDIAN 0010 - GGT_MEAN 0010 - GGT_MIN 0010 - GGT_MAX 0010 - GGT_DIFF 0010 - GLUCOSE_MEDIAN 0010 - GLUCOSE_MEAN 0010 - GLUCOSE_MIN 0010 - GLUCOSE_MAX 0010 - GLUCOSE_DIFF 0010 - HEMATOCRITE_MEDIAN 0010 - HEMATOCRITE_MEAN 0010 - HEMATOCRITE_MIN 0010 - HEMATOCRITE_MAX 0010 - HEMATOCRITE_DIFF 0010 - HEMOGLOBIN_MEDIAN 0010 - HEMOGLOBIN_MEAN 0010 - HEMOGLOBIN_MIN 0010 - HEMOGLOBIN_MAX 0010 - HEMOGLOBIN_DIFF 0010 - INR_MEDIAN 0010 - INR_MEAN 0010 - INR_MIN 0010 - INR_MAX 0010 - INR_DIFF 0010 - LACTATE_MEDIAN 0010 - LACTATE_MEAN 0010 - LACTATE_MIN 0010 - LACTATE_MAX 0010 - LACTATE_DIFF 0010 - LEUKOCYTES_MEDIAN 0010 - LEUKOCYTES_MEAN 0010 - LEUKOCYTES_MIN 0010 - LEUKOCYTES_MAX 0010 - LEUKOCYTES_DIFF 0010 - LINFOCITOS_MEDIAN 0010 - LINFOCITOS_MEAN 0010 - LINFOCITOS_MIN 0010 - LINFOCITOS_MAX 0010 - LINFOCITOS_DIFF 0010 - NEUTROPHILES_MEDIAN 0010 - NEUTROPHILES_MEAN 0010 - NEUTROPHILES_MIN 0010 - NEUTROPHILES_MAX 0010 - NEUTROPHILES_DIFF 0010 - P02_ARTERIAL_MEDIAN 0010 - P02_ARTERIAL_MEAN 0010 - P02_ARTERIAL_MIN 0010 - P02_ARTERIAL_MAX 0010 - P02_ARTERIAL_DIFF 0010 - P02_VENOUS_MEDIAN 0010 - P02_VENOUS_MEAN 0010 - P02_VENOUS_MIN 0010 - P02_VENOUS_MAX 0010 - P02_VENOUS_DIFF 0010 - PC02_ARTERIAL_MEDIAN 0010 - PC02_ARTERIAL_MEAN 0010 - PC02_ARTERIAL_MIN 0010 - PC02_ARTERIAL_MAX 0010 - PC02_ARTERIAL_DIFF 0010 - PC02_VENOUS_MEDIAN 0010 - PC02_VENOUS_MEAN 0010 - PC02_VENOUS_MIN 0010 - PC02_VENOUS_MAX 0010 - PC02_VENOUS_DIFF 0010 - PCR_MEDIAN 0010 - PCR_MEAN 0010 - PCR_MIN 0010 - PCR_MAX 0010 - PCR_DIFF 0010 - PH_ARTERIAL_MEDIAN 0010 - PH_ARTERIAL_MEAN 0010 - PH_ARTERIAL_MIN 0010 - PH_ARTERIAL_MAX 0010 - PH_ARTERIAL_DIFF 0010 - PH_VENOUS_MEDIAN 0010 - PH_VENOUS_MEAN 0010 - PH_VENOUS_MIN 0010 - PH_VENOUS_MAX 0010 - PH_VENOUS_DIFF 0010 - PLATELETS_MEDIAN 0010 - PLATELETS_MEAN 0010 - PLATELETS_MIN 0010 - PLATELETS_MAX 0010 - PLATELETS_DIFF 0010 - POTASSIUM_MEDIAN 0010 - POTASSIUM_MEAN 0010 - POTASSIUM_MIN 0010 - POTASSIUM_MAX 0010 - POTASSIUM_DIFF 0010 - SAT02_ARTERIAL_MEDIAN 0010 - SAT02_ARTERIAL_MEAN 0010 - SAT02_ARTERIAL_MIN 0010 - SAT02_ARTERIAL_MAX 0010 - SAT02_ARTERIAL_DIFF 0010 - SAT02_VENOUS_MEDIAN 0010 - SAT02_VENOUS_MEAN 0010 - SAT02_VENOUS_MIN 0010 - SAT02_VENOUS_MAX 0010 - SAT02_VENOUS_DIFF 0010 - SODIUM_MEDIAN 0010 - SODIUM_MEAN 0010 - SODIUM_MIN 0010 - SODIUM_MAX 0010 - SODIUM_DIFF 0010 - TGO_MEDIAN 0010 - TGO_MEAN 0010 - TGO_MIN 0010 - TGO_MAX 0010 - TGO_DIFF 0010 - TGP_MEDIAN 0010 - TGP_MEAN 0010 - TGP_MIN 0010 - TGP_MAX 0010 - TGP_DIFF 0010 - TTPA_MEDIAN 0010 - TTPA_MEAN 0010 - TTPA_MIN 0010 - TTPA_MAX 0010 - TTPA_DIFF 0010 - UREA_MEDIAN 0010 - UREA_MEAN 0010 - UREA_MIN 0010 - UREA_MAX 0010 - UREA_DIFF 0010 - DIMER_MEDIAN 0010 - DIMER_MEAN 0010 - DIMER_MIN 0010 - DIMER_MAX 0010 - DIMER_DIFF 0005 - BLOODPRESSURE_DIASTOLIC_MEAN 0005 - BLOODPRESSURE_SISTOLIC_MEAN 0005 - HEART_RATE_MEAN 0005 - RESPIRATORY_RATE_MEAN 0005 - TEMPERATURE_MEAN 0005 - OXYGEN_SATURATION_MEAN 0005 - BLOODPRESSURE_DIASTOLIC_MEDIAN 0005 - BLOODPRESSURE_SISTOLIC_MEDIAN 0005 - HEART_RATE_MEDIAN 0005 - RESPIRATORY_RATE_MEDIAN 0005 - TEMPERATURE_MEDIAN 0005 - OXYGEN_SATURATION_MEDIAN 0005 - BLOODPRESSURE_DIASTOLIC_MIN 0005 - BLOODPRESSURE_SISTOLIC_MIN 0005 - HEART_RATE_MIN 0005 - RESPIRATORY_RATE_MIN 0005 - TEMPERATURE_MIN 0005 - OXYGEN_SATURATION_MIN 0005 - BLOODPRESSURE_DIASTOLIC_MAX 0005 - BLOODPRESSURE_SISTOLIC_MAX 0005 - HEART_RATE_MAX 0005 - RESPIRATORY_RATE_MAX 0005 - TEMPERATURE_MAX 0005 - OXYGEN_SATURATION_MAX 0005 - BLOODPRESSURE_DIASTOLIC_DIFF 0005 - BLOODPRESSURE_SISTOLIC_DIFF 0005 - HEART_RATE_DIFF 0005 - RESPIRATORY_RATE_DIFF 0005 - TEMPERATURE_DIFF 0005 - OXYGEN_SATURATION_DIFF 0005 - BLOODPRESSURE_DIASTOLIC_DIFF_REL 0005 - BLOODPRESSURE_SISTOLIC_DIFF_REL 0005 - HEART_RATE_DIFF_REL 0005 - RESPIRATORY_RATE_DIFF_REL 0005 - TEMPERATURE_DIFF_REL 0005 - OXYGEN_SATURATION_DIFF_REL 0000 - WINDOW 0000 - ICU ###Markdown Foi escolhido aleatóriamente uma frature que tem 10 valores NaN, nesse caso a ```UREA_MEDIAN```, e foi feita uma query para selecionar os dados NaN dessa coluna.É possível observar que existem 2 VISITAS que apresentam NaN, a ID **199** e a ID **287**, isso afirma a hipótese de ser multiplos das 5 janelas.![alt text](../img/NaN_2_VisitId.png "Timeline_Example_Best") ###Code df_1_without_nan.query('UREA_MEDIAN.isnull()', engine='python') ###Output _____no_output_____ ###Markdown **```df_2_without_nan : pd.DataFrame```** é a ```variável df_1_without_nan``` após o drop dos index 199 e 287, que continham valores NaN ###Code df_2_without_nan = df_1_without_nan.drop(df_1_without_nan.query('UREA_MEDIAN.isnull()', engine='python').index) print(f'A quantidade de valores NaN após a remocção dos dois IDs de visita é: {df_2_without_nan.isna().sum().sum()}') ###Output A quantidade de valores NaN após a remocção dos dois IDs de visita é: 0 ###Markdown -----------------Retornar ao [Sumário](sumario) Valores repetidosEssa etapa procura *features* ou colunas com um único valor repetido em todas as linhas ou observações.Uma feature com um único valor repetido em todas as linhas é desnecessária para o modelo, uma vez que não existe variação, logo, se não serve para o modelo de predição, deve ser removido do DataFrame.A função ```plot_features_with_same_value_for_all_observations``` busca por colunas nessa condição, e plota o nome das colunas.Existem 36 resultados relacionados com **exames de sangue**, e note que o total de colunas repetidas é 36. Isso porque o exame de sangue é coletado apenas uma vez por dia, não tendo assim um novo exame para se calcular a **diferença** entre a medição anterior e a atual, ou seja, os resultados relacionados com o exame de sangue com o sulfixo **_DIFF** podem ser removidos do DataFrame sem impactos negativos para a modelagem. ###Code plot_features_with_same_value_for_all_observations(df_2_without_nan) ###Output Nome das Colunas: -------------------- ALBUMIN_DIFF BE_ARTERIAL_DIFF BE_VENOUS_DIFF BIC_ARTERIAL_DIFF BIC_VENOUS_DIFF BILLIRUBIN_DIFF BLAST_DIFF CALCIUM_DIFF CREATININ_DIFF FFA_DIFF GGT_DIFF GLUCOSE_DIFF HEMATOCRITE_DIFF HEMOGLOBIN_DIFF INR_DIFF LACTATE_DIFF LEUKOCYTES_DIFF LINFOCITOS_DIFF NEUTROPHILES_DIFF P02_ARTERIAL_DIFF P02_VENOUS_DIFF PC02_ARTERIAL_DIFF PC02_VENOUS_DIFF PCR_DIFF PH_ARTERIAL_DIFF PH_VENOUS_DIFF PLATELETS_DIFF POTASSIUM_DIFF SAT02_ARTERIAL_DIFF SAT02_VENOUS_DIFF SODIUM_DIFF TGO_DIFF TGP_DIFF TTPA_DIFF UREA_DIFF DIMER_DIFF -------------------- Total: 36 ###Markdown A função ```drop_features_with_same_value_for_all_observations``` remove do DataFrame essas colunas apontadas pela célula anterior.**```df_3_without_same_value_col : pd.DataFrame```** é o DataFrame após a remoção das colunas com valores repetidos. ###Code df_3_without_same_value_col = drop_features_with_same_value_for_all_observations(df_2_without_nan, False) ###Output Total of dropped columns: 36 ###Markdown Confirmação de que todas as colunas com valores repetidos foram excluidas com sucesso. ###Code plot_features_with_same_value_for_all_observations(df_3_without_same_value_col) ###Output Nome das Colunas: -------------------- -------------------- Total: 0 ###Markdown -----------------Retornar ao [Sumário](sumario) Features duplicadasEssa etapa procura por *features* ou colunas que tem os valores idênticos a uma ou mais colunas.Features duplicadas são desnecessárias para o modelo de regressão, uma vez que não trazem nenhuma informação nova para o modelo, e devem ser removidas do DataFrame.A função ```plot_duplicated_features``` vai mostrar quais *features* são duplicadas. quando existem duas ou mais *features* identicas, a função deixa a primeira e retorna todas as demais, indicando quais devem ser excluidas. ###Code plot_duplicated_features(df_3_without_same_value_col) ###Output Nome das Colunas: -------------------- ALBUMIN_MEAN ALBUMIN_MIN ALBUMIN_MAX BE_ARTERIAL_MEAN BE_ARTERIAL_MIN BE_ARTERIAL_MAX BE_VENOUS_MEAN BE_VENOUS_MIN BE_VENOUS_MAX BIC_ARTERIAL_MEAN BIC_ARTERIAL_MIN BIC_ARTERIAL_MAX BIC_VENOUS_MEAN BIC_VENOUS_MIN BIC_VENOUS_MAX BILLIRUBIN_MEAN BILLIRUBIN_MIN BILLIRUBIN_MAX BLAST_MEAN BLAST_MIN BLAST_MAX CALCIUM_MEAN CALCIUM_MIN CALCIUM_MAX CREATININ_MEAN CREATININ_MIN CREATININ_MAX FFA_MEAN FFA_MIN FFA_MAX GGT_MEAN GGT_MIN GGT_MAX GLUCOSE_MEAN GLUCOSE_MIN GLUCOSE_MAX HEMATOCRITE_MEAN HEMATOCRITE_MIN HEMATOCRITE_MAX HEMOGLOBIN_MEAN HEMOGLOBIN_MIN HEMOGLOBIN_MAX INR_MEAN INR_MIN INR_MAX LACTATE_MEAN LACTATE_MIN LACTATE_MAX LEUKOCYTES_MEAN LEUKOCYTES_MIN LEUKOCYTES_MAX LINFOCITOS_MEAN LINFOCITOS_MIN LINFOCITOS_MAX NEUTROPHILES_MEAN NEUTROPHILES_MIN NEUTROPHILES_MAX P02_ARTERIAL_MEAN P02_ARTERIAL_MIN P02_ARTERIAL_MAX P02_VENOUS_MEAN P02_VENOUS_MIN P02_VENOUS_MAX PC02_ARTERIAL_MEAN PC02_ARTERIAL_MIN PC02_ARTERIAL_MAX PC02_VENOUS_MEAN PC02_VENOUS_MIN PC02_VENOUS_MAX PCR_MEAN PCR_MIN PCR_MAX PH_ARTERIAL_MEAN PH_ARTERIAL_MIN PH_ARTERIAL_MAX PH_VENOUS_MEAN PH_VENOUS_MIN PH_VENOUS_MAX PLATELETS_MEAN PLATELETS_MIN PLATELETS_MAX POTASSIUM_MEAN POTASSIUM_MIN POTASSIUM_MAX SAT02_ARTERIAL_MEAN SAT02_ARTERIAL_MIN SAT02_ARTERIAL_MAX SAT02_VENOUS_MEAN SAT02_VENOUS_MIN SAT02_VENOUS_MAX SODIUM_MEAN SODIUM_MIN SODIUM_MAX TGO_MEAN TGO_MIN TGO_MAX TGP_MEAN TGP_MIN TGP_MAX TTPA_MEAN TTPA_MIN TTPA_MAX UREA_MEAN UREA_MIN UREA_MAX DIMER_MEAN DIMER_MIN DIMER_MAX -------------------- Total: 108 ###Markdown Existem então **108** *features* duplicadas nesse DataFrame. Se é possivel observar que **108** é um multiplo de **36**, ou seja, os dados de sangue, por possuírem um único valor por visita de paciente (uma única medição de sangue por paciente), isso faz com que as características como: ```MIN```, ```MAX```, ```MEAN``` e ```MEDIAN``` sejam iguais. ###Code single_value_array = [5] print(f'Média: {np.mean(single_value_array)}\nMediana: {np.median(single_value_array)}\nMínimo: {np.min(single_value_array)}\nMáximo: {np.max(single_value_array)}') ###Output Média: 5.0 Mediana: 5.0 Mínimo: 5 Máximo: 5 ###Markdown É necessário então remover as colunas duplicadas, e para tal, foi criada a função ```drop_duplicated_features```, que elimina do DataFrame as features duplicadas.**```df_4_without_duplicated_features : pd.DataFrame```** é o DataFrame contendo os valores do ```df_3_without_same_value_col``` com as features duplicadas removidas. ###Code df_4_without_duplicated_features = drop_duplicated_features(df_3_without_same_value_col, False) ###Output Total dropped columns: 108 ###Markdown Confirmando que essas colunas foram removidas: ###Code plot_duplicated_features(df_4_without_duplicated_features) ###Output Nome das Colunas: -------------------- -------------------- Total: 0 ###Markdown -----------------Retornar ao [Sumário](sumario) Nome features exame de sangue Caso dos nomesCompunham o *raw data* 36 exames de sangue, como: ```[ALBUMIN, BE_ARTERIAL, BE_VENOUS, BIC_ARTERIAL]```.Cada exame tinham 5 features associadas, sendo ```[MEDIAN, MEAN, MIN, MAX e DIFF]```.Como o exame de sangue é feito somente uma vez, o valor da ```DIFF``` é irrelevante, bem como os valores de ```MAX```, ```MIN``` e ```MEAN``` que são iguais ao valor de ```MEDIAN```.Após todo o processamento que fizemos até aqui, a sobraram no DataFrame 36 features relacionadas com o exame de sangue, só que com o sulfixo ```_MEDIAN```. O problema é que é inapropriado apontar a Mediana de um conjunto contendo um único valor. Isso poderia confundir uma pessoa que está por avaliar o modelo, por esse motivo, os sulfixos ```_MEDIAN``` serão removidos dos nomes das features relacionadas com o exame de sangue. ###Code df_5_renamed = rename_portion_of_columns(df_4_without_duplicated_features, 13, (13+36), '_MEDIAN', '') df_5_renamed.columns[13: (13+36)] ###Output _____no_output_____ ###Markdown -----------------Retornar ao [Sumário](sumario) Seleção da Janela de tempoPara saber mais detalhes sobre a estratégia de seleção das janelas de tempo, veja o documento:[0.0_understanding_the_data.md](0.0_understanding_the_data.md), no tópico: **Quanto antes, melhor**. Removendo Entrada direta na UTIOs pacientes que entraram para o hospital e foram encaminhados diretamente para a UTI (contendo a Janela de tempo 0 - 2h com ICU == 1) serão removidas do DataFrame.**```df_6_icu_on_first_window: pd.DataFrame```** é a variável que vai conter os dados do DataFrame com após a remoção dos valores com ICU == 1 na primeira janela temporal. Para isso, será utilizada a função: ```drop_patient_moved_to_icu_on_first_window```. ###Code plot_patient_moved_to_icu_on_first_window(df_5_renamed) df_6_icu_on_first_window = drop_patient_moved_to_icu_on_first_window(df_5_renamed) ###Output Total dropped visit id: 32 ###Markdown confirmando se os casos de pacientes que entraram diretamente na UTI foram removidos ###Code plot_patient_moved_to_icu_on_first_window(df_6_icu_on_first_window) ###Output PATIENT_VISIT_IDENTIFIER: -------------------- -------------------- Total: 0 ###Markdown Seleção da Janela com ICU futuroFiltrar do ```df_6_icu_on_first_window``` os resultados relacionados com a primeira janela temporal de cada visita (0 - 2h).Precisa-se também manipular os resultados da ```ICU``` de maneira a indicar 1 se o paciente foi direcionado para a UTI em qualquer uma das demais janelas temporais.**```df_7_cleaned: pd.DataFrame```** é a variável que vai conter os dados limpos, após a seleção de janelas. A seleção de janelas será feita agrupando o DataFrame por ```PATIENT_VISIT_IDENTIFIER``` e aplicando a função ```prepare_window``` que serve justamente para selecionar a janela 1 e alterar o ICU para 1 na Janela 1 se em qualquer outra janela desse paciente o valor for 1. ###Code df_7_cleaned = df_6_icu_on_first_window.groupby("PATIENT_VISIT_IDENTIFIER", as_index=False).apply(prepare_window)\ .reset_index().drop(['level_0', 'level_1'], axis=1) df_7_cleaned.head() ###Output _____no_output_____ ###Markdown -----------------Retornar ao [Sumário](sumario) Salva dados limposApós a limpeza dos dados, é importante salvar esse arquivo no projeto, para que não seja necessário passar por todas essas etapas para produzir um arquivo com um DataFrame limpo.O DataFrame: ```df_7_cleaned``` será salvo como csv na pasta ```../data/interim/```. Isso porque os dados podem ainda sofrer alterações na etapa de análise das *features*. ###Code df_7_cleaned.to_csv('../data/interim/df_7_cleaned.csv') ###Output _____no_output_____ ###Markdown -----------------Retornar ao [Sumário](sumario) Conclusão -----------------Retornar ao [Sumário](sumario) ###Code df_7_cleaned.shape ###Output _____no_output_____
exercises/.ipynb_checkpoints/feature_sets-checkpoint.ipynb
###Markdown Copyright 2017 Google LLC. ###Code # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # https://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. ###Output _____no_output_____ ###Markdown Feature Sets **Learning Objective:** Create a minimal set of features that performs just as well as a more complex feature set So far, we've thrown all of our features into the model. Models with fewer features use fewer resources and are easier to maintain. Let's see if we can build a model on a minimal set of housing features that will perform equally as well as one that uses all the features in the data set. SetupAs before, let's load and prepare the California housing data. ###Code from __future__ import print_function import math from IPython import display from matplotlib import cm from matplotlib import gridspec from matplotlib import pyplot as plt import numpy as np import pandas as pd from sklearn import metrics import tensorflow as tf from tensorflow.python.data import Dataset tf.logging.set_verbosity(tf.logging.ERROR) pd.options.display.max_rows = 10 pd.options.display.float_format = '{:.1f}'.format california_housing_dataframe = pd.read_csv("https://download.mlcc.google.com/mledu-datasets/california_housing_train.csv", sep=",") california_housing_dataframe = california_housing_dataframe.reindex( np.random.permutation(california_housing_dataframe.index)) def preprocess_features(california_housing_dataframe): """Prepares input features from California housing data set. Args: california_housing_dataframe: A Pandas DataFrame expected to contain data from the California housing data set. Returns: A DataFrame that contains the features to be used for the model, including synthetic features. """ selected_features = california_housing_dataframe[ ["latitude", "longitude", "housing_median_age", "total_rooms", "total_bedrooms", "population", "households", "median_income"]] processed_features = selected_features.copy() # Create a synthetic feature. processed_features["rooms_per_person"] = ( california_housing_dataframe["total_rooms"] / california_housing_dataframe["population"]) return processed_features def preprocess_targets(california_housing_dataframe): """Prepares target features (i.e., labels) from California housing data set. Args: california_housing_dataframe: A Pandas DataFrame expected to contain data from the California housing data set. Returns: A DataFrame that contains the target feature. """ output_targets = pd.DataFrame() # Scale the target to be in units of thousands of dollars. output_targets["median_house_value"] = ( california_housing_dataframe["median_house_value"] / 1000.0) return output_targets # Choose the first 12000 (out of 17000) examples for training. training_examples = preprocess_features(california_housing_dataframe.head(12000)) training_targets = preprocess_targets(california_housing_dataframe.head(12000)) # Choose the last 5000 (out of 17000) examples for validation. validation_examples = preprocess_features(california_housing_dataframe.tail(5000)) validation_targets = preprocess_targets(california_housing_dataframe.tail(5000)) # Double-check that we've done the right thing. print("Training examples summary:") display.display(training_examples.describe()) print("Validation examples summary:") display.display(validation_examples.describe()) print("Training targets summary:") display.display(training_targets.describe()) print("Validation targets summary:") display.display(validation_targets.describe()) ###Output Training examples summary: ###Markdown Task 1: Develop a Good Feature Set**What's the best performance you can get with just 2 or 3 features?**A **correlation matrix** shows pairwise correlations, both for each feature compared to the target and for each feature compared to other features.Here, correlation is defined as the [Pearson correlation coefficient](https://en.wikipedia.org/wiki/Pearson_product-moment_correlation_coefficient). You don't have to understand the mathematical details for this exercise.Correlation values have the following meanings: * `-1.0`: perfect negative correlation * `0.0`: no correlation * `1.0`: perfect positive correlation ###Code correlation_dataframe = training_examples.copy() correlation_dataframe["target"] = training_targets["median_house_value"] correlation_dataframe.corr() ###Output _____no_output_____ ###Markdown Features that have strong positive or negative correlations with the target will add information to our model. We can use the correlation matrix to find such strongly correlated features.We'd also like to have features that aren't so strongly correlated with each other, so that they add independent information.Use this information to try removing features. You can also try developing additional synthetic features, such as ratios of two raw features.For convenience, we've included the training code from the previous exercise. ###Code def construct_feature_columns(input_features): """Construct the TensorFlow Feature Columns. Args: input_features: The names of the numerical input features to use. Returns: A set of feature columns """ return set([tf.feature_column.numeric_column(my_feature) for my_feature in input_features]) def my_input_fn(features, targets, batch_size=1, shuffle=True, num_epochs=None): """Trains a linear regression model. Args: features: pandas DataFrame of features targets: pandas DataFrame of targets batch_size: Size of batches to be passed to the model shuffle: True or False. Whether to shuffle the data. num_epochs: Number of epochs for which data should be repeated. None = repeat indefinitely Returns: Tuple of (features, labels) for next data batch """ # Convert pandas data into a dict of np arrays. features = {key:np.array(value) for key,value in dict(features).items()} # Construct a dataset, and configure batching/repeating. ds = Dataset.from_tensor_slices((features,targets)) # warning: 2GB limit ds = ds.batch(batch_size).repeat(num_epochs) # Shuffle the data, if specified. if shuffle: ds = ds.shuffle(10000) # Return the next batch of data. features, labels = ds.make_one_shot_iterator().get_next() return features, labels def train_model( learning_rate, steps, batch_size, training_examples, training_targets, validation_examples, validation_targets): """Trains a linear regression model. In addition to training, this function also prints training progress information, as well as a plot of the training and validation loss over time. Args: learning_rate: A `float`, the learning rate. steps: A non-zero `int`, the total number of training steps. A training step consists of a forward and backward pass using a single batch. batch_size: A non-zero `int`, the batch size. training_examples: A `DataFrame` containing one or more columns from `california_housing_dataframe` to use as input features for training. training_targets: A `DataFrame` containing exactly one column from `california_housing_dataframe` to use as target for training. validation_examples: A `DataFrame` containing one or more columns from `california_housing_dataframe` to use as input features for validation. validation_targets: A `DataFrame` containing exactly one column from `california_housing_dataframe` to use as target for validation. Returns: A `LinearRegressor` object trained on the training data. """ periods = 10 steps_per_period = steps / periods # Create a linear regressor object. my_optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate) my_optimizer = tf.contrib.estimator.clip_gradients_by_norm(my_optimizer, 5.0) linear_regressor = tf.estimator.LinearRegressor( feature_columns=construct_feature_columns(training_examples), optimizer=my_optimizer ) # Create input functions. training_input_fn = lambda: my_input_fn(training_examples, training_targets["median_house_value"], batch_size=batch_size) predict_training_input_fn = lambda: my_input_fn(training_examples, training_targets["median_house_value"], num_epochs=1, shuffle=False) predict_validation_input_fn = lambda: my_input_fn(validation_examples, validation_targets["median_house_value"], num_epochs=1, shuffle=False) # Train the model, but do so inside a loop so that we can periodically assess # loss metrics. print("Training model...") print("RMSE (on training data):") training_rmse = [] validation_rmse = [] for period in range (0, periods): # Train the model, starting from the prior state. linear_regressor.train( input_fn=training_input_fn, steps=steps_per_period, ) # Take a break and compute predictions. training_predictions = linear_regressor.predict(input_fn=predict_training_input_fn) training_predictions = np.array([item['predictions'][0] for item in training_predictions]) validation_predictions = linear_regressor.predict(input_fn=predict_validation_input_fn) validation_predictions = np.array([item['predictions'][0] for item in validation_predictions]) # Compute training and validation loss. training_root_mean_squared_error = math.sqrt( metrics.mean_squared_error(training_predictions, training_targets)) validation_root_mean_squared_error = math.sqrt( metrics.mean_squared_error(validation_predictions, validation_targets)) # Occasionally print the current loss. print(" period %02d : %0.2f" % (period, training_root_mean_squared_error)) # Add the loss metrics from this period to our list. training_rmse.append(training_root_mean_squared_error) validation_rmse.append(validation_root_mean_squared_error) print("Model training finished.") # Output a graph of loss metrics over periods. plt.ylabel("RMSE") plt.xlabel("Periods") plt.title("Root Mean Squared Error vs. Periods") plt.tight_layout() plt.plot(training_rmse, label="training") plt.plot(validation_rmse, label="validation") plt.legend() return linear_regressor ###Output _____no_output_____ ###Markdown Spend 5 minutes searching for a good set of features and training parameters. Then check the solution to see what we chose. Don't forget that different features may require different learning parameters. ###Code # # Your code here: add your features of choice as a list of quoted strings. # minimal_features = [ ] assert minimal_features, "You must select at least one feature!" minimal_training_examples = training_examples[minimal_features] minimal_validation_examples = validation_examples[minimal_features] # # Don't forget to adjust these parameters. # train_model( learning_rate=0.001, steps=500, batch_size=5, training_examples=minimal_training_examples, training_targets=training_targets, validation_examples=minimal_validation_examples, validation_targets=validation_targets) ###Output Training model... RMSE (on training data): period 00 : 228.83 period 01 : 221.17 period 02 : 213.61 period 03 : 206.16 period 04 : 198.84 period 05 : 191.66 period 06 : 184.63 period 07 : 177.78 period 08 : 171.13 period 09 : 164.70 Model training finished. ###Markdown SolutionClick below for a solution. ###Code minimal_features = [ "median_income", "latitude", ] minimal_training_examples = training_examples[minimal_features] minimal_validation_examples = validation_examples[minimal_features] _ = train_model( learning_rate=0.01, steps=500, batch_size=5, training_examples=minimal_training_examples, training_targets=training_targets, validation_examples=minimal_validation_examples, validation_targets=validation_targets) ###Output Training model... RMSE (on training data): period 00 : 165.75 period 01 : 126.77 period 02 : 116.79 period 03 : 115.96 period 04 : 115.84 period 05 : 114.99 period 06 : 114.37 period 07 : 115.16 period 08 : 113.80 period 09 : 112.85 Model training finished. ###Markdown Task 2: Make Better Use of LatitudePlotting `latitude` vs. `median_house_value` shows that there really isn't a linear relationship there.Instead, there are a couple of peaks, which roughly correspond to Los Angeles and San Francisco. ###Code plt.scatter(training_examples["latitude"], training_targets["median_house_value"]) ###Output _____no_output_____ ###Markdown **Try creating some synthetic features that do a better job with latitude.**For example, you could have a feature that maps `latitude` to a value of `|latitude - 38|`, and call this `distance_from_san_francisco`.Or you could break the space into 10 different buckets. `latitude_32_to_33`, `latitude_33_to_34`, etc., each showing a value of `1.0` if `latitude` is within that bucket range and a value of `0.0` otherwise.Use the correlation matrix to help guide development, and then add them to your model if you find something that looks good.What's the best validation performance you can get? ###Code # # YOUR CODE HERE: Train on a new data set that includes synthetic features based on latitude. # ###Output _____no_output_____ ###Markdown SolutionClick below for a solution. Aside from `latitude`, we'll also keep `median_income`, to compare with the previous results.We decided to bucketize the latitude. This is fairly straightforward in Pandas using `Series.apply`. ###Code def select_and_transform_features(source_df): LATITUDE_RANGES = zip(range(32, 44), range(33, 45)) selected_examples = pd.DataFrame() selected_examples["median_income"] = source_df["median_income"] for r in LATITUDE_RANGES: selected_examples["latitude_%d_to_%d" % r] = source_df["latitude"].apply( lambda l: 1.0 if l >= r[0] and l < r[1] else 0.0) return selected_examples selected_training_examples = select_and_transform_features(training_examples) selected_validation_examples = select_and_transform_features(validation_examples) _ = train_model( learning_rate=0.01, steps=500, batch_size=5, training_examples=selected_training_examples, training_targets=training_targets, validation_examples=selected_validation_examples, validation_targets=validation_targets) ###Output Training model... RMSE (on training data): period 00 : 227.63 period 01 : 217.43 period 02 : 207.31 period 03 : 197.31 period 04 : 187.41 period 05 : 177.65 period 06 : 168.05 period 07 : 158.65 period 08 : 149.45 period 09 : 140.52 Model training finished.
test_m1_madg.ipynb
###Markdown Dependencies ###Code !nvidia-smi !jupyter notebook list %env CUDA_VISIBLE_DEVICES=1 %matplotlib inline %load_ext autoreload %autoreload 2 import time from pathlib import Path import numpy as np import matplotlib.pyplot as plt import torch import torch.nn as nn import torch.optim as optim import torchvision import torchvision.transforms as transforms from models import tiramisu from models import tiramisu_bilinear from models import tiramisu_m1 from models import unet from datasets import deepglobe from datasets import madg from datasets import joint_transforms import utils.imgs import utils.training as train_utils # tensorboard from torch.utils.tensorboard import SummaryWriter ###Output _____no_output_____ ###Markdown Dataset- Download the DeepGlobe dataset from https://competitions.codalab.org/competitions/18467. Place it in datasets/deepglobe/dataset/train,test,valid- Download the Massachusetts Road Dataset from https://www.cs.toronto.edu/~vmnih/data/. Combine the training, validation, and test sets, process with `crop_dataset.ipynb` and place the output in datasets/maroads/dataset/map,sat- Run `combine_datasets.ipynb` to combine the two and output to datasets/madg ###Code run = "expM.1.madg.4" DEEPGLOBE_PATH = Path('datasets/', 'deepglobe/dataset') MADG_PATH = Path('datasets/', 'madg/dataset') RESULTS_PATH = Path('.results/') WEIGHTS_PATH = Path('.weights/') / run RUNS_PATH = Path('.runs/') RESULTS_PATH.mkdir(exist_ok=True) WEIGHTS_PATH.mkdir(exist_ok=True) RUNS_PATH.mkdir(exist_ok=True) batch_size = 1 # TODO: Should be `MAX_BATCH_PER_CARD * torch.cuda.device_count()` (which in this case is 1 assuming max of 1 batch per card) # resize = joint_transforms.JointRandomCrop((300, 300)) # normalize = transforms.Normalize(mean=deepglobe.mean, std=deepglobe.std) # normalize = transforms.Normalize(mean=madg.mean, std=madg.std) train_joint_transformer = transforms.Compose([ # resize, joint_transforms.JointRandomHorizontalFlip(), joint_transforms.JointRandomVerticalFlip(), joint_transforms.JointRandomRotate() ]) train_dset = madg.Madg(MADG_PATH, 'train', joint_transform=train_joint_transformer, transform=transforms.Compose([ # transforms.ColorJitter(brightness=.4,contrast=.4,saturation=.4), transforms.ToTensor(), # normalize, ])) train_loader = torch.utils.data.DataLoader( train_dset, batch_size=batch_size, shuffle=True) resize_joint_transformer = None val_dset = madg.Madg( MADG_PATH, 'valid', joint_transform=resize_joint_transformer, transform=transforms.Compose([ transforms.ToTensor(), # normalize ])) val_loader = torch.utils.data.DataLoader( val_dset, batch_size=batch_size, shuffle=False) test_dset = madg.Madg( MADG_PATH, 'test', joint_transform=resize_joint_transformer, transform=transforms.Compose([ transforms.ToTensor(), # normalize ])) test_loader = torch.utils.data.DataLoader( test_dset, batch_size=batch_size, shuffle=False) print("Train: %d" %len(train_loader.dataset)) print("Val: %d" %len(val_loader.dataset.imgs)) print("Test: %d" %len(test_loader.dataset.imgs)) # print("Classes: %d" % len(train_loader.dataset.classes)) print((iter(train_loader))) inputs, targets = next(iter(train_loader)) print("Inputs: ", inputs.size()) print("Targets: ", targets.size()) utils.imgs.view_image(inputs[0]) # utils.imgs.view_image(targets[0]) utils.imgs.view_annotated(targets[0]) print(inputs[0].max(),inputs[0].min()) print(targets[0].max(),targets[0].min()) ###Output Train: 4771 Val: 1182 Test: 1182 <torch.utils.data.dataloader._SingleProcessDataLoaderIter object at 0x7f42100832d0> Inputs: torch.Size([1, 3, 1024, 1024]) Targets: torch.Size([1, 1024, 1024]) ###Markdown Train ###Code LR = 1e-4 LR_DECAY = 0.995 DECAY_EVERY_N_EPOCHS = 1 N_EPOCHS = 1000 torch.cuda.manual_seed(0) from utils.bceloss import dice_bce_loss from loss.BCESSIM import BCESSIM model = tiramisu_m1.FCDenseNetSmall(n_classes=1, dropout_rate=0.2).cuda() optimizer = torch.optim.RMSprop(model.parameters(), lr=LR, weight_decay=1e-4) # criterion = dice_bce_loss() criterion = BCESSIM() # summary(model, input_size=inputs[0].shape) start_epoch = 0 !ls -l {WEIGHTS_PATH/'latest.th'} # start_epoch = train_utils.load_weights(model, (WEIGHTS_PATH/'latest.th')) + 1 start_epoch = train_utils.load_weights(model, (WEIGHTS_PATH/'weights-212-0.194-0.528.pth')) + 1 print("Starting from epoch", start_epoch) # Writer will output to ./runs/ directory by default writer = SummaryWriter(log_dir=(RUNS_PATH.as_posix() + "/" + "run" + str(run) + "/")) from torch.autograd import Variable debug_max_size=None # break # errors. Used to stop "run all" for epoch in range(start_epoch, N_EPOCHS+1): since = time.time() # ### Train ### # trn_loss, trn_err = train_utils.train( # model, train_loader, optimizer, criterion, epoch, debug_max_size=debug_max_size) # print('Epoch {:d}\nTrain - Loss: {:.4f}, Acc: {:.4f}'.format( # epoch, trn_loss, 1-trn_err)) # time_elapsed = time.time() - since # print('Train Time {:.0f}m {:.0f}s'.format( # time_elapsed // 60, time_elapsed % 60)) # ### Validation ### # val_loss, val_err, val_iou = train_utils.test(model, val_loader, criterion, epoch, debug_max_size=debug_max_size) # print('Tes - Loss: {:.4f} | Jacc: {:.4f} , Err: {:.4f}'.format(val_loss, val_iou, val_err)) # time_elapsed = time.time() - since # print('Total Time {:.0f}m {:.0f}s\n'.format( # time_elapsed // 60, time_elapsed % 60)) ### Test ### test_loss, test_err, test_iou = train_utils.test(model, test_loader, criterion, epoch, debug_max_size=debug_max_size) print('Tes - Loss: {:.4f} | Jacc: {:.4f} , Err: {:.4f}'.format(test_loss, test_iou, test_err)) time_elapsed = time.time() - since print('Total Time {:.0f}m {:.0f}s\n'.format( time_elapsed // 60, time_elapsed % 60)) # ### Checkpoint ### # train_utils.save_weights(model, epoch, val_loss, val_err, WEIGHTS_PATH=WEIGHTS_PATH) # # Log on tensorboard # writer.add_scalar('Loss/train', trn_loss, epoch) # writer.add_scalar('Loss/val', val_loss, epoch) # writer.add_scalar('Error/train', trn_err, epoch) # writer.add_scalar('Error/val', val_err, epoch) # writer.add_scalar('Accuracy/train', 1-trn_err, epoch) # writer.add_scalar('Accuracy/val', val_iou, epoch) # for param_group in optimizer.param_groups: # writer.add_scalar('Params/learning_rage', param_group['lr'], epoch) # # writer.add_scalar('params/learning_rate', optimizer.lr, epoch) # # writer.add_scalar('Params/no_optim', no_optim, epoch) # # log a sample image # # sample_images = [0,1030,281,623,636,655,1028,1353,2222,2224] # sample_images = [0,1030,281,623,636,655,1028,1000,1001,1002] # for i in sample_images: # inputs, targets, pred, loss, err, iou = train_utils.get_sample_predictions(model, val_loader, n=1, criterion=criterion, idx=i) # raw = model(inputs.cuda()).cpu() # img = torchvision.utils.make_grid(torch.stack([ # inputs[0], # targets[0].unsqueeze(0).expand(3,-1,-1).float(), # pred[0].unsqueeze(0).expand(3,-1,-1).float(), # raw[0].expand(3,-1,-1).float() # ]), normalize=True) # writer.add_image('sample_pred/val/' + str(i), img, epoch) # start_epoch = epoch ###Output /opt/anaconda3/lib/python3.7/site-packages/torch/nn/functional.py:2479: UserWarning: Default upsampling behavior when mode=bilinear is changed to align_corners=False since 0.4.0. Please specify align_corners=True if the old behavior is desired. See the documentation of nn.Upsample for details. "See the documentation of nn.Upsample for details.".format(mode)) ###Markdown Debug ###Code stats = train_utils.view_sample_predictions(model, val_loader, n=1, criterion=criterion) print("loss", "error", "jaccard") print(stats) # !pip install torchsummary from torchsummary import summary summary(model, input_size=inputs[0].shape) ###Output _____no_output_____
EVDemandModel_EVScenarios/RunningModel/supplement_scenarios.ipynb
###Markdown Supplementary Paper ScenariosThis notebook runs the scenarios used in the paper's Supplementary Information.Developed by Siobhan Powell, 2021. Updated in 2022. ###Code import pandas as pd import matplotlib.pyplot as plt import boto3 import numpy as np import pickle import time from speech_classes import SPEECh from speech_classes import SPEEChGeneralConfiguration from speech_classes import LoadProfile from speech_classes import Plotting from speech_classes import DataSetConfigurations ###Output _____no_output_____ ###Markdown Number of Home Chargers in each scenario: ###Code home_chargers = pd.DataFrame(np.zeros((4, 11)), index=['UniversalHome', 'HighHome', 'LowHome_HighWork', 'LowHome_LowWork'], columns=['CA', 'OR', 'WA', 'ID', 'MT', 'WY', 'NV', 'UT', 'CO', 'NM', 'AZ']) for scenario_name in ['UniversalHome', 'HighHome', 'LowHome_HighWork', 'LowHome_LowWork']: print(scenario_name) for state in ['CA', 'OR', 'WA', 'ID', 'MT', 'WY', 'NV', 'UT', 'CO', 'NM', 'AZ']: data = DataSetConfigurations(data_set='CP') speech = SPEECh(data=data, penetration_level=1.0, outside_california=True, states=[state]) speech.pa_ih(scenario=scenario_name) speech.pg_multiple_regions(region_type='State', region_value_list=[state]) tot = 0 for key in speech.p_abe_data.index: if 'home_l2' in key: tot += speech.p_abe_data.loc[key, 'p_abe'] home_chargers.loc[scenario_name, state] = tot * speech.num_evs home_chargers home_chargers['Total'] = home_chargers.sum(axis=1) home_chargers['Total'] ###Output _____no_output_____ ###Markdown Large Battery Case ###Code def run_100p_wecc_largebattery(scenario_name, remove_timers, utility_region, save_string, date, tz_aware=True): for weekday_string in ['weekday', 'weekend']: wecc_tot_evs = 0 state_list = ['CA', 'OR', 'WA', 'ID', 'MT', 'WY', 'NV', 'UT', 'CO', 'NM', 'AZ'] time_zones = {'CA':0, 'OR':0, 'WA':0, 'ID':1, 'MT':1, 'WY':1, 'NV':0, 'UT':1, 'CO':1, 'NM':1, 'AZ':1} state_results = {} total_load_dict = {key:np.zeros((1440,)) for key in ['Residential L1', 'Residential L2', 'MUD L2', 'Workplace L2', 'Public L2', 'Public DCFC']} total_load_segments = np.zeros((1440, 6)) for state in state_list: print('----------'+state+'----------') data = DataSetConfigurations(data_set='CP') speech = SPEECh(data=data, penetration_level=1.0, outside_california=True, states=[state]) speech.pa_ih(scenario=scenario_name) # Large Batteries Only: speech.pb_i(scenario='Equal') speech.pg_multiple_regions(region_type='State', region_value_list=[state]) config = SPEEChGeneralConfiguration(speech, remove_timers=remove_timers, utility_region=utility_region) config.run_all(verbose=False, weekday=weekday_string) state_results[state] = {'Speech':speech, 'Config':config} if tz_aware: if time_zones[state] == 0: for key in total_load_dict.keys(): total_load_dict[key] += config.total_load_dict[key] total_load_segments += config.total_load_segments else: # Put into California time for key in total_load_dict.keys(): tmp = np.copy(config.total_load_dict[key]) tmp2 = np.zeros((1440,)) tmp2[np.arange(0, 1440-60)] = tmp[np.arange(60, 1440)] tmp2[np.arange(1440-60, 1440)] = tmp[np.arange(0, 60)] total_load_dict[key] += np.copy(tmp2) tmp = np.copy(config.total_load_segments) tmp2 = np.zeros((1440,6)) tmp2[np.arange(0, 1440-60), :] = tmp[np.arange(60, 1440), :] tmp2[np.arange(1440-60, 1440), :] = tmp[np.arange(0, 60), :] total_load_segments += np.copy(tmp2) else: for key in total_load_dict.keys(): total_load_dict[key] += config.total_load_dict[key] total_load_segments += config.total_load_segments print('Total EVs: ', config.num_total_drivers) wecc_tot_evs += config.num_total_drivers if weekday_string == 'weekday': pd.DataFrame(config.total_load_dict).to_csv('Outputs/Supplement/'+save_string+'_'+str(state)+'_'+date+'.csv') else: pd.DataFrame(config.total_load_dict).to_csv('Outputs/Supplement/'+save_string+'_weekend_'+str(state)+'_'+date+'.csv') if weekday_string == 'weekday': pd.DataFrame(total_load_dict).to_csv('Outputs/Supplement/'+save_string+'_WECC_'+date+'.csv') else: pd.DataFrame(total_load_dict).to_csv('Outputs/Supplement/'+save_string+'_weekend_WECC_'+date+'.csv') print('Total EVs in WECC: ', wecc_tot_evs) return date = '20220506' run_100p_wecc_largebattery('UniversalHome', True, 'PGE', 'UniversalHome_LargeBatteryOnly_100p_NoTimers', date) run_100p_wecc_largebattery('HighHome', True, 'PGE', 'HighHome_LargeBatteryOnly_100p_NoTimers', date) run_100p_wecc_largebattery('LowHome_HighWork', True, 'PGE', 'LowHome_HighWork_LargeBatteryOnly_100p_NoTimers', date) run_100p_wecc_largebattery('LowHome_LowWork', True, 'PGE', 'LowHome_LowWork_LargeBatteryOnly_100p_NoTimers', date) ###Output ----------CA---------- Total EVs: 24234832 ----------OR---------- Total EVs: 2923172 ----------WA---------- Total EVs: 5280998 ----------ID---------- Total EVs: 1269502 ----------MT---------- Total EVs: 863108 ----------WY---------- Total EVs: 478703 ----------NV---------- Total EVs: 1879178 ----------UT---------- Total EVs: 1951222 ----------CO---------- Total EVs: 3977177 ----------NM---------- Total EVs: 1407939 ----------AZ---------- Total EVs: 4374941 Total EVs in WECC: 48640772 ----------CA---------- Total EVs: 24234832 ----------OR---------- Total EVs: 2923172 ----------WA---------- Total EVs: 5280998 ----------ID---------- Total EVs: 1269502 ----------MT---------- Total EVs: 863108 ----------WY---------- Total EVs: 478703 ----------NV---------- Total EVs: 1879178 ----------UT---------- Total EVs: 1951222 ----------CO---------- Total EVs: 3977177 ----------NM---------- Total EVs: 1407939 ----------AZ---------- Total EVs: 4374941 Total EVs in WECC: 48640772 ----------CA---------- ###Markdown Fast Charging Case ###Code data = DataSetConfigurations(data_set='CP') speech = SPEECh(data=data, penetration_level=1.0, outside_california=True, states=['OR']) speech.pa_ih(scenario='HighHome') speech.pg_multiple_regions(region_type='State', region_value_list=['OR']) config = SPEEChGeneralConfiguration(speech, remove_timers=True, utility_region='PGE') config.run_all(verbose=False, weekday='weekday') plots = Plotting(speech, config) plots.plot_single(config.total_load_segments, config.total_load_dict, save_str=None) import copy data = DataSetConfigurations(data_set='CP') speech = SPEECh(data=data, penetration_level=1.0, outside_california=True, states=['OR']) speech.pa_ih(scenario='HighHome') speech.pg_multiple_regions(region_type='State', region_value_list=['OR']) config = SPEEChGeneralConfiguration(speech, remove_timers=True, utility_region='PGE') for g in range(data.ng): for weekday in ['weekday', 'weekend']: config.group_configs[g].segment_session_numbers[weekday]['public_l3'] += config.group_configs[g].segment_session_numbers[weekday]['public_l2'] config.group_configs[g].segment_session_numbers[weekday]['public_l2'] = 0 if config.group_configs[g].segment_session_numbers[weekday]['public_l3'] > 0: if 'public_l3' not in config.group_configs[g].segment_gmms[weekday].keys(): # copy gmm from another group that has most fast charging within the same energy bin if g <= 20: target_g = 11 elif g <= 34: target_g = 27 elif g <= 48: target_g = 41 elif g <= 68: target_g = 55 elif g <= 93: target_g = 71 elif g <= 114: target_g = 94 else: target_g = 115 config.group_configs[g].segment_gmms[weekday]['public_l3'] = copy.deepcopy(config.group_configs[target_g].segment_gmms[weekday]['public_l3']) config.run_all(verbose=False, weekday='weekday') plots = Plotting(speech, config) plots.plot_single(config.total_load_segments, config.total_load_dict, save_str=None) def run_100p_wecc_fastcharging(scenario_name, remove_timers, utility_region, save_string, date, tz_aware=True): for weekday_string in ['weekday', 'weekend']: wecc_tot_evs = 0 state_list = ['CA', 'OR', 'WA', 'ID', 'MT', 'WY', 'NV', 'UT', 'CO', 'NM', 'AZ'] time_zones = {'CA':0, 'OR':0, 'WA':0, 'ID':1, 'MT':1, 'WY':1, 'NV':0, 'UT':1, 'CO':1, 'NM':1, 'AZ':1} state_results = {} total_load_dict = {key:np.zeros((1440,)) for key in ['Residential L1', 'Residential L2', 'MUD L2', 'Workplace L2', 'Public L2', 'Public DCFC']} total_load_segments = np.zeros((1440, 6)) for state in state_list: print('----------'+state+'----------') data = DataSetConfigurations(data_set='CP') speech = SPEECh(data=data, penetration_level=1.0, outside_california=True, states=[state]) speech.pa_ih(scenario=scenario_name) speech.pg_multiple_regions(region_type='State', region_value_list=[state]) config = SPEEChGeneralConfiguration(speech, remove_timers=remove_timers, utility_region=utility_region) # switch public l2 to fast charging for g in range(data.ng): for weekday in ['weekday', 'weekend']: config.group_configs[g].segment_session_numbers[weekday]['public_l3'] += config.group_configs[g].segment_session_numbers[weekday]['public_l2'] config.group_configs[g].segment_session_numbers[weekday]['public_l2'] = 0 if config.group_configs[g].segment_session_numbers[weekday]['public_l3'] > 0: if 'public_l3' not in config.group_configs[g].segment_gmms[weekday].keys(): # copy gmm from another group that has most fast charging within the same energy bin if g <= 20: target_g = 11 elif g <= 34: target_g = 27 elif g <= 48: target_g = 41 elif g <= 68: target_g = 55 elif g <= 93: target_g = 71 elif g <= 114: target_g = 94 else: target_g = 115 config.group_configs[g].segment_gmms[weekday]['public_l3'] = copy.deepcopy(config.group_configs[target_g].segment_gmms[weekday]['public_l3']) config.run_all(verbose=False, weekday=weekday_string) state_results[state] = {'Speech':speech, 'Config':config} if tz_aware: if time_zones[state] == 0: for key in total_load_dict.keys(): total_load_dict[key] += config.total_load_dict[key] total_load_segments += config.total_load_segments else: # Put into California time for key in total_load_dict.keys(): tmp = np.copy(config.total_load_dict[key]) tmp2 = np.zeros((1440,)) tmp2[np.arange(0, 1440-60)] = tmp[np.arange(60, 1440)] tmp2[np.arange(1440-60, 1440)] = tmp[np.arange(0, 60)] total_load_dict[key] += np.copy(tmp2) tmp = np.copy(config.total_load_segments) tmp2 = np.zeros((1440,6)) tmp2[np.arange(0, 1440-60), :] = tmp[np.arange(60, 1440), :] tmp2[np.arange(1440-60, 1440), :] = tmp[np.arange(0, 60), :] total_load_segments += np.copy(tmp2) else: for key in total_load_dict.keys(): total_load_dict[key] += config.total_load_dict[key] total_load_segments += config.total_load_segments print('Total EVs: ', config.num_total_drivers) wecc_tot_evs += config.num_total_drivers if weekday_string == 'weekday': pd.DataFrame(config.total_load_dict).to_csv('Outputs/Supplement/'+save_string+'_'+str(state)+'_'+date+'.csv') else: pd.DataFrame(config.total_load_dict).to_csv('Outputs/Supplement/'+save_string+'_weekend_'+str(state)+'_'+date+'.csv') if weekday_string == 'weekday': pd.DataFrame(total_load_dict).to_csv('Outputs/Supplement/'+save_string+'_WECC_'+date+'.csv') else: pd.DataFrame(total_load_dict).to_csv('Outputs/Supplement/'+save_string+'_weekend_WECC_'+date+'.csv') print('Total EVs in WECC: ', wecc_tot_evs) return date = '20220506' run_100p_wecc_fastcharging('UniversalHome', True, 'PGE', 'UniversalHome_FastCharging_100p_NoTimers', date) run_100p_wecc_fastcharging('HighHome', True, 'PGE', 'HighHome_FastCharging_100p_NoTimers', date) run_100p_wecc_fastcharging('LowHome_HighWork', True, 'PGE', 'LowHome_HighWork_FastCharging_100p_NoTimers', date) run_100p_wecc_fastcharging('LowHome_LowWork', True, 'PGE', 'LowHome_LowWork_FastCharging_100p_NoTimers', date) ###Output ----------CA---------- Total EVs: 24234832 ----------OR---------- Total EVs: 2923172 ----------WA---------- Total EVs: 5280998 ----------ID---------- Total EVs: 1269502 ----------MT---------- Total EVs: 863108 ----------WY---------- Total EVs: 478703 ----------NV---------- Total EVs: 1879178 ----------UT---------- Total EVs: 1951222 ----------CO---------- Total EVs: 3977177 ----------NM---------- Total EVs: 1407939 ----------AZ---------- Total EVs: 4374941 Total EVs in WECC: 48640772 ----------CA---------- Total EVs: 24234832 ----------OR---------- Total EVs: 2923172 ----------WA---------- Total EVs: 5280998 ----------ID---------- Total EVs: 1269502 ----------MT---------- Total EVs: 863108 ----------WY---------- Total EVs: 478703 ----------NV---------- Total EVs: 1879178 ----------UT---------- Total EVs: 1951222 ----------CO---------- Total EVs: 3977177 ----------NM---------- Total EVs: 1407939 ----------AZ---------- Total EVs: 4374941 Total EVs in WECC: 48640772 ----------CA---------- Total EVs: 24234832 ----------OR---------- Total EVs: 2923172 ----------WA---------- Total EVs: 5280998 ----------ID---------- Total EVs: 1269502 ----------MT---------- Total EVs: 863108 ----------WY---------- Total EVs: 478703 ----------NV---------- Total EVs: 1879178 ----------UT---------- Total EVs: 1951222 ----------CO---------- Total EVs: 3977177 ----------NM---------- Total EVs: 1407939 ----------AZ---------- Total EVs: 4374941 Total EVs in WECC: 48640772 ----------CA---------- Total EVs: 24234832 ----------OR---------- Total EVs: 2923172 ----------WA---------- Total EVs: 5280998 ----------ID---------- Total EVs: 1269502 ----------MT---------- Total EVs: 863108 ----------WY---------- Total EVs: 478703 ----------NV---------- Total EVs: 1879178 ----------UT---------- Total EVs: 1951222 ----------CO---------- Total EVs: 3977177 ----------NM---------- Total EVs: 1407939 ----------AZ---------- Total EVs: 4374941 Total EVs in WECC: 48640772 ----------CA---------- Total EVs: 24234832 ----------OR---------- Total EVs: 2923172 ----------WA---------- Total EVs: 5280998 ----------ID---------- Total EVs: 1269502 ----------MT---------- Total EVs: 863108 ----------WY---------- Total EVs: 478703 ----------NV---------- Total EVs: 1879178 ----------UT---------- Total EVs: 1951222 ----------CO---------- Total EVs: 3977177 ----------NM---------- Total EVs: 1407939 ----------AZ---------- Total EVs: 4374941 Total EVs in WECC: 48640772 ----------CA---------- Total EVs: 24234832 ----------OR---------- Total EVs: 2923172 ----------WA---------- Total EVs: 5280998 ----------ID---------- Total EVs: 1269502 ----------MT---------- Total EVs: 863108 ----------WY---------- Total EVs: 478703 ----------NV---------- Total EVs: 1879178 ----------UT---------- Total EVs: 1951222 ----------CO---------- Total EVs: 3977177 ----------NM---------- Total EVs: 1407939 ----------AZ---------- Total EVs: 4374941 Total EVs in WECC: 48640772 ----------CA---------- Total EVs: 24234832 ----------OR---------- Total EVs: 2923172 ----------WA---------- Total EVs: 5280998 ----------ID---------- Total EVs: 1269502 ----------MT---------- Total EVs: 863108 ----------WY---------- Total EVs: 478703 ----------NV---------- Total EVs: 1879178 ----------UT---------- Total EVs: 1951222 ----------CO---------- Total EVs: 3977177 ----------NM---------- Total EVs: 1407939 ----------AZ---------- Total EVs: 4374941 Total EVs in WECC: 48640772 ----------CA---------- Total EVs: 24234832 ----------OR---------- Total EVs: 2923172 ----------WA---------- Total EVs: 5280998 ----------ID---------- Total EVs: 1269502 ----------MT---------- Total EVs: 863108 ----------WY---------- Total EVs: 478703 ----------NV---------- Total EVs: 1879178 ----------UT---------- Total EVs: 1951222 ----------CO---------- Total EVs: 3977177 ----------NM---------- Total EVs: 1407939 ----------AZ---------- Total EVs: 4374941 Total EVs in WECC: 48640772
seminar_skript/Regression_Techniques.ipynb
###Markdown Lineare RegressionIn der nachfolgenden Zelle werden zuerst Daten geladen, die zur Veranschaulichung der linearen Regression dienen.Anschliessend wird ein lineares Modell mit Hilfe der der Klasse Lineare Regression aus `sklearn.linear_model` gerechnet. Die Vorhersage (d.h. die Geradengleichung) ergibt sich aus den Koeffizienten durch $y = a + bX$. ###Code from sklearn.linear_model import LinearRegression import numpy as np import matplotlib.pyplot as plt y = np.load('/home/martin/python/fhnw_lecture/scripts/regression_y.pickle.npy') X = np.load('/home/martin/python/fhnw_lecture/scripts/regression_X.pickle.npy') model = LinearRegression() model.fit(X, y) y_hat = model.coef_ * X + model.intercept_ ###Output _____no_output_____ ###Markdown Warum wird für $\mathbf{X}$ immer ein Grossbuchstabe verwendet und für $\mathbf{y}$ ein kleiner Buchstabe ?Die Matrix der Variablen X wird gross geschrieben, da in Matrix-Notation Matrizen immer mit grossen Buchstaben bezeichnet werden, Vektoren - so wie die abhängige Variable y - werden mit kleinen Buchstaben benannt. ###Code f = plt.figure(figsize=(4, 4), dpi=120) plt.title(label='regression line, residues', fontdict={'fontsize':20}) axes = f.add_subplot(111) axes.plot(X, y, 'ro', X, y_hat) #axes = plt.gca() axes.set_ylim([np.min(y)-5, np.max(y) +5]) for i in range(len(y)): plt.plot((X[i, 0], X[i, 0]), (y[i], y_hat[i])) axes.set_xlabel('X') axes.set_ylabel('Y') axes.annotate('$y$', xy=(X[-3, 0], y[-3, 0]), xycoords='data', xytext=(X[-3, 0] - 1.5, y[-3, 0] + 1), textcoords='data', size = 20, arrowprops=dict(arrowstyle="->")) axes.annotate('$\hat{y}$', xy=(X[-3, 0], y_hat[-3, 0]), xycoords='data', xytext=(X[-3, 0] - 1.5, y_hat[-3, 0] + 1), textcoords='data', size = 20, arrowprops=dict(arrowstyle="->")) axes.annotate('$\hat{y} = a + bX$', xy=(X[3, 0] + 0.5, model.coef_ * (X[3, 0] + 0.5) + model.intercept_), xycoords='data', xytext=(X[3, 0] + 0.5, 55), textcoords='data', horizontalalignment = 'center', size = 20, arrowprops=dict(arrowstyle="->")) plt.show() #plt.close('all') ###Output _____no_output_____ ###Markdown Der Plot zeigt die berechnete Regressionsgerade, sowie die Abweichungen (die Fehler) der wirklichen Messwerte von dieser Geraden. Diese Abweichungen werden als __Residuen__ bezeichnet, weil es der Anteil der gemessenen Werte ist, der “übrig bleibt”, d.h. nicht durch das Modell erklärt werden kann. Vorhergesagte Variablen werden meist mit einem Dach (Hut) bezeichnet, sowie $\hat{y}$. Analytische Herleitung der Parameter der Linearen RegressionAllgemein kann man den Nullpunkt einer quadratischen Funktion bestimmen, indem man ihre erste Ableitung gleich $0$ setzt. Die erste Ableitung gibt die Steigung der Funktion an. In der Physik ist dies of die Beschleunigung. Die Steigung ist am Minimum der Funktion schliesslich $0$. Man beachte, dass quadratische Funktionen immer nur einen Maximalwert haben können.Nachfolgend ist dieser Sachverhalt für die quadratische Funktion $f(x) = (x-1)^2$ dargestellt. Die Ableitung$2x-2$ ist ebenfalls eingetragen. Bei dem Minimum der Funktion ist die erste Ableitung gleich $0$ (die Stelle an der der Funktionsgraph, der der ersten Ableitung und die rote, horizontale Linie sich schneiden). ###Code Image('../images/first_derivative.png', height= 280, width=280) # <img alt="taken from homepage of 20 newsgroups" caption="The different categories of the newsgroup posts" # id="20_newsgroups" src="../images/first_derivative.png" width="320" height="320"> ###Output _____no_output_____ ###Markdown Die Parameter einer linearen Regression können analytisch berechnet werden. Dazu wird der quadrierte Fehler $(y_i-\hat{y}_i)^2$ über alle Messwerte aufsummiert. Diese Summe wird nach den Parametern abgeleitet und gleich $0$ gesetzt. Somit erhält man die Stelle an der die quadratische Funktion keine Steigung (erste Ableitung ist Steigung) hat. Weil eine quadratische Funktion als einzige Nullstelle der Steigung ein Minimum hat, erhalten wir somit die Parameter an dem Minimum unserer quadratischen Fehlerfunktion. derivative of the error term $(y - \hat{y})^2$:* für $\hat{y}$ können wir auch schreiben: $a + b\cdot x$, dies ist die Vorhersage mit Hilfe der Regression-Gerade (der Geraden-Gleichung):$$\sum_i^{n}(y_i - \hat{y_i})^2 = \sum_i^{n}[y_i - (a + b\cdot x_i)]^{2}$$* wir leiten diese Fehler-Funktion nach $a$ ab und setzen diese erste Ableitung gleich $0$ (Hierbei wird die Kettenregel verwendet):\begin{align*}\frac{\delta \sum_i^{n}(y_i - \hat{y_i})^2}{\delta a} = -2\sum_i^{n}y_i + 2b\sum_i^{n}x_i + 2na =& 0\\2na =& 2\sum_i^{n}y_i - 2b\sum_i^{n}x_i\\ a =& \frac{2\sum_i^{n}y_i}{2n} - \frac{2b\sum_i^{n}x_i}{2n}\end{align*}* die Summe über alle $x_i$ geteilt durch $n$ -- die Anzahl aller Beobachtungen -- ergibt den Mittelwert $\bar{x}$, gleiches gilt für $\bar{y}$:$$a = \bar{y} - b\bar{x}$$* die Lösung für $b$ ergibt sich analog; hier ersetzen wir $a$ mit obigen Ergebnis und erhalten:$$ b = \frac{\frac{1}{n}\sum_i^n(x_i - \bar{x})(y_i - \bar{y})}{\frac{1}{n}\sum_i^n (x_i - \bar{x})^2} = \frac{\text{cov}_{xy}}{\text{var}_x}$$* Vereinfacht ist die Former: Kovarianz der beiden Variablen $x$ und $y$ geteilt durch die Varianz von $x$.Nachfolgend wird demonstriert, wie die hergeleiteten Formeln, in python angewendet dieselben Parameter-Schätzer ergeben wie die aus der Klasse `LineareRegression` aus `sklearn.linear_model`. Dies soll einfach nur demonstrieren, dass die alles ganz leicht zu rechnen ist und keiner komplizierten Algorithmen bedarf. ###Code # we can easily verify these results print(f'the parameter b is the coefficient of the linear model {model.coef_}') print(f'the parameter a is called the intercept of the model because it indicates\n where the regression line intercepts the y-axis at x=0 {model.intercept_}') cov_xy =(1/X.shape[0]) * np.dot((X - np.mean(X)).T,y - np.mean(y))[0][0] var_x = (1/X.shape[0]) * np.dot((X - np.mean(X)).T,X - np.mean(X))[0][0] b = cov_xy/var_x a = np.mean(y)-b*np.mean(X) print(f'\nour self-computed b parameter is: {b}') print(f'our self-computed a parameter is: {a}') ###Output the parameter b is the coefficient of the linear model [[8.07912445]] the parameter a is called the intercept of the model because it indicates where the regression line intercepts the y-axis at x=0 [-8.49032154] our self-computed b parameter is: 8.079124453577005 our self-computed a parameter is: -8.490321540681798 ###Markdown multivariate case: more than one x variableFür Multivariate Lineare Regression kann die Schreibweise mit Matrizen zusammengefasst werden. Dafür kann es lohnend sein, sich die Matrizen-Multiplikation noch einmal kurz anzusehen. \begin{align*} y_1&=a+b_1\cdot x_{11}+b_2\cdot x_{21}+\cdots + b_p\cdot x_{p1}\\ y_2&=a+b_1\cdot x_{12}+b_2\cdot x_{22}+\cdots + b_p\cdot x_{p2}\\ \ldots& \ldots\\ y_i&=a+b_1\cdot x_{1i}+b_2\cdot x_{2i}+\cdots + b_p\cdot x_{pi}\\\end{align*}\begin{equation*} \begin{bmatrix} y_1\\ y_2\\ . \\ . \\ . \\ y_i \end{bmatrix} = a+ \begin{bmatrix} x_{11} & x_{21} & x_{31} & \ldots & x_{p1}\\ x_{12} & x_{22} & x_{32} & \ldots & x_{p2}\\ \ldots&\ldots&\ldots&\ldots&\ldots\\ \ldots&\ldots&\ldots&\ldots&\ldots\\ \ldots&\ldots&\ldots&\ldots&\ldots\\ x_{1i} & x_{2i} & x_{3i} & \ldots & x_{pi}\\ \end{bmatrix} \cdot \begin{bmatrix} b_1\\ b_2\\ .\\ .\\ .\\ b_p \end{bmatrix}\end{equation*} Den konstanten inercept Term ($a$) können wir mit in den Vektor der Parameter $\mathbf{b}$ aufnehmen, indem wir in $\mathbf{X}$ eine Einser-Spalte hinzufügen. Somit wird die Schreibweise sehr kompakt und der intercept $a$ wird nicht mehr explizit aufgeführt: \begin{equation*} \begin{bmatrix} y_1\\ y_2\\ . \\ . \\ . \\ y_i \end{bmatrix} = \begin{bmatrix} 1& x_{11} & x_{21} & x_{31} & \ldots & x_{p1}\\ 1 & x_{12} & x_{22} & x_{32} & \ldots & x_{p2}\\ &\ldots&\ldots&\ldots&\ldots&\ldots\\ &\ldots&\ldots&\ldots&\ldots&\ldots\\ 1& x_{1i} & x_{2i} & x_{3i} & \ldots & x_{pi} \end{bmatrix} \cdot \begin{bmatrix} a\\ b_1\\ b_2\\ .\\ .\\ b_p \end{bmatrix} \end{equation*} In Matrizen-Schreibweise können wir jetzt einfach schreiben:$\mathbf{y} = \mathbf{X}\mathbf{b}$ derivation of $\mathbf{\text{b}}$ for the matrix notationAnschliessend wird die Berechnung der Parameter der Multivariaten Regression in Matrizen-Schreibweise erläutert. Konzeptionell ist dies nicht vom univariaten Fall verschieden. Diese Formel wird nur hergeleitet um demonstrieren zu können, wie das Ergebnis der expliziten Berechnung in Python mit dem aus der sklearn Klasse `LinearRegression` übereinstimmt. * we expand the error term: \begin{align*} \text{min}=&(\mathbf{y}-\hat{\mathbf{y}})^2=(\mathbf{y}-\mathbf{X}\mathbf{b})'(\mathbf{y}-\mathbf{X}\mathbf{b})=\\ &(\mathbf{y}'-\mathbf{b}'\mathbf{X}')(\mathbf{y}-\mathbf{X}\mathbf{b})=\\ &\mathbf{y}'\mathbf{y}-\mathbf{b}'\mathbf{X}'\mathbf{y}-\mathbf{y}' \mathbf{X}\mathbf{b}+\mathbf{b}'\mathbf{X}'\mathbf{X}\mathbf{b}=\\ &\mathbf{y}'\mathbf{y}-2\mathbf{b}'\mathbf{X}'\mathbf{y}+\mathbf{b}'\mathbf{X}' \mathbf{X}\mathbf{b}\\ \end{align*} * derivative of the error term with respect to $\mathbf{b}$* we set the result equal to zero and solve for $\mathbf{b}$ \begin{align*} \frac{\delta}{\delta \mathbf{b}}=&-2\mathbf{X}'\mathbf{y}+2\mathbf{X}'\mathbf{X}\mathbf{b}=0\\ 2\mathbf{X}'\mathbf{X}\mathbf{b}=&2\mathbf{X}'\mathbf{y}\\ \mathbf{b}=&(\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\mathbf{y}\quad \end{align*} Hierbei bedarf es der Inversion des Kreuzproduktes der Variablen-Matrix $(\mathbf{X}'\mathbf{X})^{-1}$. Die Matrizen-Inversion ist für grosse Anzahl von Variablen mathematisch sehr aufwändig und kann unter Umständen zu Ungenauigkeiten führen. In der Vergangenheit wurde viel an Algorithmen geforscht um die Inversion schneller und stabiler zu machen. Oftmals stehen Fehlermeldungen in Zusammenhang mit diesem Berechnungsschritt. Polynomial regression as an example for more than one variableUm einfach Multivariate Lineare Regression an einem Beispiel zeigen zu können wird die quadratische Regression (ein Spezial-Fall der Multivariaten Regression) eingeführt. Eine neue Variable entsteht durch das Quadrieren der bisherigen univiaraten Variable x. Das Praktische ist, dass sich der Sachverhalt der Multivariaten Regression noch immer sehr schön 2-dimensional darstellen lässt. $y = a + b_1 x + b_2 x^2$Hier ist zu beachten:* wir haben jetzt zwei Variablen und können folglich unsere Formel in Matrizen-Schreibweise anwenden* mehr Variablen führen hoffentlich zu einem besseren Modell* durch den quadratischen Term ist die resultierende Regressions-Funktion keine Gerade mehr.__Der Ausdruck "linear" in Linearer Regression bedeutet dass die Funktion linear in den Parametern $a, \mathbf{b}_\mathbf{1}, \mathbf{b}_\mathbf{2}$ ist. Für alle Werte einer Variablen $\mathbf{x_1}$ gilt der gleiche Parameter $\mathbf{b_1}$.Es bedeutet nicht, dass die Regressions-Funktion durch eine gerade Linie gegeben ist!__Nachfolgend fügen wir die weitere Variable durch Quadrieren der bisherigen Variable hinzu und berechnen abermals das Lineare Modell aus `sklearn.linear_model`. ###Code from numpy.linalg import inv # polynomial y = np.load('/home/martin/python/fhnw_lecture/scripts/regression_y.pickle.npy') X = np.load('/home/martin/python/fhnw_lecture/scripts/regression_X.pickle.npy') # underdetermined, ill-posed: infinitely many solutions X = np.c_[X, X**2] # the x (small x) is just for plotting purpose model.fit(X, y) y_hat = np.dot(x , model.coef_.T) + model.intercept_ f = plt.figure(figsize=(5, 5), dpi=100) plt.title(label='quadratic regression', fontdict={'fontsize':20}) axes = f.add_subplot(111) axes.plot(X[:,0], y, 'ro', x[:,0], y_hat.reshape((-1,))) #axes = plt.gca() axes.set_ylim([np.min(y)-5, np.max(y) +5]) ###Output _____no_output_____ ###Markdown Jetzt berechnen wir die Parameter der Multiplen Linearen Regression mit Hilfe der hergeleiteten Formeln. Hierfür fügen wir zu den bisherigen Variablen $x$ und $x^2$ noch eine Einser-Spalte für den intercpet ein. `np.dot` berechnet das dot-product zweier Variablen. Um das Kreuzprodukt von $\mathbf{X}$ berechnen zu können, muss eine der beiden Matrizen transponiert werden. Dies geschieht durch `.T`.`inv` invertiert das Kreuzprodukt.`coefs = np.dot(np.dot(inv(np.dot(X_intercept.T,X_intercept)),X_intercept.T),y)` ist gleichbedeutend mit:\begin{equation*}\mathbf{b}=(\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\mathbf{y}\end{equation*} ###Code # again we can compare the parameters of the model with those resulting from # our derived equation: # b=(X'X)^{-1} X'y from numpy.linalg import inv # first we have to add the intercept into our X-Variable; we rename it X_intercept X_intercept = np.c_[np.ones(X.shape[0]), X] coefs = np.dot(np.dot(inv(np.dot(X_intercept.T,X_intercept)),X_intercept.T),y) print(f'the parameter b is the coefficient of the linear model {model.coef_}') print(f'the parameter a is called the intercept of the model because it indicates\n where the regression line intercepts the y-axis at x=0 {model.intercept_}') print(f'our coefs already include the intercept: {coefs}') ###Output the parameter b is the coefficient of the linear model [[-12.14930516 1.68570247]] the parameter a is called the intercept of the model because it indicates where the regression line intercepts the y-axis at x=0 [35.33794262] our coefs already include the intercept: [[ 35.33794262] [-12.14930516] [ 1.68570247]] ###Markdown OverfittingNun wird diese Vorgehensweise für weitere Terme höherer Ordnung angewendet. Graphisch lässt sich zeigen, dass die Anpassung des Modells an die Daten immer besser wird, die Vorhersage für __neue Datenpunkte__ aber sehr schlecht sein dürfte. Das Polynom hat an vielen Stellen Schlenker und absurde Kurven eingebaut. Dies ist ein erstes Beispiel für __“overfitting”__. Einen ‘perfekten’ fit erhält man, wenn man genausoviele Paramter (10 Steigunskoeffizienten + intercept) hat wie Daten-Messpunkte. The important points to note here:* the fit to our empirical y-values gets better* at the same time, the regression line starts behaving strangly* the predictions made by the regression line in between the empirical y-values are grossly wrong: this is an example of __overfitting__ ###Code y = np.load('/home/martin/python/fhnw_lecture/scripts/regression_y.pickle.npy') X = np.load('/home/martin/python/fhnw_lecture/scripts/regression_X.pickle.npy') # underdetermined, ill-posed: infinitely many solutions X = np.c_[X, X**2, X**3, X**4, X**5, X**6, X**7, X**8, X**9] x = np.arange(-1, 12, 0.05).reshape((-1, 1)) x = np.c_[x, x**2, x**3, x**4, x**5, x**6, x**7, x**8, x**9] model.fit(X, y) y_hat = np.dot(x , model.coef_.T) + model.intercept_ f = plt.figure(figsize=(5, 5), dpi=100) plt.title(label='regression line for polynome of 9th degree', fontdict={'fontsize':20}) axes = f.add_subplot(111) axes.plot(X[:,0], y, 'ro', x[:,0], y_hat.reshape((-1,))) #axes = plt.gca() axes.set_ylim([np.min(y)-10, np.max(y) +10]) ###Output _____no_output_____ ###Markdown perfect fit: as many variables as data samplesA perfect fit is possible as is demonstrated next. We have as many variables (terms derived from x) as observations (data points). So for each data point we have a variable to accommodate it.__Note__, that a perfect fit is achieved with 10 variables + intercept. The intercept is also a parameter and in this case the number of observations $n$ equals the number of variables $p$, i.e. $p=n$. ###Code y = np.load('/home/martin/python/fhnw_lecture/scripts/regression_y.pickle.npy') X = np.load('/home/martin/python/fhnw_lecture/scripts/regression_X.pickle.npy') # underdetermined, ill-posed: infinitely many solutions X = np.c_[X, X**2, X**3, X**4, X**5, X**6, X**7, X**8, X**9, X**10] x = np.arange(-1, 12, 0.05).reshape((-1, 1)) x = np.c_[x, x**2, x**3, x**4, x**5, x**6, x**7, x**8, x**9, x**10] model.fit(X, y) y_hat = np.dot(x , model.coef_.T) + model.intercept_ print(f'the intercept and the coefficients are: {model.intercept_}, {model.coef_}') f = plt.figure(figsize=(5, 5), dpi=100) plt.title(label='regression line for polynome of 10th degree', fontdict={'fontsize':20}) axes = f.add_subplot(111) axes.plot(X[:,0], y, 'ro', x[:,0], y_hat.reshape((-1,))) #axes = plt.gca() axes.set_ylim([np.min(y)-10, np.max(y) +20]) ###Output _____no_output_____ ###Markdown What happens if we have more variables than data points?Gibt es mehr Parameter als Datenpunkte, existieren unendlich viele Lösungen und das Problem ist nicht mehr eindeutig lösbar. Früher gelang die Inversion des Kreuzproduktes der Variablen $\mathbf{X}'\mathbf{X}$ nicht. Mittlerweile gibt es Näherungsverfahren, die dennoch Ergebnisse liefern - wenn auch sehr Ungenaue.Mittlerweile gibt es aber mathematische Näherungsverfahren die es ermöglichen auch singuläre Matrizen zu invertieren.`numpy` verwendet hierfür die sogenannte LU-decomposition.One way to see in python that the solution is erroneous is to use the `scipy.linalg.solve` package and solve for the matix S that solves $(\mathbf{X}'\mathbf{X})^{-1} \mathbf{S} = \mathbf{I}$. $\mathbf{I}$ is called the eye-matrix wih 1s in the diagonale and zeros otherwise:$$\mathbf{I}=\left[\begin{array}{ccc} 1 & \cdots & 0 \\ \vdots & \ddots & \vdots \\ 0 & \cdots & 1\end{array}\right]$$Die entscheidende Zeile im nachfolgenden Code ist:`S = solve(inv(np.dot(X.T, X)), np.eye(13))`Sie besagt: gib mir die Matrix $\mathbf{S}$, die multipliziert mit $(\mathbf{X}'\mathbf{X})^{-1}$ die Matrix $\mathbf{I}$ gibt.Für unseren Fall von mehr Variablen als Beobachtungspunkten werden wir gewarnt, dass das Ergebnis falsch sein könnte. Mit älteren Mathematik- oder Statistik-Programmen ist dies überhaupt nicht möglich. ###Code warnings.filterwarnings("default") from numpy.linalg import inv from scipy.linalg import solve model = LinearRegression() y = np.load('/home/martin/python/fhnw_lecture/scripts/regression_y.pickle.npy') X = np.load('/home/martin/python/fhnw_lecture/scripts/regression_X.pickle.npy') # underdetermined, ill-posed: infinitely many solutions X = np.c_[X, X**2, X**3, X**4, X**5, X**6, X**7, X**8, X**9, X**10, X**11, X**12, X**13] # this should give at least a warning, because matrix inversion as done above is not possible # any more, due to singular covariance matrix [X'X] model.fit(X, y) #y_hat = np.dot(x , model.coef_.T) + model.intercept_ S = solve(inv(np.dot(X.T, X)), np.eye(13)) ###Output /home/martin/miniconda3/lib/python3.6/site-packages/ipykernel_launcher.py:15: LinAlgWarning: Ill-conditioned matrix (rcond=3.8573e-21): result may not be accurate. from ipykernel import kernelapp as app ###Markdown statistical package RIn der statistischen Programmiersprache R wird keine Warnung herausgegeben. Es werden einfach nur soviele Koeffizienten (intercept ist auch ein Koeffizient) berechnet, wie möglich ist. Alle weiteren Koeffizienten sind `NA`. ###Code warnings.filterwarnings("ignore") Image("../images/R_inverse_example.png") # <img alt="taken from homepage of 20 newsgroups" caption="The different categories of the newsgroup posts" id="20_newsgroups" src="../images/R_inverse_example.png" width="640" height="640"> ###Output _____no_output_____ ###Markdown Dealing with overfittingWie wir gesehen haben tendiert klassische Lineare Regression zu 'overfitting' sobald es wenige Datenpunkte gibt und mehrere Koeffizienten berechnet werden. Eine Lösung für dieses Problem ist, die Koeffizienten $b_1, b_2, b_3, \ldots$ kleiner zu machen. Dies kann erreicht werden, wenn der Fehler der Regression mit grösseren Koeffizienten auch grösser wird. Um nun das Minimum der Fehlerfunktion zu finden ist ein probates Mittel, die Koeffizienten kleiner zu machen und somit implizit 'overfitting' zu verhindern.Parameter können jetzt nur noch sehr gross werden, wenn dadurch gleichzeitig der Fehler stark reduziert werden kann.Nachfolgend wird ein Strafterm ('penalty') für grosse Parameter eingeführt. Im Falle der Ridge-Regression gehen die Koeffizienten quadriert in die Fehlerfunktion mit ein. Der Gewichtungsfaktor $\lambda$ bestimmt die Höhe des Strafterms und ist ein zusätzlicher Parameter für den -- je nach Datensatz -- ein optimaler Wert gefunden werden muss. Ridge regressionRemember this formula:\begin{equation*}\sum_i^{n}(y_i - \hat{y_i})^2 = \sum_i^{n}[y_i - (a + b\cdot x_i)]^{2}\end{equation*}To make the error term bigger, we could simply add $\lambda\cdot b^2$ to the error:\begin{equation*}\sum_i^{n}(y_i - \hat{y_i})^2 + \lambda b^2= \sum_i^{n}[y_i - (a + b\cdot x_i)]^{2}+ \lambda b^2\end{equation*}The parameter $\lambda$ is for scaling the amount of shrinkage.Die beiden Ausdrücke \begin{equation}\sum_i^{n}[y_i - (a + b\cdot x_i)]^{2}\label{eq:fehler}\end{equation} und \begin{equation}\lambda b^2\label{eq:ridge_error}\end{equation} sind wie Antagonisten. Der Koeffizient $b$ darf nur gross werden, wenn er es vermag $\eqref{eq:fehler}$ stark zu verkleinern, so dass der Zugewinn in $\eqref{eq:fehler}$ den Strafterm in $\eqref{eq:ridge_error}$ überwiegt.For two variables we can write:\begin{equation*}\sum_i^{n}(y_i - \hat{y_i})^2 + \lambda b_1^2 + \lambda b_2^2= \sum_i^{n}[y_i - (a + b_1\cdot x_{i1} + b_2\cdot x_{i2})]^{2}+ \lambda b_1^2 + \lambda b_2^2\end{equation*}And in matrix notation for an arbitrary number of variables:\begin{align*} \text{min}=&(\mathbf{y}-\hat{\mathbf{y}})^2 + \lambda \mathbf{b}^2=(\mathbf{y}-\mathbf{X}\mathbf{b})'(\mathbf{y}-\mathbf{X}\mathbf{b}) + \lambda \mathbf{b}'\mathbf{b}\end{align*} Interessanterweise gibt es für diesen Fall ebenfalls eine exakte analytische Lösung. Allerdings haben wir den intercept Koeffizienten $a$ mit in $\mathbf{b}$ aufgenommen und die zusätzliche Spalte mit lauter Einsern in $\mathbf{X}$ hinzugefügt. Wenn wir nun $\lambda \mathbf{b}'\mathbf{b}$ berechnen, den quadrierten Strafterm für den Parametervektor, dann würden wir auch $a$ bestrafen. Die Rolle von $a$ ist aber, die Höhenlage der Regressionsfunktion zu definieren (die Stelle an der die Funktion die y-Achse schneidet).Der intercept $a$ kann allerdings aus der Gleichung genommen werden, wenn die Variablen vorher standardisiert werden (Mittelwert $\bar{x} = 0$ und $\bar{y} = 0$). Jetzt verschwindet $a$ von ganz allein, wenn wir die standardisierten Mittelwerte in die Gleichung für $a$ einfügen:\begin{equation*}a=\bar{y} - b\bar{x} = 0 - b\cdot 0 = 0\end{equation*}Nun muss $a$ nicht mehr berücksichtigt werden und die Lösung für $\mathbf{b}$ ergibt sich zu:\begin{equation*}\hat{\mathbf{b}} = (\mathbf{X}'\mathbf{X} + \lambda\mathbf{I})^{-1}\mathbf{X}'\mathbf{y}\end{equation*}Nach Hastie et al., wurde dieses Verfahren ursprünglich verwendet um 'rank deficiency' Probleme zu beheben. Wenn Die Spalten oder Zeilen einer Matrix nicht lineare unabhängig sind, so hat die Matrix nicht vollen Rang. Beispielsweise kann sich eine Spalte durch Addition anderer Spalten ergeben. In diesem Fall funktionierte die Matrix Inversion nicht zufriedenstellend. Als Lösung hat man gefunden, dass es ausreichend ist, einen kleinen positiven Betrag zu den Diagonal-Elementen der Matrix zu addieren.Dies wird nachfolgend in einem numerischen Beispiel gezeigt: - `np.c_` fügt die einzelenne Variablen zu einer Matrix zusammen - `np.dot(X.T, X)` ist das bekannte Kreuzprodukt der transponierten Matrix $\mathbf{X'}$ und $\mathbf{X}$ - `np.linalg.matrix_rank` gibt uns den Rang der Matrix - `np.eye(7) * 2` erstellt eine Diagonal-Matrix mit 2 in der Diagonalen und 0 überall sonst ###Code warnings.filterwarnings("ignore") y = np.load('/home/martin/python/fhnw_lecture/scripts/regression_y.pickle.npy') X = np.load('/home/martin/python/fhnw_lecture/scripts/regression_X.pickle.npy') X_6 = np.c_[X, X**2, X**3, X**4, X**5, X**6] print(f'With 6 variables (polynom of 6th degree), the rank of the quare matrix\n is '\ + f'{np.linalg.matrix_rank(np.dot(X_6.T, X_6))}') X_7 = np.c_[X, X**2, X**3, X**4, X**5, X**6, X**7] print(f'With 7 variables (polynom of 7th degree), the rank of the quare matrix\n is '\ + f'{np.linalg.matrix_rank(np.dot(X_7.T, X_7))}') print(f'By adding a small amount to the diagonal of the matrix, it is of full rank\n again: '\ + f'{np.linalg.matrix_rank(np.dot(X_7.T, X_7) + np.eye(7) * 2)}') ## you can see how small this amount is, by having a glimpse on the diagonal elements: print('\nto see how small the added amount in reality is, we display the diagonal elements:') np.diag(np.dot(X_7.T, X_7)) ###Output With 6 variables (polynom of 6th degree), the rank of the quare matrix is 6 With 7 variables (polynom of 7th degree), the rank of the quare matrix is 6 By adding a small amount to the diagonal of the matrix, it is of full rank again: 7 to see how small the added amount in reality is, we display the diagonal elements: ###Markdown example of ridge regressionNext, we will apply ridge regression as implemented in the python `sklearn` library and compare the results to the linear algebra solution. Note, that we have to center the variables.* we can center $\mathbf{X}$ and $\mathbf{y}$ and display the result in the centered coordinate system* or we can center $\mathbf{X}$ and add the mean of $\mathbf{y}$ to the predicted values to display the result in the original coordinate system. This approaches allows for an easy comparison to the overfitted resultDie Zeile `Xc = X - np.mean(X, axis=0)` standardisiert die Variablen auf den Mittelwert von 0 ###Code from sklearn.linear_model import Ridge y = np.load('/home/martin/python/fhnw_lecture/scripts/regression_y.pickle.npy') X = np.load('/home/martin/python/fhnw_lecture/scripts/regression_X.pickle.npy') X = np.c_[X, X**2, X**3, X**4, X**5, X**6, X**7] # here is the necessary standardization: Xc = X - np.mean(X, axis=0) # for plotting purpose x = np.arange(-1, 12, 0.05).reshape((-1, 1)) x = np.c_[x, x**2, x**3, x**4, x**5, x**6, x**7] xc = x -np.mean(x, axis = 0) # the result as obtained from the sklearn library model = Ridge(alpha=2, fit_intercept=False) model.fit(Xc, y) print(f'the parameters from the sklearn library:\n'\ + f'{model.coef_}') # the analytical result as discussed above inverse = np.linalg.inv(np.dot(np.transpose(Xc), Xc) + np.eye(Xc.shape[1]) * 2) Xy = np.dot(np.transpose(Xc),y) params = np.dot(inverse, Xy) print(f'the parameters as obtained from the analytical solution:\n' + f'{np.transpose(params)}') params_ridge = params # here we add the mean of y to the predictions to display results in original coord. system y_hat = np.dot(xc , params) + np.mean(y) f = plt.figure(figsize=(5, 5), dpi=100) plt.title(label='ridge regression for polynome of 7th degree and $\lambda=2$', fontdict={'fontsize':15}) axes = f.add_subplot(111) axes.plot(X[:,0], y, 'ro') axes.plot( x[:,0], y_hat.reshape((-1,)), 'b-', label='ridge regression') #axes = plt.gca() axes.set_ylim([np.min(y)-10, np.max(y) +20]) # now the overfitted solution from sklearn.linear_model import LinearRegression modelLR = LinearRegression() modelLR.fit(X, y) y_overfitted = np.dot(x , modelLR.coef_.T) + modelLR.intercept_ axes.plot(x[:,0], y_overfitted, 'y--', label='unregularized regression') leg = axes.legend() ###Output _____no_output_____ ###Markdown LassoAlternativ zu einem quadratischen Strafterm $b^2$ könnte man auch den absoluten Wert nehmen $|b|$. In diesem Fall erhält man die sog.~Lasso Regression; $\lambda\cdot |b|$ wird zum Vorhersage-Fehler addiert:$$\sum_i^{n}(y_i - \hat{y_i})^2 + \lambda |b|= \sum_i^{n}[y_i - (a + b\cdot x_i)]^{2}+ \lambda |b|$$Für zwei Variablen würde man folglich schreiben:$$\sum_i^{n}(y_i - \hat{y_i})^2 + \lambda |b_1| + \lambda |b_2|= \sum_i^{n}[y_i - (a + b_1\cdot x_{i1} + b_2\cdot x_{i2})]^{2}+ \lambda |b_1| + \lambda |b_2|$$ Leider gibt es im Gegesatz zur Ridge Regression keine eindeutige analytische Lösung um die Koeffizienten der Lasso Regression zu erhalten. Hier kommen iterative Verfahren zum Einsatz, wie wir sie in Session 2 kennen lernen werden. Vergleich der Koeffizienten der Lasso Regression mit denen der Ridge RegressionNext, we will apply lasso regression as implemented in the python sklearn library and compare the results to the unconstraint regression results.As before, we have to center the variables (-> see discussion above) ###Code import numpy as np from sklearn.linear_model import Lasso y = np.load('/home/martin/python/fhnw_lecture/scripts/regression_y.pickle.npy') X = np.load('/home/martin/python/fhnw_lecture/scripts/regression_X.pickle.npy') X = np.c_[X, X**2, X**3, X**4, X**5, X**6, X**7] Xc = X - np.mean(X, axis=0) # for plotting purpose x = np.arange(-1, 12, 0.05).reshape((-1, 1)) x = np.c_[x, x**2, x**3, x**4, x**5, x**6, x**7] xc = x -np.mean(x, axis = 0) # the result as obtained from the sklearn library model = Lasso(alpha=2, fit_intercept=False) model.fit(Xc, y) params_lasso = model.coef_ # comparison of parameters ridge vs. lasso: print(f'the parameters of the ridge regression:\n'\ + f'{np.transpose(params_ridge)}') print(f'the parameters of the lasso regression:\n'\ + f'{params_lasso}') ###Output the parameters of the ridge regression: [[-1.96523119e-01 -6.47914004e-01 -9.37247118e-01 1.55320112e-01 3.20681203e-02 -6.80277139e-03 3.08899915e-04]] the parameters of the lasso regression: [-0.00000000e+00 -1.27169261e+00 2.49755651e-01 7.47152651e-04 -5.77539403e-04 -2.73002774e-05 1.76588437e-06] ###Markdown Ridge Regression tendiert dazu alle Koeffizienten im gleichen Mass zu verkleinern. Lasso führt oft zu Lösungen, bei denen einige Koeffizienten ganz zu $0$ konvergiert sind. Wenn man die Ergebnisse im obigen Beispiel betrachtet, fällt einem auf dass für Lasso eigentlich nur zwei Koeffizienten verschieden von $0$ sind (for $X^2$ and $X^3$).Die Werte alle anderen Koeffizienten sind kleiner als $0.000747 = 7.47\text{e}-04$. ###Code y_hat = np.dot(xc, model.coef_.reshape((-1,1))) + np.mean(y) f = plt.figure(figsize=(5, 5), dpi=100) plt.title(label='lasso regression for polynome of 7th degree and $\lambda=2$', fontdict={'fontsize':15}) axes = f.add_subplot(111) axes.plot(X[:,0], y, 'ro') axes.plot( x[:,0], y_hat.reshape((-1,)), 'b-', label='lasso regression') #axes = plt.gca() axes.set_ylim([np.min(y)-10, np.max(y) +20]) # now the overfitted solution from sklearn.linear_model import LinearRegression modelLR = LinearRegression() modelLR.fit(X, y) y_overfitted = np.dot(x , modelLR.coef_.T) + modelLR.intercept_ axes.plot(x[:,0], y_overfitted, 'y--', label='unregularized regression') leg = axes.legend() ###Output _____no_output_____ ###Markdown the difference between ridge and lassoIn der folgenden graphischen Darstellung haben die __wahren Koeffizienten__ die Werte $b_1=1.5,\quad b_2=0.5$. Für ein grid aus beliebigen Werten für $b_1$ und $b_2$ wird der __mean squared error__ (MSE) berechnet und der Fehler als Kontur graphisch dargestellt. Wie man sieht, wird der Fehler umso geringer, je näher die Koeffizienten im grid an den wahren Koeffizienten liegen.Als nächstes werden alle Koeffizienten-Kombinationen aus $b_1$ und $b_2$ eingetragen, deren Strafterm ($b_1^2 + b_2^2$im Falle von Ridge und $b_1 + b_2$ im Falle von Lasso) den Wert von $1.0$ nicht übersteigt. Die Lösung, die den __wahren Koeffizienten__ am nähesten ist, wird jeweils durch einen Punkt eingezeichnet.Hierbei sieht man, dass sich die besten Lösungen von Ridge auf einem Halbkreis bewegen, die von Lasso auf einem Dreieck. An der Stelle, an der die Lasso-Lösung der eigentlichen Lösung (b=1.5, b2=0.5) am Nähesten ist, ist ein Parameter ($b_2$) fast $0$. Das zeigt die Tendenz von Lasso, einige Parameter gegen $0$ zu schrumpfen. Dieses Verhalten kann man sich zum Beispiel bei Variablen-Selektion zu Nutzen machen. ###Code # generation of random data set: X1 = np.random.normal(loc = 1.0, scale = 0.8, size = 100) X2 = np.random.normal(loc = 0.5, scale = 1.2, size = 100) beta1 = 1.5 beta2 = 0.5 Y = beta1 * X1 + beta2 * X2 X = np.c_[X1, X2] # test with linear regression from sklearn.linear_model import LinearRegression model = LinearRegression() model.fit(X, Y) model.intercept_ # essentiall zero model.coef_ # essentially 0.2 and 0.5 #print(f'the model parameters from data generation could be recovered: {model.coef_}') # make regular grid of values for b_1 and b_2 b1 = np.linspace(beta1 - 0.9, beta1 + 0.9, 100) b2 = np.linspace(beta2 - 0.9, beta2 + 0.9, 100) bb1, bb2 = np.meshgrid(b1, b2) # compute MSE-error Yhat = bb1.reshape(-1, 1) * X1.reshape(1, -1) + bb2.reshape(-1, 1) * X2.reshape(1, -1) errors = np.square(Yhat - Y.reshape(1, -1)) error = np.sum(errors, axis = 1)/len(Y) error_to_plot = error.reshape(bb1.shape) # plot MSE-error contour f = plt.figure(figsize=(5, 5), dpi=100) plt.title(label='minimal errors with penalties', fontdict={'fontsize':13}) axes = f.add_subplot(111) cp = plt.contour(bb1, bb2, error_to_plot) plt.clabel(cp, inline=1, fontsize=10) axes.set_xlabel('b1') axes.set_ylabel('b2') axes.set_ylim([np.min(b2)-0.5, np.max(b2) + 0.5]) axes.set_xlim([np.min(b1)-0.5, np.max(b1) + 0.5]) # plot optimal solution axes.scatter(beta1, beta2, s = 20) axes.annotate('$\hat{b}$', xy=(beta1 , beta2 + 0.1), xycoords='data', horizontalalignment = 'center', size = 20) # all ridge solutions with a penalty budget of 1 constraint_error = 1.0 values = np.linspace(0, 1.0, 100) constraint_l2 = np.sqrt(constraint_error - values**2) axes.plot(values, constraint_l2, 'y-', label = 'ridge') axes.plot(-values, constraint_l2, 'y-') axes.plot(values, -constraint_l2, 'y-') # all lasso solutions with a penalty budget of 1 constraint_l1 = constraint_error -values axes.plot(values, constraint_l1, 'r-', label = 'lasso') axes.plot(-values, constraint_l1, 'r-') axes.plot(values, -constraint_l1, 'r-') # best ridge solution with penalty budget of 1 Yhat_ridge = np.concatenate((values, values)).reshape(-1,1) * X1.reshape(1, -1) + \ np.concatenate((constraint_l2, -constraint_l2)).reshape(-1,1) * X2.reshape(1, -1) errors_ridge = np.square(Yhat_ridge - Y.reshape(1, -1)) error_ridge = np.sum(errors_ridge, axis = 1)/len(Y) index_ridge = np.where(error_ridge ==np.amin(error_ridge))[0][0] axes.scatter(np.concatenate((values, values))[index_ridge], np.concatenate((constraint_l2, -constraint_l2))[index_ridge], s=20, c='y') # best lasso solution with penalty budget of 1 Yhat_lasso = np.concatenate((values, values)).reshape(-1,1) * X1.reshape(1, -1) + \ np.concatenate((constraint_l1, -constraint_l1)).reshape(-1,1) * X2.reshape(1, -1) errors_lasso = np.square(Yhat_lasso - Y.reshape(1, -1)) error_lasso = np.sum(errors_lasso, axis = 1)/len(Y) index_lasso = np.where(error_lasso ==np.amin(error_lasso))[0][0] axes.scatter(np.concatenate((values, values))[index_lasso], np.concatenate((constraint_l1, -constraint_l1))[index_lasso], s=20, c='r') legs = axes.legend() plt.show() print(f'optimal coefficients of the ridge solution: {np.concatenate((values, values))[index_ridge]}'\ f' and {np.concatenate((constraint_l2, -constraint_l2))[index_ridge]}') print(f'optimal coefficients of the lasso solution: {np.concatenate((values, values))[index_lasso]}'\ f' and {np.concatenate((constraint_l1, -constraint_l1))[index_lasso]}') ###Output _____no_output_____ ###Markdown ElasticNetAus der Physik kommend werden die Strafterme von Ridge und Lasso als $\text{L}_2$ und $\text{L}_1$ bezeichnet. Eigentlich ist die $\text{L}_2$-Norm die Quadratwurzel der Summe der quadrierten Elemente eines Vectors und die $\text{L}_1$-Norm nur die Summe der Vektorelemente.ElasticNet ist ein lineares Regressions-Verfahren, in welches sowohl die regularization-terms von Lasso ($\text{L}_1$), als auch von Ridge ($\text{L}_2$) eingehen. Hier gibt es nicht nur einen $\lambda$-Paramter, der das Ausmass von regularization bestimmt, sondern einen zusätzlichen Parameter $\alpha$, der das Verhältnis von $\text{L}_1$ und $\text{L}_2$ regularization angibt.Weil Ridge Regression und Lasso die Koeffizienten sehr unterschiedlich regulieren, ist als Kompromiss die Kombination aus beiden Methoden sehr beliebt geworden. \begin{equation*}\lambda\sum_j (\alpha b_j^2 + (1-\alpha)|b_j|)\end{equation*}Die Interpretation der beiden paramter $\lambda$ und $\alpha$ ist wie folgt: - $\lambda$ bestimmt das generelle Mass an regularisation - $\alpha$ gibt das Verhältnis an, mit dem diese beiden Strafterme indie regularisation einfliessen sollenIm Übungs-Notebook zu den Boston house-prices werden wir ElasticNet verwenden. InteractionInteraktionen sind ein weiteres wichtiges Konzept in der linearen Regression. Hier ist der Effekt einer Variablen auf die abhängige Variable $y$ abhängig von dem Wert einer anderen Variable. In unterem Beispiel versuchen wir die Wahrscheinlichkeit zu modellieren, dass eine Person ein Haus kauft. Natürlich ist das monatliche Einkommen eine wichtige Variable und desto höher dieses, desto wahrscheinlicher auch, dass besagte Person ein Haus kauft. Eine andere wichtige Variable ist der Zivilstand. Verheiratet Personen mit Kindern im Haushalt tendieren stark zu Hauskauf, besonders wenn das monatliche Einkommen hoch ist. Auf der anderen Seite werden Singles, auch wenn sie ein hohes Einkommen haben, eher nicht zum Hauskauf tendieren.Wir sehen also, die Variable "monatliches Einkommen" __interagiert__ mit der Variable "Zivilstand": ###Code import numpy as np from statsmodels.graphics.factorplots import interaction_plot import pandas as pd income = np.random.randint(0, 2, size = 80) # low vs high marital = np.random.randint(1, 4, size = 80) # single, married, married & kids probability = np.random.rand(80) + income * np.random.rand(80) * marital probability = (probability - np.min(probability)) probability = probability/np.max(probability) marital = pd.Series(marital) marital.replace(to_replace = {1:'single', 2:'married', 3:'marrid w kids'}, inplace =True) income = pd.Series(income) income.replace(to_replace = {0:'low', 1:'high'}, inplace = True) fig = interaction_plot(income, marital, probability, colors=['mediumorchid', 'cyan', 'fuchsia'], ms=10, xlabel='income', ylabel='probability of buying a house', legendtitle='marital status') ###Output _____no_output_____ ###Markdown Das obige Beispiel beinhaltete kategorielle Variablen. Beispiele wie diese trifft man oft im Bereich der Varianzanalysen (ANOVA) an.Interaktions-Effekte bestehen aber auch für kontinuierliche Variablen. In diesem Fall ist es aber etwas komplizierter die Effekte zu visualisieren.Wir werden jetzt unseren eigenen Datensatz so erzeugen, dass er einen deutlichen Interaktions-Effekt aufweist. Damit der Effekt zwischen 2 kontinuierlichen Variablen überhaut in 2D dargestellt werden kann, musss eine der beiden Variablen wieder diskretisiert werden, d.h. wir müssen für sie wieder Kategorien bilden.Im nächsten Rechenbeispiel versuchen wir dann, die Parameter, die zur Generierung der Daten gedient haben mit einer Linearen-Regressions-Analyse wieder zu finden.Die Daten wurden nach folgendem Modell generiert:\begin{equation*}y = 2\cdot x + -2\cdot m + -7\cdot (x\cdot m) + \text{np.random.normal(loc = 0, scale = 4, size = n)}\end{equation*}`np.random.normal(loc=0, scale=4, size=n)` ist der Random-Error-Term, den wir hinzufügen, damit die Daten nicht alle auf einer Lienie liegen. `loc=0` besagt, dass der Mittelwert unseres zufälligen Fehlers $0$ ist, `scale=4`, dass die Varianz der Werte $4$ ist und `size=n` gibt die Anzahl der zu generierenden zufälligen Werte anFolgliche haben wir also die Koeffizienten: - $b_x = 2$ - $b_m = -2$ - $b_{x\cdot m} = -7$ ###Code import seaborn as sns n = 500 x = np.random.uniform(size=n) m = np.random.normal(loc = 0.5, scale = 1, size = n) # lin effects + interaction + random error y = 2*x + -2*m + -7*(x*m) + np.random.normal(loc = 0, scale = 4, size = n) newM = pd.cut(m, bins=3, labels = ['small', 'average', 'large']) toy = pd.DataFrame({'x' : x, 'y' : y, 'moderator' : newM}) sns.lmplot(x="x", y="y", hue="moderator", data=toy); ###Output _____no_output_____ ###Markdown Interaktions-Terme können gebildet werden, indem man zwei Variablen elemente-weise miteinander multipliziert.Durch die Hinzuname weiterer Terme sollte die Modell-Anpassung eigentlich besser werden - besonders wenn ein starker Interaktionsterm in den Daten vorliegt, so wie wir ihn eingebaut haben.Vergleichen wir die Koeffizienten, so wie sie im Linearen-Modell gefunden werden mit denen, die zur Erzeugung unseres Datensatzes gedient haben. Gar nicht schlecht, oder? Die zufälligen Fehler mit der grossen Varianz sorgen natürlich dafür, dass sie dennoch von den 'generating parameters' verschieden sind. ###Code from sklearn.linear_model import LinearRegression model = LinearRegression() X = np.c_[x, m] model.fit(X, y) y_hat = model.intercept_ + np.dot(X, model.coef_) print(f'without considering the interaction, the mse is: {np.mean((y-y_hat)**2)}') X = np.c_[x, m, x * m] model.fit(X, y) y_hat = model.intercept_ + np.dot(X, model.coef_) print(f'considering the interaction, the mse drops to: {np.mean((y-y_hat)**2)}') print(f'\nthe coefficients are given by {model.coef_}; compare these values\n to the values '\ + f'we used for generating the data') ###Output without considering the interaction, the mse is: 20.59039561666012 considering the interaction, the mse drops to: 16.932358091695946 the coefficients are given by [ 1.32323881 -2.15215978 -7.05045543]; compare these values to the values we used for generating the data ###Markdown some considerationsDie Überlegung hier veranschaulicht, dass es schon bei moderater Variablen-Anzahl sehr viele mögliche Interaktions-Terme gibt. Für die normale Lineare Regression würde die grosse Anzahl dieser Terme zum Verhängnis werden, weil dann wieder der Fall eintreten könnte indem wir die Daten overfitten oder gar mehr Variablen als Beobachtunge zur Verfügung stehen. Auch in diesem Fall kann auf die vorgestellten^ Regularisierungs-Verfahren (ElasticNet, Ridge und Lasso) zurückgegriffen werden:Nehmen wir an, wir haben ein data-set mit 70 verschiedenen Variablen. Weil wir nichts über die Beziehungen der Variablen zur abhängigen Variable $y$ noch über die Beziehungen der Variablen untereinander wissen, sind wir geneigt eine Menge zusätzlicher 'features' für unser Modell zu erzeugen:* wir können 70 quadratische Terme hinzufügen ($x_j^2$)* wir können 70 kubische Terme aufnehmen ($x_j^3$)* wir können auch $\binom{70}{2} = 2415$ Interaktionen erster Ordnung zwischen den 70 Variablen annehmen* anstatt dessen könnte wir auch die Interaktions-Terme der 210 (70 Variablen + 70 quadratische Terme + 70 kubische Terme) Variablen mit aufnhemne: $\binom{210}{2} = 21945$* neben quadratisch und kubischen Termen gibt es auch viele andere linearisierende Transformation, die unter Umständen zu besseren ergebnissen führen wie beispielsweise die log-Transformation. Im praktischen Beipiel des Bosten house-prices data-Sets werden wir die `box-cox-Transformation` kennen lernen.Wie wir gesehen haben, kann die Anzahl möglicher Variablen sehr schnell wachsen, wenn man alle Effekte berücksichtigt, die ausschlaggebend sein könnten. Manchmal existieren sogar Interaktionseffekte zweiter Ordnund, d.h. drei Variablen sind dann daran beteiligt. Würden wir alle möglichen Variablen berücksichtigen, die sich derart bilden lassen, dann würde dies auch bei grossen Daten-Sets zu ausgeprägten 'overfitting' führen. __Aus diesem Grund wurden die regularization techniques wie das ElasticNet und seine Komponenten, die Ridge Regression und die Lasso Regression eingeführt__. Wie zuversichtlich sind wir hinsichtlich unserer Modell-VorhersagenSelten werden wir mit unserem Modell genau die Koeffizienten schätzen können, die in der gesamten Population (alle Daten, die wir erheben könnten) anzutreffen sind. Viel öfter ist unsere Stichprobe nicht repräsentativ für die gesamte Population oder sie ist schlicht zu klein und zufällige, normalverteilte Fehler in unseren Daten beeinflussen die Schätzung der Koeffizienten. Dies umsomehr, desto mehr Variablen wir in user Modell aufnehmen.Wie können wir nun die Güte unserer Schätzung beurteilen? Hier sind mindestens zwei verschieden Fragen denkbar:* Wie sicher sind wir mit Hinblick auf die geschätzen Koeffizienten $\mathbf{b}$?. Diese Frage ist besonders für Wissenschaftler wichtig, da die Antwort dafür ausschlaggebend ist, ob eine Hypothese beibehalten oder verworfen werden muss.* Wie sicher sind wir uns bezüglich einzelner Vorhersagen. Dies spielt die grösste Rolle im Machine Learning Umfeld, da wir das trainierte Modell gerne in unsere Business-Abläufe integrieren würden.Diese beiden Fragestellungen lassen sich mit Hinblick auf die Regression auch wie folgt formulieren: * Wie sehr ist die 'mean response', die Vorhersage unsere Regressions-Funktion von der Stichprobe abhängig. Variiert Erstere sehr stark und umfasst unter Umständen sogar den Wert $0$, dann können diese Effekte (Koeffizienten) nicht interpretiert werden. * Wie sehr können Beobachtungen $y$ für eine gegebene Kombination von Variablen-Werten in $\mathbf{X}$ variieren? Ist diese Variation sehr gross, so werden wir auch grosse Fehler in unseren Business-Process einbauen Recap of assumptions underlying regressionDies sind Linearität (der Zusammenhang einer Variablen und der abhängigen Variablen ist linear, d.h. der selbe Steigungsparamter gilt für alle Bereiche der Variablen), Homoskedastizität (die Fehler der Regression -- die Residuen -- sind in allen Bereichen von X normal verteilt mit gleicher Varianz) und Normalität der Residuen bei gegebenem Wert von X.Diese Voraussetzungen sind in vielen Fällen nicht erfüllt und auch bekannterweise verletzt. * __Linearity__: Die Regression-Funktion ist eine gute Annäherung für die Beziehung zwischen $\mathbf{X}$ and $\mathbf{y}$, d.h. ist ein quadratischer Trend in den Daten und wir haben keine quadratischen Effekte in das Modell aufgenommen, so sind die Annahmen nicht erfüllt. Die Linearität besagt nämlich, dass für den Zusammenhang einer Variablen $x$ und der abhängigen Variablen $y$ der selbe Steigungs-Koeffizient $b_x$ für all Bereich für $x$ gelten muss. Ansonsten hat das Modell einen __bias__, es schätzt einen Koeffizienten systematisch falsch.* __Homoscedasticity__: Die Varianz unseres Vorhersagefehlers (Residuen) ist für alle Bereiche einer Variablen $x$ identisch.* __Normality__: Die Werte der abhängigen Variablen $\mathbf{y}$ sind für einen gegeben Wert von $\mathbf{x}$ normal verteilt: $\mathbf{y}|\mathbf{x} \sim N(\mu, \sigma)$In der nächsten Graphik werden die Voraussetzungen der linearen Regression veranschaulicht:Image taken from [here](https://janhove.github.io/analysis/2019/04/11/assumptions-relevance) ###Code Image('../images/homoscedasticity.png') ###Output _____no_output_____ ###Markdown Now, with respect to our confidence need:1. __Vohersage Intervall (prediction interval)__: Dies ist das Intervall, in welchem mit (1-$\alpha$)% Wahrscheinlichkeit die beobachteten $y$ Werte zu unseren vorhergesagten Werten $\hat{y}$ liegen. Dieses Intervall ist symmetrisch um die Regressionsfunktion - was natürlich aus den Voraussetzungen der linearen Regression folgt. Der Standardfehler der Vorhersage ist gegeben durch:\begin{equation*}\hat{\sigma}_e = \sqrt{\frac{1}{N-(p+1)}\sum_i^N e_i^2}, \end{equation*}hier ist $p$ die Anzahl der Parameter im Modell (der zusätzliche Parameter $+1$ kommt vom intercept); $e_i$ sind die Vorhersage-Fehler, die Residuen, also die Differenz aus unseren vorhergesagten $\hat{y}_i$ und den beobachteten Werten $y_i$. Das Konfidenz-Intervall ergibt sich zu:\begin{equation*} CI_i = y_i \pm t_{1-\alpha/2, N-p} \cdot \hat{\sigma}_e.\end{equation*}Hier ist $t_{1-\alpha/2, N-p}$ der Wert der Student-t-Verteilung für das Konfidenzlevel von $1-\alpha/2$ und $N-p$ Freiheitsgraden. Der Wert von $\alpha$ gibt an, wie sehr wir uns mit dem Konfidenzintervall gegen falsche Entscheidungen absichern wollen. Wollen wir beispielseweise mit 95% Sicherheit den Bereich angeben können, in dem die beobachteten Werte liegen, dann müssen wir als untere Konfidenzgrenze den Wert bestimmen unterhalb dessen nur mit einer Wahrscheinlichkeit von 2.5% die beobachteten Werte liegen und als obere Konfidenzgrenze den Wert unterhalb dessen mit einer Wahrscheinlichkeit von 97.5% die beobachteten Werte liegen. So machen wir nur in 5% aller Fälle einen Fehler, $\alpha = 0.05$ und weil das Konfidenzintervall symmetrisch ist benötigen wir den Wert $1-\alpha/2$ damit wir von beiden Enden 2.5% abschneiden. 2. __Mean Prediction Confidence interval__: In ähnlicher Weise können wir ein Konfidenzintervall für unsere durchschnittliche Vorhersage $\hat{\bar{y}}$ bestimmen. Wir erinnern uns, dass die Regressions-Funktion unsere Vorhersage ist und die Daten um diese normal verteilt sein sollten. Weil unsere Stichprobe aber nur eine Momentaufnahme eines Auschnitts aller möglichen Werte ist, die wir erheben könnten, wird die Regressions-Funktion je nach Stichprobe variieren. Das Konfidenz-Intervall gibt an, in welchem Bereich die Regressions-Funktion mit grosser Wahrscheinlichkeit liegen würde, könnten wir alle Daten erfassen (die gesamte Population). Das Konfidenzintervall ist nicht für alle Werte von $x$ gleich weit. Dort wo wenige Messwerte vorliegen kann der genaue Verlauf schlechter geschätzt werden als dort wo wir eine breitere Datenbasis für die Schätzung haben. Nahe dem Mittelwert von $x$, also bei $\bar{x}$ sollte unsere Schätzung immer genauer sein als nahe den Extremwerten. Natürlich gehen wir wieder von normalverteilten $x$ Werten aus. 3. __CI for regression coefficients__: Auch diese Intervall ist schwierig zu bestimmen. Es gibt die obere und untere Grenze für unsere Regressions-Koeffizienten $\mathbf{b}$. Die Interpretation dieser Koeffizienten findet vor allem in der Wissenschaft statt. Es kann getestet werden, ob ein a priori postulierter Effekt tatsächlich vorliegt oder nicht. Umfasst das Konfidenzintervall für einen Koeffizienten $b$ den Wert Null, so kann nicht ausgeschlossen werden, dass der Effekt in der Stichprobe nur rein zufällig zu Stande kommt. Beispielsweise könnte folgende Fragestellung hiermit untersucht werden: "Hat die Schliessung von Schulen und Universitäten einen signifikanten Einfluss auf die Reproduktions-Zahl $R_0$ oder nicht". Dies ist typerischweise nicht die Art von Fragestellung, mit der sich Data Scientists beschäftigen. Im nachfolgenden Beispiel sehen wir die typische Ausgabe eines klassischen, statistischen Ansatzes. In der Mitte sehen wir die Konfidenz-Intervalle für die Regressions-Koeffizienten, `const` (intercept) und `x1`, d.h. der Koeffizient der Variablen $x_1$, also $b_1$. Der intercept ist nicht signifikant, weil das Konfidenzinervall ($\left[-43.351, 26.370\right]$) den Wert $0$ umfasst. Der Koeffizient $b_1$ für die Variable $x_1$ ist aber signifikant von $0$ verschieden. Sein Konfidenzintervall ist $\left[2.939, 13.219\right]$.Bei grossem Interesse für klassische statistische Modelle kann ich für python diese [Quelle](http://web.vu.lt/mif/a.buteikis/wp-content/uploads/PE_Book/3-7-UnivarPredict.html) empfehlen. ###Code import statsmodels.api as sm # data example y = np.load('/home/martin/python/fhnw_lecture/scripts/regression_y.pickle.npy') X = np.load('/home/martin/python/fhnw_lecture/scripts/regression_X.pickle.npy') # the x (small x) is just for plotting purpose x = np.arange(1, 12, 0.05).reshape((-1, 1)) x_intercept = np.c_[np.ones(x.shape[0]), x] X_intercept = np.c_[np.ones(X.shape[0]), X] ols_result_lin = sm.OLS(y, X_intercept).fit() y_hat_lin = ols_result_lin.get_prediction(x_intercept) dt_lin = y_hat_lin.summary_frame() mean_lin = dt_lin['mean'] meanCIs_lin = dt_lin[['mean_ci_lower', 'mean_ci_upper']] obsCIs_lin = dt_lin[['obs_ci_lower', 'obs_ci_upper']] print(ols_result_lin.summary()) # beta-coefficients ### figure for linear plot f = plt.figure(figsize=(5, 5), dpi=100) plt.title(label='linear regression', fontdict={'fontsize':20}) axes = f.add_subplot(111) axes.plot(X_intercept[:,1], y, 'ro') axes.plot(x_intercept[:, 1], mean_lin.values.reshape((-1,)), color = "red", label = "regression line") axes.plot(x_intercept[:, 1], obsCIs_lin.iloc[:, 0], color = "darkgreen", linestyle = "--", label = "Predictions interval (1.)") axes.plot(x_intercept[:, 1], obsCIs_lin.iloc[:, 1], color = "darkgreen", linestyle = "--") axes.plot(x_intercept[:, 1], meanCIs_lin.iloc[:, 0], color = "blue", linestyle = "--", label = "Mean Prediction CI (2.)") axes.plot(x_intercept[:, 1], meanCIs_lin.iloc[:, 1], color = "blue", linestyle = "--") axes.legend() axes.set_ylim([np.min(y)-10, np.max(y) +10]) ###Output _____no_output_____ ###Markdown Als nächstes berechnen wir die Konfidenzintervalle für die Regression mit einem quadratischen Term.Hierbei fällt auf, dass zwar jetzt der quadratische Term `x2` signifikant ist (Intervall $\left[0.247, 3.125\right]$), nicht mehr aber der `x1` Term. ###Code X_intercept_quad = np.c_[X_intercept, X**2] # for plotting: x = np.arange(1, 12, 0.05).reshape((-1, 1)) x_intercept_quad = np.c_[np.ones(x.shape[0]), x, x**2] ols_result_quad = sm.OLS(y, X_intercept_quad).fit() y_hat_quad = ols_result_quad.get_prediction(x_intercept_quad) dt_quad = y_hat_quad.summary_frame() mean_quad = dt_quad['mean'] meanCIs_quad = dt_quad[['mean_ci_lower', 'mean_ci_upper']] obsCIs_quad = dt_quad[['obs_ci_lower', 'obs_ci_upper']] print(ols_result_quad.summary()) ### figure for linear plot f = plt.figure(figsize=(5, 5), dpi=100) plt.title(label='regression with quadratic term', fontdict={'fontsize':20}) axes = f.add_subplot(111) axes.plot(X_intercept_quad[:,1], y, 'ro') axes.plot(x_intercept_quad[:, 1], mean_quad.values.reshape((-1,)), color = "red", label = "regression line") axes.plot(x_intercept_quad[:, 1], obsCIs_quad.iloc[:, 0], color = "darkgreen", linestyle = "--", label = "Predictions interval (1.)") axes.plot(x_intercept_quad[:, 1], obsCIs_quad.iloc[:, 1], color = "darkgreen", linestyle = "--") axes.plot(x_intercept_quad[:, 1], meanCIs_quad.iloc[:, 0], color = "blue", linestyle = "--", label = "Mean Prediction CI (2.)") axes.plot(x_intercept[:, 1], meanCIs_quad.iloc[:, 1], color = "blue", linestyle = "--") axes.legend() axes.set_ylim([np.min(y)-10, np.max(y) +10]) ###Output _____no_output_____ ###Markdown BootstrapDie Daten, mit denen ein Data Scientist normalerweise arbeitet, erfüllen meist nie die Voraussetzungen der Linearen Regression. Deshalb können wir auch die Theory zu den Konfidenzintervallen nicht anwenden - schliesslich beruht sie auf den Annahmen wie normalverteilte Daten.Eine robuste, parameter-freie Alternative ist der __Bootstrap__. Gewissernahmen ziehen wir uns an den eigenen Haaren aus dem Schlamassel: - Wir betrachten unsere Stichprobe als die Gesamtheit (Population) der Daten. - Nun ziehen wir wiederholt und mit Zurücklegen neue Stichproben aus dieser Stichprobe. - Für jede dieser Stichproben wird das Modell angepasst und die relevanten Statistiken werden gespeichert. - Abschliessend finden wir in unseren gespeicherten Statistiken das 2.5% Quantil (der Wert, unter dem nur 2.5% der Beobachtungen liegen) und das 97.5% Quantil (der Wert über dem nur noch 2.5% der Beobachtungen liegen). Diese Werte teilen wir als untere und obere Grenze des Konfidenz-Intervalls mit, bei einem Konfidenz-Level von $\alpha=0.05$.Nachfolgendes Code-Beispiel veranschaulicht den Vorgang:* `sampler = (choices(indices, k = len(indices)) for i in range(200))` erzeugt einen Generator, der 200 Mal eine Zufallsstichprobe zieht.* `np.percentile(np.array([Lasso(alpha=2, fit_intercept=True).fit(X[drew,:], y[drew, :]).predict(x).tolist() for drew in sampler]), [2.5, 97.5], axis = 0)` iteriert über den Generator und passt insgesammt 200 mal das Modell an und macht eine Vorhersage für kontinuierliche x-Werte im Bereich von 1 bis 12. Diese Vorhersagen werden in einem numpy-array (`np.array`) gespeichert und zu Schluss die Funktion `np.percentile` auf die 200 Vorhersagen angewendet. Somit erhalten wir für den x-Bereich von 1 bis 12 die Intervall-Grenzen für die mean-preditction, d.h. die Regressions-Funktion ###Code from random import choices from sklearn.linear_model import Lasso import warnings warnings.filterwarnings('ignore') y = np.load('/home/martin/python/fhnw_lecture/scripts/regression_y.pickle.npy') X = np.load('/home/martin/python/fhnw_lecture/scripts/regression_X.pickle.npy') #X = np.c_[np.ones(X.shape[0]), X, X**2, X**3, X**4] X = np.c_[X, X**2, X**3, X**4] x = np.arange(1, 12, 0.05).reshape((-1, 1)) #x = np.c_[np.ones(x.shape[0]), x, x**2, x**3, x**4] x = np.c_[x, x**2, x**3, x**4] indices = np.arange(0, X.shape[0]) drew = choices(indices, k=len(indices)) sampler = (choices(indices, k = len(indices)) for i in range(200)) CIS = np.percentile(np.array([Lasso(alpha=2, fit_intercept=True).fit(X[drew,:], y[drew, :])\ .predict(x).tolist() for drew in sampler]), [2.5, 97.5], axis = 0) # x is 220 long model = Lasso(alpha=2, fit_intercept=True) model.fit(X, y) y_hat = model.predict(x) f = plt.figure(figsize=(5, 5), dpi=100) plt.title(label='lasso regression for polynome of 4th degree and $\lambda=2$', fontdict={'fontsize':15}) axes = f.add_subplot(111) axes.plot(X[:,0], y, 'ro') axes.plot( x[:,0], y_hat.reshape((-1,)), 'b-', label='lasso regression') axes.plot(x[:, 0], CIS[0, :], color = "cyan", linestyle = "--", label = "Mean Prediction CI") axes.plot(x[:, 0], CIS[1, :], color = "cyan", linestyle = "--") axes.legend() ###Output _____no_output_____ ###Markdown Extension: logistic regression and the GLMEs gibt andere Modelle, die eng verwandt mit der hier besprochenen Linearen Regression sind. Das Prominenteste unter ihnen ist die __Logistische Regression__. Diese Modell gehört zu dem "__Verallgemeinerten Linearen Modell__" (im engl. __generalized lineare model__ (GLM)). Diese Modelle dürfen nicht mit dem "__Allgemeinen Linearen Modell__" (im engl. __general linear model__) verwechselt werden. Letzteres parametrisiert eine Varianzanalyse als ein lineares Modell mit Dummy-Variablen.Das Verallgemeinerte Lineare Modell erweitert die Lineare Regression um Modelle, deren Fehler nicht normalverteilt sind.[Dieser Artikel](https://en.wikipedia.org/wiki/Generalized_linear_modelConfusion_with_general_linear_models) in der Wikipedia gibt weitere Auskunft. exponential family of distributionsAus der Perspektive der Modernen Statistik beinhaltet das Verallgemeinerte Lineare Modell verschiedene Lineare Modelle, unter anderem das der klassischen linearen Regression. Eine Verteilung, die in der "exponential family" von Verteilungen ist, kann immer folgendermassen geschrieben werden:\begin{equation}f(y| \theta) = \exp\left(\frac{y \theta + b(\theta)}{\Phi} + c(y, \Phi)\right),\end{equation}wobei $\theta$ als Kanonischer Parameter bezeichnet wird, welcher eine Funktion von $\mu$ ist dem Mittel. Diese Funktion wird als Kanonische Link-Funktion bezeichnet. Wie wir später an einem Beispiel sehen werden, ist es genau diese Funktion welche die Beziehung zwischen der abhängigen Variablen und den unabhängigen Variablen linearisiert.Der Vollständigkeit halber: $b(\theta)$ ist eine Funktion des Kanonischen Parameters und ist somit ebenfalls von $\mu$ abhängig. $\Phi$ wird als Streuungsparameter bezeichnet und $c(y, \Phi)$ ist eine Funktion, die sowohl von beobachteten Daten wie auch dem Streuungsparameter abhängig ist. Normalverteilung\begin{eqnarray*}f(y| \mu, \sigma) =& (2\pi \sigma^2)^{-\frac{1}{2}} \exp\left(-\frac{1}{2}\frac{y^2 -2y\mu + \mu^2}{\sigma^2}\right) \\ =&\quad \exp \left(\frac{y\mu -\frac{\mu^2}{2}}{\sigma^2} - \frac{1}{2}\left(\frac{y^2}{\sigma^2} + \log(2\pi\sigma^2\right)\right),\quad \text{wobei}\end{eqnarray*}$\mu = \theta(\mu)$, d.h. $\mu$ ist der Kanonische Parameter und die Link-Funktion ist die Identitäts-Funktion. Der Mittelwert kann also ohne weitere Transformation direkt modelliert werden, so wie wir es in der klassischen Linearen Regression machen.Der Streuungsparameter $\Phi$ ist durch $\sigma^2$, die Varianz gegeben. Dies ist die klassische Lineare Regression normalverteilter Variablen Poisson distributionDie Poisson-Verteilung gehört ebenfalls der exponential family von Verteilungen an:\begin{eqnarray*}f(y| \mu) =& \frac{\mu^{y} e^{-\mu}}{y!} = \mu^y e^{-\mu}\frac{1}{y!}\\=& \quad\exp\left(y \log(\mu) - \mu - \log(y!)\right), \quad\text{where}\end{eqnarray*}Die Link-Funktion ist hier $\log(\mu)$. Beachte bitte, dass die Poisson-Verteilung keinen Streuungsparameter besitzt. Bernoulli distribution $\Rightarrow$ logistic regressionZuguter Letzte, die Bernoulli Verteilung, von der wir die Logistische Regression ableiten können.Die Bernoulli Verteilung eignet sich um binäre Ereignisse zu modellieren, die sich gegenseitig ausschliessen. Ein klassisches Beispiel ist der wiederholte Münzwurf. Die Wahrscheinlichkeit für 'Kopf' wird mit $\pi$ bezeichnet, dir für 'Zahl' mit $(1-\pi)$. Hiermit lässt sich die Wahrscheinlichkeit berechnen, mit einer fairen Münze bei 10 Würfen eine bestimmte Sequenz mit genau 7 Mal 'Kopf' zu erhalten:\begin{equation}\pi^7 (1-\pi)^3 = 0.5^7 0.5^3 = 0.5^{10} = 0.0009765625\end{equation}__Vorsicht__, wenn wir die Wahrscheinlichkeit für Sequenzen mit genau 7 Mal Kopf berechnen wollen, benötigen wir noch den Binomial-Koeffizienten, der uns die Anzahl an möglichen Sequenzen mit 7 Mal 'Kopf' angibt.Jetzt zeige ich, wie wir die Bernoulli Verteilung so umschreiben können, dass man ihre Zugehörigkeit zur exponential family von Verteilungen erkennt:\begin{eqnarray*}f(y |\pi) =& \pi^y (1-\pi)^{1-y} = \exp\left(y \log(\pi) + (1-y) \log(1-\pi)\right)\\= & \quad \exp\left(y \log(\pi) + \log(1-\pi) - y\log(1-\pi)\right)\\=&\quad \exp\left(y\log(\frac{\pi}{1-\pi}) + \log(1-\pi)\right),\quad\text{wobei}\end{eqnarray*}sich die Link-Funktion zu $\log(\frac{\pi}{1-\pi})$ ergibt. Diese Funktion wird auch als Logit-Funktion bezeichnet. Die Umkehrfunktion der Logit-Funktion ist die __Logistische Funktion__. This functionis also called the logit function whose reverse function is the logisticfunction. Es ist also die Logit-Funktion, die als lineare Kombination der unabhängigen Variablen modelliert wird.$\log(\frac{\pi}{1-\pi}) = a + b_{1}x_1 + \ldots + b_jx_j$. Wenn wir den rechten Teil dieser Gleichung in die Logistische Funktion einsetzen erhalten wir die geschätzten Wahrscheinlichkeiten:\begin{equation}P(y=1 |x) = \frac{\exp(a + b_{1}x_1 + \ldots + b_jx_j)}{1 + \exp(a + b_{1}x_1 + \ldots + b_jx_j)}.\end{equation} Somit haben wir also gezeigt, dass das klassische Lineare Regressions-Modell nur ein Spezialfall einer grossen Anzahl von Modellen ist, deren Verteilungen alle in der exponential family enthalten sind. (Für eine vollständigere Abhandlung dieses Themas:https://en.wikipedia.org/wiki/Generalized\_linear\_model.) GLMNETIn der statistischen Programmiersprache R gibt es eine library die 'glmnet' genannt ist. Dieses Packet implementiert das ElasticNet für das Verallgemeinerte Lineare Modell und nicht nur für die klassische Lineare Regression.https://web.stanford.edu/~hastie/glmnet/glmnet_alpha.htmlEs gibt auch ein python package welches den exakt gleichen Fortran-Code verwendet: __glmnet-python__.Es gibt ein paar kleine Unterschiede zu der Version von ElasticNet wie sie in `scikit-learn` implementiert isthttps://pypi.org/project/glmnet-python/ Neural NetworkEs ist auch möglich Neuronale Netzwerke unter dem Blickwinkel der Linearen Regression zu betrachten. Ein Netzwerk mit nur einer Eingabe-Schicht und einem Neuron wird als Perceptron bezeichnet. Die Aktivierungs-Funktion dieses Neurons ist entweder die Identitäts-Funktion, so wie in der klassischen Linearen Regression oder die Logistische Funktion wie in der Logistischen Regression. In letzterem Fall soll das Perceptron Wahrscheinlichkeiten für binäre Ereignisse bestimmen. ###Code # Image('../images/Regression_as_NN.png') Image("../images/NN_class_reg.png",height=520, width=520) ###Output _____no_output_____ ###Markdown classical linear regressionIm Jargon der neural network community werden unsere $b$-Koeffizienten als __Gewichte__ bezeichnet. Der intercept $\alpha$ heisst __bias__.Erinnert Euch, dass wir den intercept $\alpha$ in den Vektor $\pmb{\beta}$ der $b$-Koeffizienten aufgenomen haben, indem wir eine Einser-Spalte in die Variablen-Matrix $\mathbf{X}$ eingefügt haben. Wir konnten also Schreiben:\begin{equation*}\mathbf{y} = \mathbf{X} \pmb{\beta}\end{equation*}In der obigen Graphik könnt ihr sehen, dass im Perceptron die Input-Variablen mit den Gewichten der Verbindungen multipliziert werden und dass der konstante Wert $\alpha$ hinzu addiert wird. Wie in der Linearen Regression werden diese Produkte dann aufsummiert.Im Kontext Neuronaler Netzwerke wird der Vektor $\pmb{\beta}$ als Netzwerk-Gewichte bezeichnet und wird mit $\mathbf{W}$ angegeben. Wir hatten gelernt, dass Vektoren mit kleinen Buchstaben bezeichnet werden. In einem richtigen Neuronalen Netz haben wir in einer Schicht viel Perceptrons nebeneinander. Alle erhalten aber den Input aus der darunter liegenden Schicht. Fügt man die Gewichts-Vektoren der einzelnen Neurone in eine Matrix zusammen, erhält man $\mathbf{W}$.Neuronale Netzwerke sind also eigentlich nur viele parallele und hintereinander geschaltete Regressionen, die sehr effizient mit Matrizen-Multiplikation gerechnet werden können. ###Code Image("../images/NN_logistic_reg.png", height=520, width=520) ###Output _____no_output_____
courses/machine_learning/deepdive/03_tensorflow/diagrams/coretensorflow.ipynb
###Markdown Diagrams for Course 3 ###Code import tensorflow as tf x = tf.constant(3) print(x) import tensorflow as tf x = tf.constant([3, 5, 7]) print(x) import tensorflow as tf x = tf.constant([[3, 5, 7], [4, 6, 8]]) print(x) import tensorflow as tf x = tf.constant([[[3, 5, 7],[4, 6, 8]], [[1, 2, 3],[4, 5, 6]] ]) print(x) import tensorflow as tf x1 = tf.constant([2, 3, 4]) x2 = tf.stack([x1, x1]) x3 = tf.stack([x2, x2, x2, x2]) x4 = tf.stack([x3, x3]) print(x1) print(x2) print(x3) print(x4) import tensorflow as tf x = tf.constant([[3, 5, 7], [4, 6, 8]]) y = x[:, 1] with tf.Session() as sess: print(y.eval()) import tensorflow as tf x = tf.constant([[3, 5, 7], [4, 6, 8]]) y = tf.reshape(x, [3, 2]) with tf.Session() as sess: print(y.eval()) import tensorflow as tf x = tf.constant([[3, 5, 7], [4, 6, 8]]) y = tf.reshape(x, [3, 2])[1, :] with tf.Session() as sess: print(y.eval()) import tensorflow as tf from tensorflow.contrib.eager.python import tfe tfe.enable_eager_execution() x = tf.constant([[3, 5, 7], [4, 6, 8]]) y = tf.reshape(x, [3, 2])[1, :] print(y) import tensorflow as tf from tensorflow.contrib.eager.python import tfe tfe.enable_eager_execution() x = tf.constant([3, 5, 7]) y = tf.constant([1, 2, 3]) print(x-y) import tensorflow as tf x = tf.constant([3, 5, 7]) y = tf.constant([1, 2, 3]) z = tf.add(x, y) with tf.Session() as sess: print(z.eval()) import tensorflow as tf x = tf.constant([3, 5, 7]) y = tf.constant([1, 2, 3]) z = tf.add(x, y) with tf.Session() as sess: print(sess.run(z)) import tensorflow as tf x = tf.constant([3, 5, 7]) y = tf.constant([1, 2, 3]) z1 = x + y z2 = x * y z3 = z2 - z1 with tf.Session() as sess: a1, a3 = sess.run([z1, z3]) print(a1) print(a3) import tensorflow as tf x = tf.constant([3, 5, 7], name="x") y = tf.constant([1, 2, 3], name="y") z1 = tf.add(x, y, name="z1") z2 = x * y z3 = z2 - z1 with tf.Session() as sess: with tf.summary.FileWriter('summaries', sess.graph) as writer: a1, a3 = sess.run([z1, z3]) !ls summaries from google.datalab.ml import TensorBoard TensorBoard().start('./summaries') from google.datalab.ml import TensorBoard TensorBoard().stop(13045) print('stopped TensorBoard') import tensorflow as tf def forward_pass(w, x): return tf.matmul(w, x) def train_loop(x, niter=5): with tf.variable_scope("model", reuse=tf.AUTO_REUSE): w = tf.get_variable("weights", shape=(1,2), # 1 x 2 matrix initializer=tf.truncated_normal_initializer(), trainable=True) preds = [] for k in range(niter): preds.append(forward_pass(w, x)) w = w + 0.1 # "gradient update" return preds with tf.Session() as sess: preds = train_loop(tf.constant([[3.2, 5.1, 7.2],[4.3, 6.2, 8.3]])) # 2 x 3 matrix tf.global_variables_initializer().run() for i in range(len(preds)): print("{}:{}".format( i, preds[i].eval() )) ###Output 0:[[ 8.568541 12.702375 17.271353]] 1:[[ 9.318541 13.8323765 18.821354 ]] 2:[[10.068541 14.962376 20.371353]] 3:[[10.818541 16.092377 21.921354]] 4:[[11.56854 17.222376 23.471352]] ###Markdown Diagrams for Course 3 ###Code import tensorflow as tf x = tf.constant(3) print x import tensorflow as tf x = tf.constant([3, 5, 7]) print x import tensorflow as tf x = tf.constant([[3, 5, 7], [4, 6, 8]]) print x import tensorflow as tf x = tf.constant([[[3, 5, 7],[4, 6, 8]], [[1, 2, 3],[4, 5, 6]] ]) print x import tensorflow as tf x1 = tf.constant([2, 3, 4]) x2 = tf.stack([x1, x1]) x3 = tf.stack([x2, x2, x2, x2]) x4 = tf.stack([x3, x3]) print x1 print x2 print x3 print x4 import tensorflow as tf x = tf.constant([[3, 5, 7], [4, 6, 8]]) y = x[:, 1] with tf.Session() as sess: print y.eval() import tensorflow as tf x = tf.constant([[3, 5, 7], [4, 6, 8]]) y = tf.reshape(x, [3, 2]) with tf.Session() as sess: print y.eval() import tensorflow as tf x = tf.constant([[3, 5, 7], [4, 6, 8]]) y = tf.reshape(x, [3, 2])[1, :] with tf.Session() as sess: print y.eval() import tensorflow as tf from tensorflow.contrib.eager.python import tfe tfe.enable_eager_execution() x = tf.constant([[3, 5, 7], [4, 6, 8]]) y = tf.reshape(x, [3, 2])[1, :] print y import tensorflow as tf from tensorflow.contrib.eager.python import tfe tfe.enable_eager_execution() x = tf.constant([3, 5, 7]) y = tf.constant([1, 2, 3]) print (x-y) import tensorflow as tf x = tf.constant([3, 5, 7]) y = tf.constant([1, 2, 3]) z = tf.add(x, y) with tf.Session() as sess: print z.eval() import tensorflow as tf x = tf.constant([3, 5, 7]) y = tf.constant([1, 2, 3]) z = tf.add(x, y) with tf.Session() as sess: print sess.run(z) import tensorflow as tf x = tf.constant([3, 5, 7]) y = tf.constant([1, 2, 3]) z1 = x + y z2 = x * y z3 = z2 - z1 with tf.Session() as sess: a1, a3 = sess.run([z1, z3]) print a1 print a3 import tensorflow as tf x = tf.constant([3, 5, 7], name="x") y = tf.constant([1, 2, 3], name="y") z1 = tf.add(x, y, name="z1") z2 = x * y z3 = z2 - z1 with tf.Session() as sess: with tf.summary.FileWriter('summaries', sess.graph) as writer: a1, a3 = sess.run([z1, z3]) !ls summaries from google.datalab.ml import TensorBoard TensorBoard().start('./summaries') from google.datalab.ml import TensorBoard TensorBoard().stop(13045) print 'stopped TensorBoard' import tensorflow as tf def forward_pass(w, x): return tf.matmul(w, x) def train_loop(x, niter=5): with tf.variable_scope("model", reuse=tf.AUTO_REUSE): w = tf.get_variable("weights", shape=(1,2), # 1 x 2 matrix initializer=tf.truncated_normal_initializer(), trainable=True) preds = [] for k in xrange(niter): preds.append(forward_pass(w, x)) w = w + 0.1 # "gradient update" return preds with tf.Session() as sess: preds = train_loop(tf.constant([[3.2, 5.1, 7.2],[4.3, 6.2, 8.3]])) # 2 x 3 matrix tf.global_variables_initializer().run() for i in xrange(len(preds)): print "{}:{}".format( i, preds[i].eval() ) ###Output 0:[[-0.53224635 -1.4080029 -2.3759441 ]] 1:[[ 0.21775389 -0.27800274 -0.82594395]] 2:[[0.96775365 0.8519969 0.72405624]] 3:[[1.7177541 1.981997 2.2740564]] 4:[[2.4677541 3.1119976 3.8240576]] ###Markdown Diagrams for Course 3 ###Code import tensorflow as tf x = tf.constant(3) print(x) import tensorflow as tf x = tf.constant([3, 5, 7]) print(x) import tensorflow as tf x = tf.constant([[3, 5, 7], [4, 6, 8]]) print(x) import tensorflow as tf x = tf.constant([[[3, 5, 7],[4, 6, 8]], [[1, 2, 3],[4, 5, 6]] ]) print(x) import tensorflow as tf x1 = tf.constant([2, 3, 4]) x2 = tf.stack([x1, x1]) x3 = tf.stack([x2, x2, x2, x2]) x4 = tf.stack([x3, x3]) print(x1) print(x2) print(x3) print(x4) import tensorflow as tf x = tf.constant([[3, 5, 7], [4, 6, 8]]) y = x[:, 1] with tf.Session() as sess: print(y.eval()) import tensorflow as tf x = tf.constant([[3, 5, 7], [4, 6, 8]]) y = tf.reshape(x, [3, 2]) with tf.Session() as sess: print(y.eval()) import tensorflow as tf x = tf.constant([[3, 5, 7], [4, 6, 8]]) y = tf.reshape(x, [3, 2])[1, :] with tf.Session() as sess: print(y.eval()) import tensorflow as tf from tensorflow.contrib.eager.python import tfe tfe.enable_eager_execution() x = tf.constant([[3, 5, 7], [4, 6, 8]]) y = tf.reshape(x, [3, 2])[1, :] print(y) import tensorflow as tf from tensorflow.contrib.eager.python import tfe tfe.enable_eager_execution() x = tf.constant([3, 5, 7]) y = tf.constant([1, 2, 3]) print(x-y) import tensorflow as tf x = tf.constant([3, 5, 7]) y = tf.constant([1, 2, 3]) z = tf.add(x, y) with tf.Session() as sess: print(z.eval()) import tensorflow as tf x = tf.constant([3, 5, 7]) y = tf.constant([1, 2, 3]) z = tf.add(x, y) with tf.Session() as sess: print(sess.run(z)) import tensorflow as tf x = tf.constant([3, 5, 7]) y = tf.constant([1, 2, 3]) z1 = x + y z2 = x * y z3 = z2 - z1 with tf.Session() as sess: a1, a3 = sess.run([z1, z3]) print(a1) print(a3) import tensorflow as tf x = tf.constant([3, 5, 7], name="x") y = tf.constant([1, 2, 3], name="y") z1 = tf.add(x, y, name="z1") z2 = x * y z3 = z2 - z1 with tf.Session() as sess: with tf.summary.FileWriter('summaries', sess.graph) as writer: a1, a3 = sess.run([z1, z3]) !ls summaries from google.datalab.ml import TensorBoard TensorBoard().start('./summaries') from google.datalab.ml import TensorBoard TensorBoard().stop(13045) print('stopped TensorBoard') import tensorflow as tf def forward_pass(w, x): return tf.matmul(w, x) def train_loop(x, niter=5): with tf.variable_scope("model", reuse=tf.AUTO_REUSE): w = tf.get_variable("weights", shape=(1,2), # 1 x 2 matrix initializer=tf.truncated_normal_initializer(), trainable=True) preds = [] for k in range(niter): preds.append(forward_pass(w, x)) w = w + 0.1 # "gradient update" return preds with tf.Session() as sess: preds = train_loop(tf.constant([[3.2, 5.1, 7.2],[4.3, 6.2, 8.3]])) # 2 x 3 matrix tf.global_variables_initializer().run() for i in range(len(preds)): print("{}:{}".format( i, preds[i].eval() )) ###Output 0:[[ 8.568541 12.702375 17.271353]] 1:[[ 9.318541 13.8323765 18.821354 ]] 2:[[10.068541 14.962376 20.371353]] 3:[[10.818541 16.092377 21.921354]] 4:[[11.56854 17.222376 23.471352]]
example/ex.ipynb
###Markdown Test and Demonstrate the use of lp_solve[click here](http://lpsolve.sourceforge.net/5.5/Python.htm) for more details. import library ###Code from lp_solve import * ###Output _____no_output_____ ###Markdown Example 1 from the lp_solve distribution ###Code f = [-1, 2] A = [[2, 1], [-4, 4]] b = [5, 5] e = [-1, -1] xint = [1, 2] [v,x,duals] = lp_solve(f,A,b,e,None,None,xint) print(v) print(x) ###Output _____no_output_____ ###Markdown Example 2 ###Code f = [50, 100] A = [[10, 5],[4, 10],[1, 1.5]] b = [2500, 2000, 450] e = [-1, -1, -1] [v,x,duals] = lp_solve(f,A,b,e) print(v) print(x) ###Output _____no_output_____ ###Markdown Example 3 ###Code f = [-40, -36] vub = [8, 10] A = [[5, 3]] b = [45] e = [1] [v,x,duals] = lp_solve(f,A,b,e,None,vub) print(v) print(x) ###Output _____no_output_____ ###Markdown Example 4 ###Code f = [10, 6, 4] A = [[1, 1, 1], [10, 4, 5], [2, 2, 6]] b = [100, 600, 300] e = [-1, -1, -1] xint = [2] [v,x,duals] = lp_solve(f,A,b,e,None,None,xint) print(v) print(x) ###Output _____no_output_____ ###Markdown Example 5 Integer programming example, page 218 of Ecker & Kupferschmid ###Code f = [-3, 7, 12] a = [[-3, 6, 8], [6, -3, 7], [-6, 3, 3]] b = [12, 8, 5] e = [-1, -1, -1] xint = [1, 2, 3] [v,x,duals] = lp_solve(f,a,b,e,None,None,xint) print(v) print(x) ###Output _____no_output_____ ###Markdown Example 6 0-1 programming example, page 228 233 of Ecker & Kupferschmid ###Code f = [-2, -3, -7, -7] a = [[1, 1, -2, -5], [-1, 2, 1, 4]] b = [2, -3] e = [1, 1] xint = [1, 2, 3, 4] vub = [1, 1, 1, 1] [v,x,duals] = lp_solve(f,a,b,e,None,vub,xint) print(v) print(x) ###Output _____no_output_____ ###Markdown Example 7 0-1 programming example, page 238 of Ecker & Kupferschmid ###Code f = [-1, -2, -3, -7, -8, -8] a = [[5, -3, 2, -3, -1, 2], [-1, 0, 2, 1, 3, -3], [1, 2, -1, 0, 5, -1]] b = [-5, -1, 3] e = [1, 1, 1] xint = [1, 2, 3, 4, 5, 6] vub = [1, 1, 1, 1, 1, 1] [v,x,duals] = lp_solve(f,a,b,e,None,vub,xint) print(v) print(x) ###Output _____no_output_____ ###Markdown ex2.lp from the lp_solve distribution ###Code f=[8, 15] a = [[10, 21], [2, 1]] b = [156, 22] e = [-1, -1] [v,x,duals] = lp_solve(f,a,b,e) print(v) print(x) ###Output _____no_output_____ ###Markdown ex3.lp from the lp_solve distribution ###Code f=[3, 13] a = [[2, 9], [11, -8]] b = [40, 82] e = [-1, -1] [v,x,duals] = lp_solve(f,a,b,e) print(v) print(x) ###Output _____no_output_____ ###Markdown ex6.lp from the lp_solve distribution ###Code f=[592, 381, 273, 55, 48, 37, 23] a = [[3534, 2356, 1767, 589, 528, 451, 304]] b = [119567] e = [-1] xint = [1, 2, 3, 4, 5, 6, 7] vub = None [v,x,duals] = lp_solve(f,a,b,e,None,vub,xint) print(v) print(x) ###Output _____no_output_____
recursive_filters/introduction.ipynb
###Markdown Realization of Recursive Filters*This jupyter notebook is part of a [collection of notebooks](../index.ipynb) on various topics of Digital Signal Processing. Please direct questions and suggestions to [[email protected]](mailto:[email protected]).* IntroductionComputing the output $y[k] = \mathcal{H} \{ x[k] \}$ of a [linear time-invariant](https://en.wikipedia.org/wiki/LTI_system_theory) (LTI) system is of central importance in digital signal processing. This is often referred to as [*filtering*](https://en.wikipedia.org/wiki/Digital_filter) of the input signal $x[k]$. We already have discussed the realization of [non-recursive filters](../nonrecursive_filters/introduction.ipynb). This section focuses on the realization of recursive filters. Recursive FiltersLinear difference equations with constant coefficients represent linear time-invariant (LTI) systems\begin{equation}\sum_{n=0}^{N} a_n \; y[k-n] = \sum_{m=0}^{M} b_m \; x[k-m]\end{equation}where $y[k] = \mathcal{H} \{ x[k] \}$ denotes the response of the system to the input signal $x[k]$, $N$ the order, $a_n$ and $b_m$ constant coefficients, respectively. Above equation can be rearranged with respect to the output signal $y[k]$ by extracting the first element ($n=0$) of the left-hand sum\begin{equation}y[k] = \frac{1}{a_0} \left( \sum_{m=0}^{M} b_m \; x[k-m] - \sum_{n=1}^{N} a_n \; y[k-n] \right)\end{equation}It is evident that the output signal $y[k]$ at time instant $k$ is given as a linear combination of past output samples $y[k-n]$ superimposed by a linear combination of the actual $x[k]$ and past $x[k-m]$ input samples. Hence, the actual output $y[k]$ is composed from the two contributions1. a [non-recursive part](../nonrecursive_filters/introduction.ipynbNon-Recursive-Filters), and2. a recursive part where a linear combination of past output samples is fed back.The impulse response of the system is given as the response of the system to a Dirac impulse at the input $h[k] = \mathcal{H} \{ \delta[k] \}$. Using above result and the properties of the discrete Dirac impulse we get\begin{equation}h[k] = \frac{1}{a_0} \left( b_k - \sum_{n=1}^{N} a_n \; h[k-n] \right)\end{equation}Due to the feedback, the impulse response will in general be of infinite length. The impulse response is termed as [infinite impulse response](https://en.wikipedia.org/wiki/Infinite_impulse_response) (IIR) and the system as recursive system/filter. Transfer FunctionApplying a $z$-transform to the left- and right-hand side of the difference equation and rearranging terms yields the transfer function $H(z)$ of the system\begin{equation}H(z) = \frac{Y(z)}{X(z)} = \frac{\sum_{m=0}^{M} b_m \; z^{-m}}{\sum_{n=0}^{N} a_n \; z^{-n}}\end{equation}The transfer function is given as a [rational function](https://en.wikipedia.org/wiki/Rational_function) in $z$. The polynominals of the numerator and denominator can be expressed alternatively by their roots as\begin{equation}H(z) = \frac{b_M}{a_N} \cdot \frac{\prod_{\mu=1}^{P} (z - z_{0\mu})^{m_\mu}}{\prod_{\nu=1}^{Q} (z - z_{\infty\nu})^{n_\nu}}\end{equation}where $z_{0\mu}$ and $z_{\infty\nu}$ denote the $\mu$-th zero and $\nu$-th pole of degree $m_\mu$ and $n_\nu$ of $H(z)$, respectively. The total number of zeros and poles is denoted by $P$ and $Q$. Due to the symmetries of the $z$-transform, the transfer function of a real-valued system $h[k] \in \mathbb{R}$ exhibits complex conjugate symmetry\begin{equation}H(z) = H^*(z^*)\end{equation}Poles and zeros are either real valued or complex conjugate pairs for real-valued systems ($b_m\in\mathbb{R}$, $a_n\in\mathbb{R}$). For the poles of a causal and stable system $H(z)$ the following condition has to hold\begin{equation}\max_{\nu} | z_{\infty\nu} | < 1\end{equation}Hence, all poles have to be located inside the unit circle $|z| = 1$. Amongst others, this implies that $M \leq N$. ExampleThe following example shows the pole/zero diagram, the magnitude and phase response, and impulse response of a recursive filter with so-called [Butterworth](https://en.wikipedia.org/wiki/Butterworth_filter) lowpass characteristic. ###Code import numpy as np import matplotlib.pyplot as plt from matplotlib.markers import MarkerStyle from matplotlib.patches import Circle import scipy.signal as sig N = 5 # order of recursive filter L = 128 # number of computed samples def zplane(z, p, title='Poles and Zeros'): "Plots zero and pole locations in the complex z-plane" ax = plt.gca() ax.plot(np.real(z), np.imag(z), 'bo', fillstyle='none', ms=10) ax.plot(np.real(p), np.imag(p), 'rx', fillstyle='none', ms=10) unit_circle = Circle((0, 0), radius=1, fill=False, color='black', ls='solid', alpha=0.9) ax.add_patch(unit_circle) ax.axvline(0, color='0.7') ax.axhline(0, color='0.7') plt.title(title) plt.xlabel(r'Re{$z$}') plt.ylabel(r'Im{$z$}') plt.axis('equal') plt.xlim((-2, 2)) plt.ylim((-2, 2)) plt.grid() # compute coefficients of recursive filter b, a = sig.butter(N, 0.2, 'low') # compute transfer function Om, H = sig.freqz(b, a) # compute impulse response k = np.arange(L) x = np.where(k == 0, 1.0, 0) h = sig.lfilter(b, a, x) # plot pole/zero-diagram plt.figure(figsize=(5, 5)) zplane(np.roots(b), np.roots(a)) # plot magnitude response plt.figure(figsize=(10, 3)) plt.plot(Om, 20 * np.log10(abs(H))) plt.xlabel(r'$\Omega$') plt.ylabel(r'$|H(e^{j \Omega})|$ in dB') plt.grid() plt.title('Magnitude response') # plot phase response plt.figure(figsize=(10, 3)) plt.plot(Om, np.unwrap(np.angle(H))) plt.xlabel(r'$\Omega$') plt.ylabel(r'$\varphi (\Omega)$ in rad') plt.grid() plt.title('Phase response') # plot impulse response (magnitude) plt.figure(figsize=(10, 3)) plt.stem(20*np.log10(np.abs(np.squeeze(h))), use_line_collection=True) plt.xlabel(r'$k$') plt.ylabel(r'$|h[k]|$ in dB') plt.grid() plt.title('Impulse response (magnitude)') ###Output _____no_output_____ ###Markdown Realization of Recursive Filters*This jupyter notebook is part of a [collection of notebooks](../index.ipynb) on various topics of Digital Signal Processing. Please direct questions and suggestions to [[email protected]](mailto:[email protected]).* IntroductionComputing the output $y[k] = \mathcal{H} \{ x[k] \}$ of a [linear time-invariant](https://en.wikipedia.org/wiki/LTI_system_theory) (LTI) system is of central importance in digital signal processing. This is often referred to as [*filtering*](https://en.wikipedia.org/wiki/Digital_filter) of the input signal $x[k]$. We already have discussed the realization of [non-recursive filters](../nonrecursive_filters/introduction.ipynb). This section focuses on the realization of recursive filters. Recursive FiltersLinear difference equations with constant coefficients represent linear time-invariant (LTI) systems\begin{equation}\sum_{n=0}^{N} a_n \; y[k-n] = \sum_{m=0}^{M} b_m \; x[k-m]\end{equation}where $y[k] = \mathcal{H} \{ x[k] \}$ denotes the response of the system to the input signal $x[k]$, $N$ the order, $a_n$ and $b_m$ constant coefficients, respectively. Above equation can be rearranged with respect to the output signal $y[k]$ by extracting the first element ($n=0$) of the left-hand sum\begin{equation}y[k] = \frac{1}{a_0} \left( \sum_{m=0}^{M} b_m \; x[k-m] - \sum_{n=1}^{N} a_n \; y[k-n] \right)\end{equation}It is evident that the output signal $y[k]$ at time instant $k$ is given as a linear combination of past output samples $y[k-n]$ superimposed by a linear combination of the actual $x[k]$ and past $x[k-m]$ input samples. Hence, the actual output $y[k]$ is composed from the two contributions1. a [non-recursive part](../nonrecursive_filters/introduction.ipynbNon-Recursive-Filters), and2. a recursive part where a linear combination of past output samples is fed back.The impulse response of the system is given as the response of the system to a Dirac impulse at the input $h[k] = \mathcal{H} \{ \delta[k] \}$. Using above result and the properties of the discrete Dirac impulse we get\begin{equation}h[k] = \frac{1}{a_0} \left( b_k - \sum_{n=1}^{N} a_n \; h[k-n] \right)\end{equation}Due to the feedback, the impulse response will in general be of infinite length. The impulse response is termed as [infinite impulse response](https://en.wikipedia.org/wiki/Infinite_impulse_response) (IIR) and the system as recursive system/filter. Transfer FunctionApplying a $z$-transform to the left- and right-hand side of the difference equation and rearranging terms yields the transfer function $H(z)$ of the system\begin{equation}H(z) = \frac{Y(z)}{X(z)} = \frac{\sum_{m=0}^{M} b_m \; z^{-m}}{\sum_{n=0}^{N} a_n \; z^{-n}}\end{equation}The transfer function is given as a [rational function](https://en.wikipedia.org/wiki/Rational_function) in $z$. The polynominals of the numerator and denominator can be expressed alternatively by their roots as\begin{equation}H(z) = \frac{b_M}{a_N} \cdot \frac{\prod_{\mu=1}^{P} (z - z_{0\mu})^{m_\mu}}{\prod_{\nu=1}^{Q} (z - z_{\infty\nu})^{n_\nu}}\end{equation}where $z_{0\mu}$ and $z_{\infty\nu}$ denote the $\mu$-th zero and $\nu$-th pole of degree $m_\mu$ and $n_\nu$ of $H(z)$, respectively. The total number of zeros and poles is denoted by $P$ and $Q$. Due to the symmetries of the $z$-transform, the transfer function of a real-valued system $h[k] \in \mathbb{R}$ exhibits complex conjugate symmetry\begin{equation}H(z) = H^*(z^*)\end{equation}Poles and zeros are either real valued or complex conjugate pairs for real-valued systems ($b_m\in\mathbb{R}$, $a_n\in\mathbb{R}$). For the poles of a causal and stable system $H(z)$ the following condition has to hold\begin{equation}\max_{\nu} | z_{\infty\nu} | < 1\end{equation}Hence, all poles have to be located inside the unit circle $|z| = 1$. Amongst others, this implies that $M \leq N$. ExampleThe following example shows the pole/zero diagram, the magnitude and phase response, and impulse response of a recursive filter with so-called [Butterworth](https://en.wikipedia.org/wiki/Butterworth_filter) lowpass characteristic. ###Code %matplotlib inline import numpy as np import matplotlib.pyplot as plt from matplotlib.markers import MarkerStyle from matplotlib.patches import Circle import scipy.signal as sig N = 5 # order of recursive filter L = 128 # number of computed samples def zplane(z, p, title='Poles and Zeros'): "Plots zero and pole locations in the complex z-plane" ax = plt.gca() ax.plot(np.real(z), np.imag(z), 'bo', fillstyle='none', ms = 10) ax.plot(np.real(p), np.imag(p), 'rx', fillstyle='none', ms = 10) unit_circle = Circle((0,0), radius=1, fill=False, color='black', ls='solid', alpha=0.9) ax.add_patch(unit_circle) ax.axvline(0, color='0.7') ax.axhline(0, color='0.7') plt.title(title) plt.xlabel(r'Re{$z$}') plt.ylabel(r'Im{$z$}') plt.axis('equal') plt.xlim((-2, 2)) plt.ylim((-2, 2)) plt.grid() # compute coefficients of recursive filter b, a = sig.butter(N, 0.2, 'low') # compute transfer function Om, H = sig.freqz(b, a) # compute impulse response k = np.arange(L) x = np.where(k==0, 1.0, 0) h = sig.lfilter(b, a, x) # plot pole/zero-diagram plt.figure(figsize=(5, 5)) zplane(np.roots(b), np.roots(a)) # plot magnitude response plt.figure(figsize=(10, 3)) plt.plot(Om, 20 * np.log10(abs(H))) plt.xlabel(r'$\Omega$') plt.ylabel(r'$|H(e^{j \Omega})|$ in dB') plt.grid() plt.title('Magnitude response') # plot phase response plt.figure(figsize=(10, 3)) plt.plot(Om, np.unwrap(np.angle(H))) plt.xlabel(r'$\Omega$') plt.ylabel(r'$\varphi (\Omega)$ in rad') plt.grid() plt.title('Phase response') # plot impulse response (magnitude) plt.figure(figsize=(10, 3)) plt.stem(20*np.log10(np.abs(np.squeeze(h)))) plt.xlabel(r'$k$') plt.ylabel(r'$|h[k]|$ in dB') plt.grid() plt.title('Impulse response (magnitude)'); ###Output _____no_output_____ ###Markdown Realization of Recursive Filters*This jupyter notebook is part of a [collection of notebooks](../index.ipynb) on various topics of Digital Signal Processing. Please direct questions and suggestions to [[email protected]](mailto:[email protected]).* IntroductionComputing the output $y[k] = \mathcal{H} \{ x[k] \}$ of a [linear time-invariant](https://en.wikipedia.org/wiki/LTI_system_theory) (LTI) system is of central importance in digital signal processing. This is often referred to as [*filtering*](https://en.wikipedia.org/wiki/Digital_filter) of the input signal $x[k]$. We already have discussed the realization of [non-recursive filters](../nonrecursive_filters/introduction.ipynb). This section focuses on the realization of recursive filters. Recursive FiltersLinear difference equations with constant coefficients represent linear-time invariant (LTI) systems\begin{equation}\sum_{n=0}^{N} a_n \; y[k-n] = \sum_{m=0}^{M} b_m \; x[k-m]\end{equation}where $y[k] = \mathcal{H} \{ x[k] \}$ denotes the response of the system to the input signal $x[k]$, $N$ the order, $a_n$ and $b_m$ constant coefficients, respectively. Above equation can be rearranged with respect to the output signal $y[k]$ by extracting the first element ($n=0$) of the left hand sum\begin{equation}y[k] = \frac{1}{a_0} \left( \sum_{m=0}^{M} b_m \; x[k-m] - \sum_{n=1}^{N} a_n \; y[k-n] \right)\end{equation}It is evident that the output signal $y[k]$ at time instant $k$ is given as a linear combination of past output samples $y[k-n]$ superimposed by a linear combination of the actual $x[k]$ and past $x[k-m]$ input samples. Hence, the actual output $y[k]$ is composed from the two contributions1. a [non-recursive part](../nonrecursive_filters/introduction.ipynbNon-Recursive-Filters), and2. a recursive part where a linear combination of past output samples is fed back.The impulse response of the system is given as the response of the system to a Dirac impulse at the input $h[k] = \mathcal{H} \{ \delta[k] \}$. Using above result and the properties of the discrete Dirac impulse we get\begin{equation}h[k] = \frac{1}{a_0} \left( b_k - \sum_{n=1}^{N} a_n \; h[k-n] \right)\end{equation}Due to the feedback, the impulse response will in general be of infinite length. The impulse response is termed as [infinite impulse response](https://en.wikipedia.org/wiki/Infinite_impulse_response) (IIR) and the system as recursive system/filter. Transfer FunctionApplying a $z$-transform to the left and right hand side of the difference equation and rearranging terms yields the transfer function $H(z)$ of the system\begin{equation}H(z) = \frac{Y(z)}{X(z)} = \frac{\sum_{m=0}^{M} b_m \; z^{-m}}{\sum_{n=0}^{N} a_n \; z^{-n}}\end{equation}The transfer function is given as a [rational function](https://en.wikipedia.org/wiki/Rational_function) in $z$. The polynominals of the numerator and denominator can expressed alternatively by their roots as\begin{equation}H(z) = \frac{b_M}{a_N} \cdot \frac{\prod_{\mu=1}^{P} (z - z_{0\mu})^{m_\mu}}{\prod_{\nu=1}^{Q} (z - z_{\infty\nu})^{n_\nu}}\end{equation}where $z_{0\mu}$ and $z_{\infty\nu}$ denote the $\mu$-th zero and $\nu$-th pole of degree $m_\mu$ and $n_\nu$ of $H(z)$, respectively. The total number of zeros and poles is denoted by $P$ and $Q$. Due to the symmetries of the $z$-transform, the transfer function of a real-valued system $h[k] \in \mathbb{R}$ exhibits complex conjugate symmetry\begin{equation}H(z) = H^*(z^*)\end{equation}Poles and zeros are either real valued or conjugate complex pairs for real-valued systems ($b_m\in\mathbb{R}$, $a_n\in\mathbb{R}$). For the poles of a causal and stable system $H(z)$ the following condition has to hold\begin{equation}\max_{\nu} | z_{\infty\nu} | < 1\end{equation}Hence all poles have to be located inside the unit circle $|z| = 1$. Amongst others, this implies that $M \leq N$. ExampleThe following example shows the pole/zero diagram, the magnitude and phase response, and impulse response of a recursive filter with so called [Butterworth](https://en.wikipedia.org/wiki/Butterworth_filter) lowpass characteristic. ###Code %matplotlib inline import numpy as np import matplotlib.pyplot as plt from matplotlib.markers import MarkerStyle from matplotlib.patches import Circle import scipy.signal as sig N = 5 # order of recursive filter def zplane(z, p): fig = plt.figure(figsize=(5,5)) ax = fig.gca() plt.hold(True) unit_circle = Circle((0,0), radius=1, fill=False, color='black', ls='solid', alpha=0.9) ax.add_patch(unit_circle) ax.axvline(0, color='0.7') ax.axhline(0, color='0.7') plt.axis('equal') plt.xlim((-2, 2)) plt.ylim((-2, 2)) plt.grid() plt.title('Poles and Zeros') plt.xlabel(r'Re{$z$}') plt.ylabel(r'Im{$z$}') ax.plot(np.real(z), np.imag(z), 'bo', fillstyle='none', ms = 10) ax.plot(np.real(p), np.imag(p), 'rx', fillstyle='none', ms = 10) plt.hold(False) # coefficients of recursive filter b, a = sig.butter(N, 0.2, 'low') # compute transfer function of filter Om, H = sig.freqz(b, a) # compute impulse response k = np.arange(128) x = np.where(k==0, 1.0, 0) h = sig.lfilter(b, a, x) # plot pole/zero-diagram zplane(np.roots(b), np.roots(a)) # plot magnitude response plt.figure(figsize=(10, 3)) plt.plot(Om, 20 * np.log10(abs(H))) plt.xlabel(r'$\Omega$') plt.ylabel(r'$|H(e^{j \Omega})|$ in dB') plt.grid() plt.title('Magnitude response') # plot phase response plt.figure(figsize=(10, 3)) plt.plot(Om, np.unwrap(np.angle(H))) plt.xlabel(r'$\Omega$') plt.ylabel(r'$\varphi (\Omega)$ in rad') plt.grid() plt.title('Phase') # plot impulse response plt.figure(figsize=(10, 3)) plt.stem(20*np.log10(np.abs(np.squeeze(h)))) plt.xlabel(r'$k$') plt.ylabel(r'$|h[k]|$ in dB') plt.grid() plt.title('Impulse response'); ###Output _____no_output_____ ###Markdown Realization of Recursive Filters*This jupyter notebook is part of a [collection of notebooks](../index.ipynb) on various topics of Digital Signal Processing. Please direct questions and suggestions to [[email protected]](mailto:[email protected]).* IntroductionComputing the output $y[k] = \mathcal{H} \{ x[k] \}$ of a [linear time-invariant](https://en.wikipedia.org/wiki/LTI_system_theory) (LTI) system is of central importance in digital signal processing. This is often referred to as [*filtering*](https://en.wikipedia.org/wiki/Digital_filter) of the input signal $x[k]$. We already have discussed the realization of [non-recursive filters](../nonrecursive_filters/introduction.ipynb). This section focuses on the realization of recursive filters. Recursive FiltersLinear difference equations with constant coefficients represent linear time-invariant (LTI) systems\begin{equation}\sum_{n=0}^{N} a_n \; y[k-n] = \sum_{m=0}^{M} b_m \; x[k-m]\end{equation}where $y[k] = \mathcal{H} \{ x[k] \}$ denotes the response of the system to the input signal $x[k]$, $N$ the order, $a_n$ and $b_m$ constant coefficients, respectively. Above equation can be rearranged with respect to the output signal $y[k]$ by extracting the first element ($n=0$) of the left-hand sum\begin{equation}y[k] = \frac{1}{a_0} \left( \sum_{m=0}^{M} b_m \; x[k-m] - \sum_{n=1}^{N} a_n \; y[k-n] \right)\end{equation}It is evident that the output signal $y[k]$ at time instant $k$ is given as a linear combination of past output samples $y[k-n]$ superimposed by a linear combination of the actual $x[k]$ and past $x[k-m]$ input samples. Hence, the actual output $y[k]$ is composed from the two contributions1. a [non-recursive part](../nonrecursive_filters/introduction.ipynbNon-Recursive-Filters), and2. a recursive part where a linear combination of past output samples is fed back.The impulse response of the system is given as the response of the system to a Dirac impulse at the input $h[k] = \mathcal{H} \{ \delta[k] \}$. Using above result and the properties of the discrete Dirac impulse we get\begin{equation}h[k] = \frac{1}{a_0} \left( b_k - \sum_{n=1}^{N} a_n \; h[k-n] \right)\end{equation}Due to the feedback, the impulse response will in general be of infinite length. The impulse response is termed as [infinite impulse response](https://en.wikipedia.org/wiki/Infinite_impulse_response) (IIR) and the system as recursive system/filter. Transfer FunctionApplying a $z$-transform to the left- and right-hand side of the difference equation and rearranging terms yields the transfer function $H(z)$ of the system\begin{equation}H(z) = \frac{Y(z)}{X(z)} = \frac{\sum_{m=0}^{M} b_m \; z^{-m}}{\sum_{n=0}^{N} a_n \; z^{-n}}\end{equation}The transfer function is given as a [rational function](https://en.wikipedia.org/wiki/Rational_function) in $z$. The polynominals of the numerator and denominator can be expressed alternatively by their roots as\begin{equation}H(z) = \frac{b_M}{a_N} \cdot \frac{\prod_{\mu=1}^{P} (z - z_{0\mu})^{m_\mu}}{\prod_{\nu=1}^{Q} (z - z_{\infty\nu})^{n_\nu}}\end{equation}where $z_{0\mu}$ and $z_{\infty\nu}$ denote the $\mu$-th zero and $\nu$-th pole of degree $m_\mu$ and $n_\nu$ of $H(z)$, respectively. The total number of zeros and poles is denoted by $P$ and $Q$. Due to the symmetries of the $z$-transform, the transfer function of a real-valued system $h[k] \in \mathbb{R}$ exhibits complex conjugate symmetry\begin{equation}H(z) = H^*(z^*)\end{equation}Poles and zeros are either real valued or complex conjugate pairs for real-valued systems ($b_m\in\mathbb{R}$, $a_n\in\mathbb{R}$). For the poles of a causal and stable system $H(z)$ the following condition has to hold\begin{equation}\max_{\nu} | z_{\infty\nu} | < 1\end{equation}Hence, all poles have to be located inside the unit circle $|z| = 1$. Amongst others, this implies that $M \leq N$. ExampleThe following example shows the pole/zero diagram, the magnitude and phase response, and impulse response of a recursive filter with so-called [Butterworth](https://en.wikipedia.org/wiki/Butterworth_filter) lowpass characteristic. ###Code import numpy as np import matplotlib.pyplot as plt from matplotlib.markers import MarkerStyle from matplotlib.patches import Circle import scipy.signal as sig %matplotlib inline N = 5 # order of recursive filter L = 128 # number of computed samples def zplane(z, p, title='Poles and Zeros'): "Plots zero and pole locations in the complex z-plane" ax = plt.gca() ax.plot(np.real(z), np.imag(z), 'bo', fillstyle='none', ms=10) ax.plot(np.real(p), np.imag(p), 'rx', fillstyle='none', ms=10) unit_circle = Circle((0, 0), radius=1, fill=False, color='black', ls='solid', alpha=0.9) ax.add_patch(unit_circle) ax.axvline(0, color='0.7') ax.axhline(0, color='0.7') plt.title(title) plt.xlabel(r'Re{$z$}') plt.ylabel(r'Im{$z$}') plt.axis('equal') plt.xlim((-2, 2)) plt.ylim((-2, 2)) plt.grid() # compute coefficients of recursive filter b, a = sig.butter(N, 0.2, 'low') # compute transfer function Om, H = sig.freqz(b, a) # compute impulse response k = np.arange(L) x = np.where(k == 0, 1.0, 0) h = sig.lfilter(b, a, x) # plot pole/zero-diagram plt.figure(figsize=(5, 5)) zplane(np.roots(b), np.roots(a)) # plot magnitude response plt.figure(figsize=(10, 3)) plt.plot(Om, 20 * np.log10(abs(H))) plt.xlabel(r'$\Omega$') plt.ylabel(r'$|H(e^{j \Omega})|$ in dB') plt.grid() plt.title('Magnitude response') # plot phase response plt.figure(figsize=(10, 3)) plt.plot(Om, np.unwrap(np.angle(H))) plt.xlabel(r'$\Omega$') plt.ylabel(r'$\varphi (\Omega)$ in rad') plt.grid() plt.title('Phase response') # plot impulse response (magnitude) plt.figure(figsize=(10, 3)) plt.stem(20*np.log10(np.abs(np.squeeze(h))), use_line_collection=True) plt.xlabel(r'$k$') plt.ylabel(r'$|h[k]|$ in dB') plt.grid() plt.title('Impulse response (magnitude)') ###Output _____no_output_____ ###Markdown Realization of Recursive Filters*This jupyter notebook is part of a [collection of notebooks](../index.ipynb) on various topics of Digital Signal Processing. Please direct questions and suggestions to [[email protected]](mailto:[email protected]).* IntroductionComputing the output $y[k] = \mathcal{H} \{ x[k] \}$ of a [linear time-invariant](https://en.wikipedia.org/wiki/LTI_system_theory) (LTI) system is of central importance in digital signal processing. This is often referred to as [*filtering*](https://en.wikipedia.org/wiki/Digital_filter) of the input signal $x[k]$. We already have discussed the realization of [non-recursive filters](../nonrecursive_filters/introduction.ipynb). This section focuses on the realization of recursive filters. Recursive FiltersLinear difference equations with constant coefficients represent linear time-invariant (LTI) systems\begin{equation}\sum_{n=0}^{N} a_n \; y[k-n] = \sum_{m=0}^{M} b_m \; x[k-m]\end{equation}where $y[k] = \mathcal{H} \{ x[k] \}$ denotes the response of the system to the input signal $x[k]$, $N$ the order, $a_n$ and $b_m$ constant coefficients, respectively. Above equation can be rearranged with respect to the output signal $y[k]$ by extracting the first element ($n=0$) of the left-hand sum\begin{equation}y[k] = \frac{1}{a_0} \left( \sum_{m=0}^{M} b_m \; x[k-m] - \sum_{n=1}^{N} a_n \; y[k-n] \right)\end{equation}It is evident that the output signal $y[k]$ at time instant $k$ is given as a linear combination of past output samples $y[k-n]$ superimposed by a linear combination of the actual $x[k]$ and past $x[k-m]$ input samples. Hence, the actual output $y[k]$ is composed from the two contributions1. a [non-recursive part](../nonrecursive_filters/introduction.ipynbNon-Recursive-Filters), and2. a recursive part where a linear combination of past output samples is fed back.The impulse response of the system is given as the response of the system to a Dirac impulse at the input $h[k] = \mathcal{H} \{ \delta[k] \}$. Using above result and the properties of the discrete Dirac impulse we get\begin{equation}h[k] = \frac{1}{a_0} \left( b_k - \sum_{n=1}^{N} a_n \; h[k-n] \right)\end{equation}Due to the feedback, the impulse response will in general be of infinite length. The impulse response is termed as [infinite impulse response](https://en.wikipedia.org/wiki/Infinite_impulse_response) (IIR) and the system as recursive system/filter. Transfer FunctionApplying a $z$-transform to the left- and right-hand side of the difference equation and rearranging terms yields the transfer function $H(z)$ of the system\begin{equation}H(z) = \frac{Y(z)}{X(z)} = \frac{\sum_{m=0}^{M} b_m \; z^{-m}}{\sum_{n=0}^{N} a_n \; z^{-n}}\end{equation}The transfer function is given as a [rational function](https://en.wikipedia.org/wiki/Rational_function) in $z$. The polynominals of the numerator and denominator can be expressed alternatively by their roots as\begin{equation}H(z) = \frac{b_M}{a_N} \cdot \frac{\prod_{\mu=1}^{P} (z - z_{0\mu})^{m_\mu}}{\prod_{\nu=1}^{Q} (z - z_{\infty\nu})^{n_\nu}}\end{equation}where $z_{0\mu}$ and $z_{\infty\nu}$ denote the $\mu$-th zero and $\nu$-th pole of degree $m_\mu$ and $n_\nu$ of $H(z)$, respectively. The total number of zeros and poles is denoted by $P$ and $Q$. Due to the symmetries of the $z$-transform, the transfer function of a real-valued system $h[k] \in \mathbb{R}$ exhibits complex conjugate symmetry\begin{equation}H(z) = H^*(z^*)\end{equation}Poles and zeros are either real valued or complex conjugate pairs for real-valued systems ($b_m\in\mathbb{R}$, $a_n\in\mathbb{R}$). For the poles of a causal and stable system $H(z)$ the following condition has to hold\begin{equation}\max_{\nu} | z_{\infty\nu} | < 1\end{equation}Hence, all poles have to be located inside the unit circle $|z| = 1$. Amongst others, this implies that $M \leq N$. ExampleThe following example shows the pole/zero diagram, the magnitude and phase response, and impulse response of a recursive filter with so-called [Butterworth](https://en.wikipedia.org/wiki/Butterworth_filter) lowpass characteristic. ###Code %matplotlib inline import numpy as np import matplotlib.pyplot as plt from matplotlib.markers import MarkerStyle from matplotlib.patches import Circle import scipy.signal as sig N = 5 # order of recursive filter L = 128 # number of computed samples def zplane(z, p, title='Poles and Zeros'): "Plots zero and pole locations in the complex z-plane" ax = plt.gca() ax.plot(np.real(z), np.imag(z), 'bo', fillstyle='none', ms = 10) ax.plot(np.real(p), np.imag(p), 'rx', fillstyle='none', ms = 10) unit_circle = Circle((0,0), radius=1, fill=False, color='black', ls='solid', alpha=0.9) ax.add_patch(unit_circle) ax.axvline(0, color='0.7') ax.axhline(0, color='0.7') plt.title(title) plt.xlabel(r'Re{$z$}') plt.ylabel(r'Im{$z$}') plt.axis('equal') plt.xlim((-2, 2)) plt.ylim((-2, 2)) plt.grid() # compute coefficients of recursive filter b, a = sig.butter(N, 0.2, 'low') # compute transfer function Om, H = sig.freqz(b, a) # compute impulse response k = np.arange(L) x = np.where(k==0, 1.0, 0) h = sig.lfilter(b, a, x) # plot pole/zero-diagram plt.figure(figsize=(5, 5)) zplane(np.roots(b), np.roots(a)) # plot magnitude response plt.figure(figsize=(10, 3)) plt.plot(Om, 20 * np.log10(abs(H))) plt.xlabel(r'$\Omega$') plt.ylabel(r'$|H(e^{j \Omega})|$ in dB') plt.grid() plt.title('Magnitude response') # plot phase response plt.figure(figsize=(10, 3)) plt.plot(Om, np.unwrap(np.angle(H))) plt.xlabel(r'$\Omega$') plt.ylabel(r'$\varphi (\Omega)$ in rad') plt.grid() plt.title('Phase response') # plot impulse response (magnitude) plt.figure(figsize=(10, 3)) plt.stem(20*np.log10(np.abs(np.squeeze(h)))) plt.xlabel(r'$k$') plt.ylabel(r'$|h[k]|$ in dB') plt.grid() plt.title('Impulse response (magnitude)'); ###Output _____no_output_____
notebooks/wandb/run-20200211_133637-lr9ii429/code/notebooks/SiameseNetChannelCharting-v1.ipynb
###Markdown We observe that a lot of information is contained on the imaginary part of the impulse. So the 16 antennas, we are gong to have 32 'Channels' for our dataset. So we will have a training batch of shape [batch_size, 32 , 100]. Siamese Neural Network ###Code import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.utils.data import Dataset, DataLoader from sklearn.model_selection import train_test_split ###Output _____no_output_____ ###Markdown Setting up the Custom Dataset ###Code # undersampling idces = np.random.randint(0, data.shape[0], int(0.3*data.shape[0])) data_undersampled = data[idces] data_undersampled.shape data_undersampled.shape # train test split train, test= train_test_split(data_undersampled) train_dataset = data_preparation.SiameseDataset(train) scaler = train_dataset.scaler_real, train_dataset.scaler_imag test_dataset = data_preparation.SiameseDataset(test, scaler) plt.figure(figsize=(20,20)) for i in range(1, 17): plt.subplot(4,4,i) plt.plot(train_dataset[0][0][i-1, :], label='1_sample') #plt.plot(train_dataset[0][1][i-1, :], label='2_sample') plt.legend() train_dataset.nb_channels() class SimpleNN(nn.Module): def __init__(self): super(SimpleNN, self).__init__() self.conv1 = nn.Conv1d(in_channels=train_dataset.nb_channels(), out_channels=128, kernel_size=16) self.conv2 = nn.Conv1d(in_channels=128, out_channels=64, kernel_size=8) self.conv3 = nn.Conv1d(in_channels=64, out_channels=16, kernel_size=4) f = data_preparation.conv1d_output_size self.features = f(f(f(train_dataset.nb_samples(),kernel_size=16), kernel_size=8), kernel_size=4) self.lin1 = nn.Linear(in_features= 16 * self.features, out_features=128) self.lin2 = nn.Linear(in_features=128, out_features=32) self.lin3 = nn.Linear(in_features=32, out_features=8) self.lin4 = nn.Linear(in_features=8, out_features=3) def forward(self, x): x = F.relu(self.conv1(x)) x = F.relu(self.conv2(x)) x = F.relu(self.conv3(x)) x = torch.flatten(x, 1) x = F.relu(self.lin1(x)) x = F.relu(self.lin2(x)) x = F.relu(self.lin3(x)) out = self.lin4(x) return out model = SimpleNN() wandb.watch(model) def loss_function(x1, x2, y1, y2): x_difference = torch.sum(torch.abs(x1 - x2), dim=[1,2]) print(x_difference) y_difference = torch.sum(torch.abs(y1 - y2), dim=[1]) print(y_difference) return torch.sum(torch.pow(x_difference - y_difference, 2)/x_difference) #x1, x2 = train_dataset[0:10][0], train_dataset[0:10][1] #y1, y2 = model(x1), model(x2) ###Output _____no_output_____ ###Markdown Training ###Code a = len(test_dataset)/len(train_dataset) batch_size = 64 train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True) test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=True) criterion = loss_function optimizer = optim.Adam(model.parameters()) wandb.log({"Batch_size": batch_size}) for e in range(80): # train loss = 0 for x1, x2 in train_loader: optimizer.zero_grad() y1, y2 = model(x1), model(x2) batch_loss = criterion(x1, x2, y1 ,y2) batch_loss.backward() optimizer.step() loss+=batch_loss #validation model.eval() val_loss = 0 for x1, x2 in test_loader: y1, y2 = model(x1), model(x2) val_loss += criterion(x1, x2, y1 ,y2) wandb.log({ "Training Loss": loss, "Validation Loss": a*val_loss, }) print(f"Epoch {e+1}, Training Loss: {a*loss}, Validation Loss: {val_loss}") ###Output _____no_output_____ ###Markdown Evaluate results ###Code example_1= test_dataset[0:1][0] example_2= test_dataset[3:4][0] example_1.shape, example_2.shape example_1_mapping, example_2_mapping = model(example_1), model(example_2) example_1_mapping, example_2_mapping loss_function(example_1, example_2, example_1_mapping, example_2_mapping) ###Output _____no_output_____
CVND_Exercises-master/1_1_Image_Representation/6_3. Average Brightness.ipynb
###Markdown Day and Night Image Classifier---The day/night image dataset consists of 200 RGB color images in two categories: day and night. There are equal numbers of each example: 100 day images and 100 night images.We'd like to build a classifier that can accurately label these images as day or night, and that relies on finding distinguishing features between the two types of images!*Note: All images come from the [AMOS dataset](http://cs.uky.edu/~jacobs/datasets/amos/) (Archive of Many Outdoor Scenes).* Import resourcesBefore you get started on the project code, import the libraries and resources that you'll need. ###Code import cv2 # computer vision library import helpers import numpy as np import matplotlib.pyplot as plt import matplotlib.image as mpimg %matplotlib inline ###Output _____no_output_____ ###Markdown Training and Testing DataThe 200 day/night images are separated into training and testing datasets. * 60% of these images are training images, for you to use as you create a classifier.* 40% are test images, which will be used to test the accuracy of your classifier.First, we set some variables to keep track of some where our images are stored: image_dir_training: the directory where our training image data is stored image_dir_test: the directory where our test image data is stored ###Code # Image data directories image_dir_training = "day_night_images/training/" image_dir_test = "day_night_images/test/" ###Output _____no_output_____ ###Markdown Load the datasetsThese first few lines of code will load the training day/night images and store all of them in a variable, `IMAGE_LIST`. This list contains the images and their associated label ("day" or "night"). For example, the first image-label pair in `IMAGE_LIST` can be accessed by index: ``` IMAGE_LIST[0][:]```. ###Code # Using the load_dataset function in helpers.py # Load training data IMAGE_LIST = helpers.load_dataset(image_dir_training) ###Output _____no_output_____ ###Markdown Construct a `STANDARDIZED_LIST` of input images and output labels.This function takes in a list of image-label pairs and outputs a **standardized** list of resized images and numerical labels. ###Code # Standardize all training images STANDARDIZED_LIST = helpers.standardize(IMAGE_LIST) ###Output _____no_output_____ ###Markdown Visualize the standardized dataDisplay a standardized image from STANDARDIZED_LIST. ###Code # Display a standardized image and its label # Select an image by index image_num = 0 selected_image = STANDARDIZED_LIST[image_num][0] selected_label = STANDARDIZED_LIST[image_num][1] # Display image and data about it plt.imshow(selected_image) print("Shape: "+str(selected_image.shape)) print("Label [1 = day, 0 = night]: " + str(selected_label)) ###Output Shape: (600, 1100, 3) Label [1 = day, 0 = night]: 1 ###Markdown Feature ExtractionCreate a feature that represents the brightness in an image. We'll be extracting the **average brightness** using HSV colorspace. Specifically, we'll use the V channel (a measure of brightness), add up the pixel values in the V channel, then divide that sum by the area of the image to get the average Value of the image. RGB to HSV conversionBelow, a test image is converted from RGB to HSV colorspace and each component is displayed in an image. ###Code # Convert and image to HSV colorspace # Visualize the individual color channels image_num = 0 test_im = STANDARDIZED_LIST[image_num][0] test_label = STANDARDIZED_LIST[image_num][1] # Convert to HSV hsv = cv2.cvtColor(test_im, cv2.COLOR_RGB2HSV) # Print image label print('Label: ' + str(test_label)) # HSV channels h = hsv[:,:,0] s = hsv[:,:,1] v = hsv[:,:,2] # Plot the original image and the three channels f, (ax1, ax2, ax3, ax4) = plt.subplots(1, 4, figsize=(20,10)) ax1.set_title('Standardized image') ax1.imshow(test_im) ax2.set_title('H channel') ax2.imshow(h, cmap='gray') ax3.set_title('S channel') ax3.imshow(s, cmap='gray') ax4.set_title('V channel') ax4.imshow(v, cmap='gray') ###Output _____no_output_____ ###Markdown --- Find the average brightness using the V channelThis function takes in a **standardized** RGB image and returns a feature (a single value) that represent the average level of brightness in the image. We'll use this value to classify the image as day or night. ###Code # Find the average Value or brightness of an image def avg_brightness(rgb_image): # Convert image to HSV hsv = cv2.cvtColor(rgb_image, cv2.COLOR_RGB2HSV) # Add up all the pixel values in the V channel sum_brightness = np.sum(hsv[:,:,2]) ## TODO: Calculate the average brightness using the area of the image # and the sum calculated above avg = 0 return avg # Testing average brightness levels # Look at a number of different day and night images and think about # what average brightness value separates the two types of images # As an example, a "night" image is loaded in and its avg brightness is displayed image_num = 190 test_im = STANDARDIZED_LIST[image_num][0] avg = avg_brightness(test_im) print('Avg brightness: ' + str(avg)) plt.imshow(test_im) ###Output _____no_output_____
gallery/general/plot_lineplot_with_legend.ipynb
###Markdown Multi-line temperature profile plot^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ ###Code import matplotlib.pyplot as plt import iris import iris.plot as iplt import iris.quickplot as qplt def main(): fname = iris.sample_data_path("air_temp.pp") # Load exactly one cube from the given file. temperature = iris.load_cube(fname) # We only want a small number of latitudes, so filter some out # using "extract". temperature = temperature.extract( iris.Constraint(latitude=lambda cell: 68 <= cell < 78) ) for cube in temperature.slices("longitude"): # Create a string label to identify this cube (i.e. latitude: value). cube_label = "latitude: %s" % cube.coord("latitude").points[0] # Plot the cube, and associate it with a label. qplt.plot(cube, label=cube_label) # Add the legend with 2 columns. plt.legend(ncol=2) # Put a grid on the plot. plt.grid(True) # Tell matplotlib not to extend the plot axes range to nicely # rounded numbers. plt.axis("tight") # Finally, show it. iplt.show() if __name__ == "__main__": main() ###Output _____no_output_____
Model Selection _ Boosting/Regression/Adjusted R2.ipynb
###Markdown Adjusted R2 ![image.png](data:image/png;base64,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) ![image.png](data:image/png;base64,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) ![image.png](data:image/png;base64,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) No formula for Adjusted R2??? Why?? ###Code from sklearn import metrics Adj R2 = 1 - (1-metrics.r2_score(y_test, y_pred))*(len(y)-1)/(len(y)-X.shape[1]-1) ###Output _____no_output_____
wrangling/.ipynb_checkpoints/Fifa19_notebook-checkpoint.ipynb
###Markdown Getting Club players ###Code def get_club_players(df_teams): df_club_list = df_teams[['Name','Position','OvervalueRatio','Overall','Potential','Wage']] sort_club_list = df_club_list.sort_values(by='OvervalueRatio', ascending=False) df_top_2_rated_players = sort_club_list.head(2) df_bottom_2_rated_players = sort_club_list.tail(2) return df_club_list, df_top_2_rated_players, df_bottom_2_rated_players df_attributes = data[['FieldPositionNum', 'Overall', 'Potential', 'Crossing', 'Finishing', 'HeadingAccuracy', 'ShortPassing', 'Volleys', 'Dribbling', 'Curve', 'FKAccuracy', 'LongPassing', 'BallControl', 'Acceleration', 'SprintSpeed', 'Agility', 'Reactions', 'Balance', 'ShotPower', 'Jumping', 'Stamina', 'Strength', 'LongShots', 'Aggression', 'Interceptions', 'Positioning', 'Vision', 'Penalties', 'Composure', 'Marking']] ###Output _____no_output_____ ###Markdown Recommender System ###Code # Create function that uses above 'suggested' variables to output x players to potentially obtain by trade # for df_attributes dataframe, see above cells def get_suggested_trades(df_teams): # player argument changed to 'club', after get_club_players refactored trades_p1 = [] # this will be the output object that club_suggested_changes receives/uses trades_p2 = [] players_wages = [] all_players, top_2, bottom_2 = get_club_players(df_teams) # df_club_list, df_top_2_players, df_bottom_2_players # looping throught 2 player names in 'top_2' for idx, player in enumerate(top_2.Name): # getting 'index' for player in 'df_teams' DF input_player_index = df_teams[df_teams['Name']==player].index.values[0] # getting the 'Overall', 'Potential', and 'Field Position Num' p_overall = df_teams.iloc[input_player_index]['Overall'] p_potential = df_teams.iloc[input_player_index]['Potential'] p_position = df_teams.iloc[input_player_index]['FieldPositionNum'] # getting 'Wage' for player in 'df_teams' DF # to be used later for 'Post-trade Leftover Wage' in returned DF input_player_wage = df_teams.iloc[input_player_index]['Wage'] players_wages.append(input_player_wage) # getting 'row' for same player in 'df_attributes' using index (No 'Name' col in 'df_attributes') player_attributes = df_attributes.iloc[input_player_index] # filtering attributes logic: filtered_attributes = df_attributes[(df_attributes['Overall'] > p_overall-10) & (df_attributes['Potential'] > p_potential-10) & (df_attributes['FieldPositionNum'] == p_position)] # use filter logic to suggest replacement players - top 5 suggested # gives DF of with all indexes and correlation ratio suggested_players = filtered_attributes.corrwith(player_attributes, axis=1) # Top 2 suggested players (most positively correlated) suggested_players = suggested_players.sort_values(ascending=False).head(6) cols = ['Name', 'Position', 'OvervalueRatio', 'Overall', 'Potential', 'Wage'] for i, corr in enumerate(suggested_players): if idx == 0: # player 1 - suggested trades trades_p1.append(data[data.index==suggested_players.index[i]][cols].values) else: # player 2 - suggested trades trades_p2.append(data[data.index==suggested_players.index[i]][cols].values) cols1 = ['Name', 'Position', 'OvervalueRatio', 'Overall', 'Potential', 'Wage'] # suggested trades DF for player 1 - dropping 1st row (most positively correlated = same as player 1) trades_p1_df = pd.DataFrame(np.row_stack(trades_p1), columns=cols1) trades_p1_df = trades_p1_df.drop(trades_p1_df.index[0]).reset_index(drop=True) # suggested trades DF for player 2 - dropping 1st row (most positively correlated = same as player 2) trades_p2_df = pd.DataFrame(np.row_stack(trades_p2), columns=cols1) trades_p2_df = trades_p2_df.drop(trades_p2_df.index[0]).reset_index(drop=True) #adding 'Post-trade Leftover Wage' column to each returned DF trades_p1_df['Post-tradeLeftoverWage'] = players_wages[0] - trades_p1_df['Wage'] trades_p2_df['Post-tradeLeftoverWage'] = players_wages[1] - trades_p2_df['Wage'] return top_2, bottom_2, trades_p1_df, trades_p2_df # See comment line inside of function just below def get_replacement_players(df_teams): '''Gets 2 lowest-rated players, and suggests four possible replacements.''' replacements_p1 = [] # this will be the output object that club_suggested_changes receives/uses replacements_p2 = [] players_wages = [] all_players, top_2, bottom_2 = get_club_players(df_teams) # df_club_list, df_top_2_players, df_bottom_2_players # looping throught 2 player names in 'top_2' for idx, player in enumerate(bottom_2.Name): # getting 'index' for player in 'df_teams' DF input_player_index = df_teams[df_teams['Name']==player].index.values[0] # getting the 'Overall', 'Potential', and 'Field Position Num' p_overall = df_teams.iloc[input_player_index]['Overall'] p_potential = df_teams.iloc[input_player_index]['Potential'] p_position = df_teams.iloc[input_player_index]['FieldPositionNum'] # getting 'Wage' for player in 'df_teams' DF # to be used later for 'Post-trade Leftover Wage' in returned DF input_player_wage = df_teams.iloc[input_player_index]['Wage'] players_wages.append(input_player_wage) # getting 'row' for same player in 'df_attributes' using index (No 'Name' col in 'df_attributes') player_attributes = df_attributes.iloc[input_player_index] # filtering weak attributes logic: filtered_weak_attributes = df_attributes[(df_attributes['Overall'] < 90) & (df_attributes['Potential'] > p_potential) & (df_attributes['Potential'] < 89) & (df_attributes['FieldPositionNum'] == p_position)] suggested_players = filtered_weak_attributes.corrwith(player_attributes, axis=1) suggested_players = suggested_players.sort_values(ascending=False).head(3) cols = ['Name', 'Position', 'OvervalueRatio', 'Overall', 'Potential', 'Wage'] for i, corr in enumerate(suggested_players): if idx == 0: # player 1 - suggested replacements replacements_p1.append(data[data.index==suggested_players.index[i]][cols].values) else: # player 2 - suggested replacements replacements_p2.append(data[data.index==suggested_players.index[i]][cols].values) cols1 = ['Name', 'Position', 'OvervalueRatio', 'OverallRating', 'PotentialRating', 'Wage'] # suggested replacements DF for player 1 - dropping 1st row (most positively correlated = same as player 1) replacements_p1_df = pd.DataFrame(np.row_stack(replacements_p1), columns=cols1) replacements_p1_df = replacements_p1_df.drop(replacements_p1_df.index[0]).reset_index(drop=True) # suggested replacements DF for player 2 - dropping 1st row (most positively correlated = same as player 2) replacements_p2_df = pd.DataFrame(np.row_stack(replacements_p2), columns=cols1) replacements_p2_df = replacements_p2_df.drop(replacements_p2_df.index[0]).reset_index(drop=True) #adding 'Post-trade Leftover Wage' column to each returned DF replacements_p1_df['Post-tradeLeftoverWage'] = players_wages[0] - replacements_p1_df['Wage'] replacements_p2_df['Post-tradeLeftoverWage'] = players_wages[1] - replacements_p2_df['Wage'] #print(replacements_p1_df, '\n') #print(replacements_p2_df) return replacements_p1_df, replacements_p2_df # All tables allplayers, top2overvalued, bottom2weak = get_club_players(df_teams) top_2, bottom_2, trades_p1_df, trades_p2_df = get_suggested_trades(df_teams) replacements_p1_df, replacements_p2_df = get_replacement_players(df_teams) # turning all tables into JSON top_2 = top_2.to_json(orient='records') bottom_2 = bottom_2.to_json(orient='records') trades_p1_df = trades_p1_df.to_json(orient='records') trades_p2_df = trades_p2_df.to_json(orient='records') replacements_p1_df = replacements_p1_df.to_json(orient='records') replacements_p2_df = replacements_p2_df.to_json(orient='records') def all_dfs_json(): json_dict = dict({'top2overvalued': top_2, 'suggestedtrades' : [trades_p1_df, trades_p2_df], 'bottom2weak': bottom_2, 'suggestedreplacements': [replacements_p1_df, replacements_p2_df]}) return json.dumps(json_dict) ###Output _____no_output_____
old/python_ver0.1/demo.ipynb
###Markdown Demo of NMF-SO and NMF-ARD-SO [1] Motoki Shiga, Kazuyoshi Tatsumi, Shunsuke Muto, Koji Tsuda, Yuta Yamamoto, Toshiyuki Mori, Takayoshi Tanji, "Sparse Modeling of EELS and EDX Spectral Imaging Data by Nonnegative Matrix Factorization", Ultramicroscopy, Vol.170, p.43-59, 2016. ###Code %matplotlib inline import numpy as np import scipy.io as sio from libnmf import NMF, NMF_SO, NMF_ARD_SO ###Output _____no_output_____ ###Markdown Generate a synthetic dataset with noise ###Code #load theoretical data of Mn3O4 without noise mat_dict = sio.loadmat('mn3o4_f2.mat') ximage = mat_dict['datar'] # focusing channel n_ch = np.arange(37-1,116); ximage = ximage[:,:,n_ch]; # # of pixels along x and y axis, # of EELS channels xdim,ydim,Nch = ximage.shape # generating pahtom data by adding gaussian noise X = np.reshape(ximage, (xdim*ydim, Nch)) scale_spect = np.max(X) s2_noise = 0.1 #noise variance X = X + np.random.randn(xdim*ydim, Nch) * s2_noise * scale_spect; X = (X + np.abs(X))/2; scale_X = np.mean(X) X = X / scale_X ###Output _____no_output_____ ###Markdown NMF-SO ###Code # define and training model nmf_so = NMF_SO(n_components=2, wo=0.05, reps=3, max_itr=100) nmf_so.fit(X, num_xy=(xdim,ydim), channel_vals=n_ch) nmf_so.imshow_component(figsize=(6, 3)) # for 2D spactrum (Spectrum Imaging) dataset # nmf.plot_component() # for 1D spactrum dataset nmf_so.plot_spectra(figsize=(6,3)) # plot decomposed spectra nmf_so.plot_object_fun(figsize=(6,3)) # plot learnig curve (object function) ###Output _____no_output_____ ###Markdown NMF-ARD-SO ###Code # define and training model nmf_ard = NMF_ARD_SO(n_components=9, wo=0.05, reps=3, max_itr=100) nmf_ard.fit(X, num_xy=(xdim,ydim), channel_vals=n_ch) nmf_ard.plot_ard() # plot learning curve with component intensities nmf_ard.imshow_component(figsize=(6, 3)) # for 2D spactrum (Spectrum Imaging) dataset nmf_ard.plot_spectra(figsize=(6,3)) # plot decomposed spectra nmf_ard.plot_object_fun(figsize=(6,3)) # #plot learnig curve (object function) ###Output _____no_output_____
section4/4-2.ipynb
###Markdown 4-2. 離散値データと分散表現 シェイクスピア・データセット> この節と次節は[tensorflowのチュートリアル](https://www.tensorflow.org/tutorials/text/text_generation?hl=ja)では同じデータを使って、以下の説明と似たことが説明されていますが、このように台詞に分けずに学習を行っています。そのようにするとGPUの処理などのコーディングがこのノートに比べ楽になりますが、このノートでは確率論的な説明を優先するためこのようなデータ形式を採用します。tensorflowを使って、シェイクスピアのデータをダウンロードできます: ###Code ''' This code is derived from https://www.tensorflow.org/tutorials/text/text_generation which is licensed under Apache 2.0 License. ''' path = tf.keras.utils.get_file('shakespeare.txt', 'https://storage.googleapis.com/download.tensorflow.org/data/shakespeare.txt') ###Output _____no_output_____ ###Markdown `get_corpus`という関数を定義しておきました。これを使うと、以下のようにデータを「戯曲の各登場人物+台詞」のリスト=`corpus`に分解できます: ###Code corpus = get_corpus(path) N = len(corpus) print("台詞数 N=%d\n"%N) for _ in range(5): i = np.random.randint(len(corpus)) print(corpus[i]) ###Output 台詞数 N=3461 Boatswain: What, must our mouths be cold? ************************************** Nurse: She's dead, deceased, she's dead; alack the day! ************************ BRUTUS: He's poor in no one fault, but stored with all. ************************ KING EDWARD IV: Huntsman, what say'st thou? wilt thou go along? **************** Clown: Fear not thou, man, thou shalt lose nothing here. *********************** ###Markdown それぞれ- `corpus = [台詞1, 台詞2, ... ]`- `台詞 = "xxxxx***"` (*は終わりを表すトークンで、文字数を 80 に揃えるために挿入、tensorflowの関数でも処理できますが、ここではpythonで書きました。)のような構成になっています。`N`はこのデータに含まれる全台詞の数です。今回はこれをデータセットとします。 この中で出現する文字を重複を除き数えると ###Code ''' This code is derived from https://www.tensorflow.org/tutorials/text/text_generation which is licensed under Apache 2.0 License. ''' chars = sorted(set(str("".join(corpus)))) print(chars); print(len(chars), "種類") ###Output ['\n', ' ', '!', '&', "'", '*', ',', '-', '.', '3', ':', ';', '?', 'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'I', 'J', 'K', 'L', 'M', 'N', 'O', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z', 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z'] 65 種類 ###Markdown のように13種類の(!, ?, ;や半角スペースや改行\nなどの)特殊文字+アルファベット26種類 $\times$ (大文字, 小文字)=52種類の計65種類からなる文字の集まりから構成されているのがわかります。これを$$\mathcal{S}_{chars} = \{\text{'$\backslash$n', ' ', '!', ..., 'A', 'B', 'C', ..., 'a', 'b', 'c', ...} \}$$と呼びましょう。各データはさっきも表示しましたが、改行コードなども明示的に表示させると: ###Code i = np.random.randint(len(corpus)) print(list(corpus[i])) ###Output ['F', 'i', 'r', 's', 't', ' ', 'W', 'a', 't', 'c', 'h', 'm', 'a', 'n', ':', '\n', "'", 'T', 'i', 's', ' ', 't', 'h', 'e', ' ', 'L', 'o', 'r', 'd', ' ', 'H', 'a', 's', 't', 'i', 'n', 'g', 's', ',', ' ', 't', 'h', 'e', ' ', 'k', 'i', 'n', 'g', "'", 's', ' ', 'c', 'h', 'i', 'e', 'f', 'e', 's', 't', ' ', 'f', 'r', 'i', 'e', 'n', 'd', '.', '\n', '*', '*', '*', '*', '*', '*', '*', '*', '*', '*', '*', '*'] ###Markdown のように、一文字ずつのリスト=各データ、となっています。 離散値データの表現 インデックスによる表現(one-hot表現)コンピュータ上で文字がどのように処理されているかですが、例えば [ASCII](https://ja.wikipedia.org/wiki/ASCII) と呼ばれる規格では、整数値にそれぞれの文字を割り振っています: ###Code ord('A'), ord('B'), ord('C'), ord('D') ###Output _____no_output_____ ###Markdown 今回はASCIIと似た戦法で、 $\mathcal{S}_{chars}$ の頭から順に 0,1,2,...,64 と呼ぶことにします: ###Code ''' This code is derived from https://www.tensorflow.org/tutorials/text/text_generation which is licensed under Apache 2.0 License. ''' char2idx = {u:i for i, u in enumerate(chars)} char2idx['A'], char2idx['B'], char2idx['C'], char2idx['D'] ###Output _____no_output_____ ###Markdown この写像の逆変換も後から使うので作っておきます。単にインデックス入力に対応した配列として $\mathcal{S}_{chars}$ を実装すればよいだけです: ###Code ''' This code is derived from https://www.tensorflow.org/tutorials/text/text_generation which is licensed under Apache 2.0 License. ''' idx2char = np.array(chars) idx2char[13], idx2char[14], idx2char[15], idx2char[16] ###Output _____no_output_____ ###Markdown $$\mathcal{S}_{chars} = \{\underbrace{\text{'$\backslash$n'}}_0, \underbrace{\text{' '}}_1, \underbrace{\text{'!'}}_2, \dots, \underbrace{\text{'A'}}_{13}, \underbrace{\text{'B'}}_{14}, \underbrace{\text{'C'}}_{15}, \dots, \underbrace{\text{'a'}}_{39}, \underbrace{\text{'b'}}_{40}, \underbrace{\text{'c'}}_{41}, \dots \}$$以後、$\mathcal{S}_{chars}$ の要素は整数とし、 $n$ で表すことにします。すると各セリフは **数値 $n$ の集まり** $${\bf n} = [n_0, n_1, \dots, n_{T-1}]$$で表すことができます。 ###Code i = np.random.randint(len(corpus)) print(corpus[i]) print([char2idx[c] for c in corpus[i]]) ###Output BRAKENBURY: Why looks your grace so heavily today? ***************************** [14, 30, 13, 23, 17, 26, 14, 33, 30, 37, 10, 0, 35, 46, 63, 1, 50, 53, 53, 49, 57, 1, 63, 53, 59, 56, 1, 45, 56, 39, 41, 43, 1, 57, 53, 1, 46, 43, 39, 60, 47, 50, 63, 1, 58, 53, 42, 39, 63, 12, 0, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5] ###Markdown コーパスを数値化した`corpus_num`を作っておきます: ###Code corpus_num = np.array([[char2idx[c] for c in n] for n in corpus]) print(corpus[0], '\n', corpus_num[0]) ###Output First Citizen: Before we proceed any further, hear me speak. ******************* [18 47 56 57 58 1 15 47 58 47 64 43 52 10 0 14 43 44 53 56 43 1 61 43 1 54 56 53 41 43 43 42 1 39 52 63 1 44 59 56 58 46 43 56 6 1 46 43 39 56 1 51 43 1 57 54 43 39 49 8 0 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5] ###Markdown 尚、同値な表現として、**one-hot表現** というものがあります。これは$$n \leftrightarrow \vec{n} = [0, \dots, 0, \overbrace{1}^{n-th}, 0, \dots, 0]$$のように、$n$ 番目の成分のみ1であるようなベクトル表現のことを指します。 2進数表現整数をベクトルとして表現したい場合、one-hot表現だけしか無いわけではありません。例えば 10進数 $n$ を 2進数表示して、桁ごとに区切る方法も考えられます。(10進数のまま桁ごとに区切ってもベクトル化できますが、あまり使われない気がします。)2進数表現はまた後ほど使うので、そこで詳しく説明します。 分散表現分散表現は、もっと幾何的にベクトル化する方法です。深層学習では分散表現は訓練対象である場合が多く、離散値から何らかの次元のベクトルへの埋め込み方$$n_t \to {\bf e}[n_t]$$が訓練対象になります。one-hot表現を使うと単なる線形変換で表現できます: ${\bf e}[n] = W \vec{n}= (W_{i n})_{i=1, 2, \dots}$ が、専用のコマンドがあります: ###Code emb = tf.keras.layers.Embedding(len(chars), 2) ###Output _____no_output_____ ###Markdown これで適当に初期化された2次元の分散表現を得ることができます: ###Code plt.figure(figsize=(10,10)) for nt in range(len(chars)): xn = emb(nt).numpy() plt.scatter(xn[0], xn[1]) plt.annotate(r'${\bf e}[%d]$=(%s)'%(nt,idx2char[nt]), (xn[0], xn[1]), size=12) plt.show() ###Output _____no_output_____ ###Markdown ここでは説明のために2次元に取りましたが、もっと高次元に埋め込む場合が多いようです。以下でニューラルネットワークを構成する際、一層目にこの分散表現を使います。 無限の猿定理[無限の猿定理](https://ja.wikipedia.org/wiki/無限の猿定理) という話があります。これは **猿が(ランダムに)タイプすることでシェイクスピアの作品を生成する** という「ゼロではない可能性」が無限回試行すれば起こり得る、という類の無限に関する話です。今回は規模を縮小して- 猿がシェイクスピアのような台詞 ${\bf n}$ を書く確率を考えてみます。これまでの理論部分で、データ生成確率 $p$ とモデル $q$ を考えていましたが、ここでも同じことで ${\bf n}=[n_0, n_1, \dots, n_{T-1}]$ を生成する確率をそれぞれ- シェイクスピア:$p({\bf n})$- 猿:$q_{monkey}({\bf n})$としましょう。更に、猿はとりあえず文脈を把握しておらず、各時刻でどの文字をタイプするかは独立だとします:時刻 $t$ で1文字 $n_t$ をタイプする確率 $q_{monkey}(n_t)$ があるとして$$q_{monkey}({\bf n}) = \prod_{t=0}^{T-1} q_{monkey}(n_t)$$ということにします(猿には少々失礼な仮定かもしれませんが)。例えば1文字打つ確率がすべての文字を当確率(つまりランダム)だとすると$$q_{monkey}(n_t) = \frac{1}{|\mathcal{S}_{chars}|} = \frac{1}{65}$$です。すると、$q_{monkey}({\bf n}) = \Big( \frac{1}{65}\Big)^{T}$です。確かに厳密に言えばゼロではないですが、これはほぼ無いと言えるでしょう。一応この **猿=完全ランダムモデル** で生成してみると: ###Code T = 100 n = np.random.choice(chars, T) print("".join([nt for nt in n])) ###Output l pq lYsLn:RmvlxPC!TFW,E;XI!' faPkOFk YUjHquVPNg.Elvwwlorq&VDvkZJN&q;lq'WVeiC&*XdBdREd&IOL.YkBvMuWS ###Markdown のようになり、シェイクスピアの戯曲を書くにはかなり遠そうなことがわかります。 ニューラルネットワークでまねてみるここでは`corpus`または`corpus_num`を、一つの台詞 ${\bf n}$ をシェイクスピアが書く確率を $p({\bf n})$ として、その確率分布から独立に $N$ 個のサンプリング$$\text{corpus}=\mathcal{D}_N = \{ {\bf n}_1, {\bf n}_2, ..., {\bf n}_N\} \sim p({\bf n})^N$$だと思っているわけですが、$p({\bf n})$を真似るモデル $q({\bf n})$ を考えるには、まず $p$ がどのような構造を持つべきか考えると良さそうです。そもそも、台詞は$${\bf n} = [n_0, n_1, \dots, n_{T-1}]$$の順に意味があるため、単に1文字出る確率 $q(n_t)$ を考えるのでは順序を捉えられそうにないのでだめでしょう。とりあえず台詞の生成は**マルコフ性**を持つ、と考えてみます。つまり$$q_{Markov}({\bf n}) = q(n_0)\prod_{t=0}^{T-2} q_{Markov}(n_{t+1}\mid n_t) $$だと考えてみます。最初の確率 $q(n_0)$ としては例えば、台詞は役名から始まるので、大文字から始まるでしょうから、大文字26種類からランダムに取るとしましょう。ここでは$q_{Markov}(n_t\mid n_{t-1})$ に注目してください。これは条件付き確率で、**丁度ニューラルネットワーク** を導入できる形式になっています。そこで$$q_{Markov}(n_{t+1}\mid n_t) = q_{\theta}(n_{t+1}\mid n_t) $$として、機械学習にかけてみましょう。ここでは分散表現を用いた以下のような構造を持つネットワーク![alt](markov.jpg)$$\left\{ \begin{array}{ll}{\bf e}_t={\color{red}{l_{emb}}}(n_t) & \color{red}{\text{Embedding}}\\{\bf h}_t = \tanh({\color{red}{l_h}}({\bf e}_t))= \tanh({\color{red}{W_h}}{\bf e}_t + {\color{red}{b_h}}) & \color{red}{\text{Dense}}\\{\bf z}_t = {\color{red}{l_z}}({\bf h}_t)={\color{red}{W_z}} {\bf h}_t + {\color{red}{b_z}} & \color{red}{\text{Dense}}\\q_{\color{red}{\theta}}(n\mid n_t) = [{\bf \color{blue}{\sigma}}({\bf z}_t)]_{n\text{-th component}} & \color{blue}{\text{Softmax}}\end{array} \right.$$を使ってみます: ###Code class Markov(Generator): # モデル設計 def __init__(self, emb_dim, hidden_dim): super(Markov, self).__init__() self.l_emb= tf.keras.layers.Embedding(len(chars), emb_dim) self.l_h = tf.keras.layers.Dense(units=hidden_dim) self.l_z = tf.keras.layers.Dense(units=len(chars)) def call(self, nt): e = self.l_emb(nt) h = tf.keras.activations.tanh(self.l_h(e)) z = self.l_z(h) return z ###Output _____no_output_____ ###Markdown これを使えば、自身のsoftmax出力=確率からのサンプリングを次の文字と思って、次々に文字が生成できます:![alt](markov2.jpg)モデルを実際に適当な初期化の元、オブジェクトとして作って文字列を生成してみると以下のような感じです: ###Code model= Markov(1,1) model.sample_from("S", num_string=100) # * が生成されたら止まります。 ###Output SOHc* ###Markdown このモデルを訓練することを考えてみましょう。 汎化誤差、経験誤差これまで通り汎化誤差は $D_{KL}(p\|q)$ として、経験誤差は$$L(\theta; \mathcal{D}_N) = \frac{1}{N} \sum_{i=1}^N (- \log q_\theta({\bf n}_i))$$ですので、1台詞サンプル ${\bf n}_i=[n_0, n_1, \dots, n_{T-1}]$ あたりの負の対数尤度を考えればそれの $N$ 平均が経験誤差となります。1サンプルあたりの負の対数尤度 $L(\theta; \{{\bf n}\})$ は$$ - \log q_\theta({\bf n}) = - \log q(n_0) \prod_{t=0}^{T-2}q_\theta(n_{t+1} \mid n_t) = {\color{blue}{L(\theta; \{{\bf n}\})}}$$です。$q(x_0)$ を後で固定した確率分布に取るので、学習パラメータ $\theta$ に依存しないため、結局 $L(\theta; \{{\bf n}\})$ を下げることが、データ ${\bf n}$ に含まれる文字の順でサンプルが生成される確率を上げるという意味になっています。$-\log$ を先に取ると積は和になりますから、$q_\theta$ が最後はsoftmax関数でモデル化していたことを思い出すと、$$ {\color{blue}{L(\theta;\{ {\bf n} \})}} = - \sum_{t=0}^{T-2} \log q_\theta(n_{t+1} \mid n_t)= - \sum_{t=0}^{T-2} \Big({\bf \log} {\bf \sigma}_{softmax}({\bf z}_t )\Big)_{{n_t}\text{-th comp}} =\sum_{t=0}^{T-2} L_{softmax}\big(n_{t+1}, {\bf \sigma}_{softmax}({\bf z}_t ) \big) $$と、各時刻毎の「次の文字とその予測の間のsoftmaxクロスエントロピーを時間で足し上げたもの」が誤差関数であることがわかります。すると、$n_{T-1}$ だけ除いた文字列 ${\bf x}$ を入力信号、 $n_0$だけ除いた文字列 ${\bf y}$ を教師信号と思って、あたかも教師あり学習かのように書くことができます:![alt](markov4.jpg) データから誤差関数の実装${\bf x}$:入力データ($n_{T-1}$ だけ除いた文字列)|${\bf y}$:教師データ($n_{0}$ だけ除いた文字列):---:|:---:$[n_0, n_1, \dots, n_{T-2}]$|$[n_1, n_2, \dots, n_{T-1}]$と見なして教師あり学習できることが分かったので、あらかじめデータを$$\mathcal{D}_N = \Big\{ {\bf n}^{(1)} , \quad {\bf n}^{(2)} , \quad\dots \Big\}$$と考えるのではなく、これと同型な$$\mathcal{D}_N = \Big\{ (\bf x^{(1)}, {\bf y}^{(1)}), \quad (\bf x^{(2)}, {\bf y}^{(2)}), \quad \dots \Big\}$$だと見なしましょう。これは以下のようなコマンドで作ることができます。 ###Code ''' This code is derived from functions prepareing `dataset` object in https://www.tensorflow.org/tutorials/text/text_generation which is licensed under Apache 2.0 License. ''' D = tf.data.Dataset.from_tensor_slices(corpus_num) f = lambda n: (n[:-1], n[1:]) D = D.map(f) for (x, y) in D.take(1): print("x=", x) print("y=", y) ###Output x= tf.Tensor( [18 47 56 57 58 1 15 47 58 47 64 43 52 10 0 14 43 44 53 56 43 1 61 43 1 54 56 53 41 43 43 42 1 39 52 63 1 44 59 56 58 46 43 56 6 1 46 43 39 56 1 51 43 1 57 54 43 39 49 8 0 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5], shape=(79,), dtype=int64) y= tf.Tensor( [47 56 57 58 1 15 47 58 47 64 43 52 10 0 14 43 44 53 56 43 1 61 43 1 54 56 53 41 43 43 42 1 39 52 63 1 44 59 56 58 46 43 56 6 1 46 43 39 56 1 51 43 1 57 54 43 39 49 8 0 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5], shape=(79,), dtype=int64) ###Markdown 誤差関数は上で説明したように、各 $t$ の誤差の和$$L({\bf y}, model({\bf x})) = \sum_{t=0}^{T-2} L_{softmax}(n_{t+1}, \underbrace{model(n_t)}_{{\bf \sigma}({\bf z}_t)})$$なのですが、今のデータ形式では ###Code model = Markov(emb_dim=256, hidden_dim=1024) for (x, y) in D.take(1): print(tf.keras.losses.sparse_categorical_crossentropy(y, model(x), from_logits=True)) # z=model(x) は logit ###Output tf.Tensor( [4.194728 4.1714563 4.168303 4.1747737 4.1414847 4.1660385 4.1127143 4.1524386 4.1641912 4.14454 4.1425385 4.1772213 4.172122 4.209669 4.2049527 4.160085 4.2067733 4.1902823 4.2153435 4.19617 4.136768 4.1741676 4.1675963 4.136768 4.18212 4.148957 4.144289 4.1870117 4.1680865 4.211338 4.1836004 4.149681 4.1699243 4.2101874 4.1624427 4.1753488 4.168456 4.1514997 4.1580896 4.1771946 4.147566 4.154824 4.12127 4.1632433 4.1716886 4.191299 4.154824 4.21504 4.113487 4.135355 4.202867 4.216572 4.136768 4.173488 4.1893177 4.1805544 4.21504 4.2035832 4.1819816 4.1828103 4.1720543 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353 4.1366353], shape=(79,), dtype=float32) ###Markdown のように、各 $t$ での $L_{softmax}$ の値を並べてベクトル化してしまいます。これの和を取る操作は、`tf.reduce_mean`を使うと良いです。(これは和ではなく平均値を取る操作ですが、同じことです): ###Code def loss_sum(y, z): return tf.reduce_mean(tf.keras.losses.sparse_categorical_crossentropy(y, z, from_logits=True)) # z=model(x) は logit model = Markov(emb_dim=256, hidden_dim=1024) for (x, y) in D.take(1): print(loss_sum(y, model(x))) ###Output tf.Tensor(4.1836224, shape=(), dtype=float32) ###Markdown 学習させる今まで同様SGDさせてみます。ミニバッチ `(X, Y)` が与えられたときのモデルの学習ステップを ###Code @tf.function def update(X, Y, model, optimizer): # 学習ステップ with tf.GradientTape() as tape: Z = model(X) loss_value = loss_sum(Y, Z) grads = tape.gradient(loss_value, model.trainable_variables) optimizer.apply_gradients(zip(grads, model.trainable_variables)) return loss_value ###Output _____no_output_____ ###Markdown として、今までと同じようにSGDでパラメータ更新してみます: ###Code %%time model = Markov(emb_dim=256, hidden_dim=1024) optimizer=tf.keras.optimizers.Adam() batch_size = 32 epoch_size = 15 loss_averages = [] for epoch in range(epoch_size): batch = D.shuffle(5000).batch(batch_size, drop_remainder=True) loss_values = [] for (X,Y) in batch: loss_value = update(X, Y, model, optimizer) loss_values.append(loss_value) loss_averages.append(np.average(loss_values)) ###Output CPU times: user 8.62 s, sys: 1.82 s, total: 10.4 s Wall time: 5.78 s ###Markdown 更新後のモデルで台詞生成を試みてみます。ここで $q(n_0)$ の設定に対応するのが、はじめの文字を指定することです: ###Code for _ in range(5): model.sample_from("A", num_string=100) ###Output AS: Whe bat me. Whr, yeawayo ds t'dareloree t nothet mes t d. What t JUMARONurcotho lliear: I r stowi ADY: * Ayod, w, w mith! Se,-is bery dsthisepl---- NGowsesangr? * AMERUMathel m ll; d aknIOLUMI rr t ncy om! * APONGEE izepru b? *
transfer_learning/(model)evaluation.ipynb
###Markdown 1. 모델 성능 평가 ###Code import glob import numpy as np import matplotlib.pyplot as plt from tensorflow.keras.preprocessing.image import load_img, img_to_array, array_to_img from tensorflow.keras.models import load_model import model_evaluation_utils as meu img_aug_cnn = load_model('cats_dogs_cnn_img_aug.h5') tl_img_aug_finetune_cnn = load_model('10-06-me.h5') ## 이미지 기본 모양 IMG_DIM = (150, 150) input_shape = (150, 150, 3) num2class_label_transformer = lambda l: ['cat' if x == 0 else 'dog' for x in l] class2num_label_transformer = lambda l: [0 if x == 'cat' else 1 for x in l] ###Output _____no_output_____ ###Markdown 샘플 테스트 이미지로 모델 예측 ###Code #sample_img_path = 'my_cat.jpg' #sample_img_path = "C:\\Users\\user\\Documents\\한국선급\\CNG_P1\\2_104449_11_11.png" #sample_img_path = 'dog_my.jpg' sample_img_path = 'tiger.jpg' sample_img = load_img(sample_img_path, target_size=IMG_DIM) sample_img_tensor = img_to_array(sample_img) sample_img_tensor = np.expand_dims(sample_img_tensor, axis=0) sample_img_tensor /= 255. print(sample_img_tensor.shape) plt.imshow(sample_img_tensor[0]) cnn_img_aug_prediction = num2class_label_transformer(img_aug_cnn.predict_classes(sample_img_tensor, verbose=0)) tlearn_cnn_finetune_img_aug_prediction = num2class_label_transformer(tl_img_aug_finetune_cnn.predict_classes(sample_img_tensor, verbose=0)) print('Predictions for our sample image:\n', '\nCNN with Img Augmentation:', cnn_img_aug_prediction, '\nPre-trained CNN with Fine-tuning & Img Augmentation (Transfer Learning):', tlearn_cnn_finetune_img_aug_prediction) img_aug_cnn.predict_proba(sample_img_tensor, verbose=0) tl_img_aug_finetune_cnn.predict_proba(sample_img_tensor, verbose=0) ###Output _____no_output_____
scripts/nrw.ipynb
###Markdown Covid-19 Sim provides two populations:"current" : The population is based on the current population (2019) "household" : The population is bases on a subsample in 2010 but with household numbers and additional persons per household ###Code age, agegroup, gender, contacts, drate, hnr, persons = cl.makepop("household",17900000) # Einlesen der realisierten Daten nrw = pd.read_excel("./nrw_dat.xlsx") nrw["Datum"] = nrw["Datum"].dt.date nrw["Meldedatum"] = nrw["Meldedatum"].dt.date ###Output _____no_output_____ ###Markdown Szenarien mit Community Attack Basis Szenario NRW mit Community AttackDie gesamte Repdouktionsetzt sich zusammen aus Infektionen innerhalb der Haushalte und Kontakt alle Haushaltsmitglieder mit der Aussenwelt. r_change ist hier jeweils der Kontaktwert mit der AussenweltDie Kontakte sind proportional zur den Kontaktraten der Altersgruppen ohne Beschränkung. ###Code day0date = datetime.date(2020, 3, 8) r_change = {} r_change['2020-01-01'] = 3 * contacts/np.mean(contacts) contacts_new = np.where(age < 20, contacts, contacts) r_change['2020-03-08'] = 1.0 * contacts/np.mean(contacts) r_change['2020-03-16'] = 0.3 * contacts/np.mean(contacts) r_change['2020-03-23'] = 0.2 * contacts/np.mean(contacts) com_attack_rate = {} com_attack_rate["2020-01-1"] = 0.5 com_attack_rate["2020-05-4"] = 0.5 state, statesum, infections, day0, rnow, args, gr = cl.sim( age, drate, nday=180, prob_icu=0.009, day0cumrep=450, mean_days_to_icu=16, mean_duration_icu=14, mean_time_to_death=21, mean_serial=7.5, std_serial=3.0, immunt0=0.0, ifr=0.003, long_term_death=False, hnr=hnr, com_attack_rate=com_attack_rate, r_change=r_change, simname="NRW Basis", datadir=".", realized=nrw, rep_delay=13, alpha=0.125, day0date=day0date) cl.plotoverview(gr, args) persgroup = np.where(persons>5,">5",persons) cl.groupresults({"Personen":persgroup},state) ###Output _____no_output_____ ###Markdown Der Anteil der Infizierten steigt deutlich mit der Haushaltsgröße. Dieses Bild sollte qualitativ recht realistisch sein. Es bestehen deutliche Hinweise, dass die Community Attack Rate 50% oder höher ist. ###Code aux = cl.groupresults({"Alter":agegroup},state) # Die CFR aus der IFR über die Dunkelziffer berechnen aux["CFR"] = aux.IFR / args["alpha"] display(aux) ## Öffnung von Kitas day0date = datetime.date(2020, 3, 8) r_change = {} r_change['2020-01-01'] = 3 * contacts/np.mean(contacts) contacts_new = np.where(age < 20, contacts, contacts) r_change['2020-03-08'] = 1.0 * contacts/np.mean(contacts) r_change['2020-03-16'] = 0.3 * contacts/np.mean(contacts) r_change['2020-03-23'] = 0.2 * contacts/np.mean(contacts) r_change['2020-04-20'] = 0.5 * contacts/np.mean(contacts) r_change['2020-05-04'] = np.where(age<6,5*r_change['2020-04-20'],r_change['2020-04-20']) com_attack_rate = {} com_attack_rate["2020-01-1"] = 0.5 com_attack_rate["2020-05-4"] = 0.5 state, statesum, infections, day0, rnow, args, gr = cl.sim( age, drate, nday=180, prob_icu=0.009, day0cumrep=450, mean_days_to_icu=16, mean_duration_icu=14, mean_time_to_death=21, mean_serial=7.5, std_serial=3.0, immunt0=0.0, ifr=0.003, long_term_death=False, hnr=hnr, com_attack_rate=com_attack_rate, r_change=r_change, simname="NRW Basis", datadir=".", realized=nrw, rep_delay=13, alpha=0.125, day0date=day0date) cl.plotoverview(gr, args) ###Output _____no_output_____ ###Markdown Der Anteil der Infizierten ist hier in den jungen Altersgruppen höher. Für die Meldefälle könnte man eine Annahme treffen, dass die Wahrscheinlichkeit einer "Meldung" ebenfalls proportional zur statistischen Sterbewahrscheinlichkeit ist. Die CFR ist hier einfach als "Dunkeziffer" * IFR berechnet. Der Alterstrend sieht plausibel aus. Lockerung in allen Bereichen ausser bei den über 60 Jährigen ###Code day0date = datetime.date(2020, 3, 8) r_change = {} r_change['2020-01-01'] = 3 * contacts/np.mean(contacts) contacts_new = np.where(age < 20, contacts, contacts) r_change['2020-03-08'] = 1.0 * contacts/np.mean(contacts) r_change['2020-03-16'] = 0.3 * contacts/np.mean(contacts) r_change['2020-03-23'] = 0.2 * contacts/np.mean(contacts) r_change['2020-05-04'] = 0.2 * np.where(age < 60, 3*contacts, contacts)/np.mean(contacts) com_attack_rate = {} com_attack_rate["2020-01-1"] = 0.5 com_attack_rate["2020-05-4"] = 0.5 state, statesum, infections, day0, rnow, args, gr = cl.sim( age, drate, nday=180, prob_icu=0.009, day0cumrep=450, mean_days_to_icu=16, mean_duration_icu=14, mean_time_to_death=21, mean_serial=7.5, std_serial=3.0, immunt0=0.0, ifr=0.003, long_term_death=False, hnr=hnr, com_attack_rate=com_attack_rate, r_change=r_change, simname="Lockerung U60", datadir=".", realized=nrw, rep_delay=13, alpha=0.125, day0date=day0date) cl.plotoverview(gr, args) ###Output _____no_output_____ ###Markdown Das effektive R steigt zwar leicht über 1 aber die Intesivbelastung ist weiter fallend, da die Maßnahmen für die ältere Bevölkerung weiter wirken. Es könnte jedoch sein, dass diese Annahme zu opimistisch ist. ###Code # Lockerung in allen Bereichen day0date = datetime.date(2020, 3, 8) r_change = {} r_change['2020-01-01'] = 3 * contacts/np.mean(contacts) contacts_new = np.where(age < 20, contacts, contacts) r_change['2020-03-08'] = 1.0 * contacts/np.mean(contacts) r_change['2020-03-16'] = 0.3 * contacts/np.mean(contacts) r_change['2020-03-23'] = 0.2 * contacts/np.mean(contacts) r_change['2020-05-04'] = 0.6 * contacts/np.mean(contacts) com_attack_rate = {} com_attack_rate["2020-01-1"] = 0.5 com_attack_rate["2020-05-4"] = 0.5 state, statesum, infections, day0, rnow, args, gr = cl.sim( age, drate, nday=180, prob_icu=0.009, day0cumrep=450, mean_days_to_icu=16, mean_duration_icu=14, mean_time_to_death=21, mean_serial=7.5, std_serial=3.0, immunt0=0.0, ifr=0.003, long_term_death=False, hnr=hnr, com_attack_rate=com_attack_rate, r_change=r_change, simname="Lockerung Alle", datadir=".", realized=nrw, rep_delay=13, alpha=0.125, day0date=day0date) cl.plotoverview(gr, args) ###Output _____no_output_____ ###Markdown Das effektive R ist ebenfalls leicht über 1, aber die Anzahl der ICU-Fälle steigt wieder stetig an. Starke Lockerung aber Reduzierung der Community Attack Rate durch frühe Tests und Quarantäne ###Code day0date = datetime.date(2020, 3, 8) r_change = {} r_change['2020-01-01'] = 3 * contacts/np.mean(contacts) contacts_new = np.where(age < 20, contacts, contacts) r_change['2020-03-08'] = 1.0 * contacts/np.mean(contacts) r_change['2020-03-16'] = 0.3 * contacts/np.mean(contacts) r_change['2020-03-23'] = 0.2 * contacts/np.mean(contacts) r_change['2020-05-04'] = 1.0 * contacts/np.mean(contacts) com_attack_rate = {} com_attack_rate["2020-01-1"] = 0.5 com_attack_rate["2020-05-4"] = 0.25 state, statesum, infections, day0, rnow, args, gr = cl.sim( age, drate, nday=180, prob_icu=0.009, day0cumrep=450, mean_days_to_icu=16, mean_duration_icu=14, mean_time_to_death=21, mean_serial=7.5, std_serial=3.0, immunt0=0.0, ifr=0.003, long_term_death=False, hnr=hnr, com_attack_rate=com_attack_rate, r_change=r_change, simname="Lockerung Alle kleinere CAR", datadir=".", realized=nrw, rep_delay=13, alpha=0.125, day0date=day0date) cl.plotoverview(gr, args) ###Output _____no_output_____ ###Markdown Test der Community Attack Rate ###Code day0date = datetime.date(2020, 3, 8) r_change = {} r_change['2020-01-01'] = 3 * contacts/np.mean(contacts) contacts_new = np.where(age < 20, contacts, contacts) r_change['2020-03-08'] = 1.0 * contacts/np.mean(contacts) r_change['2020-03-16'] = 0.3 * contacts/np.mean(contacts) r_change['2020-03-23'] = 0.2 * contacts/np.mean(contacts) r_change['2020-05-04'] = 0.2 * contacts/np.mean(contacts) com_attack_rate = {} com_attack_rate["2020-01-1"] = 0.5 com_attack_rate["2020-05-4"] = 0.1 state, statesum, infections, day0, rnow, args, gr = cl.sim( age, drate, nday=180, prob_icu=0.009, day0cumrep=450, mean_days_to_icu=16, mean_duration_icu=14, mean_time_to_death=21, mean_serial=7.5, std_serial=3.0, immunt0=0.0, ifr=0.003, long_term_death=False, hnr=hnr, com_attack_rate=com_attack_rate, r_change=r_change, simname="Test Community Attack", datadir=".", realized=nrw, rep_delay=13, alpha=0.125, day0date=day0date) cl.plotoverview(gr, args) ###Output _____no_output_____ ###Markdown Szenarien ohne Community Attack Basisszenario NRW ###Code age, agegroup, gender, contacts, drate, hnr, persons = cl.makepop("current",17900000) day0date = datetime.date(2020, 3, 8) r_change = {} r_change['2020-01-01'] = 4.0 * contacts/np.mean(contacts) r_change['2020-03-08'] = 0.24*8 * contacts/np.mean(contacts) r_change['2020-03-16'] = 0.15*8 * contacts/np.mean(contacts) r_change['2020-03-23'] = 0.10*6 * contacts/np.mean(contacts) r_change['2020-05-04'] = 0.6 * contacts/np.mean(contacts) com_attack_rate = {} com_attack_rate["2020-01-1"] = 0.0 state, statesum, infections, day0, rnow, args, gr = cl.sim( age, drate, nday=180, prob_icu=0.009, day0cumrep=450, mean_days_to_icu=16, mean_duration_icu=14, mean_time_to_death=21, mean_serial=7.5, std_serial=3.0, immunt0=0.0, ifr=0.003, long_term_death=False, hnr=None, com_attack_rate=com_attack_rate, r_change=r_change, simname="Basis NRW ohne Community Attack", datadir=".", realized=nrw, rep_delay=13, alpha=0.125, day0date=day0date) cl.plotoverview(gr, args) import pkg_resources pkg_resources.get_distribution("covid19sim").version cl.plotoverview(gr, args) gr["Wochentag"] = [x.weekday() for x in gr.Datum] gr["WE"] = np.where(gr.Wochentag > 4, "WE", "WT") fig = make_subplots(rows=2, cols=2) fig.add_trace(go.Scatter(x=gr["Datum"], y=gr["Erwartete Neu-Meldefälle"], mode="lines", name="Erwartete Neu-Meldefälle"), row=1, col=1) fig.add_trace(go.Scatter(x=gr[gr.WE == "WE"]["Datum"], y=gr[gr.WE == "WE"]["RKI Neu-Meldefälle"], name="RKI Neu-Meldefälle (WE)", mode="markers"), row=1, col=1) fig.add_trace(go.Scatter(x=gr[gr.WE == "WT"]["Datum"], y=gr[gr.WE == "WT"]["RKI Neu-Meldefälle"], name="RKI Neu-Meldefälle (WT)", mode="markers"), row=1, col=1) fig.add_trace(go.Scatter(x=gr["Datum"], y=gr["Erwartete Gesamt-Meldefälle"], name="Erwartete Gesamt-Meldefälle", mode="lines"), row=2, col=1) fig.add_trace(go.Scatter(x=gr["Datum"], y=gr["RKI Gesamt-Meldefälle"], name="RKI Gesamt-Meldefälle", mode="lines"), row=2, col=1) fig.add_trace(go.Scatter(x=gr["Datum"], y=gr["Erwartete Tote"], name="Erwartete Tote", mode="lines"), row=1, col=2) fig.add_trace(go.Scatter(x=gr["Datum"], y=gr["IST Tote gesamt"], name="Ist Tote gesamt", mode="lines"), row=1, col=2) fig.add_trace(go.Scatter(x=gr["Datum"], y=gr["ICU"], name="Erwartete Intensiv", mode="lines"), row=2, col=2) fig.add_trace(go.Scatter(x=gr["Datum"], y=gr["Ist Intensiv"], name="IST Intensiv", mode="lines"), row=2, col=2) fig.update_layout(legend_orientation="h", title=args["simname"]) plot(fig, filename=os.path.join(args["datadir"], args["simname"] + "_overview.html"), auto_open=False, auto_play=False) fig.show() fig = make_subplots(rows=1, cols=1) fig.add_trace(go.Scatter(x=gr["Datum"], y=gr["Reproduktionszahl"], name="R effektiv", mode="lines"), row=1, col=1) fig.add_trace(go.Scatter(x=gr["Datum"], y=gr["R extern"], name="R extern", mode="lines"), row=1, col=1) plot(fig, filename=os.path.join(args["datadir"], args["simname"] + "_reproduction.html"), auto_open=False, auto_play=False) plot(fig) day0date = datetime.date(2020, 3, 8) r_change = {} r_change['2020-01-01'] = 4.0 * contacts/np.mean(contacts) r_change['2020-03-08'] = 0.24*8 * contacts/np.mean(contacts) r_change['2020-03-16'] = 0.15*8 * contacts/np.mean(contacts) r_change['2020-03-23'] = 0.9 * contacts/np.mean(contacts) r_change['2020-05-04'] = 0.9 * contacts/np.mean(contacts) com_attack_rate = {} com_attack_rate["2020-01-1"] = 0.0 state, statesum, infections, day0, rnow, args, gr = cl.sim( age, drate, nday=180, prob_icu=0.009, day0cumrep=450, mean_days_to_icu=16, mean_duration_icu=14, mean_time_to_death=21, mean_serial=7.5, std_serial=3.0, immunt0=0.0, ifr=0.003, long_term_death=False, hnr=None, com_attack_rate=com_attack_rate, r_change=r_change, simname="Basis NRW ohne Community Attack", datadir=".", realized=nrw, rep_delay=13, alpha=0.125, day0date=day0date) cl.plotoverview(gr, args) ###Output _____no_output_____
Drawing/OldDrawing/Shapes-Copy5.ipynb
###Markdown Draw a ShapesAttempt at programmatically drawing shapes.All units in mm. ```1``` = ```1 mm```. ###Code s = GCode.Shapes.Square() s p = GCode.Program() p p.generate_gcode() p cnc = GRBL.GRBL(port="/dev/cnc_3018") cnc.reset() cnc.status cnc.home() cnc.status o0 = 10 def origin_calc(lines): if len(lines) == 0: o = o0 else: o = np.sum(list(map(lambda s: s.len_side, lines))) + o0 return np.array([o, o]) sides = [0.1, 0.25, 0.5, 0.75, 1, 2.5, 5, 10, 20, 25.4, 25.4 / 4] np.sum(sides) + o0 lines = list() for side in sides: _ = GCode.Shapes.Square(len_side=side, origin=origin_calc(lines)) lines.append(_) lines prog = GCode.Program() prog.generate_gcode() prog.buffer cnc.run(prog) for line in lines: break line.__repr__() line.buffer cnc.status cnc.run(line) ###Output _____no_output_____
hpc_ai/ai_science_climate/English/python/jupyter_notebook/Start_Here.ipynb
###Markdown Welcome to AI for Science BootcampThe objective of this bootcamp is to give an introduction to application of Artificial Intelligence (AI) algorithms in Science ( High Performance Computing(HPC) Simulations ). This bootcamp will introduce participants to fundamentals of AI and how those can be applied to different HPC simulation domains. The following contents will be covered during the Bootcamp :- [CNN Primer and Keras 101](Intro_to_DL/Part_2.ipynb)- [Tropical Cyclone Intensity Estimation using Deep Convolution Neural Networks.](Tropical_Cyclone_Intensity_Estimation/The_Problem_Statement.ipynb) Quick GPU CheckBefore moving forward let us check if Tensorflow backend is able to see and use GPU ###Code # Import Necessary Libraries from __future__ import absolute_import, division, print_function, unicode_literals # TensorFlow and tf.keras import tensorflow as tf from tensorflow import keras # Helper libraries import numpy as np import matplotlib.pyplot as plt print(tf.__version__) tf.test.gpu_device_name() ###Output _____no_output_____
Programming Assignments/Course 5: Sequence Models/Emojify_v2a.ipynb
###Markdown Emojify! Welcome to the second assignment of Week 2. You are going to use word vector representations to build an Emojifier. Have you ever wanted to make your text messages more expressive? Your emojifier app will help you do that. So rather than writing:>"Congratulations on the promotion! Let's get coffee and talk. Love you!" The emojifier can automatically turn this into:>"Congratulations on the promotion! 👍 Let's get coffee and talk. ☕️ Love you! ❤️"* You will implement a model which inputs a sentence (such as "Let's go see the baseball game tonight!") and finds the most appropriate emoji to be used with this sentence (⚾️). Using word vectors to improve emoji lookups* In many emoji interfaces, you need to remember that ❤️ is the "heart" symbol rather than the "love" symbol. * In other words, you'll have to remember to type "heart" to find the desired emoji, and typing "love" won't bring up that symbol.* We can make a more flexible emoji interface by using word vectors!* When using word vectors, you'll see that even if your training set explicitly relates only a few words to a particular emoji, your algorithm will be able to generalize and associate additional words in the test set to the same emoji. * This works even if those additional words don't even appear in the training set. * This allows you to build an accurate classifier mapping from sentences to emojis, even using a small training set. What you'll build1. In this exercise, you'll start with a baseline model (Emojifier-V1) using word embeddings.2. Then you will build a more sophisticated model (Emojifier-V2) that further incorporates an LSTM. Updates If you were working on the notebook before this update...* The current notebook is version "2a".* You can find your original work saved in the notebook with the previous version name ("v2") * To view the file directory, go to the menu "File->Open", and this will open a new tab that shows the file directory. List of updates* sentence_to_avg * Updated instructions. * Use separate variables to store the total and the average (instead of just `avg`). * Additional hint about how to initialize the shape of `avg` vector.* sentences_to_indices * Updated preceding text and instructions, added additional hints.* pretrained_embedding_layer * Additional instructions to explain how to implement each step.* Emoify_V2 * Modifies instructions to specify which parameters are needed for each Keras layer. * Remind users of Keras syntax. * Explanation of how to use the layer object that is returned by `pretrained_embedding_layer`. * Provides sample Keras code.* Spelling, grammar and wording corrections. Let's get started! Run the following cell to load the package you are going to use. ###Code import numpy as np from emo_utils import * import emoji import matplotlib.pyplot as plt %matplotlib inline ###Output _____no_output_____ ###Markdown 1 - Baseline model: Emojifier-V1 1.1 - Dataset EMOJISETLet's start by building a simple baseline classifier. You have a tiny dataset (X, Y) where:- X contains 127 sentences (strings).- Y contains an integer label between 0 and 4 corresponding to an emoji for each sentence. **Figure 1**: EMOJISET - a classification problem with 5 classes. A few examples of sentences are given here. Let's load the dataset using the code below. We split the dataset between training (127 examples) and testing (56 examples). ###Code X_train, Y_train = read_csv('data/train_emoji.csv') X_test, Y_test = read_csv('data/tesss.csv') maxLen = len(max(X_train, key=len).split()) ###Output _____no_output_____ ###Markdown Run the following cell to print sentences from X_train and corresponding labels from Y_train. * Change `idx` to see different examples. * Note that due to the font used by iPython notebook, the heart emoji may be colored black rather than red. ###Code for idx in range(10): print(X_train[idx], label_to_emoji(Y_train[idx])) ###Output never talk to me again 😞 I am proud of your achievements 😄 It is the worst day in my life 😞 Miss you so much ❤️ food is life 🍴 I love you mum ❤️ Stop saying bullshit 😞 congratulations on your acceptance 😄 The assignment is too long 😞 I want to go play ⚾ ###Markdown 1.2 - Overview of the Emojifier-V1In this part, you are going to implement a baseline model called "Emojifier-v1". **Figure 2**: Baseline model (Emojifier-V1). Inputs and outputs* The input of the model is a string corresponding to a sentence (e.g. "I love you). * The output will be a probability vector of shape (1,5), (there are 5 emojis to choose from).* The (1,5) probability vector is passed to an argmax layer, which extracts the index of the emoji with the highest probability. One-hot encoding* To get our labels into a format suitable for training a softmax classifier, lets convert $Y$ from its current shape $(m, 1)$ into a "one-hot representation" $(m, 5)$, * Each row is a one-hot vector giving the label of one example. * Here, `Y_oh` stands for "Y-one-hot" in the variable names `Y_oh_train` and `Y_oh_test`: ###Code Y_oh_train = convert_to_one_hot(Y_train, C = 5) Y_oh_test = convert_to_one_hot(Y_test, C = 5) ###Output _____no_output_____ ###Markdown Let's see what `convert_to_one_hot()` did. Feel free to change `index` to print out different values. ###Code idx = 50 print(f"Sentence '{X_train[50]}' has label index {Y_train[idx]}, which is emoji {label_to_emoji(Y_train[idx])}", ) print(f"Label index {Y_train[idx]} in one-hot encoding format is {Y_oh_train[idx]}") ###Output Sentence 'I missed you' has label index 0, which is emoji ❤️ Label index 0 in one-hot encoding format is [ 1. 0. 0. 0. 0.] ###Markdown All the data is now ready to be fed into the Emojify-V1 model. Let's implement the model! 1.3 - Implementing Emojifier-V1As shown in Figure 2 (above), the first step is to:* Convert each word in the input sentence into their word vector representations.* Then take an average of the word vectors. * Similar to the previous exercise, we will use pre-trained 50-dimensional GloVe embeddings. Run the following cell to load the `word_to_vec_map`, which contains all the vector representations. ###Code word_to_index, index_to_word, word_to_vec_map = read_glove_vecs('../../readonly/glove.6B.50d.txt') ###Output _____no_output_____ ###Markdown You've loaded:- `word_to_index`: dictionary mapping from words to their indices in the vocabulary - (400,001 words, with the valid indices ranging from 0 to 400,000)- `index_to_word`: dictionary mapping from indices to their corresponding words in the vocabulary- `word_to_vec_map`: dictionary mapping words to their GloVe vector representation.Run the following cell to check if it works. ###Code word = "cucumber" idx = 289846 print("the index of", word, "in the vocabulary is", word_to_index[word]) print("the", str(idx) + "th word in the vocabulary is", index_to_word[idx]) ###Output the index of cucumber in the vocabulary is 113317 the 289846th word in the vocabulary is potatos ###Markdown **Exercise**: Implement `sentence_to_avg()`. You will need to carry out two steps:1. Convert every sentence to lower-case, then split the sentence into a list of words. * `X.lower()` and `X.split()` might be useful. 2. For each word in the sentence, access its GloVe representation. * Then take the average of all of these word vectors. * You might use `numpy.zeros()`. Additional Hints* When creating the `avg` array of zeros, you'll want it to be a vector of the same shape as the other word vectors in the `word_to_vec_map`. * You can choose a word that exists in the `word_to_vec_map` and access its `.shape` field. * Be careful not to hard code the word that you access. In other words, don't assume that if you see the word 'the' in the `word_to_vec_map` within this notebook, that this word will be in the `word_to_vec_map` when the function is being called by the automatic grader. * Hint: you can use any one of the word vectors that you retrieved from the input `sentence` to find the shape of a word vector. ###Code # GRADED FUNCTION: sentence_to_avg def sentence_to_avg(sentence, word_to_vec_map): """ Converts a sentence (string) into a list of words (strings). Extracts the GloVe representation of each word and averages its value into a single vector encoding the meaning of the sentence. Arguments: sentence -- string, one training example from X word_to_vec_map -- dictionary mapping every word in a vocabulary into its 50-dimensional vector representation Returns: avg -- average vector encoding information about the sentence, numpy-array of shape (50,) """ ### START CODE HERE ### # Step 1: Split sentence into list of lower case words (≈ 1 line) #print("sentence = " + sentence) words = sentence.lower().split() #words = [i.lower() for i in sentence.split()] #print("words lenght = " + str(len(words))) # Initialize the average word vector, should have the same shape as your word vectors. avg = np.zeros((50,)) # <Gema> one dimensional array with 50 elements (we use pretrained 50-dimensional GloVe embeddings) # Step 2: average the word vectors. You can loop over the words in the list "words". for w in words: avg += word_to_vec_map[w] avg = avg/len(words) ### END CODE HERE ### return avg avg = sentence_to_avg("Morrocan couscous is my favorite dish", word_to_vec_map) print("avg = \n", avg) ###Output avg = [-0.008005 0.56370833 -0.50427333 0.258865 0.55131103 0.03104983 -0.21013718 0.16893933 -0.09590267 0.141784 -0.15708967 0.18525867 0.6495785 0.38371117 0.21102167 0.11301667 0.02613967 0.26037767 0.05820667 -0.01578167 -0.12078833 -0.02471267 0.4128455 0.5152061 0.38756167 -0.898661 -0.535145 0.33501167 0.68806933 -0.2156265 1.797155 0.10476933 -0.36775333 0.750785 0.10282583 0.348925 -0.27262833 0.66768 -0.10706167 -0.283635 0.59580117 0.28747333 -0.3366635 0.23393817 0.34349183 0.178405 0.1166155 -0.076433 0.1445417 0.09808667] ###Markdown **Expected Output**:```Pythonavg =[-0.008005 0.56370833 -0.50427333 0.258865 0.55131103 0.03104983 -0.21013718 0.16893933 -0.09590267 0.141784 -0.15708967 0.18525867 0.6495785 0.38371117 0.21102167 0.11301667 0.02613967 0.26037767 0.05820667 -0.01578167 -0.12078833 -0.02471267 0.4128455 0.5152061 0.38756167 -0.898661 -0.535145 0.33501167 0.68806933 -0.2156265 1.797155 0.10476933 -0.36775333 0.750785 0.10282583 0.348925 -0.27262833 0.66768 -0.10706167 -0.283635 0.59580117 0.28747333 -0.3366635 0.23393817 0.34349183 0.178405 0.1166155 -0.076433 0.1445417 0.09808667]``` ModelYou now have all the pieces to finish implementing the `model()` function. After using `sentence_to_avg()` you need to:* Pass the average through forward propagation* Compute the cost* Backpropagate to update the softmax parameters**Exercise**: Implement the `model()` function described in Figure (2). * The equations you need to implement in the forward pass and to compute the cross-entropy cost are below:* The variable $Y_{oh}$ ("Y one hot") is the one-hot encoding of the output labels. $$ z^{(i)} = W . avg^{(i)} + b$$$$ a^{(i)} = softmax(z^{(i)})$$$$ \mathcal{L}^{(i)} = - \sum_{k = 0}^{n_y - 1} Y_{oh,k}^{(i)} * log(a^{(i)}_k)$$**Note** It is possible to come up with a more efficient vectorized implementation. For now, let's use nested for loops to better understand the algorithm, and for easier debugging.We provided the function `softmax()`, which was imported earlier. ###Code # GRADED FUNCTION: model def model(X, Y, word_to_vec_map, learning_rate = 0.01, num_iterations = 400): """ Model to train word vector representations in numpy. Arguments: X -- input data, numpy array of sentences as strings, of shape (m, 1) Y -- labels, numpy array of integers between 0 and 7, numpy-array of shape (m, 1) word_to_vec_map -- dictionary mapping every word in a vocabulary into its 50-dimensional vector representation learning_rate -- learning_rate for the stochastic gradient descent algorithm num_iterations -- number of iterations Returns: pred -- vector of predictions, numpy-array of shape (m, 1) W -- weight matrix of the softmax layer, of shape (n_y, n_h) b -- bias of the softmax layer, of shape (n_y,) """ np.random.seed(1) # Define number of training examples m = Y.shape[0] # number of training examples n_y = 5 # number of classes n_h = 50 # dimensions of the GloVe vectors # Initialize parameters using Xavier initialization W = np.random.randn(n_y, n_h) / np.sqrt(n_h) b = np.zeros((n_y,)) # Convert Y to Y_onehot with n_y classes Y_oh = convert_to_one_hot(Y, C = n_y) # Optimization loop for t in range(num_iterations): # Loop over the number of iterations for i in range(m): # Loop over the training examples ### START CODE HERE ### (≈ 4 lines of code) # Average the word vectors of the words from the i'th training example avg = sentence_to_avg(X[i], word_to_vec_map) # Forward propagate the avg through the softmax layer z = np.dot(W,avg)+b a = softmax(z) # Compute cost using the i'th training label's one hot representation and "A" (the output of the softmax) cost = -np.sum(Y_oh[i]*np.log(a)) #cost = -np.sum(np.multiply(Y_oh[i], np.log(a))) ### END CODE HERE ### # Compute gradients dz = a - Y_oh[i] dW = np.dot(dz.reshape(n_y,1), avg.reshape(1, n_h)) db = dz # Update parameters with Stochastic Gradient Descent W = W - learning_rate * dW b = b - learning_rate * db if t % 100 == 0: print("Epoch: " + str(t) + " --- cost = " + str(cost)) pred = predict(X, Y, W, b, word_to_vec_map) return pred, W, b print(X_train.shape) print(Y_train.shape) print(np.eye(5)[Y_train.reshape(-1)].shape) print(X_train[0]) print(type(X_train)) Y = np.asarray([5,0,0,5, 4, 4, 4, 6, 6, 4, 1, 1, 5, 6, 6, 3, 6, 3, 4, 4]) print(Y.shape) X = np.asarray(['I am going to the bar tonight', 'I love you', 'miss you my dear', 'Lets go party and drinks','Congrats on the new job','Congratulations', 'I am so happy for you', 'Why are you feeling bad', 'What is wrong with you', 'You totally deserve this prize', 'Let us go play football', 'Are you down for football this afternoon', 'Work hard play harder', 'It is suprising how people can be dumb sometimes', 'I am very disappointed','It is the best day in my life', 'I think I will end up alone','My life is so boring','Good job', 'Great so awesome']) print(X.shape) print(np.eye(5)[Y_train.reshape(-1)].shape) print(type(X_train)) ###Output (132,) (132,) (132, 5) never talk to me again <class 'numpy.ndarray'> (20,) (20,) (132, 5) <class 'numpy.ndarray'> ###Markdown Run the next cell to train your model and learn the softmax parameters (W,b). ###Code pred, W, b = model(X_train, Y_train, word_to_vec_map) print(pred) ###Output Epoch: 0 --- cost = 1.95204988128 Accuracy: 0.348484848485 Epoch: 100 --- cost = 0.0797181872601 Accuracy: 0.931818181818 Epoch: 200 --- cost = 0.0445636924368 Accuracy: 0.954545454545 Epoch: 300 --- cost = 0.0343226737879 Accuracy: 0.969696969697 [[ 3.] [ 2.] [ 3.] [ 0.] [ 4.] [ 0.] [ 3.] [ 2.] [ 3.] [ 1.] [ 3.] [ 3.] [ 1.] [ 3.] [ 2.] [ 3.] [ 2.] [ 3.] [ 1.] [ 2.] [ 3.] [ 0.] [ 2.] [ 2.] [ 2.] [ 1.] [ 4.] [ 3.] [ 3.] [ 4.] [ 0.] [ 3.] [ 4.] [ 2.] [ 0.] [ 3.] [ 2.] [ 2.] [ 3.] [ 4.] [ 2.] [ 2.] [ 0.] [ 2.] [ 3.] [ 0.] [ 3.] [ 2.] [ 4.] [ 3.] [ 0.] [ 3.] [ 3.] [ 3.] [ 4.] [ 2.] [ 1.] [ 1.] [ 1.] [ 2.] [ 3.] [ 1.] [ 0.] [ 0.] [ 0.] [ 3.] [ 4.] [ 4.] [ 2.] [ 2.] [ 1.] [ 2.] [ 0.] [ 3.] [ 2.] [ 2.] [ 0.] [ 3.] [ 3.] [ 1.] [ 2.] [ 1.] [ 2.] [ 2.] [ 4.] [ 3.] [ 3.] [ 2.] [ 4.] [ 0.] [ 0.] [ 3.] [ 3.] [ 3.] [ 3.] [ 2.] [ 0.] [ 1.] [ 2.] [ 3.] [ 0.] [ 2.] [ 2.] [ 2.] [ 3.] [ 2.] [ 2.] [ 2.] [ 4.] [ 1.] [ 1.] [ 3.] [ 3.] [ 4.] [ 1.] [ 2.] [ 1.] [ 1.] [ 3.] [ 1.] [ 0.] [ 4.] [ 0.] [ 3.] [ 3.] [ 4.] [ 4.] [ 1.] [ 4.] [ 3.] [ 0.] [ 2.]] ###Markdown **Expected Output** (on a subset of iterations): **Epoch: 0** cost = 1.95204988128 Accuracy: 0.348484848485 **Epoch: 100** cost = 0.0797181872601 Accuracy: 0.931818181818 **Epoch: 200** cost = 0.0445636924368 Accuracy: 0.954545454545 **Epoch: 300** cost = 0.0343226737879 Accuracy: 0.969696969697 Great! Your model has pretty high accuracy on the training set. Lets now see how it does on the test set. 1.4 - Examining test set performance * Note that the `predict` function used here is defined in emo_util.spy. ###Code print("Training set:") pred_train = predict(X_train, Y_train, W, b, word_to_vec_map) print('Test set:') pred_test = predict(X_test, Y_test, W, b, word_to_vec_map) ###Output Training set: Accuracy: 0.977272727273 Test set: Accuracy: 0.857142857143 ###Markdown **Expected Output**: **Train set accuracy** 97.7 **Test set accuracy** 85.7 * Random guessing would have had 20% accuracy given that there are 5 classes. (1/5 = 20%).* This is pretty good performance after training on only 127 examples. The model matches emojis to relevant wordsIn the training set, the algorithm saw the sentence >"*I love you*" with the label ❤️. * You can check that the word "adore" does not appear in the training set. * Nonetheless, lets see what happens if you write "*I adore you*." ###Code X_my_sentences = np.array(["i adore you", "i love you", "funny lol", "lets play with a ball", "food is ready", "not feeling happy"]) Y_my_labels = np.array([[0], [0], [2], [1], [4],[3]]) pred = predict(X_my_sentences, Y_my_labels , W, b, word_to_vec_map) print_predictions(X_my_sentences, pred) ###Output Accuracy: 0.833333333333 i adore you ❤️ i love you ❤️ funny lol 😄 lets play with a ball ⚾ food is ready 🍴 not feeling happy 😄 ###Markdown Amazing! * Because *adore* has a similar embedding as *love*, the algorithm has generalized correctly even to a word it has never seen before. * Words such as *heart*, *dear*, *beloved* or *adore* have embedding vectors similar to *love*. * Feel free to modify the inputs above and try out a variety of input sentences. * How well does it work? Word ordering isn't considered in this model* Note that the model doesn't get the following sentence correct:>"not feeling happy" * This algorithm ignores word ordering, so is not good at understanding phrases like "not happy." Confusion matrix* Printing the confusion matrix can also help understand which classes are more difficult for your model. * A confusion matrix shows how often an example whose label is one class ("actual" class) is mislabeled by the algorithm with a different class ("predicted" class). ###Code print(Y_test.shape) print(' '+ label_to_emoji(0)+ ' ' + label_to_emoji(1) + ' ' + label_to_emoji(2)+ ' ' + label_to_emoji(3)+' ' + label_to_emoji(4)) print(pd.crosstab(Y_test, pred_test.reshape(56,), rownames=['Actual'], colnames=['Predicted'], margins=True)) plot_confusion_matrix(Y_test, pred_test) ###Output (56,) ❤️ ⚾ 😄 😞 🍴 Predicted 0.0 1.0 2.0 3.0 4.0 All Actual 0 6 0 0 1 0 7 1 0 8 0 0 0 8 2 2 0 16 0 0 18 3 1 1 2 12 0 16 4 0 0 1 0 6 7 All 9 9 19 13 6 56 ###Markdown What you should remember from this section- Even with a 127 training examples, you can get a reasonably good model for Emojifying. - This is due to the generalization power word vectors gives you. - Emojify-V1 will perform poorly on sentences such as *"This movie is not good and not enjoyable"* - It doesn't understand combinations of words. - It just averages all the words' embedding vectors together, without considering the ordering of words. **You will build a better algorithm in the next section!** 2 - Emojifier-V2: Using LSTMs in Keras: Let's build an LSTM model that takes word **sequences** as input!* This model will be able to account for the word ordering. * Emojifier-V2 will continue to use pre-trained word embeddings to represent words.* We will feed word embeddings into an LSTM.* The LSTM will learn to predict the most appropriate emoji. Run the following cell to load the Keras packages. ###Code import numpy as np np.random.seed(0) from keras.models import Model from keras.layers import Dense, Input, Dropout, LSTM, Activation from keras.layers.embeddings import Embedding from keras.preprocessing import sequence from keras.initializers import glorot_uniform np.random.seed(1) ###Output Using TensorFlow backend. ###Markdown 2.1 - Overview of the modelHere is the Emojifier-v2 you will implement: **Figure 3**: Emojifier-V2. A 2-layer LSTM sequence classifier. 2.2 Keras and mini-batching * In this exercise, we want to train Keras using mini-batches. * However, most deep learning frameworks require that all sequences in the same mini-batch have the **same length**. * This is what allows vectorization to work: If you had a 3-word sentence and a 4-word sentence, then the computations needed for them are different (one takes 3 steps of an LSTM, one takes 4 steps) so it's just not possible to do them both at the same time. Padding handles sequences of varying length* The common solution to handling sequences of **different length** is to use padding. Specifically: * Set a maximum sequence length * Pad all sequences to have the same length. Example of padding* Given a maximum sequence length of 20, we could pad every sentence with "0"s so that each input sentence is of length 20. * Thus, the sentence "I love you" would be represented as $(e_{I}, e_{love}, e_{you}, \vec{0}, \vec{0}, \ldots, \vec{0})$. * In this example, any sentences longer than 20 words would have to be truncated. * One way to choose the maximum sequence length is to just pick the length of the longest sentence in the training set. 2.3 - The Embedding layer* In Keras, the embedding matrix is represented as a "layer".* The embedding matrix maps word indices to embedding vectors. * The word indices are positive integers. * The embedding vectors are dense vectors of fixed size. * When we say a vector is "dense", in this context, it means that most of the values are non-zero. As a counter-example, a one-hot encoded vector is not "dense."* The embedding matrix can be derived in two ways: * Training a model to derive the embeddings from scratch. * Using a pretrained embedding Using and updating pre-trained embeddings* In this part, you will learn how to create an [Embedding()](https://keras.io/layers/embeddings/) layer in Keras* You will initialize the Embedding layer with the GloVe 50-dimensional vectors. * In the code below, we'll show you how Keras allows you to either train or leave fixed this layer. * Because our training set is quite small, we will leave the GloVe embeddings fixed instead of updating them. Inputs and outputs to the embedding layer* The `Embedding()` layer's input is an integer matrix of size **(batch size, max input length)**. * This input corresponds to sentences converted into lists of indices (integers). * The largest integer (the highest word index) in the input should be no larger than the vocabulary size.* The embedding layer outputs an array of shape (batch size, max input length, dimension of word vectors).* The figure shows the propagation of two example sentences through the embedding layer. * Both examples have been zero-padded to a length of `max_len=5`. * The word embeddings are 50 units in length. * The final dimension of the representation is `(2,max_len,50)`. **Figure 4**: Embedding layer Prepare the input sentences**Exercise**: * Implement `sentences_to_indices`, which processes an array of sentences (X) and returns inputs to the embedding layer: * Convert each training sentences into a list of indices (the indices correspond to each word in the sentence) * Zero-pad all these lists so that their length is the length of the longest sentence. Additional Hints* Note that you may have considered using the `enumerate()` function in the for loop, but for the purposes of passing the autograder, please follow the starter code by initializing and incrementing `j` explicitly. ###Code for idx, val in enumerate(["I", "like", "learning"]): print(idx,val) # GRADED FUNCTION: sentences_to_indices def sentences_to_indices(X, word_to_index, max_len): """ Converts an array of sentences (strings) into an array of indices corresponding to words in the sentences. The output shape should be such that it can be given to `Embedding()` (described in Figure 4). Arguments: X -- array of sentences (strings), of shape (m, 1) word_to_index -- a dictionary containing the each word mapped to its index max_len -- maximum number of words in a sentence. You can assume every sentence in X is no longer than this. Returns: X_indices -- array of indices corresponding to words in the sentences from X, of shape (m, max_len) """ m = X.shape[0] # number of training examples ### START CODE HERE ### # Initialize X_indices as a numpy matrix of zeros and the correct shape (≈ 1 line) X_indices = np.zeros((m, max_len)) for i in range(m): # loop over training examples # Convert the ith training sentence in lower case and split is into words. You should get a list of words. #print("X[i] = " + X[i]) sentence_words =X[i].lower().split() #print("sentence_words lenght = " + str(len(sentence_words))) # Initialize j to 0 j = 0 # Loop over the words of sentence_words for w in sentence_words: # Set the (i,j)th entry of X_indices to the index of the correct word. #print("w = " + w) X_indices[i, j] = word_to_index[w] # Increment j to j + 1 j = j+1 #j += 1 ### END CODE HERE ### return X_indices ###Output _____no_output_____ ###Markdown Run the following cell to check what `sentences_to_indices()` does, and check your results. ###Code X1 = np.array(["funny lol", "lets play baseball", "food is ready for you"]) X1_indices = sentences_to_indices(X1,word_to_index, max_len = 5) print("X1 =", X1) print("X1_indices =\n", X1_indices) ###Output X1 = ['funny lol' 'lets play baseball' 'food is ready for you'] X1_indices = [[ 155345. 225122. 0. 0. 0.] [ 220930. 286375. 69714. 0. 0.] [ 151204. 192973. 302254. 151349. 394475.]] ###Markdown **Expected Output**:```PythonX1 = ['funny lol' 'lets play baseball' 'food is ready for you']X1_indices = [[ 155345. 225122. 0. 0. 0.] [ 220930. 286375. 69714. 0. 0.] [ 151204. 192973. 302254. 151349. 394475.]]``` Build embedding layer* Let's build the `Embedding()` layer in Keras, using pre-trained word vectors. * The embedding layer takes as input a list of word indices. * `sentences_to_indices()` creates these word indices.* The embedding layer will return the word embeddings for a sentence. **Exercise**: Implement `pretrained_embedding_layer()` with these steps:1. Initialize the embedding matrix as a numpy array of zeros. * The embedding matrix has a row for each unique word in the vocabulary. * There is one additional row to handle "unknown" words. * So vocab_len is the number of unique words plus one. * Each row will store the vector representation of one word. * For example, one row may be 50 positions long if using GloVe word vectors. * In the code below, `emb_dim` represents the length of a word embedding.2. Fill in each row of the embedding matrix with the vector representation of a word * Each word in `word_to_index` is a string. * word_to_vec_map is a dictionary where the keys are strings and the values are the word vectors.3. Define the Keras embedding layer. * Use [Embedding()](https://keras.io/layers/embeddings/). * The input dimension is equal to the vocabulary length (number of unique words plus one). * The output dimension is equal to the number of positions in a word embedding. * Make this layer's embeddings fixed. * If you were to set `trainable = True`, then it will allow the optimization algorithm to modify the values of the word embeddings. * In this case, we don't want the model to modify the word embeddings.4. Set the embedding weights to be equal to the embedding matrix. * Note that this is part of the code is already completed for you and does not need to be modified. ###Code # GRADED FUNCTION: pretrained_embedding_layer def pretrained_embedding_layer(word_to_vec_map, word_to_index): """ Creates a Keras Embedding() layer and loads in pre-trained GloVe 50-dimensional vectors. Arguments: word_to_vec_map -- dictionary mapping words to their GloVe vector representation. word_to_index -- dictionary mapping from words to their indices in the vocabulary (400,001 words) Returns: embedding_layer -- pretrained layer Keras instance """ vocab_len = len(word_to_index) + 1 # adding 1 to fit Keras embedding (requirement) emb_dim = word_to_vec_map["cucumber"].shape[0] # define dimensionality of your GloVe word vectors (= 50) ### START CODE HERE ### # Initialize the embedding matrix as a numpy array of zeros of shape (vocab_len, dimensions of word vectors = emb_dim) emb_matrix =np.zeros((vocab_len,emb_dim)) # Set each row "index" of the embedding matrix to be the word vector representation of the "index"th word of the vocabulary for word, index in word_to_index.items(): emb_matrix[index, :] = word_to_vec_map[word] # Define Keras embedding layer with the correct output/input sizes, make it trainable. Use Embedding(...). Make sure to set trainable=False. # <Gema> The Embedding layer is defined as the first hidden layer of a network. #It must specify 3 arguments: #input_dim: This is the size of the vocabulary in the text data. For example, if your data is integer encoded to values between 0-10, then the size of the vocabulary would be 11 words. #output_dim: This is the size of the vector space in which words will be embedded. It defines the size of the output vectors from this layer for each word. embedding_layer = Embedding(vocab_len,emb_dim,trainable=False) # <Gema>The output of the Embedding layer is a 2D vector with one embedding for each word in the input sequence of words (input document). ### END CODE HERE ### # Build the embedding layer, it is required before setting the weights of the embedding layer. Do not modify the "None". embedding_layer.build((None,)) # Set the weights of the embedding layer to the embedding matrix. Your layer is now pretrained. embedding_layer.set_weights([emb_matrix]) return embedding_layer embedding_layer = pretrained_embedding_layer(word_to_vec_map, word_to_index) print("weights[0][1][3] =", embedding_layer.get_weights()[0][1][3]) ###Output weights[0][1][3] = -0.3403 ###Markdown **Expected Output**:```Pythonweights[0][1][3] = -0.3403``` 2.3 Building the Emojifier-V2Lets now build the Emojifier-V2 model. * You feed the embedding layer's output to an LSTM network. **Figure 3**: Emojifier-v2. A 2-layer LSTM sequence classifier. **Exercise:** Implement `Emojify_V2()`, which builds a Keras graph of the architecture shown in Figure 3. * The model takes as input an array of sentences of shape (`m`, `max_len`, ) defined by `input_shape`. * The model outputs a softmax probability vector of shape (`m`, `C = 5`). * You may need to use the following Keras layers: * [Input()](https://keras.io/layers/core/input) * Set the `shape` and `dtype` parameters. * The inputs are integers, so you can specify the data type as a string, 'int32'. * [LSTM()](https://keras.io/layers/recurrent/lstm) * Set the `units` and `return_sequences` parameters. * [Dropout()](https://keras.io/layers/core/dropout) * Set the `rate` parameter. * [Dense()](https://keras.io/layers/core/dense) * Set the `units`, * Note that `Dense()` has an `activation` parameter. For the purposes of passing the autograder, please do not set the activation within `Dense()`. Use the separate `Activation` layer to do so. * [Activation()](https://keras.io/activations/). * You can pass in the activation of your choice as a lowercase string. * [Model](https://keras.io/models/model/) Set `inputs` and `outputs`. Additional Hints* Remember that these Keras layers return an object, and you will feed in the outputs of the previous layer as the input arguments to that object. The returned object can be created and called in the same line.```Python How to use Keras layers in two lines of codedense_object = Dense(units = ...)X = dense_object(inputs) How to use Keras layers in one line of codeX = Dense(units = ...)(inputs)```* The `embedding_layer` that is returned by `pretrained_embedding_layer` is a layer object that can be called as a function, passing in a single argument (sentence indices).* Here is some sample code in case you're stuck```Pythonraw_inputs = Input(shape=(maxLen,), dtype='int32')preprocessed_inputs = ... some pre-processingX = LSTM(units = ..., return_sequences= ...)(processed_inputs)X = Dropout(rate = ..., )(X)...X = Dense(units = ...)(X)X = Activation(...)(X)model = Model(inputs=..., outputs=...)...``` ###Code # GRADED FUNCTION: Emojify_V2 def Emojify_V2(input_shape, word_to_vec_map, word_to_index): """ Function creating the Emojify-v2 model's graph. Arguments: input_shape -- shape of the input, usually (max_len,) word_to_vec_map -- dictionary mapping every word in a vocabulary into its 50-dimensional vector representation word_to_index -- dictionary mapping from words to their indices in the vocabulary (400,001 words) Returns: model -- a model instance in Keras """ ### START CODE HERE ### # Define sentence_indices as the input of the graph, it should be of shape input_shape and dtype 'int32' (as it contains indices). # <Gema>first, let's define an string model that will encode sentences into 50-dimensional vectors. sentence_indices = Input(shape=input_shape,dtype='int32') # <Gema> The model takes as input an array of sentences of shape (m, max_len, ) defined by input_shape # Create the embedding layer pretrained with GloVe Vectors (≈1 line) embedding_layer = pretrained_embedding_layer(word_to_vec_map, word_to_index) # Propagate sentence_indices through your embedding layer, you get back the embeddings embeddings = embedding_layer(sentence_indices) # Propagate the embeddings through an LSTM layer with 128-dimensional hidden state # Be careful, the returned output should be a batch of sequences. # <Gema> https://machinelearningmastery.com/return-sequences-and-return-states-for-lstms-in-keras/ X = LSTM(units=128, return_sequences=True)(embeddings) # Add dropout with a probability of 0.5 X = Dropout(0.5)(X) # Propagate X trough another LSTM layer with 128-dimensional hidden state # Be careful, the returned output should be a single hidden state, not a batch of sequences. X = LSTM(units=128,return_sequences=False)(X) # Add dropout with a probability of 0.5 X = Dropout(0.5)(X) # Propagate X through a Dense layer with softmax activation to get back a batch of 5-dimensional vectors. X = Dense(5)(X) # Add a softmax activation X = Activation('softmax')(X) # Create Model instance which converts sentence_indices into X. model = Model(inputs=sentence_indices, outputs=X) ### END CODE HERE ### return model ###Output _____no_output_____ ###Markdown Run the following cell to create your model and check its summary. Because all sentences in the dataset are less than 10 words, we chose `max_len = 10`. You should see your architecture, it uses "20,223,927" parameters, of which 20,000,050 (the word embeddings) are non-trainable, and the remaining 223,877 are. Because our vocabulary size has 400,001 words (with valid indices from 0 to 400,000) there are 400,001\*50 = 20,000,050 non-trainable parameters. ###Code model = Emojify_V2((maxLen,), word_to_vec_map, word_to_index) model.summary() ###Output _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= input_1 (InputLayer) (None, 10) 0 _________________________________________________________________ embedding_2 (Embedding) (None, 10, 50) 20000050 _________________________________________________________________ lstm_1 (LSTM) (None, 10, 128) 91648 _________________________________________________________________ dropout_1 (Dropout) (None, 10, 128) 0 _________________________________________________________________ lstm_2 (LSTM) (None, 128) 131584 _________________________________________________________________ dropout_2 (Dropout) (None, 128) 0 _________________________________________________________________ dense_1 (Dense) (None, 5) 645 _________________________________________________________________ activation_1 (Activation) (None, 5) 0 ================================================================= Total params: 20,223,927 Trainable params: 223,877 Non-trainable params: 20,000,050 _________________________________________________________________ ###Markdown As usual, after creating your model in Keras, you need to compile it and define what loss, optimizer and metrics your are want to use. Compile your model using `categorical_crossentropy` loss, `adam` optimizer and `['accuracy']` metrics: ###Code model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy']) ###Output _____no_output_____ ###Markdown It's time to train your model. Your Emojifier-V2 `model` takes as input an array of shape (`m`, `max_len`) and outputs probability vectors of shape (`m`, `number of classes`). We thus have to convert X_train (array of sentences as strings) to X_train_indices (array of sentences as list of word indices), and Y_train (labels as indices) to Y_train_oh (labels as one-hot vectors). ###Code X_train_indices = sentences_to_indices(X_train, word_to_index, maxLen) Y_train_oh = convert_to_one_hot(Y_train, C = 5) ###Output _____no_output_____ ###Markdown Fit the Keras model on `X_train_indices` and `Y_train_oh`. We will use `epochs = 50` and `batch_size = 32`. ###Code model.fit(X_train_indices, Y_train_oh, epochs = 50, batch_size = 32, shuffle=True) ###Output Epoch 1/50 132/132 [==============================] - 0s - loss: 1.6083 - acc: 0.1970 Epoch 2/50 132/132 [==============================] - 0s - loss: 1.5322 - acc: 0.2955 Epoch 3/50 132/132 [==============================] - 0s - loss: 1.5008 - acc: 0.3258 Epoch 4/50 132/132 [==============================] - 0s - loss: 1.4384 - acc: 0.3561 Epoch 5/50 132/132 [==============================] - 0s - loss: 1.3469 - acc: 0.4545 Epoch 6/50 132/132 [==============================] - 0s - loss: 1.2331 - acc: 0.5076 Epoch 7/50 132/132 [==============================] - 0s - loss: 1.1758 - acc: 0.4470 Epoch 8/50 132/132 [==============================] - 0s - loss: 1.0539 - acc: 0.5758 Epoch 9/50 132/132 [==============================] - 0s - loss: 0.8765 - acc: 0.7121 Epoch 10/50 132/132 [==============================] - 0s - loss: 0.8228 - acc: 0.6970 Epoch 11/50 132/132 [==============================] - 0s - loss: 0.7027 - acc: 0.7500 Epoch 12/50 132/132 [==============================] - 0s - loss: 0.6004 - acc: 0.8030 Epoch 13/50 132/132 [==============================] - 0s - loss: 0.4932 - acc: 0.8333 Epoch 14/50 132/132 [==============================] - 0s - loss: 0.5094 - acc: 0.8333 - ETA: 0s - loss: 0.5157 - acc: 0.828 Epoch 15/50 132/132 [==============================] - 0s - loss: 0.4786 - acc: 0.8258 Epoch 16/50 132/132 [==============================] - 0s - loss: 0.3540 - acc: 0.8636 Epoch 17/50 132/132 [==============================] - 0s - loss: 0.3902 - acc: 0.8636 Epoch 18/50 132/132 [==============================] - 0s - loss: 0.6484 - acc: 0.8106 Epoch 19/50 132/132 [==============================] - 0s - loss: 0.5179 - acc: 0.8182 Epoch 20/50 132/132 [==============================] - 0s - loss: 0.3960 - acc: 0.8409 Epoch 21/50 132/132 [==============================] - 0s - loss: 0.4723 - acc: 0.8182 Epoch 22/50 132/132 [==============================] - 0s - loss: 0.3892 - acc: 0.8636 Epoch 23/50 132/132 [==============================] - 0s - loss: 0.3795 - acc: 0.8561 Epoch 24/50 132/132 [==============================] - 0s - loss: 0.3056 - acc: 0.9091 Epoch 25/50 132/132 [==============================] - 0s - loss: 0.3489 - acc: 0.8864 Epoch 26/50 132/132 [==============================] - 0s - loss: 0.2422 - acc: 0.9394 Epoch 27/50 132/132 [==============================] - 0s - loss: 0.3179 - acc: 0.8864 Epoch 28/50 132/132 [==============================] - 0s - loss: 0.2402 - acc: 0.9318 Epoch 29/50 132/132 [==============================] - 0s - loss: 0.3943 - acc: 0.8712 Epoch 30/50 132/132 [==============================] - 0s - loss: 0.2677 - acc: 0.9091 Epoch 31/50 132/132 [==============================] - 0s - loss: 0.2955 - acc: 0.8864 Epoch 32/50 132/132 [==============================] - 0s - loss: 0.2040 - acc: 0.9318 Epoch 33/50 132/132 [==============================] - 0s - loss: 0.2124 - acc: 0.9470 Epoch 34/50 132/132 [==============================] - 0s - loss: 0.1566 - acc: 0.9621 Epoch 35/50 132/132 [==============================] - 0s - loss: 0.1635 - acc: 0.9621 Epoch 36/50 132/132 [==============================] - 0s - loss: 0.1874 - acc: 0.9394 Epoch 37/50 132/132 [==============================] - 0s - loss: 0.1776 - acc: 0.9470 Epoch 38/50 132/132 [==============================] - 0s - loss: 0.2140 - acc: 0.9394 Epoch 39/50 132/132 [==============================] - 0s - loss: 0.1389 - acc: 0.9621 Epoch 40/50 132/132 [==============================] - 0s - loss: 0.1530 - acc: 0.9545 Epoch 41/50 132/132 [==============================] - 0s - loss: 0.0870 - acc: 0.9848 Epoch 42/50 132/132 [==============================] - 0s - loss: 0.0799 - acc: 0.9773 Epoch 43/50 132/132 [==============================] - 0s - loss: 0.0801 - acc: 0.9848 Epoch 44/50 132/132 [==============================] - 0s - loss: 0.0492 - acc: 0.9924 Epoch 45/50 132/132 [==============================] - 0s - loss: 0.0787 - acc: 0.9848 Epoch 46/50 132/132 [==============================] - 0s - loss: 0.1068 - acc: 0.9773 Epoch 47/50 132/132 [==============================] - 0s - loss: 0.1492 - acc: 0.9470 - ETA: 0s - loss: 0.1518 - acc: 0.945 Epoch 48/50 132/132 [==============================] - 0s - loss: 0.3031 - acc: 0.9242 Epoch 49/50 132/132 [==============================] - 0s - loss: 0.1150 - acc: 0.9773 Epoch 50/50 132/132 [==============================] - 0s - loss: 0.1831 - acc: 0.9394 ###Markdown Your model should perform around **90% to 100% accuracy** on the training set. The exact accuracy you get may be a little different. Run the following cell to evaluate your model on the test set. ###Code X_test_indices = sentences_to_indices(X_test, word_to_index, max_len = maxLen) Y_test_oh = convert_to_one_hot(Y_test, C = 5) loss, acc = model.evaluate(X_test_indices, Y_test_oh) print() print("Test accuracy = ", acc) ###Output 32/56 [================>.............] - ETA: 0s Test accuracy = 0.821428562914 ###Markdown You should get a test accuracy between 80% and 95%. Run the cell below to see the mislabelled examples. ###Code # This code allows you to see the mislabelled examples C = 5 y_test_oh = np.eye(C)[Y_test.reshape(-1)] X_test_indices = sentences_to_indices(X_test, word_to_index, maxLen) pred = model.predict(X_test_indices) for i in range(len(X_test)): x = X_test_indices num = np.argmax(pred[i]) if(num != Y_test[i]): print('Expected emoji:'+ label_to_emoji(Y_test[i]) + ' prediction: '+ X_test[i] + label_to_emoji(num).strip()) ###Output Expected emoji:😄 prediction: she got me a nice present ❤️ Expected emoji:😞 prediction: work is hard 😄 Expected emoji:😞 prediction: This girl is messing with me ❤️ Expected emoji:🍴 prediction: any suggestions for dinner 😄 Expected emoji:❤️ prediction: I love taking breaks 😞 Expected emoji:😄 prediction: you brighten my day ❤️ Expected emoji:😄 prediction: will you be my valentine ❤️ Expected emoji:🍴 prediction: See you at the restaurant 😄 Expected emoji:😞 prediction: go away ⚾ Expected emoji:🍴 prediction: I did not have breakfast ❤️ ###Markdown Now you can try it on your own example. Write your own sentence below. ###Code # Change the sentence below to see your prediction. Make sure all the words are in the Glove embeddings. x_test = np.array(['not feeling happy']) X_test_indices = sentences_to_indices(x_test, word_to_index, maxLen) print(x_test[0] +' '+ label_to_emoji(np.argmax(model.predict(X_test_indices)))) ###Output not feeling happy 😞
MNIST/mnist_cnn.ipynb
###Markdown **CNN for the MNIST Dataset** **Import libraries** ###Code import numpy from keras.datasets import mnist from keras.models import Sequential from keras.layers import Dense from keras.layers import Dropout from keras.layers import Flatten from keras.layers.convolutional import Convolution2D from keras.layers.convolutional import MaxPooling2D from keras.utils import np_utils from keras import backend as K K.tensorflow_backend.set_image_dim_ordering('th') seed = 7 numpy.random.seed(seed) ###Output _____no_output_____ ###Markdown **Load the Dataset** ###Code (X_train, y_train), (X_test, y_test) = mnist.load_data() ###Output _____no_output_____ ###Markdown **reshape to be [samples][channels][width][height]** ###Code X_train = X_train.reshape(X_train.shape[0], 1, 28, 28).astype('float32') X_test = X_test.reshape(X_test.shape[0], 1, 28, 28).astype('float32') ###Output _____no_output_____ ###Markdown **normalize inputs from 0-255 to 0-1** ###Code X_train = X_train / 255 X_test = X_test / 255 ###Output _____no_output_____ ###Markdown **one hot encode outputs** ###Code y_train = np_utils.to_categorical(y_train) y_test = np_utils.to_categorical(y_test) num_classes = y_test.shape[1] ###Output _____no_output_____ ###Markdown **define a simple CNN model** ###Code def baseline_model(): # create model model = Sequential() model.add(Convolution2D(32, 5, 5, input_shape=(1, 28, 28), activation='relu')) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.2)) model.add(Flatten()) model.add(Dense(128, activation='relu')) model.add(Dense(num_classes, activation='softmax')) # Compile model model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy']) return model ###Output _____no_output_____ ###Markdown **Build and Fit the model** ###Code model = baseline_model() model.fit(X_train, y_train, validation_data=(X_test, y_test), nb_epoch=10, batch_size=200, verbose=1) ###Output _____no_output_____ ###Markdown **Final evaluation of the model** ###Code scores = model.evaluate(X_test, y_test, verbose=0) print("CNN Error: %.2f%%" % (100-scores[1]*100)) ###Output _____no_output_____
Fundamentals of Data Analysis - Assignments .ipynb
###Markdown Fundamentals of Data Analysis - AssignmentsThis file contains the output of each assignment as part of the Fundamentals of Data Analysis module. OverviewThis module will contain four different assingments to be completed* [Assignment 1 Counts](t1)* [Assignment 2 Diceroll](t2)* [Assignment 3 Coin Flip](t3)* [Assignment 4 Simpson's Paradox](t4)*** Assignment 1 Counts Write a Python function called counts that takes a list asinput and returns a dictionary of unique items in the list as keys and the number oftimes each item appears as values. So, the input ['A', 'A', 'B', 'C', 'A']should have output {'A': 3, 'B': 1, 'C': 1} Refences Assignment 1[1] Tutorialspoint, Counting Frequencies in a list ising dictionary in Python; https://www.tutorialspoint.com/counting-the-frequencies-in-a-list-using-dictionary-in-python[2] Kite, How to count a frequency in Python; https://www.kite.com/python/answers/how-to-count-item-frequency-in-python[3] W3schools, Python List count() Method https://www.w3schools.com/python/ref_list_count.asp[4] GMIT lecture Video, For loops, Ian McLoughlin; https://web.microsoftstream.com/video/8492c53c-a684-4da9-a2c5-bce1d5c367a9[5] Adding item to a dictionary; https://www.w3schools.com/python/python_dictionaries.asp Lists for Function Test ###Code #List 1 - upper case letters list1 = ['A','A','B','C','A'] #List 2 - lower case letters list2 = ['a','a','b','c','a'] #list 3 - mix of lower and upper case list3 = ['A','a','B','c','C','A'] #List 4 - mix of upper case leters, lower case letter, string symbols, integers, floats, words list4 = [1,2,3,4.56,"22","22","@@","@@","@@",'###','hello', 'hello',1255,1255.0] #List 5 - integers list5 = [1,1,1,1,3,4,5,6,7,7,7,7,7,7,7,9,9,9,9,9,9,9,9,9] #List 6 - floats list6 = [0.5654,0,5654,.11,1.23,1.23,1.23] #List 7 - Names list7 = ['Conor','Conor','Leo','Leo','Leo','Nancy','Rose','Connie','Aideen','Aideen'] #List 8 - Joined Lists list8 = list1 + list5 ###Output _____no_output_____ ###Markdown Counts Function ###Code #https://stackoverflow.com/questions/1801668/convert-a-python-list-with-strings-all-to-lowercase-or-uppercase #https://www.python-course.eu/python3_lambda.php #count function version 2 - case sensitivity def counts(l): """ A function to count the items in a list and returns a Dictionary with Keys and Counts """ #var st (string) - lambda function to change all items to a string #required for integer and float values with a list st = list(map(lambda x: str(x),l)) #var cs (case senstitive) - Lambda function to change all items to lowercase cs = list(map(lambda x: x.lower(),st)) #variable to create an empty dictionary frequency = {} #a for loop to loop through list items for list_item in cs: #loop through list and add items to dict with a count #list item dictionay key...........l.count dictionary values #.title to capitalize the Key value of dictionary frequency[list_item.title()] = cs.count(list_item) #print output return frequency ###Output _____no_output_____ ###Markdown Function testTest the function with different types of lists ###Code #Test1 print("Test 1: Uppercase Letters") print(counts(list1)) print(" ") #Test2 print("Test 2: Lowercase Letters") print(counts(list2)) print(" ") #Test3 print("Test 3: mix of lower and upper case:") print(counts(list3)) print(" ") #Test4 print("Test 4: mix of upper case leters, lower case letter, string symbols, integers, floats, words") print(counts(list4)) print(" ") #Test5 print("Test 5:integers ") print(counts(list5)) print(" ") #Test6 print("Test 6: floating numbers") print(counts(list6)) print("") #Test7 print("Test 7: words, names etc.") print(counts(list7)) print(" ") print('Test 8: Joined Lists') #Test8 print(counts(list8)) ###Output Test 1: Uppercase Letters {'A': 3, 'B': 1, 'C': 1} Test 2: Lowercase Letters {'A': 3, 'B': 1, 'C': 1} Test 3: mix of lower and upper case: {'A': 3, 'B': 1, 'C': 2} Test 4: mix of upper case leters, lower case letter, string symbols, integers, floats, words {'1': 1, '2': 1, '3': 1, '4.56': 1, '22': 2, '@@': 3, '###': 1, 'Hello': 2, '1255': 1, '1255.0': 1} Test 5:integers {'1': 4, '3': 1, '4': 1, '5': 1, '6': 1, '7': 7, '9': 9} Test 6: floating numbers {'0.5654': 1, '0': 1, '5654': 1, '0.11': 1, '1.23': 3} Test 7: words, names etc. {'Conor': 2, 'Leo': 3, 'Nancy': 1, 'Rose': 1, 'Connie': 1, 'Aideen': 2} Test 8: Joined Lists {'A': 3, 'B': 1, 'C': 1, '1': 4, '3': 1, '4': 1, '5': 1, '6': 1, '7': 7, '9': 9} ###Markdown *** Assignment 2 Counts Write a Python function called dicerolls that simulatesrolling dice. Your function should take two parameters: the number of dice k andthe number of times to roll the dice n. The function should simulate randomlyrolling k dice n times, keeping track of each total face value. It should then returna dictionary with the number of times each possible total face value occurred. So,calling the function as diceroll(k=2, n=1000) should return a dictionary like: {'2': 19, '3': 50, '4': 82} References Assignment 2[6] Numpy documentation, numpy.random.generator; https://numpy.org/doc/stable/reference/random/generator.htmlnumpy.random.Generator https://numpy.org/doc/stable/reference/random/generated/numpy.random.Generator.integers.htmlnumpy.random.Generator.integers[7]Numpy documentation,numpy.random.integers; https://numpy.org/doc/stable/reference/random/generated/numpy.random.Generator.integers.htmlnumpy.random.Generator.integers[8] Stackoverflow, dice simulation; https://stackoverflow.com/questions/33069476/simulating-rolling-2-dice-in-python[9] Stackoverflow, sorting dictionaries in python; https://stackoverflow.com/questions/20577840/python-dictionary-sorting-in-descending-order-based-on-values[10] Stackoverflow, plotting with dictionary data; https://stackoverflow.com/questions/53431971/plotting-histogram-on-python-with-dictionary[11] W3schools, nested loops; https://www.w3schools.com/python/gloss_python_for_nested.asp[12] Quoara, finding possible outcomes of each dice; https://www.quora.com/If-4-dice-are-rolled-what-is-the-probability-of-getting-a-sum-of-5 Diceroll function ###Code import numpy as np import matplotlib.pyplot as plt #import operator rng = np.random.default_rng() #initial blank function def diceroll(k,n): """ A function to simulate 2 dice being thrown 1000 time and out the reults of the total face value in a dictionary """ #variable to create an empty dictionary resultdict = {} #list for dice numbers rolled result = [] #dice counter for number of dice dice = 0 #for loop through number of throws for i in range(n): #for loop for number of dice used for i in range(k): #random number from numy integer function, += used in loop #integer function return random integer including 1 and 6. endpoint = True makes 6 inclusive dice += rng.integers(1,6,endpoint=True) #apped random intger number result.append(dice) #create dictionary key and values resultdict[dice] = result.count(dice) #reset of count of dice to loop through again dice = 0 #s = dict(sorted(resultdict.items(),reverse=False)) #return a dictionary with keys sorted in ascending value return dict(resultdict.items())#,reverse=False)) sorted ###Output _____no_output_____ ###Markdown Plotting the Function ###Code #variables for each simulation # one dice a thousand time rollone = diceroll(1,1000) #two dice a thousand times rolltwo = diceroll(2,1000) #three dice a thiusan times rollthree = diceroll(3,1000) #four dice a thousand times rollfour = diceroll(4,1000) # extracting dictionary key and value for plotting labels1, values1 = zip(*rollone.items()) labels2, values2 = zip(*rolltwo.items()) labels3, values3 = zip(*rollthree.items()) labels4, values4 = zip(*rollfour.items()) ###Output _____no_output_____ ###Markdown Bar charts to illustrate the use of different number of dice used in the function ###Code #for plot display sizing # ref https://stackoverflow.com/questions/36367986/how-to-make-inline-plots-in-jupyter-notebook-larger plt.rcParams['figure.figsize'] = [18, 14] # plot style #ref Dr Ian Mcloughlin lectures plt.style.use('ggplot') #sub[;ot title] plt.suptitle("Plots of the different number of Dice Rolled",fontsize=24 ) #subplots 2 rows, 2 columns plt.subplot(2,2,1) #barplot plt.bar(labels1,values1,color='blue',alpha=0.5) #title plt.title("Plot 1 1: One Dice Rolled 1000 Times") #ylabel plt.ylabel('Sum of Values') #xlabel plt.xlabel('Dice Totals') plt.subplot(2,2,2) #barplot plt.bar(labels2,values2,color='blue',alpha=0.5) #title plt.title("Plot 2: Two Dice Rolled 1000 Times") #ylabel plt.ylabel('Sum of Values') #xlabel plt.xlabel('Dice Totals') plt.subplot(2,2,3) #barplot plt.bar(labels3,values3,color='blue',alpha=0.5) #title plt.title("Plot 3: Three Dice Rolled 1000 Times") #ylabel plt.ylabel('Sum of Values') #xlabel plt.xlabel('Dice Totals') plt.subplot(2,2,4) #barplot plt.bar(labels4,values4,color='blue',alpha=0.5) #title plt.title("Plot 4: Four Dice Rolled 1000 Times") #ylabel plt.ylabel('Sum of Values') #xlabel plt.xlabel('Dice Totals'); ###Output _____no_output_____ ###Markdown Possible out outcomes depending on number of dice used We can observe from the above plot how the number of dice used changes the distribuiton.For one dice used there are 6 possible outcomes, for two there are 36, for 4 there are 1296!!!If we take rolling a total value of 2, in plot 1 there is a 1 in 6 chance of getting two. Using two dice this is 1 in 36. Using more than two dice you can't get a value of two.The more dice you use, there are more possible combinations to get different totals. For two dice, there are more combinations of rolling a total of 7 than any other total value.1+6, 2+5, 3+4, 4+3, 5+2, 6+1 are all the different combinations that produce a total value of 7. We can see this clearly illustrated in plot 2.In the code below we can see the different possible outcomes for each total depending on how many dice are thrown. The number of outcomes determine the % probability.$P(E) = \frac{n(E)}{n(S)}$For example: For Two Dice the probabilty of rolling a total of 7* n(E) = 6* n(S) = 36 (6**2)* P(E) = 6/36 which is approx *17%* We can create a dictionary using our Counts function to see the different number of possible outcomes for creating different totals depending on how many dice are used. ###Code #creating a dictionary using our Counts function to see the different number of possible outcomes for creating different totals #depending on how many dice are used #empty lists to append each dice number range onedicepo = [] twodicepo = [] threedicepo = [] fourdicepo = [] #nested for loops for 4 different dice for d1 in range(1,7,1): #append first range in the first list onedicepo.append(d1) for d2 in range(1,7,1): #variable combines two dice together twodice = d1 + d2 twodicepo.append(twodice) for d3 in range(1,7,1): threedice = d1+d2+d3 threedicepo.append(threedice) for d4 in range(1,7,1): fourdice = d1+d2+d3+d4 fourdicepo.append(fourdice) #Utilising the counts function we created from Assingment 1 print("Possible Outcomes of Totals for 1 Dice") print(counts(onedicepo)) print(" ") print("Possible Outcomes of Totals for 2 Dice") print(counts(twodicepo)) print(" ") print("Possible Outcomes of Totals for 3 Dice") print(counts(threedicepo)) print(" ") print("Possible Outcomes of Totals for 4 Dice") print(counts(fourdicepo)) # extracting dictionary key and value for plotting labels_five, values_five = zip(*counts(twodicepo).items()) #subplots title plt.suptitle("Comparison of Two Dice Rolled 1000 Times V Possible Outcomes",fontsize=18 ) plt.subplot(2,1,1) #barplot color blue, transparency .50 plt.bar(labels2,values2,color='b',alpha=0.5) #title plt.title("Two Dice Rolled 1000 Times") #ylabel plt.ylabel('Sum of Values') #xlabel plt.xlabel('Dice Totals') plt.subplot(2,1,2) #barplot, color green, transparency .50 plt.bar(labels_five,values_five,color='green',alpha=0.5) #title plt.title("Possible Outcomes of Two Dice Totals") #ylabel plt.ylabel('Possible Outcomes') #xlabel plt.xlabel('Dice Totals'); ###Output _____no_output_____ ###Markdown We can observe from the above plots for two dice rolled 1000 times and the possible outcome totals for two dice that their distribution is very similar!*** Assignment 3 Coin Flip Write some python code thatsimulates flipping a coin 100 times. Then run this code 1,000 times, keeping trackof the number of heads in each of the 1,000 simulations. Select an appropriateplot to depict the resulting list of 1,000 numbers, showing that it roughly followsa bell-shaped curve. References Assignment 3 References Assignment 3[13] Numpy random choice; https://www.sharpsightlabs.com/blog/numpy-random-choice/ [14] Numpy Documentation, Simple random data, numpy.choice; https://numpy.org/doc/stable/reference/random/generated/numpy.random.Generator.choice.htmlnumpy.random.Generator.choice[15] Numpy Documentation, Distributions, binomial; https://numpy.org/doc/stable/reference/random/generated/numpy.random.Generator.binomial.htmlnumpy.random.Generator.binomial[16] Seaborn, Distplots; https://seaborn.pydata.org/generated/seaborn.distplot.html[17] Seaborn, visualising distributions; https://seaborn.pydata.org/tutorial/distributions.html Task 3 - Coin FlipFor this task we will use numpy binomial distribution function as part of the numpy random sub-package.A binomial distribution can be thought of as simply the probability of a SUCCESS or FAILURE outcome in an experiment or survey that is repeated multiple times. The binomial is a type of distribution that has two possible outcomes (the prefix “bi” means two, or twice)[12](r12).The probability density for the Binomial Distribution is:$ P(N) = \binom{n}{N}p^N(1-p)^{n-N} $For example if a fair coin is flipped once, the results must be either Heads or tails (True or False). If one dice is rolled once the results has to be either 1, 2, 3 ,4 ,5 or 6. The probability of one number is 1/6If a new medicine is introduced to cure a disease, it either works or not.The Binominal Distribution is a discrete version of the Normal Distribution. The more trials used in the Binominal Distribution the more in becomes a Normal Distribution.The general prerequisites of binomial distribution are:* The number of observations *n* is fixed.* Two potential outcomes, it either happens (Success) or it doesn't (Failure)* Each observation is independent from the next.* Probability of *success* p is the same of each outcome Binominal function*binomial(n, p, size=None)**Draw samples from a binomial distribution* [13](r13) Parameters of the function* n: must be greater or equal than 0.* p: float value between 0 - 1 inclusive* size: test size of running the sample* will return an array the size of the input for the *size* parameter Flipp a coin 100 Times ###Code #import seaborn package for plotting coin toss results import seaborn as sns #p value in function parameter success = .5 #size is array size size = 1000 #different variables for the amount of flips observed ten_flips = rng.binomial(10,success,size) one_thousand_flips = rng.binomial(1000,success,size) ten_thousand_flips = rng.binomial(10000,success,size) one_hundred_flips = rng.binomial(100,success,size) one_thousand_flips = rng.binomial(1000,success,size) ten_thousand_flips = rng.binomial(10000,success,size) print(counts(one_thousand_flips)) #title of subplots plt.suptitle('Coin Flip Comparision',fontsize=22) plt.subplot(2,2,1) #dist plot with KDE set as False sns.distplot(ten_flips,kde=False,label= f"100 Flips $p$={success}",color='blue') #legend plt.legend() #title plt.title('100 Coin Flips',Fontsize=18) plt.subplot(2,2,2) #distplot with KDE set as False sns.distplot(one_hundred_flips,kde=False,label= f"100 Flips $p$={success}",color='orange') #legend plt.legend() #title plt.title('100 Coin Flips',Fontsize=18) plt.subplot(2,2,3) #distplot wih KDE set as False sns.distplot(one_thousand_flips,kde=False,label= f"1000 Flips $p$={success}",color='green' ) #legend plt.legend() #title plt.title(f'1000 Coin Flips',Fontsize=18) plt.subplot(2,2,4) #distplot with KDE set as False sns.distplot(ten_thousand_flips,kde=False,label= f"10000 Flips$p$={success}",color='red' ) #legend plt.legend() #title plt.title(f'10000 Coin Flips',Fontsize=18); #plot title plt.title('Coin Flip Density Comparision',fontsize=16) #Kernel Density (kde) plots for 10, 100 and 1000 flips sns.distplot(ten_flips,hist=False,kde_kws = {'shade': True, 'linewidth': 3},label= f"10 Flips $p$={success}") sns.distplot(one_hundred_flips,hist=False,kde_kws = {'shade': True, 'linewidth': 3},label= f"100 Flips $p$={success}") sns.distplot(one_thousand_flips,hist=False,kde_kws = {'shade': True, 'linewidth': 3},label= f"1000 Flips $p$={success}",color='green'); ###Output _____no_output_____ ###Markdown As we can observe from the above plots, the more times the coin is flipped the more that data becomes normally distributed. The chances of a small number of heads or a large number of heads is very small, The chances of getting an average amount of heads is in the middle.*** Assignment 4 Simpson's Paradox Simpson’s paradox is a well-known statistical paradoxwhere a trend evident in a number of groups reverses when the groups are combinedinto one big data set. Use numpy to create four data sets, each with an x arrayand a corresponding y array, to demonstrate Simpson’s paradox. You mightcreate your x arrays using numpy.linspace and create the y array for eachx using notation like y = a * x + b where you choose the a and b for eachx , y pair to demonstrate the paradox. References Assignment 4[18] Bittannica.com, Simpson's paradox; https://www.britannica.com/topic/Simpsons-paradox[19] Brilliant.org, simpson's paradox; https://brilliant.org/wiki/simpsons-paradox/[20] Towardsdatascience.com, How to prove two opposite arguments using one dataset; https://towardsdatascience.com/simpsons-paradox-how-to-prove-two-opposite-arguments-using-one-dataset-1c9c917f5ff9[21] Seaborn, lmplot; https://seaborn.pydata.org/generated/seaborn.lmplot.html[22] Seaborn, relplot, https://seaborn.pydata.org/generated/seaborn.relplot.html [23] Seaborn, relplot, https://seaborn.pydata.org/generated/seaborn.relplot.html[24] Youtube, minutephysics, Simpson's Paradox; https://www.youtube.com/watch?v=ebEkn-BiW5k[25] Youtube, Dr. Trefor Bazett,How SIMPSON'S PARADOX explains weird COVID19 statistics; https://www.youtube.com/watch?v=t-Ci3FosqZs[26] numpy, polyfit function; https://numpy.org/doc/stable/reference/generated/numpy.polyfit.html Task 4 Simpson's Paradox*Simpson's paradox (or Yule-Simpson effect ), in statistics, is a phenomenon where one particular trend shown in groups of data is reversed when the groups are combined together. In order to interpret data, correctly understanding and identifying this paradox is of critical importance.* [19]Four different datsets will be created to illustrate the paradox. These for datasets will be then combined.Each dataset will effected by the same random noise, the slope increase will be the same for each group, but the linspace values and interept values will differ.For each dataset, numpy's *polyfit* function will be aplied to illustrate the *best fit* line for each group of data.This will then be al plotted to visually illusrate Simpson's paradox. The slope of each dataset will also be printed to show that thecombined dataset will have a negative slope. Creating 4 variables ###Code #noise variable using normal distribution function #mean of 10 #stdev of 15 for more noise noise = rng.normal(10,15,100) #numbers from 0 to 5 slit up into 100 segements x1 = np.linspace(0,5,100) #slope is x increased by 10 for every point with an intercept of 100 plus normally distributed noise y1 = 10* x1 + 100+noise #ordinary least square polyfit on the dataset coeffs1 = np.polyfit(x1,y1,1) #numbers from 5 to 10 slit up into 100 segements x2 = np.linspace(5,10,100) #slope is x increased by 10 for every point with an intercept of 75 plus normally distributed noise y2 = 10* x1 + 75+noise #ordinary least square polyfit on the dataset coeffs2 = np.polyfit(x2,y2,1) #numbers from 10 to 15 slit up into 100 segements x3 = np.linspace(10,15,100)#15,20 #slope is x increased by 10 for every point with an intercept of 50 plus normally distributed noise y3 = 10* x1 + 50+noise #ordinary least square polyfit on the dataset coeffs3 = np.polyfit(x3,y3,1) #numbers from 15 to 20 slit up into 100 segements x4 = np.linspace(15,20,100) #slope is x increased by 10 for every point with an intercept of 25 plus normally distributed noise y4 = 10* x1 + 25+noise #ordinary least square polyfit on the dataset coeffs4 = np.polyfit(x4,y4,1) ###Output _____no_output_____ ###Markdown Combing the datasetsThe four different datsets from above are concatenated together to make one overall dataset. The *best fit* line will be created using numpy's polyfit function and we will plot the cobined dataset and the *best fit* line. ###Code #combines varaible #numpy concatenate function x= np.concatenate([x1,x2,x3,x4]) y = np.concatenate([y1,y2,y3,y4]) #coefficient of combined data combine_coeffs = np.polyfit(x,y,1) print('The slope of the cobined dataet is:',combine_coeffs[0]) ###Output The slope of the cobined dataet is: -4.032717380112218 ###Markdown Plotting Combined DatasetThe below plot we can see that the slope is negative an is displaying a negative *best fit* line. ###Code #plotting scattered data plt.plot(x,y,'.',color='black',label='Combined Dataset') #plotting coeffeicent values plt.plot(x,combine_coeffs[0] *x + combine_coeffs[1],'--',color='black',linewidth=4, label = 'Best Fit line: Combined Dataset') plt.title("Combined Dataset",fontsize=18) plt.legend(); ###Output _____no_output_____ ###Markdown However if we show the datasets indiviually we can ilustrate Simpson's paradox. Hence, Simson's paradox occurs when the trend in a group of data reverses when the groups are combined.In the plot below, each group has a different trend to the the combined dataset. ###Code #plotting scattered data plt.plot(x1,y1,'.',label='Dataset 1',color='b') plt.plot(x2,y2,'.',label='Dataset 2',color='orange') plt.plot(x3,y3,'.',label='Dataset 3',color='green') plt.plot(x4,y4,'.',label='Dataset 4',color='red') #plotting coeffeicent values plt.plot(x1,coeffs1[0] *x1 + coeffs1[1],color='blue',label = 'Best Fit line: Dataset 1') plt.plot(x2,coeffs2[0] *x2 + coeffs2[1],color='orange',label = 'Best Fit line: Dataset 2') plt.plot(x3,coeffs3[0] *x3 + coeffs3[1],color='green',label = 'Best Fit line: Dataset 3') plt.plot(x4,coeffs4[0] *x4 + coeffs4[1],color='red', label = 'Best Fit line: Dataset 4') plt.plot(x,combine_coeffs[0] *x + combine_coeffs[1],'--',color='black',linewidth=4, label = 'Best Fit line: Combined Dataset') #title plt.title("Example of Simpson's Paradox",fontsize=18) #legend plt.legend(); ###Output _____no_output_____ ###Markdown The slope for each group of data is positive but when combined becomes negative. This also shows the presence of Simpson's paradox. ###Code print('Slope for each variable') print('Slope Dataset 1:',coeffs1[0]) print('Slope Dataset 2:',coeffs2[0]) print('Slope Dataset 3:',coeffs3[0]) print('Slope Dataset 4:',coeffs4[0]) print('Slope Combined Dataset:',combine_coeffs[0]) ###Output Slope for each variable Slope Dataset 1: 10.189210248930946 Slope Dataset 2: 10.189210248930943 Slope Dataset 3: 10.18921024893094 Slope Dataset 4: 10.189210248930946 Slope Combined Dataset: -4.032717380112218
turkiye-student-evaluation_generic/Turkiye-Student-Evaluation-Data-Set.ipynb
###Markdown Turkiye-Student-Evaluation-Data-Set Problem Statement:- In this project we are basically going to perform clustering on the given data and these clusters will signify different categories of students based on the marks, content, course and other features. The clustering algorithm used in the project is K-Means Clustering. The aim is to cluster the data on the basis of the given features, which will ultimately cluster together the student's with similar performance. Attribute Information:-The dataset has 5820 instances with 33 attributes. The description of each column is given below:instr: Instructor's identifier; values taken from {1,2,3}class: Course code (descriptor); values taken from {1-13}repeat: Number of times the student is taking this course; values taken from {0,1,2,3,...}attendance: Code of the level of attendance; values from {0, 1, 2, 3, 4}difficulty: Level of difficulty of the course as perceived by the student; values taken from {1,2,3,4,5}Q1: The semester course content, teaching method and evaluation system were provided at the start.Q2: The course aims and objectives were clearly stated at the beginning of the period.Q3: The course was worth the amount of credit assigned to it.Q4: The course was taught according to the syllabus announced on the first day of class.Q5: The class discussions, homework assignments, applications and studies were satisfactory.Q6: The textbook and other courses resources were sufficient and up to date.Q7: The course allowed field work, applications, laboratory, discussion and other studies.Q8: The quizzes, assignments, projects and exams contributed to helping the learning.Q9: I greatly enjoyed the class and was eager to actively participate during the lectures.Q10: My initial expectations about the course were met at the end of the period or year.Q11: The course was relevant and beneficial to my professional development.Q12: The course helped me look at life and the world with a new perspective.Q13: The Instructor's knowledge was relevant and up to date.Q14: The Instructor came prepared for classes.Q15: The Instructor taught in accordance with the announced lesson plan.Q16: The Instructor was committed to the course and was understandable.Q17: The Instructor arrived on time for classes.Q18: The Instructor has a smooth and easy to follow delivery/speech.Q19: The Instructor made effective use of class hours.Q20: The Instructor explained the course and was eager to be helpful to students.Q21: The Instructor demonstrated a positive approach to students.Q22: The Instructor was open and respectful of the views of students about the course.Q23: The Instructor encouraged participation in the course.Q24: The Instructor gave relevant homework assignments/projects, and helped/guided students.Q25: The Instructor responded to questions about the course inside and outside of the course.Q26: The Instructor's evaluation system (midterm and final questions, projects, assignments, etc.) effectively measured the course objectives.Q27: The Instructor provided solutions to exams and discussed them with students.Q28: The Instructor treated all students in a right and objective manner.Q1-Q28 are all Likert-type, meaning that the values are taken from {1,2,3,4,5} We will be implementing the following steps to achieve the final result:-1. Importing the necessary Libraries.2. Importing the dataset.3. Exploratory Data Analysis4. Performing feature engineering i.e. modifying existing variables and creating new ones for analysis.5. Building The model.6. Visualising the results. Step-1: Importing the Libraries ###Code # import numpy as np for processing data # import pandas as pd for importing the data and working with data # import matplotlib.pyplot as plt for visualisation # import seaborn as sns for data visualization # from sklearn.preprocessing import StandardScaler for data preprocessing # from sklearn.preprocessing import normalize for normalizing your data # from sklearn.decomposition import PCA for dimensionality reduction # from sklearn.cluster import KMeans for implementing the clustering algorithm ###Output _____no_output_____ ###Markdown Step-2: Importing the dataset ###Code # Use pd.read_csv('filename.csv') to import the necessary file ###Output _____no_output_____ ###Markdown Seeing the dataset: ###Code # using data.head() we can see the dataset ###Output _____no_output_____ ###Markdown Step-3: Exploratory Data AnalysisIt is common for data scientists to spend a majority of their time exploring and cleaning data, but approaching this as an opportunity to invest in your model (instead of viewing it as just another chore on your to-do list) will yield big dividends later on in the data science process.Performing thorough exploratory data analysis (EDA) and cleaning the dataset are not only essential steps, but also a great opportunity to lay the foundation for a strong machine learning model. Seeing the shape and size of the dataset ###Code # Using df.shape() and df.size() will give you the shape and size of the dataset ###Output _____no_output_____ ###Markdown Describing the dataset ###Code # Using df.describe() will describe the dataset ###Output _____no_output_____ ###Markdown Seeing the non null values ###Code # Using df.info() we can see number of non null values present ###Output _____no_output_____ ###Markdown Seeing the list of columns ###Code # Using df.columns we can see the list of all the columns in the dataset ###Output _____no_output_____ ###Markdown Seeing the null values in the dataset ###Code # Using df.isnull().sum will give you the null values ###Output _____no_output_____ ###Markdown Visualising the data using seaborn libraryThere are different types of plots like bar plot, box plot, scatter plot etc.Scatter plot is very useful when we are analyzing the relation ship between 2 features on x and y axis.In seaborn library we have pairplot function which is very useful to scatter plot all the features at once instead of plotting them individually. ###Code # Using sns.pairplot(df) we can visualize relationship between the features ###Output _____no_output_____ ###Markdown Visualizing the Heat Map ###Code # Using sns.heatmap(dataset.corr(),annot=True) we plot the heat map ###Output _____no_output_____ ###Markdown Step-4: Performing Feature EngineeringWe will be performing the following 3 steps:1.Standard Scaler2.Normalization3.Principal Component Analysis Standard Scaler:Standardization of a dataset is a common requirement for many machine learning estimators: they might behave badly if the individual features do not more or less look like standard normally distributed data (e.g. Gaussian with 0 mean and unit variance).For instance many elements used in the objective function of a learning algorithm (such as the RBF kernel of Support Vector Machines or the L1 and L2 regularizers of linear models) assume that all features are centered around 0 and have variance in the same order. If a feature has a variance that is orders of magnitude larger that others, it might dominate the objective function and make the estimator unable to learn from other features correctly as expectedStandardize features by removing the mean and scaling to unit variance.The standard score of a sample x is calculated as:z = (x - u) / swhere u is the mean of the training samples, and s is the standard deviation of the training samples. ###Code # to implement the standard scaler first create an object for StandardScaler # now perform scaler.fit() to fit the data # now perform scaler.transform() to get the scaled data ###Output _____no_output_____ ###Markdown Normalization:Normalization is used to scale the data of an attribute so that it falls in a smaller range, such as -1.0 to 1.0 or 0.0 to 1.0. It is generally useful for classification algorithms.Normalization is generally required when we are dealing with attributes on a different scale, otherwise, it may lead to a dilution in effectiveness of an important equally important attribute(on lower scale) because of other attribute having values on larger scale.In simple words, when multiple attributes are there but attributes have values on different scales, this may lead to poor data models while performing data mining operations. So they are normalized to bring all the attributes on the same scale. ###Code # now implement the normalization to normalize the data ###Output _____no_output_____ ###Markdown Principal Component AnalysisPrincipal Components Analysis is an unsupervised learning class of statistical techniques used to explain data in high dimension using smaller number of variables called the principal components.Assuming we have a set X made up of n measurements each represented by a set of p features, X1, X2, … , Xp. If we want to plot this data in a 2-dimensional plane, we can plot n measurements using two features at a time. If the number of features are more than 3 or four then plotting this in two dimension will be a challenge as the number of plots would be p(p-1)/2 which would be hard to plot.We would like to visualize this data in two dimension without losing information contained in the data. This is what PCA allows us to do. ###Code # so to implement PCA first we need to create an object for PCA and also need to mention that how many dimensions we need finally # now do pca.fit(data) to fit the data # now do pca.transform(data) to transform the higher-dimensionality data to lower dimensions # now after implementing pca, use pd.DataFrame(data) to convert the new data into a Data Frame else ###Output _____no_output_____ ###Markdown Step-5: Building The modelThe algorithm works as follows:1. First we initialize k points, called means, randomly.2. We categorize each item to its closest mean and we update the mean’s coordinates, which are the averages of the items categorized in that mean so far.3. We repeat the process for a given number of iterations and at the end, we have our clusters.So, basically we will be following two steps:-1. Implementing the Elbow method which we will return the optimal value of clusters to be formed.2. We will implement K-Means algorithm to create the clusters. Implementing Elbow MethodIn cluster analysis, the elbow method is a heuristic used in determining the number of clusters in a data set. The method consists of plotting the explained variation as a function of the number of clusters, and picking the elbow of the curve as the number of clusters to use. The same method can be used to choose the number of parameters in other data-driven models, such as the number of principal components to describe a data set.A fundamental step for any unsupervised algorithm is to determine the optimal number of clusters into which the data may be clustered. The Elbow Method is one of the most popular methods to determine this optimal value of k.To determine the optimal number of clusters, we have to select the value of k at the “elbow” ie the point after which the distortion/inertia start decreasing in a linear fashion. ###Code # to implement the elbow method first create an empty list name it as wcss # now initiate a for loop ranging between (1,11) and implement k-means clustering for every i number of clusters # keep appending the empty list with the kmeans.inertia_ values # now plot the graph between the range(1,11) and the wcss # from the plot we determine the optimal value of k ###Output _____no_output_____ ###Markdown Implementing K-Means:K-means clustering is a method of vector quantization, originally from signal processing, that aims to partition n observations into k clusters in which each observation belongs to the cluster with the nearest mean (cluster centers or cluster centroid), serving as a prototype of the cluster. This results in a partitioning of the data space into Voronoi cells. It is popular for cluster analysis in data mining. k-means clustering minimizes within-cluster variances (squared Euclidean distances), but not regular Euclidean distances, which would be the more difficult Weber problem: the mean optimizes squared errors, whereas only the geometric median minimizes Euclidean distances. For instance, Better Euclidean solutions can be found using k-medians and k-medoids.The above algorithm in pseudocode: Initialize k means with random valuesFor a given number of iterations: Iterate through itemsUse kmeans with different number of clustersAppend the list with kmeans.inertia_ values ###Code # now with the optimal k value implement the K-Means Clustering Algorithm # now using kmeans.fit_predict to predict that which data belong to which cluster ###Output _____no_output_____ ###Markdown Step-6: Visualising the result ###Code # to visualise the final result use plt.scatter() with respective arguments to view the created clusters # also plot the centroids of the respective clusters using kmeans.cluster_centers_ # finally you will be able to look at the result using plt.show() ###Output _____no_output_____
Classification/K_nearest_neighbors.ipynb
###Markdown K-Nearest Neighbors (K-NN) Importing the libraries ###Code import numpy as np import matplotlib.pyplot as plt import pandas as pd ###Output _____no_output_____ ###Markdown Importing the dataset ###Code dataset = pd.read_csv('Social_Network_Ads.csv') x = dataset.iloc[:, :-1].values y = dataset.iloc[:, -1].values ###Output _____no_output_____ ###Markdown Splitting the dataset into the Training set and Test set ###Code from sklearn.model_selection import train_test_split x_train, x_test, y_train, y_test = train_test_split(x, y, test_size = 0.25, random_state =0) print(x_test) ###Output [[ 30 87000] [ 38 50000] [ 35 75000] [ 30 79000] [ 35 50000] [ 27 20000] [ 31 15000] [ 36 144000] [ 18 68000] [ 47 43000] [ 30 49000] [ 28 55000] [ 37 55000] [ 39 77000] [ 20 86000] [ 32 117000] [ 37 77000] [ 19 85000] [ 55 130000] [ 35 22000] [ 35 47000] [ 47 144000] [ 41 51000] [ 47 105000] [ 23 28000] [ 49 141000] [ 28 87000] [ 29 80000] [ 37 62000] [ 32 86000] [ 21 88000] [ 37 79000] [ 57 60000] [ 37 53000] [ 24 58000] [ 18 52000] [ 22 81000] [ 34 43000] [ 31 34000] [ 49 36000] [ 27 88000] [ 41 52000] [ 27 84000] [ 35 20000] [ 43 112000] [ 27 58000] [ 37 80000] [ 52 90000] [ 26 30000] [ 49 86000] [ 57 122000] [ 34 25000] [ 35 57000] [ 34 115000] [ 59 88000] [ 45 32000] [ 29 83000] [ 26 80000] [ 49 28000] [ 23 20000] [ 32 18000] [ 60 42000] [ 19 76000] [ 36 99000] [ 19 26000] [ 60 83000] [ 24 89000] [ 27 58000] [ 40 47000] [ 42 70000] [ 32 150000] [ 35 77000] [ 22 63000] [ 45 22000] [ 27 89000] [ 18 82000] [ 42 79000] [ 40 60000] [ 53 34000] [ 47 107000] [ 58 144000] [ 59 83000] [ 24 55000] [ 26 35000] [ 58 38000] [ 42 80000] [ 40 75000] [ 59 130000] [ 46 41000] [ 41 60000] [ 42 64000] [ 37 146000] [ 23 48000] [ 25 33000] [ 24 84000] [ 27 96000] [ 23 63000] [ 48 33000] [ 48 90000] [ 42 104000]] ###Markdown Feature Scaling ###Code from sklearn.preprocessing import StandardScaler sc = StandardScaler() x_train = sc.fit_transform(x_train) x_test = sc.transform(x_test) print(x_train) ###Output [[ 0.58164944 -0.88670699] [-0.60673761 1.46173768] [-0.01254409 -0.5677824 ] [-0.60673761 1.89663484] [ 1.37390747 -1.40858358] [ 1.47293972 0.99784738] [ 0.08648817 -0.79972756] [-0.01254409 -0.24885782] [-0.21060859 -0.5677824 ] [-0.21060859 -0.19087153] [-0.30964085 -1.29261101] [-0.30964085 -0.5677824 ] [ 0.38358493 0.09905991] [ 0.8787462 -0.59677555] [ 2.06713324 -1.17663843] [ 1.07681071 -0.13288524] [ 0.68068169 1.78066227] [-0.70576986 0.56295021] [ 0.77971394 0.35999821] [ 0.8787462 -0.53878926] [-1.20093113 -1.58254245] [ 2.1661655 0.93986109] [-0.01254409 1.22979253] [ 0.18552042 1.08482681] [ 0.38358493 -0.48080297] [-0.30964085 -0.30684411] [ 0.97777845 -0.8287207 ] [ 0.97777845 1.8676417 ] [-0.01254409 1.25878567] [-0.90383437 2.27354572] [-1.20093113 -1.58254245] [ 2.1661655 -0.79972756] [-1.39899564 -1.46656987] [ 0.38358493 2.30253886] [ 0.77971394 0.76590222] [-1.00286662 -0.30684411] [ 0.08648817 0.76590222] [-1.00286662 0.56295021] [ 0.28455268 0.07006676] [ 0.68068169 -1.26361786] [-0.50770535 -0.01691267] [-1.79512465 0.35999821] [-0.70576986 0.12805305] [ 0.38358493 0.30201192] [-0.30964085 0.07006676] [-0.50770535 2.30253886] [ 0.18552042 0.04107362] [ 1.27487521 2.21555943] [ 0.77971394 0.27301877] [-0.30964085 0.1570462 ] [-0.01254409 -0.53878926] [-0.21060859 0.1570462 ] [-0.11157634 0.24402563] [-0.01254409 -0.24885782] [ 2.1661655 1.11381995] [-1.79512465 0.35999821] [ 1.86906873 0.12805305] [ 0.38358493 -0.13288524] [-1.20093113 0.30201192] [ 0.77971394 1.37475825] [-0.30964085 -0.24885782] [-1.6960924 -0.04590581] [-1.00286662 -0.74174127] [ 0.28455268 0.50496393] [-0.11157634 -1.06066585] [-1.10189888 0.59194336] [ 0.08648817 -0.79972756] [-1.00286662 1.54871711] [-0.70576986 1.40375139] [-1.29996338 0.50496393] [-0.30964085 0.04107362] [-0.11157634 0.01208048] [-0.30964085 -0.88670699] [ 0.8787462 -1.3505973 ] [-0.30964085 2.24455257] [ 0.97777845 1.98361427] [-1.20093113 0.47597078] [-1.29996338 0.27301877] [ 1.37390747 1.98361427] [ 1.27487521 -1.3505973 ] [-0.30964085 -0.27785096] [-0.50770535 1.25878567] [-0.80480212 1.08482681] [ 0.97777845 -1.06066585] [ 0.28455268 0.30201192] [ 0.97777845 0.76590222] [-0.70576986 -1.49556302] [-0.70576986 0.04107362] [ 0.48261718 1.72267598] [ 2.06713324 0.18603934] [-1.99318916 -0.74174127] [-0.21060859 1.40375139] [ 0.38358493 0.59194336] [ 0.8787462 -1.14764529] [-1.20093113 -0.77073441] [ 0.18552042 0.24402563] [ 0.77971394 -0.30684411] [ 2.06713324 -0.79972756] [ 0.77971394 0.12805305] [-0.30964085 0.6209365 ] [-1.00286662 -0.30684411] [ 0.18552042 -0.3648304 ] [ 2.06713324 2.12857999] [ 1.86906873 -1.26361786] [ 1.37390747 -0.91570013] [ 0.8787462 1.25878567] [ 1.47293972 2.12857999] [-0.30964085 -1.23462472] [ 1.96810099 0.91086794] [ 0.68068169 -0.71274813] [-1.49802789 0.35999821] [ 0.77971394 -1.3505973 ] [ 0.38358493 -0.13288524] [-1.00286662 0.41798449] [-0.01254409 -0.30684411] [-1.20093113 0.41798449] [-0.90383437 -1.20563157] [-0.11157634 0.04107362] [-1.59706014 -0.42281668] [ 0.97777845 -1.00267957] [ 1.07681071 -1.20563157] [-0.01254409 -0.13288524] [-1.10189888 -1.52455616] [ 0.77971394 -1.20563157] [ 0.97777845 2.07059371] [-1.20093113 -1.52455616] [-0.30964085 0.79489537] [ 0.08648817 -0.30684411] [-1.39899564 -1.23462472] [-0.60673761 -1.49556302] [ 0.77971394 0.53395707] [-0.30964085 -0.33583725] [ 1.77003648 -0.27785096] [ 0.8787462 -1.03167271] [ 0.18552042 0.07006676] [-0.60673761 0.8818748 ] [-1.89415691 -1.40858358] [-1.29996338 0.59194336] [-0.30964085 0.53395707] [-1.00286662 -1.089659 ] [ 1.17584296 -1.43757673] [ 0.18552042 -0.30684411] [ 1.17584296 -0.74174127] [-0.30964085 0.07006676] [ 0.18552042 2.09958685] [ 0.77971394 -1.089659 ] [ 0.08648817 0.04107362] [-1.79512465 0.12805305] [-0.90383437 0.1570462 ] [-0.70576986 0.18603934] [ 0.8787462 -1.29261101] [ 0.18552042 -0.24885782] [-0.4086731 1.22979253] [-0.01254409 0.30201192] [ 0.38358493 0.1570462 ] [ 0.8787462 -0.65476184] [ 0.08648817 0.1570462 ] [-1.89415691 -1.29261101] [-0.11157634 0.30201192] [-0.21060859 -0.27785096] [ 0.28455268 -0.50979612] [-0.21060859 1.6067034 ] [ 0.97777845 -1.17663843] [-0.21060859 1.63569655] [ 1.27487521 1.8676417 ] [-1.10189888 -0.3648304 ] [-0.01254409 0.04107362] [ 0.08648817 -0.24885782] [-1.59706014 -1.23462472] [-0.50770535 -0.27785096] [ 0.97777845 0.12805305] [ 1.96810099 -1.3505973 ] [ 1.47293972 0.07006676] [-0.60673761 1.37475825] [ 1.57197197 0.01208048] [-0.80480212 0.30201192] [ 1.96810099 0.73690908] [-1.20093113 -0.50979612] [ 0.68068169 0.27301877] [-1.39899564 -0.42281668] [ 0.18552042 0.1570462 ] [-0.50770535 -1.20563157] [ 0.58164944 2.01260742] [-1.59706014 -1.49556302] [-0.50770535 -0.53878926] [ 0.48261718 1.83864855] [-1.39899564 -1.089659 ] [ 0.77971394 -1.37959044] [-0.30964085 -0.42281668] [ 1.57197197 0.99784738] [ 0.97777845 1.43274454] [-0.30964085 -0.48080297] [-0.11157634 2.15757314] [-1.49802789 -0.1038921 ] [-0.11157634 1.95462113] [-0.70576986 -0.33583725] [-0.50770535 -0.8287207 ] [ 0.68068169 -1.37959044] [-0.80480212 -1.58254245] [-1.89415691 -1.46656987] [ 1.07681071 0.12805305] [ 0.08648817 1.51972397] [-0.30964085 0.09905991] [ 0.08648817 0.04107362] [-1.39899564 -1.3505973 ] [ 0.28455268 0.07006676] [-0.90383437 0.38899135] [ 1.57197197 -1.26361786] [-0.30964085 -0.74174127] [-0.11157634 0.1570462 ] [-0.90383437 -0.65476184] [-0.70576986 -0.04590581] [ 0.38358493 -0.45180983] [-0.80480212 1.89663484] [ 1.37390747 1.28777882] [ 1.17584296 -0.97368642] [ 1.77003648 1.83864855] [-0.90383437 -0.24885782] [-0.80480212 0.56295021] [-1.20093113 -1.5535493 ] [-0.50770535 -1.11865214] [ 0.28455268 0.07006676] [-0.21060859 -1.06066585] [ 1.67100423 1.6067034 ] [ 0.97777845 1.78066227] [ 0.28455268 0.04107362] [-0.80480212 -0.21986468] [-0.11157634 0.07006676] [ 0.28455268 -0.19087153] [ 1.96810099 -0.65476184] [-0.80480212 1.3457651 ] [-1.79512465 -0.59677555] [-0.11157634 0.12805305] [ 0.28455268 -0.30684411] [ 1.07681071 0.56295021] [-1.00286662 0.27301877] [ 1.47293972 0.35999821] [ 0.18552042 -0.3648304 ] [ 2.1661655 -1.03167271] [-0.30964085 1.11381995] [-1.6960924 0.07006676] [-0.01254409 0.04107362] [ 0.08648817 1.05583366] [-0.11157634 -0.3648304 ] [-1.20093113 0.07006676] [-0.30964085 -1.3505973 ] [ 1.57197197 1.11381995] [-0.80480212 -1.52455616] [ 0.08648817 1.8676417 ] [-0.90383437 -0.77073441] [-0.50770535 -0.77073441] [-0.30964085 -0.91570013] [ 0.28455268 -0.71274813] [ 0.28455268 0.07006676] [ 0.08648817 1.8676417 ] [-1.10189888 1.95462113] [-1.6960924 -1.5535493 ] [-1.20093113 -1.089659 ] [-0.70576986 -0.1038921 ] [ 0.08648817 0.09905991] [ 0.28455268 0.27301877] [ 0.8787462 -0.5677824 ] [ 0.28455268 -1.14764529] [-0.11157634 0.67892279] [ 2.1661655 -0.68375498] [-1.29996338 -1.37959044] [-1.00286662 -0.94469328] [-0.01254409 -0.42281668] [-0.21060859 -0.45180983] [-1.79512465 -0.97368642] [ 1.77003648 0.99784738] [ 0.18552042 -0.3648304 ] [ 0.38358493 1.11381995] [-1.79512465 -1.3505973 ] [ 0.18552042 -0.13288524] [ 0.8787462 -1.43757673] [-1.99318916 0.47597078] [-0.30964085 0.27301877] [ 1.86906873 -1.06066585] [-0.4086731 0.07006676] [ 1.07681071 -0.88670699] [-1.10189888 -1.11865214] [-1.89415691 0.01208048] [ 0.08648817 0.27301877] [-1.20093113 0.33100506] [-1.29996338 0.30201192] [-1.00286662 0.44697764] [ 1.67100423 -0.88670699] [ 1.17584296 0.53395707] [ 1.07681071 0.53395707] [ 1.37390747 2.331532 ] [-0.30964085 -0.13288524] [ 0.38358493 -0.45180983] [-0.4086731 -0.77073441] [-0.11157634 -0.50979612] [ 0.97777845 -1.14764529] [-0.90383437 -0.77073441] [-0.21060859 -0.50979612] [-1.10189888 -0.45180983] [-1.20093113 1.40375139]] ###Markdown Training the K-NN model on the Training set ###Code from sklearn.neighbors import KNeighborsClassifier classifier = KNeighborsClassifier(n_neighbors = 5, metric = 'minkowski', p =2) classifier.fit(x_train, y_train) ###Output _____no_output_____ ###Markdown Predicting a new result ###Code print(classifier.predict(sc.transform([[30, 87000]]))) ###Output [0] ###Markdown Predicting the Test set results ###Code y_pred = classifier.predict(x_test) print(np.concatenate((y_pred.reshape(len(y_pred), 1), y_test.reshape(len(y_test), 1)), 1)) ###Output [[0 0] [0 0] [0 0] [0 0] [0 0] [0 0] [0 0] [1 1] [0 0] [1 0] [0 0] [0 0] [0 0] [0 0] [0 0] [1 0] [0 0] [0 0] [1 1] [0 0] [0 0] [1 1] [0 0] [1 1] [0 0] [1 1] [0 0] [0 0] [0 0] [0 0] [0 0] [0 1] [1 1] [0 0] [0 0] [0 0] [0 0] [0 0] [0 0] [1 1] [0 0] [0 0] [0 0] [0 0] [1 1] [0 0] [0 0] [1 1] [0 0] [1 1] [1 1] [0 0] [0 0] [1 0] [1 1] [1 1] [0 0] [0 0] [1 1] [0 0] [0 0] [1 1] [0 0] [1 1] [0 0] [1 1] [0 0] [0 0] [0 0] [0 0] [1 1] [0 0] [0 0] [1 1] [0 0] [0 0] [0 0] [0 0] [1 1] [1 1] [1 1] [1 0] [0 0] [0 0] [1 1] [0 1] [0 0] [1 1] [1 1] [0 0] [0 0] [1 1] [0 0] [0 0] [0 0] [0 1] [0 0] [1 1] [1 1] [1 1]] ###Markdown Making the Confusion Matrix ###Code from sklearn.metrics import confusion_matrix, accuracy_score cm = confusion_matrix(y_test, y_pred) print(cm) accuracy_score(y_test, y_pred) ###Output [[64 4] [ 3 29]] ###Markdown Visualising the Training set results ###Code from matplotlib.colors import ListedColormap X_set, y_set = sc.inverse_transform(x_train), y_train X1, X2 = np.meshgrid(np.arange(start = X_set[:, 0].min() - 10, stop = X_set[:, 0].max() + 10, step = 1), np.arange(start = X_set[:, 1].min() - 1000, stop = X_set[:, 1].max() + 1000, step = 1)) plt.contourf(X1, X2, classifier.predict(sc.transform(np.array([X1.ravel(), X2.ravel()]).T)).reshape(X1.shape), alpha = 0.75, cmap = ListedColormap(('red', 'green'))) plt.xlim(X1.min(), X1.max()) plt.ylim(X2.min(), X2.max()) for i, j in enumerate(np.unique(y_set)): plt.scatter(X_set[y_set == j, 0], X_set[y_set == j, 1], c = ListedColormap(('red', 'green'))(i), label = j) plt.title('K-NN (Training set)') plt.xlabel('Age') plt.ylabel('Estimated Salary') plt.legend() plt.show() ###Output *c* argument looks like a single numeric RGB or RGBA sequence, which should be avoided as value-mapping will have precedence in case its length matches with *x* & *y*. Please use the *color* keyword-argument or provide a 2-D array with a single row if you intend to specify the same RGB or RGBA value for all points. *c* argument looks like a single numeric RGB or RGBA sequence, which should be avoided as value-mapping will have precedence in case its length matches with *x* & *y*. Please use the *color* keyword-argument or provide a 2-D array with a single row if you intend to specify the same RGB or RGBA value for all points. ###Markdown Visualising the Test set results ###Code from matplotlib.colors import ListedColormap X_set, y_set = sc.inverse_transform(x_test), y_test X1, X2 = np.meshgrid(np.arange(start = X_set[:, 0].min() - 10, stop = X_set[:, 0].max() + 10, step = 1), np.arange(start = X_set[:, 1].min() - 1000, stop = X_set[:, 1].max() + 1000, step = 1)) plt.contourf(X1, X2, classifier.predict(sc.transform(np.array([X1.ravel(), X2.ravel()]).T)).reshape(X1.shape), alpha = 0.75, cmap = ListedColormap(('red', 'green'))) plt.xlim(X1.min(), X1.max()) plt.ylim(X2.min(), X2.max()) for i, j in enumerate(np.unique(y_set)): plt.scatter(X_set[y_set == j, 0], X_set[y_set == j, 1], c = ListedColormap(('red', 'green'))(i), label = j) plt.title('K-NN (Test set)') plt.xlabel('Age') plt.ylabel('Estimated Salary') plt.legend() plt.show() ###Output *c* argument looks like a single numeric RGB or RGBA sequence, which should be avoided as value-mapping will have precedence in case its length matches with *x* & *y*. Please use the *color* keyword-argument or provide a 2-D array with a single row if you intend to specify the same RGB or RGBA value for all points. *c* argument looks like a single numeric RGB or RGBA sequence, which should be avoided as value-mapping will have precedence in case its length matches with *x* & *y*. Please use the *color* keyword-argument or provide a 2-D array with a single row if you intend to specify the same RGB or RGBA value for all points.
dev/08_vision_core.ipynb
###Markdown Core vision> Basic image opening/processing functionality Helpers ###Code #export imagenet_stats = ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) cifar_stats = ([0.491, 0.482, 0.447], [0.247, 0.243, 0.261]) mnist_stats = ([0.131], [0.308]) im = Image.open(TEST_IMAGE).resize((30,20)) #export if not hasattr(Image,'_patched'): _old_sz = Image.Image.size.fget @patch_property def size(x:Image.Image): return Tuple(_old_sz(x)) Image._patched = True #export @patch_property def n_px(x: Image.Image): return x.size[0] * x.size[1] ###Output _____no_output_____ ###Markdown `Image.n_px`> `Image.n_px` (property)Number of pixels in image ###Code test_eq(im.n_px, 30*20) #export @patch_property def shape(x: Image.Image): return x.size[1],x.size[0] ###Output _____no_output_____ ###Markdown `Image.shape`> `Image.shape` (property)Image (height,width) tuple (NB: opposite order of `Image.size()`, same order as numpy array and pytorch tensor) ###Code test_eq(im.shape, (20,30)) #export @patch_property def aspect(x: Image.Image): return x.size[0]/x.size[1] ###Output _____no_output_____ ###Markdown `Image.aspect`> `Image.aspect` (property)Aspect ratio of the image, i.e. `width/height` ###Code test_eq(im.aspect, 30/20) #export @patch def reshape(x: Image.Image, h, w, resample=0): "`resize` `x` to `(w,h)`" return x.resize((w,h), resample=resample) show_doc(Image.Image.reshape) test_eq(im.reshape(12,10).shape, (12,10)) #export @patch def resize_max(x: Image.Image, resample=0, max_px=None, max_h=None, max_w=None): "`resize` `x` to `max_px`, or `max_h`, or `max_w`" h,w = x.shape if max_px and x.n_px>max_px: h,w = Tuple(h,w).mul(math.sqrt(max_px/x.n_px)) if max_h and h>max_h: h,w = (max_h ,max_h*w/h) if max_w and w>max_w: h,w = (max_w*h/w,max_w ) return x.reshape(round(h), round(w), resample=resample) test_eq(im.resize_max(max_px=20*30).shape, (20,30)) test_eq(im.resize_max(max_px=300).n_px, 294) test_eq(im.resize_max(max_px=500, max_h=10, max_w=20).shape, (10,15)) test_eq(im.resize_max(max_px=500, max_h=14, max_w=15).shape, (10,15)) test_eq(im.resize_max(max_px=300, max_h=10, max_w=25).shape, (10,15)) show_doc(Image.Image.resize_max) #TODO function to resize_max all images in a path (optionally recursively) and save them somewhere (same relative dirs if recursive) ###Output _____no_output_____ ###Markdown Basic types This section regroups the basic types used in vision with the transform that create objects of those types. ###Code # TODO: docs #export def load_image(fn, mode=None, **kwargs): "Open and load a `PIL.Image` and convert to `mode`" im = Image.open(fn, **kwargs) im.load() im = im._new(im.im) return im.convert(mode) if mode else im #export class PILBase(Image.Image, metaclass=BypassNewMeta): _bypass_type=Image.Image _show_args = {'cmap':'viridis'} _open_args = {'mode': 'RGB'} @classmethod def create(cls, fn, **kwargs)->None: "Open an `Image` from path `fn`" if isinstance(fn,Tensor): fn = fn.numpy() if isinstance(fn,ndarray): return cls(Image.fromarray(fn)) return cls(load_image(fn, **merge(cls._open_args, kwargs))) def show(self, ctx=None, **kwargs): "Show image using `merge(self._show_args, kwargs)`" return show_image(self, ctx=ctx, **merge(self._show_args, kwargs)) #export class PILImage(PILBase): pass #export class PILImageBW(PILImage): _show_args,_open_args = {'cmap':'Greys'},{'mode': 'L'} im = PILImage.create(TEST_IMAGE) test_eq(type(im), PILImage) test_eq(im.mode, 'RGB') im.resize((64,64)) ax = im.show(figsize=(1,1)) test_fig_exists(ax) #export class PILMask(PILBase): _open_args,_show_args = {'mode':'L'},{'alpha':0.5, 'cmap':'tab20'} im = PILMask.create(TEST_IMAGE) test_eq(type(im), PILMask) test_eq(im.mode, 'L') #export OpenMask = Transform(PILMask.create) OpenMask.loss_func = CrossEntropyLossFlat(axis=1) PILMask.create = OpenMask ###Output _____no_output_____ ###Markdown Images ###Code mnist = untar_data(URLs.MNIST_TINY) fns = get_image_files(mnist) mnist_fn = TEST_IMAGE_BW timg = Transform(PILImageBW.create) mnist_img = timg(mnist_fn) test_eq(mnist_img.size, (28,28)) assert isinstance(mnist_img, PILImageBW) mnist_img ###Output _____no_output_____ ###Markdown Segmentation masks ###Code camvid = untar_data(URLs.CAMVID_TINY) fns = get_image_files(camvid/'images') cam_fn = fns[0] mask_fn = camvid/'labels'/f'{cam_fn.stem}_P{cam_fn.suffix}' cam_img = PILImage.create(cam_fn) test_eq(cam_img.size, (128,96)) tmask = Transform(PILMask.create) mask = tmask(mask_fn) test_eq(type(mask), PILMask) test_eq(mask.size, (128,96)) _,axs = plt.subplots(1,3, figsize=(12,3)) cam_img.show(ctx=axs[0], title='image') mask.show(alpha=1, ctx=axs[1], vmin=1, vmax=30, title='mask') cam_img.show(ctx=axs[2], title='superimposed') mask.show(ctx=axs[2], vmin=1, vmax=30); ###Output _____no_output_____ ###Markdown Points ###Code # export class TensorPoint(TensorBase): "Basic type for points in an image" _show_args = dict(s=10, marker='.', c='r') @classmethod def create(cls, t, sz=None)->None: "Convert an array or a list of points `t` to a `Tensor`" return cls(tensor(t).view(-1, 2).float(), sz=sz) def show(self, ctx=None, **kwargs): if 'figsize' in kwargs: del kwargs['figsize'] x = self.view(-1,2) ctx.scatter(x[:, 0], x[:, 1], **{**self._show_args, **kwargs}) return ctx #export TensorPointCreate = Transform(TensorPoint.create) TensorPointCreate.loss_func = MSELossFlat() TensorPoint.create = TensorPointCreate ###Output _____no_output_____ ###Markdown Points are expected to come as an array/tensor of shape `(n,2)` or as a list of lists with two elements. Unless you change the defaults in `PointScaler` (see later on), coordinates should go from 0 to width/height, with the first one being the column index (so from 0 to width) and the second one being the row index (so from 0 to height).> Note: This is differnt from the usual indeixing convention for arrays in numpy or in PyTorch, but it's the way points are expected by matplotlib or the internal functions in PyTorch like `F.grid_sample`. ###Code pnt_img = TensorImage(mnist_img.resize((28,35))) pnts = np.array([[0,0], [0,35], [28,0], [28,35], [9, 17]]) tfm = Transform(TensorPoint.create) tpnts = tfm(pnts) test_eq(tpnts.shape, [5,2]) test_eq(tpnts.dtype, torch.float32) ctx = pnt_img.show(figsize=(1,1), cmap='Greys') tpnts.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Bounding boxes ###Code # export def get_annotations(fname, prefix=None): "Open a COCO style json in `fname` and returns the lists of filenames (with maybe `prefix`) and labelled bboxes." annot_dict = json.load(open(fname)) id2images, id2bboxes, id2cats = {}, collections.defaultdict(list), collections.defaultdict(list) classes = {o['id']:o['name'] for o in annot_dict['categories']} for o in annot_dict['annotations']: bb = o['bbox'] id2bboxes[o['image_id']].append([bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]]) id2cats[o['image_id']].append(classes[o['category_id']]) id2images = {o['id']:ifnone(prefix, '') + o['file_name'] for o in annot_dict['images'] if o['id'] in id2bboxes} ids = list(id2images.keys()) return [id2images[k] for k in ids], [(id2bboxes[k], id2cats[k]) for k in ids] #hide #TODO explain and/or simplify this coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') annots = json.load(open(coco/'train.json')) test_eq(images, [k['file_name'] for k in annots['images']]) for _ in range(5): idx = random.randint(0, len(images)-1) fn = images[idx] i = 0 while annots['images'][i]['file_name'] != fn: i+=1 img_id = annots['images'][i]['id'] bbs = [ann for ann in annots['annotations'] if ann['image_id'] == img_id] i2o = {k['id']:k['name'] for k in annots['categories']} lbls = [i2o[bb['category_id']] for bb in bbs] bboxes = [bb['bbox'] for bb in bbs] bboxes = [[bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]] for bb in bboxes] test_eq(lbl_bbox[idx], [bboxes, lbls]) # export from matplotlib import patches, patheffects def _draw_outline(o, lw): o.set_path_effects([patheffects.Stroke(linewidth=lw, foreground='black'), patheffects.Normal()]) def _draw_rect(ax, b, color='white', text=None, text_size=14, hw=True, rev=False): lx,ly,w,h = b if rev: lx,ly,w,h = ly,lx,h,w if not hw: w,h = w-lx,h-ly patch = ax.add_patch(patches.Rectangle((lx,ly), w, h, fill=False, edgecolor=color, lw=2)) _draw_outline(patch, 4) if text is not None: patch = ax.text(lx,ly, text, verticalalignment='top', color=color, fontsize=text_size, weight='bold') _draw_outline(patch,1) # export class TensorBBox(TensorPoint): "Basic type for a tensor of bounding boxes in an image" @classmethod def create(cls, x, sz=None)->None: return cls(tensor(x).view(-1, 4).float(), sz=sz) def show(self, ctx=None, **kwargs): x = self.view(-1,4) for b in x: _draw_rect(ctx, b, hw=False, **kwargs) return ctx ###Output _____no_output_____ ###Markdown Bounding boxes are expected to come as tuple with an array/tensor of shape `(n,4)` or as a list of lists with four elements adn a list of corresponding labels. Unless you change the defaults in `BBoxScaler` (see later on), coordinates for each bounding box should go from 0 to height/width, with the following convetion: top, left, bottom, right.> Note: We use the same convention as for points with y axis being before x. ###Code # export class LabeledBBox(Tuple): "Basic type for a list of bounding boxes in an image" def show(self, ctx=None, **kwargs): for b,l in zip(self.bbox, self.lbl): if l != '#na#': ctx = retain_type(b, self.bbox).show(ctx=ctx, text=l) return ctx bbox,lbl = add_props(lambda i,self: self[i]) coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') idx=2 coco_fn,bbox = coco/'train'/images[idx],lbl_bbox[idx] coco_img = timg(coco_fn) tbbox = LabeledBBox(TensorBBox(bbox[0]), bbox[1]) ctx = coco_img.show(figsize=(3,3), cmap='Greys') tbbox.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Basic Transforms Unless specifically metioned, all the following transforms can be used as single-item transforms (in one of the list in the `tfms` you pass to a `TfmdDS` or a `Datasource`) or tuple transform (in the `tuple_tfms` you pass to a `TfmdDS` or a `Datasource`). The safest way that will work accross applications is to always use them as `tuple_tfms`. For instance, if you have points or bounding boxes as targets and use `Resize` as a single-item transform, when you get to `PointScaler` or `BBoxScaler` (which are tuple transforms) you won't have the correct size of the image to properly scale your points. ###Code # export def image2tensor(img): "Transform image to byte tensor in `c*h*w` dim order." res = tensor(img) if res.dim()==2: res = res.unsqueeze(-1) return res.permute(2,0,1) # export PILImage ._tensor_cls = TensorImage PILImageBW._tensor_cls = TensorImageBW PILMask ._tensor_cls = TensorMask #export @ToTensor def encodes(self, o:PILBase): return o._tensor_cls(image2tensor(o)) @ToTensor def encodes(self, o:PILMask): return o._tensor_cls(image2tensor(o)[0]) ###Output _____no_output_____ ###Markdown Any data augmentation transform that runs on PIL Images must be run before this transform. ###Code tfm = ToTensor() print(tfm) print(type(mnist_img)) print(type(tfm(mnist_img))) tfm = ToTensor() test_eq(tfm(mnist_img).shape, (1,28,28)) test_eq(type(tfm(mnist_img)), TensorImageBW) test_eq(tfm(mask).shape, (96,128)) test_eq(type(tfm(mask)), TensorMask) ###Output _____no_output_____ ###Markdown Let's confirm we can pipeline this with `PILImage.create`. ###Code pipe_img = Pipeline([PILImageBW.create, ToTensor()]) img = pipe_img(mnist_fn) test_eq(type(img), TensorImageBW) pipe_img.show(img, figsize=(1,1)); def _cam_lbl(x): return mask_fn cam_tds = DataSource([cam_fn], [[PILImage.create, ToTensor()], [_cam_lbl, PILMask.create, ToTensor()]]) show_at(cam_tds, 0); ###Output _____no_output_____ ###Markdown To work with data augmentation, and in particular the `grid_sample` method, points need to be represented with coordinates going from -1 to 1 (-1 being top or left, 1 bottom or right), which will be done unless you pass `do_scale=False`. We also need to make sure they are following our convention of points being x,y coordinates, so pass along `y_first=True` if you have your data in an y,x format to add a flip.> Warning: This transform needs to run on the tuple level, before any transform that changes the image size. ###Code #export def _scale_pnts(y, sz, do_scale=True, y_first=False): if y_first: y = y.flip(1) res = y * 2/tensor(sz).float() - 1 if do_scale else y return TensorPoint(res, sz=sz) def _unscale_pnts(y, sz): return TensorPoint((y+1) * tensor(sz).float()/2, sz=sz) #export class PointScaler(Transform): "Scale a tensor representing points" order = 1 def __init__(self, do_scale=True, y_first=False): self.do_scale,self.y_first = do_scale,y_first def _grab_sz(self, x): self.sz = [x.shape[-1], x.shape[-2]] if isinstance(x, Tensor) else x.size return x def _get_sz(self, x): sz = getattr(x, '_meta', {}).get('sz', None) assert sz is not None or self.sz is not None, "Size could not be inferred, pass it in the init of your TensorPoint with `sz=...`" return self.sz if sz is None else sz def setup(self, dl): its = dl.do_item(0) for t in its: if isinstance(t, TensorPoint): self.c = t.numel() def encodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) def decodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) def encodes(self, x:TensorPoint): return _scale_pnts(x, self._get_sz(x), self.do_scale, self.y_first) def decodes(self, x:TensorPoint): return _unscale_pnts(x, self._get_sz(x)) ###Output _____no_output_____ ###Markdown To work with data augmentation, and in particular the `grid_sample` method, points need to be represented with coordinates going from -1 to 1 (-1 being top or left, 1 bottom or right), which will be done unless you pass `do_scale=False`. We also need to make sure they are following our convention of points being x,y coordinates, so pass along `y_first=True` if you have your data in an y,x format to add a flip.> Note: This transform automatically grabs the sizes of the images it sees before a `TensorPoint` object and embeds it in them. For this to work, those images need to be before any points in the order of your final tuple. If you don't have such images, you need to embed the size of the corresponding image when creating a `TensorPoint` by passing it with `sz=...`. ###Code def _pnt_lbl(x): return TensorPoint.create(pnts) def _pnt_open(fn): return PILImage(PILImage.create(fn).resize((28,35))) pnt_tds = DataSource([mnist_fn], [_pnt_open, [_pnt_lbl]]) pnt_tdl = TfmdDL(pnt_tds, bs=1, after_item=[PointScaler(), ToTensor()]) pnt_tdl.after_item.c x,y = pnt_tdl.one_batch() #Scaling and flipping properly done #NB: we added a point earlier at (9,17); formula below scales to (-1,1) coords test_close(y[0], tensor([[-1., -1.], [-1., 1.], [1., -1.], [1., 1.], [9/14-1, 17/17.5-1]])) a,b = pnt_tdl.decode_batch((x,y))[0] test_eq(b, tensor(pnts).float()) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorPoint) test_eq(type(a), TensorImage) test_eq(type(b), TensorPoint) pnt_tdl.show_batch(figsize=(2,2), cmap='Greys'); #export class BBoxLabeler(Transform): def setup(self, dl): self.vocab = dl.vocab def before_call(self): self.bbox,self.lbls = None,None def decode (self, x, **kwargs): self.bbox,self.lbls = None,None return self._call('decodes', x, **kwargs) def decodes(self, x:TensorMultiCategory): self.lbls = [self.vocab[a] for a in x] return x if self.bbox is None else LabeledBBox(self.bbox, self.lbls) def decodes(self, x:TensorBBox): self.bbox = x return self.bbox if self.lbls is None else LabeledBBox(self.bbox, self.lbls) #export #LabeledBBox can be sent in a tl with MultiCategorize (depending on the order of the tls) but it is already decoded. @MultiCategorize def decodes(self, x:LabeledBBox): return x #export @PointScaler def encodes(self, x:TensorBBox): pnts = self.encodes(TensorPoint(x.view(-1,2), sz=x._meta.get('sz', None))) return TensorBBox(pnts.view(-1, 4), sz=x._meta.get('sz', None)) @PointScaler def decodes(self, x:TensorBBox): pnts = self.decodes(TensorPoint(x.view(-1,2), sz=x._meta.get('sz', None))) return TensorBBox(pnts.view(-1, 4), sz=x._meta.get('sz', None)) def _coco_bb(x): return TensorBBox.create(bbox[0]) def _coco_lbl(x): return bbox[1] coco_tds = DataSource([coco_fn], [PILImage.create, [_coco_bb], [_coco_lbl, MultiCategorize(add_na=True)]], n_inp=1) coco_tdl = TfmdDL(coco_tds, bs=1, after_item=[BBoxLabeler(), PointScaler(), ToTensor()]) x,y,z = coco_tdl.one_batch() test_close(y[0], -1+tensor(bbox[0])/64) test_eq(z[0], tensor([1,1,1])) a,b,c = coco_tdl.decode_batch((x,y,z))[0] test_close(b, tensor(bbox[0]).float()) test_eq(c.bbox, b) test_eq(c.lbl, bbox[1]) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorBBox) test_eq(type(z), TensorMultiCategory) test_eq(type(a), TensorImage) test_eq(type(b), TensorBBox) test_eq(type(c), LabeledBBox) coco_tdl.show_batch(); #hide #test other direction works too coco_tds = DataSource([coco_fn], [PILImage.create, [_coco_lbl, MultiCategorize(add_na=True)], [_coco_bb]]) coco_tdl = TfmdDL(coco_tds, bs=1, after_item=[BBoxLabeler(), PointScaler(), ToTensor()]) x,y,z = coco_tdl.one_batch() test_close(z[0], -1+tensor(bbox[0])/64) test_eq(y[0], tensor([1,1,1])) a,b,c = coco_tdl.decode_batch((x,y,z))[0] test_eq(b, bbox[1]) test_close(c.bbox, tensor(bbox[0]).float()) test_eq(c.lbl, b) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorMultiCategory) test_eq(type(z), TensorBBox) test_eq(type(a), TensorImage) test_eq(type(b), MultiCategory) test_eq(type(c), LabeledBBox) ###Output _____no_output_____ ###Markdown Export - ###Code #hide from local.notebook.export import notebook2script notebook2script(all_fs=True) ###Output Converted 00_test.ipynb. Converted 01_core_foundation.ipynb. Converted 01a_core_utils.ipynb. Converted 01b_core_dispatch.ipynb. Converted 01c_core_transform.ipynb. Converted 02_core_script.ipynb. Converted 03_torchcore.ipynb. Converted 03a_layers.ipynb. Converted 04_data_load.ipynb. Converted 05_data_core.ipynb. Converted 06_data_transforms.ipynb. Converted 07_data_block.ipynb. Converted 08_vision_core.ipynb. Converted 09_vision_augment.ipynb. Converted 09a_vision_data.ipynb. Converted 10_pets_tutorial.ipynb. Converted 11_vision_models_xresnet.ipynb. Converted 12_optimizer.ipynb. Converted 13_learner.ipynb. Converted 13a_metrics.ipynb. Converted 14_callback_schedule.ipynb. Converted 14a_callback_data.ipynb. Converted 15_callback_hook.ipynb. Converted 15a_vision_models_unet.ipynb. Converted 16_callback_progress.ipynb. Converted 17_callback_tracker.ipynb. Converted 18_callback_fp16.ipynb. Converted 19_callback_mixup.ipynb. Converted 20_interpret.ipynb. Converted 20a_distributed.ipynb. Converted 21_vision_learner.ipynb. Converted 22_tutorial_imagenette.ipynb. Converted 23_tutorial_transfer_learning.ipynb. Converted 30_text_core.ipynb. Converted 31_text_data.ipynb. Converted 32_text_models_awdlstm.ipynb. Converted 33_text_models_core.ipynb. Converted 34_callback_rnn.ipynb. Converted 35_tutorial_wikitext.ipynb. Converted 36_text_models_qrnn.ipynb. Converted 37_text_learner.ipynb. Converted 38_tutorial_ulmfit.ipynb. Converted 40_tabular_core.ipynb. Converted 41_tabular_model.ipynb. Converted 42_tabular_rapids.ipynb. Converted 50_data_block_examples.ipynb. Converted 60_medical_imaging.ipynb. Converted 65_medical_text.ipynb. Converted 70_callback_wandb.ipynb. Converted 71_callback_tensorboard.ipynb. Converted 90_notebook_core.ipynb. Converted 91_notebook_export.ipynb. Converted 92_notebook_showdoc.ipynb. Converted 93_notebook_export2html.ipynb. Converted 94_notebook_test.ipynb. Converted 95_index.ipynb. Converted 96_data_external.ipynb. Converted 97_utils_test.ipynb. Converted notebook2jekyll.ipynb. Converted xse_resnext.ipynb. ###Markdown Core vision> Basic image opening/processing functionality Helpers ###Code #export imagenet_stats = ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) cifar_stats = ([0.491, 0.482, 0.447], [0.247, 0.243, 0.261]) mnist_stats = ([0.131], [0.308]) im = Image.open(TEST_IMAGE).resize((30,20)) #export if not hasattr(Image,'_patched'): _old_sz = Image.Image.size.fget @patch_property def size(x:Image.Image): return Tuple(_old_sz(x)) Image._patched = True #export @patch_property def n_px(x: Image.Image): return x.size[0] * x.size[1] ###Output _____no_output_____ ###Markdown `Image.n_px`> `Image.n_px` (property)Number of pixels in image ###Code test_eq(im.n_px, 30*20) #export @patch_property def shape(x: Image.Image): return x.size[1],x.size[0] ###Output _____no_output_____ ###Markdown `Image.shape`> `Image.shape` (property)Image (height,width) tuple (NB: opposite order of `Image.size()`, same order as numpy array and pytorch tensor) ###Code test_eq(im.shape, (20,30)) #export @patch_property def aspect(x: Image.Image): return x.size[0]/x.size[1] ###Output _____no_output_____ ###Markdown `Image.aspect`> `Image.aspect` (property)Aspect ratio of the image, i.e. `width/height` ###Code test_eq(im.aspect, 30/20) #export @patch def reshape(x: Image.Image, h, w, resample=0): "`resize` `x` to `(w,h)`" return x.resize((w,h), resample=resample) show_doc(Image.Image.reshape) test_eq(im.reshape(12,10).shape, (12,10)) #export @patch def resize_max(x: Image.Image, resample=0, max_px=None, max_h=None, max_w=None): "`resize` `x` to `max_px`, or `max_h`, or `max_w`" h,w = x.shape if max_px and x.n_px>max_px: h,w = Tuple(h,w).mul(math.sqrt(max_px/x.n_px)) if max_h and h>max_h: h,w = (max_h ,max_h*w/h) if max_w and w>max_w: h,w = (max_w*h/w,max_w ) return x.reshape(round(h), round(w), resample=resample) test_eq(im.resize_max(max_px=20*30).shape, (20,30)) test_eq(im.resize_max(max_px=300).n_px, 294) test_eq(im.resize_max(max_px=500, max_h=10, max_w=20).shape, (10,15)) test_eq(im.resize_max(max_px=500, max_h=14, max_w=15).shape, (10,15)) test_eq(im.resize_max(max_px=300, max_h=10, max_w=25).shape, (10,15)) show_doc(Image.Image.resize_max) #TODO function to resize_max all images in a path (optionally recursively) and save them somewhere (same relative dirs if recursive) ###Output _____no_output_____ ###Markdown Basic types This section regroups the basic types used in vision with the transform that create objects of those types. ###Code # TODO: docs #export def load_image(fn, mode=None, **kwargs): "Open and load a `PIL.Image` and convert to `mode`" im = Image.open(fn, **kwargs) im.load() im = im._new(im.im) return im.convert(mode) if mode else im #export class PILBase(Image.Image, metaclass=BypassNewMeta): _bypass_type=Image.Image _show_args = {'cmap':'viridis'} _open_args = {'mode': 'RGB'} @classmethod def create(cls, fn, **kwargs)->None: "Open an `Image` from path `fn`" if isinstance(fn,Tensor): fn = fn.numpy() if isinstance(fn,ndarray): return cls(Image.fromarray(fn)) return cls(load_image(fn, **merge(cls._open_args, kwargs))) def show(self, ctx=None, **kwargs): "Show image using `merge(self._show_args, kwargs)`" return show_image(self, ctx=ctx, **merge(self._show_args, kwargs)) #export class PILImage(PILBase): pass #export class PILImageBW(PILImage): _show_args,_open_args = {'cmap':'Greys'},{'mode': 'L'} im = PILImage.create(TEST_IMAGE) test_eq(type(im), PILImage) test_eq(im.mode, 'RGB') im.resize((64,64)) ax = im.show(figsize=(1,1)) test_fig_exists(ax) #export class PILMask(PILBase): _open_args,_show_args = {'mode':'L'},{'alpha':0.5, 'cmap':'tab20'} im = PILMask.create(TEST_IMAGE) test_eq(type(im), PILMask) test_eq(im.mode, 'L') #export OpenMask = Transform(PILMask.create) OpenMask.loss_func = CrossEntropyLossFlat(axis=1) PILMask.create = OpenMask ###Output _____no_output_____ ###Markdown Images ###Code mnist = untar_data(URLs.MNIST_TINY) fns = get_image_files(mnist) mnist_fn = TEST_IMAGE_BW timg = Transform(PILImageBW.create) mnist_img = timg(mnist_fn) test_eq(mnist_img.size, (28,28)) assert isinstance(mnist_img, PILImageBW) mnist_img ###Output _____no_output_____ ###Markdown Segmentation masks ###Code camvid = untar_data(URLs.CAMVID_TINY) fns = get_image_files(camvid/'images') cam_fn = fns[0] mask_fn = camvid/'labels'/f'{cam_fn.stem}_P{cam_fn.suffix}' cam_img = PILImage.create(cam_fn) test_eq(cam_img.size, (128,96)) tmask = Transform(PILMask.create) mask = tmask(mask_fn) test_eq(type(mask), PILMask) test_eq(mask.size, (128,96)) _,axs = plt.subplots(1,3, figsize=(12,3)) cam_img.show(ctx=axs[0], title='image') mask.show(alpha=1, ctx=axs[1], vmin=1, vmax=30, title='mask') cam_img.show(ctx=axs[2], title='superimposed') mask.show(ctx=axs[2], vmin=1, vmax=30); ###Output _____no_output_____ ###Markdown Points ###Code # export class TensorPoint(TensorBase): "Basic type for points in an image" _show_args = dict(s=10, marker='.', c='r') @classmethod def create(cls, t, sz=None)->None: "Convert an array or a list of points `t` to a `Tensor`" return cls(tensor(t).view(-1, 2).float(), sz=sz) def show(self, ctx=None, **kwargs): if 'figsize' in kwargs: del kwargs['figsize'] x = self.view(-1,2) ctx.scatter(x[:, 0], x[:, 1], **{**self._show_args, **kwargs}) return ctx #export TensorPointCreate = Transform(TensorPoint.create) TensorPointCreate.loss_func = MSELossFlat() TensorPoint.create = TensorPointCreate ###Output _____no_output_____ ###Markdown Points are expected to come as an array/tensor of shape `(n,2)` or as a list of lists with two elements. Unless you change the defaults in `PointScaler` (see later on), coordinates should go from 0 to width/height, with the first one being the column index (so from 0 to width) and the second one being the row index (so from 0 to height).> Note: This is differnt from the usual indeixing convention for arrays in numpy or in PyTorch, but it's the way points are expected by matplotlib or the internal functions in PyTorch like `F.grid_sample`. ###Code pnt_img = TensorImage(mnist_img.resize((28,35))) pnts = np.array([[0,0], [0,35], [28,0], [28,35], [9, 17]]) tfm = Transform(TensorPoint.create) tpnts = tfm(pnts) test_eq(tpnts.shape, [5,2]) test_eq(tpnts.dtype, torch.float32) ctx = pnt_img.show(figsize=(1,1), cmap='Greys') tpnts.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Bounding boxes ###Code # export def get_annotations(fname, prefix=None): "Open a COCO style json in `fname` and returns the lists of filenames (with maybe `prefix`) and labelled bboxes." annot_dict = json.load(open(fname)) id2images, id2bboxes, id2cats = {}, collections.defaultdict(list), collections.defaultdict(list) classes = {o['id']:o['name'] for o in annot_dict['categories']} for o in annot_dict['annotations']: bb = o['bbox'] id2bboxes[o['image_id']].append([bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]]) id2cats[o['image_id']].append(classes[o['category_id']]) id2images = {o['id']:ifnone(prefix, '') + o['file_name'] for o in annot_dict['images'] if o['id'] in id2bboxes} ids = list(id2images.keys()) return [id2images[k] for k in ids], [(id2bboxes[k], id2cats[k]) for k in ids] #hide #TODO explain and/or simplify this coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') annots = json.load(open(coco/'train.json')) test_eq(images, [k['file_name'] for k in annots['images']]) for _ in range(5): idx = random.randint(0, len(images)-1) fn = images[idx] i = 0 while annots['images'][i]['file_name'] != fn: i+=1 img_id = annots['images'][i]['id'] bbs = [ann for ann in annots['annotations'] if ann['image_id'] == img_id] i2o = {k['id']:k['name'] for k in annots['categories']} lbls = [i2o[bb['category_id']] for bb in bbs] bboxes = [bb['bbox'] for bb in bbs] bboxes = [[bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]] for bb in bboxes] test_eq(lbl_bbox[idx], [bboxes, lbls]) # export from matplotlib import patches, patheffects def _draw_outline(o, lw): o.set_path_effects([patheffects.Stroke(linewidth=lw, foreground='black'), patheffects.Normal()]) def _draw_rect(ax, b, color='white', text=None, text_size=14, hw=True, rev=False): lx,ly,w,h = b if rev: lx,ly,w,h = ly,lx,h,w if not hw: w,h = w-lx,h-ly patch = ax.add_patch(patches.Rectangle((lx,ly), w, h, fill=False, edgecolor=color, lw=2)) _draw_outline(patch, 4) if text is not None: patch = ax.text(lx,ly, text, verticalalignment='top', color=color, fontsize=text_size, weight='bold') _draw_outline(patch,1) # export class TensorBBox(TensorPoint): "Basic type for a tensor of bounding boxes in an image" @classmethod def create(cls, x, sz=None)->None: return cls(tensor(x).view(-1, 4).float(), sz=sz) def show(self, ctx=None, **kwargs): x = self.view(-1,4) for b in x: _draw_rect(ctx, b, hw=False, **kwargs) return ctx ###Output _____no_output_____ ###Markdown Bounding boxes are expected to come as tuple with an array/tensor of shape `(n,4)` or as a list of lists with four elements adn a list of corresponding labels. Unless you change the defaults in `BBoxScaler` (see later on), coordinates for each bounding box should go from 0 to height/width, with the following convetion: top, left, bottom, right.> Note: We use the same convention as for points with y axis being before x. ###Code # export class LabeledBBox(Tuple): "Basic type for a list of bounding boxes in an image" def show(self, ctx=None, **kwargs): for b,l in zip(self.bbox, self.lbl): if l != '#na#': ctx = retain_type(b, self.bbox).show(ctx=ctx, text=l) return ctx bbox,lbl = add_props(lambda i,self: self[i]) coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') idx=2 coco_fn,bbox = coco/'train'/images[idx],lbl_bbox[idx] coco_img = timg(coco_fn) tbbox = LabeledBBox(TensorBBox(bbox[0]), bbox[1]) ctx = coco_img.show(figsize=(3,3), cmap='Greys') tbbox.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Basic Transforms Unless specifically metioned, all the following transforms can be used as single-item transforms (in one of the list in the `tfms` you pass to a `TfmdDS` or a `Datasource`) or tuple transform (in the `tuple_tfms` you pass to a `TfmdDS` or a `Datasource`). The safest way that will work accross applications is to always use them as `tuple_tfms`. For instance, if you have points or bounding boxes as targets and use `Resize` as a single-item transform, when you get to `PointScaler` or `BBoxScaler` (which are tuple transforms) you won't have the correct size of the image to properly scale your points. ###Code # export def image2tensor(img): "Transform image to byte tensor in `c*h*w` dim order." res = tensor(img) if res.dim()==2: res = res.unsqueeze(-1) return res.permute(2,0,1) # export PILImage ._tensor_cls = TensorImage PILImageBW._tensor_cls = TensorImageBW PILMask ._tensor_cls = TensorMask #export @ToTensor def encodes(self, o:PILBase): return o._tensor_cls(image2tensor(o)) @ToTensor def encodes(self, o:PILMask): return o._tensor_cls(image2tensor(o)[0]) ###Output _____no_output_____ ###Markdown Any data augmentation transform that runs on PIL Images must be run before this transform. ###Code tfm = ToTensor() print(tfm) print(type(mnist_img)) print(type(tfm(mnist_img))) tfm = ToTensor() test_eq(tfm(mnist_img).shape, (1,28,28)) test_eq(type(tfm(mnist_img)), TensorImageBW) test_eq(tfm(mask).shape, (96,128)) test_eq(type(tfm(mask)), TensorMask) ###Output _____no_output_____ ###Markdown Let's confirm we can pipeline this with `PILImage.create`. ###Code pipe_img = Pipeline([PILImageBW.create, ToTensor()]) img = pipe_img(mnist_fn) test_eq(type(img), TensorImageBW) pipe_img.show(img, figsize=(1,1)); def _cam_lbl(x): return mask_fn cam_tds = DataSource([cam_fn], [[PILImage.create, ToTensor()], [_cam_lbl, PILMask.create, ToTensor()]]) show_at(cam_tds, 0); ###Output _____no_output_____ ###Markdown To work with data augmentation, and in particular the `grid_sample` method, points need to be represented with coordinates going from -1 to 1 (-1 being top or left, 1 bottom or right), which will be done unless you pass `do_scale=False`. We also need to make sure they are following our convention of points being x,y coordinates, so pass along `y_first=True` if you have your data in an y,x format to add a flip.> Warning: This transform needs to run on the tuple level, before any transform that changes the image size. ###Code #export def _scale_pnts(y, sz, do_scale=True, y_first=False): if y_first: y = y.flip(1) res = y * 2/tensor(sz).float() - 1 if do_scale else y return TensorPoint(res, sz=sz) def _unscale_pnts(y, sz): return TensorPoint((y+1) * tensor(sz).float()/2, sz=sz) #export class PointScaler(Transform): "Scale a tensor representing points" order = 1 def __init__(self, do_scale=True, y_first=False): self.do_scale,self.y_first = do_scale,y_first def _grab_sz(self, x): self.sz = [x.shape[-1], x.shape[-2]] if isinstance(x, Tensor) else x.size return x def _get_sz(self, x): sz = getattr(x, '_meta', {}).get('sz', None) assert sz is not None or self.sz is not None, "Size could not be inferred, pass it in the init of your TensorPoint with `sz=...`" return self.sz if sz is None else sz def setup(self, dl): its = dl.do_item(0) for t in its: if isinstance(t, TensorPoint): self.c = t.numel() def encodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) def decodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) def encodes(self, x:TensorPoint): return _scale_pnts(x, self._get_sz(x), self.do_scale, self.y_first) def decodes(self, x:TensorPoint): return _unscale_pnts(x, self._get_sz(x)) ###Output _____no_output_____ ###Markdown To work with data augmentation, and in particular the `grid_sample` method, points need to be represented with coordinates going from -1 to 1 (-1 being top or left, 1 bottom or right), which will be done unless you pass `do_scale=False`. We also need to make sure they are following our convention of points being x,y coordinates, so pass along `y_first=True` if you have your data in an y,x format to add a flip.> Note: This transform automatically grabs the sizes of the images it sees before a `TensorPoint` object and embeds it in them. For this to work, those images need to be before any points in the order of your final tuple. If you don't have such images, you need to embed the size of the corresponding image when creating a `TensorPoint` by passing it with `sz=...`. ###Code def _pnt_lbl(x): return TensorPoint.create(pnts) def _pnt_open(fn): return PILImage(PILImage.create(fn).resize((28,35))) pnt_tds = DataSource([mnist_fn], [_pnt_open, [_pnt_lbl]]) pnt_tdl = TfmdDL(pnt_tds, bs=1, after_item=[PointScaler(), ToTensor()]) pnt_tdl.after_item.c x,y = pnt_tdl.one_batch() #Scaling and flipping properly done #NB: we added a point earlier at (9,17); formula below scales to (-1,1) coords test_close(y[0], tensor([[-1., -1.], [-1., 1.], [1., -1.], [1., 1.], [9/14-1, 17/17.5-1]])) a,b = pnt_tdl.decode_batch((x,y))[0] test_eq(b, tensor(pnts).float()) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorPoint) test_eq(type(a), TensorImage) test_eq(type(b), TensorPoint) pnt_tdl.show_batch(figsize=(2,2), cmap='Greys'); #export class BBoxLabeler(Transform): def setup(self, dl): self.vocab = dl.vocab def before_call(self): self.bbox,self.lbls = None,None def decode (self, x, **kwargs): self.bbox,self.lbls = None,None return self._call('decodes', x, **kwargs) def decodes(self, x:TensorMultiCategory): self.lbls = [self.vocab[a] for a in x] return x if self.bbox is None else LabeledBBox(self.bbox, self.lbls) def decodes(self, x:TensorBBox): self.bbox = x return self.bbox if self.lbls is None else LabeledBBox(self.bbox, self.lbls) #export #LabeledBBox can be sent in a tl with MultiCategorize (depending on the order of the tls) but it is already decoded. @MultiCategorize def decodes(self, x:LabeledBBox): return x #export @PointScaler def encodes(self, x:TensorBBox): pnts = self.encodes(TensorPoint(x.view(-1,2), sz=x._meta.get('sz', None))) return TensorBBox(pnts.view(-1, 4), sz=x._meta.get('sz', None)) @PointScaler def decodes(self, x:TensorBBox): pnts = self.decodes(TensorPoint(x.view(-1,2), sz=x._meta.get('sz', None))) return TensorBBox(pnts.view(-1, 4), sz=x._meta.get('sz', None)) def _coco_bb(x): return TensorBBox.create(bbox[0]) def _coco_lbl(x): return bbox[1] coco_tds = DataSource([coco_fn], [PILImage.create, [_coco_bb], [_coco_lbl, MultiCategorize(add_na=True)]], n_inp=1) coco_tdl = TfmdDL(coco_tds, bs=1, after_item=[BBoxLabeler(), PointScaler(), ToTensor()]) x,y,z = coco_tdl.one_batch() test_close(y[0], -1+tensor(bbox[0])/64) test_eq(z[0], tensor([1,1,1])) a,b,c = coco_tdl.decode_batch((x,y,z))[0] test_close(b, tensor(bbox[0]).float()) test_eq(c.bbox, b) test_eq(c.lbl, bbox[1]) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorBBox) test_eq(type(z), TensorMultiCategory) test_eq(type(a), TensorImage) test_eq(type(b), TensorBBox) test_eq(type(c), LabeledBBox) coco_tdl.show_batch(); #hide #test other direction works too coco_tds = DataSource([coco_fn], [PILImage.create, [_coco_lbl, MultiCategorize(add_na=True)], [_coco_bb]]) coco_tdl = TfmdDL(coco_tds, bs=1, after_item=[BBoxLabeler(), PointScaler(), ToTensor()]) x,y,z = coco_tdl.one_batch() test_close(z[0], -1+tensor(bbox[0])/64) test_eq(y[0], tensor([1,1,1])) a,b,c = coco_tdl.decode_batch((x,y,z))[0] test_eq(b, bbox[1]) test_close(c.bbox, tensor(bbox[0]).float()) test_eq(c.lbl, b) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorMultiCategory) test_eq(type(z), TensorBBox) test_eq(type(a), TensorImage) test_eq(type(b), MultiCategory) test_eq(type(c), LabeledBBox) ###Output _____no_output_____ ###Markdown Export - ###Code #hide from nbdev.export import notebook2script notebook2script() ###Output Converted 00_test.ipynb. Converted 01_core_foundation.ipynb. Converted 01a_core_utils.ipynb. Converted 01b_core_dispatch.ipynb. Converted 01c_core_transform.ipynb. Converted 02_core_script.ipynb. Converted 03_torchcore.ipynb. Converted 03a_layers.ipynb. Converted 04_data_load.ipynb. Converted 05_data_core.ipynb. Converted 06_data_transforms.ipynb. Converted 07_data_block.ipynb. Converted 08_vision_core.ipynb. Converted 09_vision_augment.ipynb. Converted 09a_vision_data.ipynb. Converted 10_pets_tutorial.ipynb. Converted 11_vision_models_xresnet.ipynb. Converted 12_optimizer.ipynb. Converted 13_learner.ipynb. Converted 13a_metrics.ipynb. Converted 14_callback_schedule.ipynb. Converted 14a_callback_data.ipynb. Converted 15_callback_hook.ipynb. Converted 15a_vision_models_unet.ipynb. Converted 16_callback_progress.ipynb. Converted 17_callback_tracker.ipynb. Converted 18_callback_fp16.ipynb. Converted 19_callback_mixup.ipynb. Converted 20_interpret.ipynb. Converted 20a_distributed.ipynb. Converted 21_vision_learner.ipynb. Converted 22_tutorial_imagenette.ipynb. Converted 23_tutorial_transfer_learning.ipynb. Converted 30_text_core.ipynb. Converted 31_text_data.ipynb. Converted 32_text_models_awdlstm.ipynb. Converted 33_text_models_core.ipynb. Converted 34_callback_rnn.ipynb. Converted 35_tutorial_wikitext.ipynb. Converted 36_text_models_qrnn.ipynb. Converted 37_text_learner.ipynb. Converted 38_tutorial_ulmfit.ipynb. Converted 40_tabular_core.ipynb. Converted 41_tabular_model.ipynb. Converted 42_tabular_rapids.ipynb. Converted 50_data_block_examples.ipynb. Converted 60_medical_imaging.ipynb. Converted 65_medical_text.ipynb. Converted 70_callback_wandb.ipynb. Converted 71_callback_tensorboard.ipynb. Converted 90_notebook_core.ipynb. Converted 91_notebook_export.ipynb. Converted 92_notebook_showdoc.ipynb. Converted 93_notebook_export2html.ipynb. Converted 94_notebook_test.ipynb. Converted 95_index.ipynb. Converted 96_data_external.ipynb. Converted 97_utils_test.ipynb. Converted notebook2jekyll.ipynb. Converted xse_resnext.ipynb. ###Markdown Core vision> Basic image opening/processing functionality Helpers ###Code #export imagenet_stats = ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) cifar_stats = ([0.491, 0.482, 0.447], [0.247, 0.243, 0.261]) mnist_stats = ([0.131], [0.308]) im = Image.open(TEST_IMAGE).resize((30,20)) #export if not hasattr(Image,'_patched'): _old_sz = Image.Image.size.fget @patch_property def size(x:Image.Image): return Tuple(_old_sz(x)) Image._patched = True #export @patch_property def n_px(x: Image.Image): return x.size[0] * x.size[1] ###Output _____no_output_____ ###Markdown `Image.n_px`> `Image.n_px` (property)Number of pixels in image ###Code test_eq(im.n_px, 30*20) #export @patch_property def shape(x: Image.Image): return x.size[1],x.size[0] ###Output _____no_output_____ ###Markdown `Image.shape`> `Image.shape` (property)Image (height,width) tuple (NB: opposite order of `Image.size()`, same order as numpy array and pytorch tensor) ###Code test_eq(im.shape, (20,30)) #export @patch_property def aspect(x: Image.Image): return x.size[0]/x.size[1] ###Output _____no_output_____ ###Markdown `Image.aspect`> `Image.aspect` (property)Aspect ratio of the image, i.e. `width/height` ###Code test_eq(im.aspect, 30/20) #export @patch def reshape(x: Image.Image, h, w, resample=0): "`resize` `x` to `(w,h)`" return x.resize((w,h), resample=resample) show_doc(Image.Image.reshape) test_eq(im.reshape(12,10).shape, (12,10)) #export @patch def resize_max(x: Image.Image, resample=0, max_px=None, max_h=None, max_w=None): "`resize` `x` to `max_px`, or `max_h`, or `max_w`" h,w = x.shape if max_px and x.n_px>max_px: h,w = Tuple(h,w).mul(math.sqrt(max_px/x.n_px)) if max_h and h>max_h: h,w = (max_h ,max_h*w/h) if max_w and w>max_w: h,w = (max_w*h/w,max_w ) return x.reshape(round(h), round(w), resample=resample) test_eq(im.resize_max(max_px=20*30).shape, (20,30)) test_eq(im.resize_max(max_px=300).n_px, 294) test_eq(im.resize_max(max_px=500, max_h=10, max_w=20).shape, (10,15)) test_eq(im.resize_max(max_px=500, max_h=14, max_w=15).shape, (10,15)) test_eq(im.resize_max(max_px=300, max_h=10, max_w=25).shape, (10,15)) show_doc(Image.Image.resize_max) #TODO function to resize_max all images in a path (optionally recursively) and save them somewhere (same relative dirs if recursive) ###Output _____no_output_____ ###Markdown Basic types This section regroups the basic types used in vision with the transform that create objects of those types. ###Code # TODO: docs #export def load_image(fn, mode=None, **kwargs): "Open and load a `PIL.Image` and convert to `mode`" im = Image.open(fn, **kwargs) im.load() im = im._new(im.im) return im.convert(mode) if mode else im #export class PILBase(Image.Image, metaclass=BypassNewMeta): _bypass_type=Image.Image default_batch_tfms = IntToFloatTensor _show_args = {'cmap':'viridis'} _open_args = {'mode': 'RGB'} @classmethod def create(cls, fn, **kwargs)->None: "Open an `Image` from path `fn`" if isinstance(fn,Tensor): fn = fn.numpy() if isinstance(fn,ndarray): return cls(Image.fromarray(fn)) return cls(load_image(fn, **merge(cls._open_args, kwargs))) def show(self, ctx=None, **kwargs): "Show image using `merge(self._show_args, kwargs)`" return show_image(self, ctx=ctx, **merge(self._show_args, kwargs)) #export class PILImage(PILBase): pass #export class PILImageBW(PILImage): _show_args,_open_args = {'cmap':'Greys'},{'mode': 'L'} im = PILImage.create(TEST_IMAGE) test_eq(type(im), PILImage) test_eq(im.mode, 'RGB') im.resize((64,64)) ax = im.show(figsize=(1,1)) test_fig_exists(ax) #export class PILMask(PILBase): _open_args,_show_args = {'mode':'L'},{'alpha':0.5, 'cmap':'tab20'} im = PILMask.create(TEST_IMAGE) test_eq(type(im), PILMask) test_eq(im.mode, 'L') #export OpenMask = Transform(PILMask.create) OpenMask.loss_func = CrossEntropyLossFlat(axis=1) PILMask.create = OpenMask ###Output _____no_output_____ ###Markdown Images ###Code mnist = untar_data(URLs.MNIST_TINY) fns = get_image_files(mnist) mnist_fn = TEST_IMAGE_BW timg = Transform(PILImageBW.create) mnist_img = timg(mnist_fn) test_eq(mnist_img.size, (28,28)) assert isinstance(mnist_img, PILImageBW) mnist_img ###Output _____no_output_____ ###Markdown Segmentation masks ###Code camvid = untar_data(URLs.CAMVID_TINY) fns = get_image_files(camvid/'images') cam_fn = fns[0] mask_fn = camvid/'labels'/f'{cam_fn.stem}_P{cam_fn.suffix}' cam_img = PILImage.create(cam_fn) test_eq(cam_img.size, (128,96)) tmask = Transform(PILMask.create) mask = tmask(mask_fn) test_eq(type(mask), PILMask) test_eq(mask.size, (128,96)) _,axs = plt.subplots(1,3, figsize=(12,3)) cam_img.show(ctx=axs[0], title='image') mask.show(alpha=1, ctx=axs[1], vmin=1, vmax=30, title='mask') cam_img.show(ctx=axs[2], title='superimposed') mask.show(ctx=axs[2], vmin=1, vmax=30); ###Output _____no_output_____ ###Markdown Points ###Code # export class TensorPoint(TensorBase): "Basic type for points in an image" _show_args = dict(s=10, marker='.', c='r') @classmethod def create(cls, t, sz=None)->None: "Convert an array or a list of points `t` to a `Tensor`" return cls(tensor(t).view(-1, 2).float(), sz=sz) def show(self, ctx=None, **kwargs): if 'figsize' in kwargs: del kwargs['figsize'] x = self.view(-1,2) ctx.scatter(x[:, 0], x[:, 1], **{**self._show_args, **kwargs}) return ctx ###Output _____no_output_____ ###Markdown Points are expected to come as an array/tensor of shape `(n,2)` or as a list of lists with two elements. Unless you change the defaults in `PointScaler` (see later on), coordinates should go from 0 to width/height, with the first one being the column index (so from 0 to width) and the second one being the row index (so from 0 to height).> Note: This is differnt from the usual indeixing convention for arrays in numpy or in PyTorch, but it's the way points are expected by matplotlib or the internal functions in PyTorch like `F.grid_sample`. ###Code pnt_img = TensorImage(mnist_img.resize((28,35))) pnts = np.array([[0,0], [0,35], [28,0], [28,35], [9, 17]]) tfm = Transform(TensorPoint.create) tpnts = tfm(pnts) test_eq(tpnts.shape, [5,2]) test_eq(tpnts.dtype, torch.float32) ctx = pnt_img.show(figsize=(1,1), cmap='Greys') tpnts.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Bounding boxes ###Code # export def get_annotations(fname, prefix=None): "Open a COCO style json in `fname` and returns the lists of filenames (with maybe `prefix`) and labelled bboxes." annot_dict = json.load(open(fname)) id2images, id2bboxes, id2cats = {}, collections.defaultdict(list), collections.defaultdict(list) classes = {o['id']:o['name'] for o in annot_dict['categories']} for o in annot_dict['annotations']: bb = o['bbox'] id2bboxes[o['image_id']].append([bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]]) id2cats[o['image_id']].append(classes[o['category_id']]) id2images = {o['id']:ifnone(prefix, '') + o['file_name'] for o in annot_dict['images'] if o['id'] in id2bboxes} ids = list(id2images.keys()) return [id2images[k] for k in ids], [(id2bboxes[k], id2cats[k]) for k in ids] #hide #TODO explain and/or simplify this coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') annots = json.load(open(coco/'train.json')) test_eq(images, [k['file_name'] for k in annots['images']]) for _ in range(5): idx = random.randint(0, len(images)-1) fn = images[idx] i = 0 while annots['images'][i]['file_name'] != fn: i+=1 img_id = annots['images'][i]['id'] bbs = [ann for ann in annots['annotations'] if ann['image_id'] == img_id] i2o = {k['id']:k['name'] for k in annots['categories']} lbls = [i2o[bb['category_id']] for bb in bbs] bboxes = [bb['bbox'] for bb in bbs] bboxes = [[bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]] for bb in bboxes] test_eq(lbl_bbox[idx], [bboxes, lbls]) # export from matplotlib import patches, patheffects def _draw_outline(o, lw): o.set_path_effects([patheffects.Stroke(linewidth=lw, foreground='black'), patheffects.Normal()]) def _draw_rect(ax, b, color='white', text=None, text_size=14, hw=True, rev=False): lx,ly,w,h = b if rev: lx,ly,w,h = ly,lx,h,w if not hw: w,h = w-lx,h-ly patch = ax.add_patch(patches.Rectangle((lx,ly), w, h, fill=False, edgecolor=color, lw=2)) _draw_outline(patch, 4) if text is not None: patch = ax.text(lx,ly, text, verticalalignment='top', color=color, fontsize=text_size, weight='bold') _draw_outline(patch,1) # export class TensorBBox(TensorPoint): "Basic type for a tensor of bounding boxes in an image" @classmethod def create(cls, x, sz=None)->None: return cls(tensor(x).view(-1, 4).float(), sz=sz) def show(self, ctx=None, **kwargs): x = self.view(-1,4) for b in x: _draw_rect(ctx, b, hw=False, **kwargs) return ctx ###Output _____no_output_____ ###Markdown Bounding boxes are expected to come as tuple with an array/tensor of shape `(n,4)` or as a list of lists with four elements adn a list of corresponding labels. Unless you change the defaults in `BBoxScaler` (see later on), coordinates for each bounding box should go from 0 to height/width, with the following convetion: top, left, bottom, right.> Note: We use the same convention as for points with y axis being before x. ###Code # export class LabeledBBox(Tuple): "Basic type for a list of bounding boxes in an image" def show(self, ctx=None, **kwargs): for b,l in zip(self.bbox, self.lbl): if l != '#na#': ctx = retain_type(b, self.bbox).show(ctx=ctx, text=l) return ctx @classmethod def create(cls, x): return cls(x) bbox,lbl = add_props(lambda i,self: self[i]) coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') idx=2 coco_fn,bbox = coco/'train'/images[idx],lbl_bbox[idx] coco_img = timg(coco_fn) tbbox = LabeledBBox(TensorBBox(bbox[0]), bbox[1]) ctx = coco_img.show(figsize=(3,3), cmap='Greys') tbbox.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Basic Transforms Unless specifically metioned, all the following transforms can be used as single-item transforms (in one of the list in the `tfms` you pass to a `TfmdDS` or a `Datasource`) or tuple transform (in the `tuple_tfms` you pass to a `TfmdDS` or a `Datasource`). The safest way that will work accross applications is to always use them as `tuple_tfms`. For instance, if you have points or bounding boxes as targets and use `ImageResizer` as a single-item transform, when you get to `PointScaler` or `BBoxScaler` (which are tuple transforms) you won't have the correct size of the image to properly scale your points. ###Code class ImageResizer(Transform): order=10 "Resize image to `size` using `resample" def __init__(self, size, resample=Image.BILINEAR): if not is_listy(size): size=(size,size) self.size,self.resample = (size[1],size[0]),resample def encodes(self, o:PILImage): return o.resize(size=self.size, resample=self.resample) def encodes(self, o:PILMask): return o.resize(size=self.size, resample=Image.NEAREST) ###Output _____no_output_____ ###Markdown `size` can either be one integer (in which case images are resized to a square) or a tuple `height,width`.> Note: This is the usual convention for arrays or in PyTorch, but it's not the usual convention for PIL Image, which use the other way round. ###Code f = ImageResizer(14) test_eq(f(mnist_img).size, (14,14)) test_eq(f(mask).size, (14,14)) f = ImageResizer((32,28)) test_eq(f(mnist_img).size, (28,32))#PIL has width first test_eq(array(f(mnist_img)).shape, (32,28))#But numpy as height first and that is our convention # export def image2tensor(img): "Transform image to byte tensor in `c*h*w` dim order." res = tensor(img) if res.dim()==2: res = res.unsqueeze(-1) return res.permute(2,0,1) # export PILImage ._tensor_cls = TensorImage PILImageBW._tensor_cls = TensorImageBW PILMask ._tensor_cls = TensorMask #export @ToTensor def encodes(self, o:PILBase): return o._tensor_cls(image2tensor(o)) @ToTensor def encodes(self, o:PILMask): return o._tensor_cls(image2tensor(o)[0]) ###Output _____no_output_____ ###Markdown Any data augmentation transform that runs on PIL Images must be run before this transform. ###Code tfm = ToTensor() print(tfm) print(type(mnist_img)) print(type(tfm(mnist_img))) tfm = ToTensor() test_eq(tfm(mnist_img).shape, (1,28,28)) test_eq(type(tfm(mnist_img)), TensorImageBW) test_eq(tfm(mask).shape, (96,128)) test_eq(type(tfm(mask)), TensorMask) ###Output _____no_output_____ ###Markdown Let's confirm we can pipeline this with `PILImage.create`. ###Code pipe_img = Pipeline([PILImageBW.create, ToTensor()]) img = pipe_img(mnist_fn) test_eq(type(img), TensorImageBW) pipe_img.show(img, figsize=(1,1)); def _cam_lbl(x): return mask_fn cam_tds = DataSource([cam_fn], [[PILImage.create, ToTensor()], [_cam_lbl, PILMask.create, ToTensor()]]) show_at(cam_tds, 0); PILMask.create.loss_func = CrossEntropyLossFlat(axis=1) ###Output _____no_output_____ ###Markdown To work with data augmentation, and in particular the `grid_sample` method, points need to be represented with coordinates going from -1 to 1 (-1 being top or left, 1 bottom or right), which will be done unless you pass `do_scale=False`. We also need to make sure they are following our convention of points being x,y coordinates, so pass along `y_first=True` if you have your data in an y,x format to add a flip.> Warning: This transform needs to run on the tuple level, before any transform that changes the image size. ###Code #export def _scale_pnts(y, sz, do_scale=True, y_first=False): if y_first: y = y.flip(1) res = y * 2/tensor(sz).float() - 1 if do_scale else y return TensorPoint(res, sz=sz) def _unscale_pnts(y, sz): return TensorPoint((y+1) * tensor(sz).float()/2, sz=sz) #export class PointScaler(Transform): "Scale a tensor representing points" order,loss_func = 1,MSELossFlat() def __init__(self, do_scale=True, y_first=False): self.do_scale,self.y_first = do_scale,y_first def _grab_sz(self, x): self.sz = [x.shape[-1], x.shape[-2]] if isinstance(x, Tensor) else x.size return x def _get_sz(self, x): sz = getattr(x, '_meta', {}).get('sz', None) assert sz is not None or self.sz is not None, "Size could not be inferred, pass it in the init of your TensorPoint with `sz=...`" return self.sz if sz is None else sz def setup(self, dl): its = dl.do_item(0) for t in its: if isinstance(t, TensorPoint): self.c = t.numel() def encodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) def decodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) def encodes(self, x:TensorPoint): return _scale_pnts(x, self._get_sz(x), self.do_scale, self.y_first) def decodes(self, x:TensorPoint): return _unscale_pnts(x, self._get_sz(x)) TensorPoint.default_item_tfms = PointScaler ###Output _____no_output_____ ###Markdown To work with data augmentation, and in particular the `grid_sample` method, points need to be represented with coordinates going from -1 to 1 (-1 being top or left, 1 bottom or right), which will be done unless you pass `do_scale=False`. We also need to make sure they are following our convention of points being x,y coordinates, so pass along `y_first=True` if you have your data in an y,x format to add a flip.> Note: This transform automatically grabs the sizes of the images it sees before a `TensorPoint` object and embeds it in them. For this to work, those images need to be before any points in the order of your final tuple. If you don't have such images, you need to embed the size of the corresponding image when creating a `TensorPoint` by passing it with `sz=...`. ###Code def _pnt_lbl(x): return TensorPoint.create(pnts) def _pnt_open(fn): return PILImage(PILImage.create(fn).resize((28,35))) pnt_tds = DataSource([mnist_fn], [_pnt_open, [_pnt_lbl]]) pnt_tdl = TfmdDL(pnt_tds, bs=1, after_item=[PointScaler(), ToTensor()]) pnt_tdl.after_item.c x,y = pnt_tdl.one_batch() #Scaling and flipping properly done #NB: we added a point earlier at (9,17); formula below scales to (-1,1) coords test_close(y[0], tensor([[-1., -1.], [-1., 1.], [1., -1.], [1., 1.], [9/14-1, 17/17.5-1]])) a,b = pnt_tdl.decode_batch((x,y))[0] test_eq(b, tensor(pnts).float()) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorPoint) test_eq(type(a), TensorImage) test_eq(type(b), TensorPoint) pnt_tdl.show_batch(figsize=(2,2), cmap='Greys'); #export class BBoxLabels(MultiCategory): create = MultiCategorize(add_na=True) default_type_tfms = None #export class BBoxLabeler(Transform): def setup(self, dl): self.vocab = dl.vocab def before_call(self): self.bbox,self.lbls = None,None def decode (self, x, **kwargs): self.bbox,self.lbls = None,None return self._call('decodes', x, **kwargs) def decodes(self, x:TensorMultiCategory): self.lbls = [self.vocab[a] for a in x] return x if self.bbox is None else LabeledBBox(self.bbox, self.lbls) def decodes(self, x:TensorBBox): self.bbox = x return self.bbox if self.lbls is None else LabeledBBox(self.bbox, self.lbls) #export BBoxLabels.default_item_tfms = BBoxLabeler #export #LabeledBBox can be sent in a tl with MultiCategorize (depending on the order of the tls) but it is already decoded. @MultiCategorize def decodes(self, x:LabeledBBox): return x #export @PointScaler def encodes(self, x:TensorBBox): pnts = self.encodes(TensorPoint(x.view(-1,2), sz=x._meta.get('sz', None))) return TensorBBox(pnts.view(-1, 4), sz=x._meta.get('sz', None)) @PointScaler def decodes(self, x:TensorBBox): pnts = self.decodes(TensorPoint(x.view(-1,2), sz=x._meta.get('sz', None))) return TensorBBox(pnts.view(-1, 4), sz=x._meta.get('sz', None)) def _coco_bb(x): return TensorBBox.create(bbox[0]) def _coco_lbl(x): return bbox[1] coco_tds = DataSource([coco_fn], [PILImage.create, [_coco_bb], [_coco_lbl, MultiCategorize(add_na=True)]], n_inp=1) coco_tdl = TfmdDL(coco_tds, bs=1, after_item=[BBoxLabeler(), PointScaler(), ToTensor()]) x,y,z = coco_tdl.one_batch() test_close(y[0], -1+tensor(bbox[0])/64) test_eq(z[0], tensor([1,1,1])) a,b,c = coco_tdl.decode_batch((x,y,z))[0] test_close(b, tensor(bbox[0]).float()) test_eq(c.bbox, b) test_eq(c.lbl, bbox[1]) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorBBox) test_eq(type(z), TensorMultiCategory) test_eq(type(a), TensorImage) test_eq(type(b), TensorBBox) test_eq(type(c), LabeledBBox) coco_tdl.show_batch(); #hide #test other direction works too coco_tds = DataSource([coco_fn], [PILImage.create, [_coco_lbl, MultiCategorize(add_na=True)], [_coco_bb]]) coco_tdl = TfmdDL(coco_tds, bs=1, after_item=[BBoxLabeler(), PointScaler(), ToTensor()]) x,y,z = coco_tdl.one_batch() test_close(z[0], -1+tensor(bbox[0])/64) test_eq(y[0], tensor([1,1,1])) a,b,c = coco_tdl.decode_batch((x,y,z))[0] test_eq(b, bbox[1]) test_close(c.bbox, tensor(bbox[0]).float()) test_eq(c.lbl, b) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorMultiCategory) test_eq(type(z), TensorBBox) test_eq(type(a), TensorImage) test_eq(type(b), MultiCategory) test_eq(type(c), LabeledBBox) ###Output _____no_output_____ ###Markdown Export - ###Code #hide from local.notebook.export import notebook2script notebook2script(all_fs=True) ###Output Converted 00_test.ipynb. Converted 01_core.ipynb. Converted 01a_utils.ipynb. Converted 01b_dispatch.ipynb. Converted 01c_transform.ipynb. Converted 02_script.ipynb. Converted 03_torch_core.ipynb. Converted 03a_layers.ipynb. Converted 04_dataloader.ipynb. Converted 05_data_core.ipynb. Converted 06_data_transforms.ipynb. Converted 07_data_block.ipynb. Converted 08_vision_core.ipynb. Converted 09_vision_augment.ipynb. Converted 09a_vision_data.ipynb. Converted 10_pets_tutorial.ipynb. Converted 11_vision_models_xresnet.ipynb. Converted 12_optimizer.ipynb. Converted 13_learner.ipynb. Converted 13a_metrics.ipynb. Converted 14_callback_schedule.ipynb. Converted 14a_callback_data.ipynb. Converted 15_callback_hook.ipynb. Converted 15a_vision_models_unet.ipynb. Converted 16_callback_progress.ipynb. Converted 17_callback_tracker.ipynb. Converted 18_callback_fp16.ipynb. Converted 19_callback_mixup.ipynb. Converted 20_interpret.ipynb. Converted 21_vision_learner.ipynb. Converted 22_tutorial_imagenette.ipynb. Converted 23_tutorial_transfer_learning.ipynb. Converted 30_text_core.ipynb. Converted 31_text_data.ipynb. Converted 32_text_models_awdlstm.ipynb. Converted 33_text_models_core.ipynb. Converted 34_callback_rnn.ipynb. Converted 35_tutorial_wikitext.ipynb. Converted 36_text_models_qrnn.ipynb. Converted 37_text_learner.ipynb. Converted 38_tutorial_ulmfit.ipynb. Converted 40_tabular_core.ipynb. Converted 41_tabular_model.ipynb. Converted 42_tabular_rapids.ipynb. Converted 50_data_block_examples.ipynb. Converted 60_medical_imaging.ipynb. Converted 65_medical_text.ipynb. Converted 90_notebook_core.ipynb. Converted 91_notebook_export.ipynb. Converted 92_notebook_showdoc.ipynb. Converted 93_notebook_export2html.ipynb. Converted 94_notebook_test.ipynb. Converted 95_index.ipynb. Converted 96_data_external.ipynb. Converted 97_utils_test.ipynb. Converted notebook2jekyll.ipynb. ###Markdown Core vision> Basic image opening/processing functionality Helpers ###Code #export imagenet_stats = ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) cifar_stats = ([0.491, 0.482, 0.447], [0.247, 0.243, 0.261]) mnist_stats = ([0.131], [0.308]) im = Image.open(TEST_IMAGE).resize((30,20)) #export if not hasattr(Image,'_patched'): _old_sz = Image.Image.size.fget @patch_property def size(x:Image.Image): return Tuple(_old_sz(x)) Image._patched = True #export @patch_property def n_px(x: Image.Image): return x.size[0] * x.size[1] ###Output _____no_output_____ ###Markdown `Image.n_px`> `Image.n_px` (property)Number of pixels in image ###Code test_eq(im.n_px, 30*20) #export @patch_property def shape(x: Image.Image): return x.size[1],x.size[0] ###Output _____no_output_____ ###Markdown `Image.shape`> `Image.shape` (property)Image (height,width) tuple (NB: opposite order of `Image.size()`, same order as numpy array and pytorch tensor) ###Code test_eq(im.shape, (20,30)) #export @patch_property def aspect(x: Image.Image): return x.size[0]/x.size[1] ###Output _____no_output_____ ###Markdown `Image.aspect`> `Image.aspect` (property)Aspect ratio of the image, i.e. `width/height` ###Code test_eq(im.aspect, 30/20) #export @patch def reshape(x: Image.Image, h, w, resample=0): "`resize` `x` to `(w,h)`" return x.resize((w,h), resample=resample) show_doc(Image.Image.reshape) test_eq(im.reshape(12,10).shape, (12,10)) #export @patch def resize_max(x: Image.Image, resample=0, max_px=None, max_h=None, max_w=None): "`resize` `x` to `max_px`, or `max_h`, or `max_w`" h,w = x.shape if max_px and x.n_px>max_px: h,w = Tuple(h,w).mul(math.sqrt(max_px/x.n_px)) if max_h and h>max_h: h,w = (max_h ,max_h*w/h) if max_w and w>max_w: h,w = (max_w*h/w,max_w ) return x.reshape(round(h), round(w), resample=resample) test_eq(im.resize_max(max_px=20*30).shape, (20,30)) test_eq(im.resize_max(max_px=300).n_px, 294) test_eq(im.resize_max(max_px=500, max_h=10, max_w=20).shape, (10,15)) test_eq(im.resize_max(max_px=500, max_h=14, max_w=15).shape, (10,15)) test_eq(im.resize_max(max_px=300, max_h=10, max_w=25).shape, (10,15)) show_doc(Image.Image.resize_max) #TODO function to resize_max all images in a path (optionally recursively) and save them somewhere (same relative dirs if recursive) ###Output _____no_output_____ ###Markdown Basic types This section regroups the basic types used in vision with the transform that create objects of those types. ###Code # TODO: docs #export def load_image(fn, mode=None, **kwargs): "Open and load a `PIL.Image` and convert to `mode`" im = Image.open(fn, **kwargs) im.load() im = im._new(im.im) return im.convert(mode) if mode else im #export class PILBase(Image.Image, metaclass=BypassNewMeta): _bypass_type=Image.Image _show_args = {'cmap':'viridis'} _open_args = {'mode': 'RGB'} @classmethod def create(cls, fn, **kwargs)->None: "Open an `Image` from path `fn`" if isinstance(fn,Tensor): fn = fn.numpy() if isinstance(fn,ndarray): return cls(Image.fromarray(fn)) return cls(load_image(fn, **merge(cls._open_args, kwargs))) def show(self, ctx=None, **kwargs): "Show image using `merge(self._show_args, kwargs)`" return show_image(self, ctx=ctx, **merge(self._show_args, kwargs)) #export class PILImage(PILBase): pass #export class PILImageBW(PILImage): _show_args,_open_args = {'cmap':'Greys'},{'mode': 'L'} im = PILImage.create(TEST_IMAGE) test_eq(type(im), PILImage) test_eq(im.mode, 'RGB') im.resize((64,64)) ax = im.show(figsize=(1,1)) test_fig_exists(ax) #export class PILMask(PILBase): _open_args,_show_args = {'mode':'L'},{'alpha':0.5, 'cmap':'tab20'} im = PILMask.create(TEST_IMAGE) test_eq(type(im), PILMask) test_eq(im.mode, 'L') #export OpenMask = Transform(PILMask.create) OpenMask.loss_func = CrossEntropyLossFlat(axis=1) PILMask.create = OpenMask ###Output _____no_output_____ ###Markdown Images ###Code mnist = untar_data(URLs.MNIST_TINY) fns = get_image_files(mnist) mnist_fn = TEST_IMAGE_BW timg = Transform(PILImageBW.create) mnist_img = timg(mnist_fn) test_eq(mnist_img.size, (28,28)) assert isinstance(mnist_img, PILImageBW) mnist_img ###Output _____no_output_____ ###Markdown Segmentation masks ###Code camvid = untar_data(URLs.CAMVID_TINY) fns = get_image_files(camvid/'images') cam_fn = fns[0] mask_fn = camvid/'labels'/f'{cam_fn.stem}_P{cam_fn.suffix}' cam_img = PILImage.create(cam_fn) test_eq(cam_img.size, (128,96)) tmask = Transform(PILMask.create) mask = tmask(mask_fn) test_eq(type(mask), PILMask) test_eq(mask.size, (128,96)) _,axs = plt.subplots(1,3, figsize=(12,3)) cam_img.show(ctx=axs[0], title='image') mask.show(alpha=1, ctx=axs[1], vmin=1, vmax=30, title='mask') cam_img.show(ctx=axs[2], title='superimposed') mask.show(ctx=axs[2], vmin=1, vmax=30); ###Output _____no_output_____ ###Markdown Points ###Code # export class TensorPoint(TensorBase): "Basic type for points in an image" _show_args = dict(s=10, marker='.', c='r') @classmethod def create(cls, t, sz=None)->None: "Convert an array or a list of points `t` to a `Tensor`" return cls(tensor(t).view(-1, 2).float(), sz=sz) def show(self, ctx=None, **kwargs): if 'figsize' in kwargs: del kwargs['figsize'] x = self.view(-1,2) ctx.scatter(x[:, 0], x[:, 1], **{**self._show_args, **kwargs}) return ctx #export TensorPointCreate = Transform(TensorPoint.create) TensorPointCreate.loss_func = MSELossFlat() TensorPoint.create = TensorPointCreate ###Output _____no_output_____ ###Markdown Points are expected to come as an array/tensor of shape `(n,2)` or as a list of lists with two elements. Unless you change the defaults in `PointScaler` (see later on), coordinates should go from 0 to width/height, with the first one being the column index (so from 0 to width) and the second one being the row index (so from 0 to height).> Note: This is differnt from the usual indeixing convention for arrays in numpy or in PyTorch, but it's the way points are expected by matplotlib or the internal functions in PyTorch like `F.grid_sample`. ###Code pnt_img = TensorImage(mnist_img.resize((28,35))) pnts = np.array([[0,0], [0,35], [28,0], [28,35], [9, 17]]) tfm = Transform(TensorPoint.create) tpnts = tfm(pnts) test_eq(tpnts.shape, [5,2]) test_eq(tpnts.dtype, torch.float32) ctx = pnt_img.show(figsize=(1,1), cmap='Greys') tpnts.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Bounding boxes ###Code # export def get_annotations(fname, prefix=None): "Open a COCO style json in `fname` and returns the lists of filenames (with maybe `prefix`) and labelled bboxes." annot_dict = json.load(open(fname)) id2images, id2bboxes, id2cats = {}, collections.defaultdict(list), collections.defaultdict(list) classes = {o['id']:o['name'] for o in annot_dict['categories']} for o in annot_dict['annotations']: bb = o['bbox'] id2bboxes[o['image_id']].append([bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]]) id2cats[o['image_id']].append(classes[o['category_id']]) id2images = {o['id']:ifnone(prefix, '') + o['file_name'] for o in annot_dict['images'] if o['id'] in id2bboxes} ids = list(id2images.keys()) return [id2images[k] for k in ids], [(id2bboxes[k], id2cats[k]) for k in ids] #hide #TODO explain and/or simplify this coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') annots = json.load(open(coco/'train.json')) test_eq(images, [k['file_name'] for k in annots['images']]) for _ in range(5): idx = random.randint(0, len(images)-1) fn = images[idx] i = 0 while annots['images'][i]['file_name'] != fn: i+=1 img_id = annots['images'][i]['id'] bbs = [ann for ann in annots['annotations'] if ann['image_id'] == img_id] i2o = {k['id']:k['name'] for k in annots['categories']} lbls = [i2o[bb['category_id']] for bb in bbs] bboxes = [bb['bbox'] for bb in bbs] bboxes = [[bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]] for bb in bboxes] test_eq(lbl_bbox[idx], [bboxes, lbls]) # export from matplotlib import patches, patheffects def _draw_outline(o, lw): o.set_path_effects([patheffects.Stroke(linewidth=lw, foreground='black'), patheffects.Normal()]) def _draw_rect(ax, b, color='white', text=None, text_size=14, hw=True, rev=False): lx,ly,w,h = b if rev: lx,ly,w,h = ly,lx,h,w if not hw: w,h = w-lx,h-ly patch = ax.add_patch(patches.Rectangle((lx,ly), w, h, fill=False, edgecolor=color, lw=2)) _draw_outline(patch, 4) if text is not None: patch = ax.text(lx,ly, text, verticalalignment='top', color=color, fontsize=text_size, weight='bold') _draw_outline(patch,1) # export class TensorBBox(TensorPoint): "Basic type for a tensor of bounding boxes in an image" @classmethod def create(cls, x, sz=None)->None: return cls(tensor(x).view(-1, 4).float(), sz=sz) def show(self, ctx=None, **kwargs): x = self.view(-1,4) for b in x: _draw_rect(ctx, b, hw=False, **kwargs) return ctx ###Output _____no_output_____ ###Markdown Bounding boxes are expected to come as tuple with an array/tensor of shape `(n,4)` or as a list of lists with four elements adn a list of corresponding labels. Unless you change the defaults in `BBoxScaler` (see later on), coordinates for each bounding box should go from 0 to height/width, with the following convetion: top, left, bottom, right.> Note: We use the same convention as for points with y axis being before x. ###Code # export class LabeledBBox(Tuple): "Basic type for a list of bounding boxes in an image" def show(self, ctx=None, **kwargs): for b,l in zip(self.bbox, self.lbl): if l != '#na#': ctx = retain_type(b, self.bbox).show(ctx=ctx, text=l) return ctx bbox,lbl = add_props(lambda i,self: self[i]) coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') idx=2 coco_fn,bbox = coco/'train'/images[idx],lbl_bbox[idx] coco_img = timg(coco_fn) tbbox = LabeledBBox(TensorBBox(bbox[0]), bbox[1]) ctx = coco_img.show(figsize=(3,3), cmap='Greys') tbbox.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Basic Transforms Unless specifically metioned, all the following transforms can be used as single-item transforms (in one of the list in the `tfms` you pass to a `TfmdDS` or a `Datasource`) or tuple transform (in the `tuple_tfms` you pass to a `TfmdDS` or a `Datasource`). The safest way that will work accross applications is to always use them as `tuple_tfms`. For instance, if you have points or bounding boxes as targets and use `Resize` as a single-item transform, when you get to `PointScaler` or `BBoxScaler` (which are tuple transforms) you won't have the correct size of the image to properly scale your points. ###Code # export def image2tensor(img): "Transform image to byte tensor in `c*h*w` dim order." res = tensor(img) if res.dim()==2: res = res.unsqueeze(-1) return res.permute(2,0,1) # export PILImage ._tensor_cls = TensorImage PILImageBW._tensor_cls = TensorImageBW PILMask ._tensor_cls = TensorMask #export @ToTensor def encodes(self, o:PILBase): return o._tensor_cls(image2tensor(o)) @ToTensor def encodes(self, o:PILMask): return o._tensor_cls(image2tensor(o)[0]) ###Output _____no_output_____ ###Markdown Any data augmentation transform that runs on PIL Images must be run before this transform. ###Code tfm = ToTensor() print(tfm) print(type(mnist_img)) print(type(tfm(mnist_img))) tfm = ToTensor() test_eq(tfm(mnist_img).shape, (1,28,28)) test_eq(type(tfm(mnist_img)), TensorImageBW) test_eq(tfm(mask).shape, (96,128)) test_eq(type(tfm(mask)), TensorMask) ###Output _____no_output_____ ###Markdown Let's confirm we can pipeline this with `PILImage.create`. ###Code pipe_img = Pipeline([PILImageBW.create, ToTensor()]) img = pipe_img(mnist_fn) test_eq(type(img), TensorImageBW) pipe_img.show(img, figsize=(1,1)); def _cam_lbl(x): return mask_fn cam_tds = DataSource([cam_fn], [[PILImage.create, ToTensor()], [_cam_lbl, PILMask.create, ToTensor()]]) show_at(cam_tds, 0); ###Output _____no_output_____ ###Markdown To work with data augmentation, and in particular the `grid_sample` method, points need to be represented with coordinates going from -1 to 1 (-1 being top or left, 1 bottom or right), which will be done unless you pass `do_scale=False`. We also need to make sure they are following our convention of points being x,y coordinates, so pass along `y_first=True` if you have your data in an y,x format to add a flip.> Warning: This transform needs to run on the tuple level, before any transform that changes the image size. ###Code #export def _scale_pnts(y, sz, do_scale=True, y_first=False): if y_first: y = y.flip(1) res = y * 2/tensor(sz).float() - 1 if do_scale else y return TensorPoint(res, sz=sz) def _unscale_pnts(y, sz): return TensorPoint((y+1) * tensor(sz).float()/2, sz=sz) #export class PointScaler(Transform): "Scale a tensor representing points" order = 1 def __init__(self, do_scale=True, y_first=False): self.do_scale,self.y_first = do_scale,y_first def _grab_sz(self, x): self.sz = [x.shape[-1], x.shape[-2]] if isinstance(x, Tensor) else x.size return x def _get_sz(self, x): sz = getattr(x, '_meta', {}).get('sz', None) assert sz is not None or self.sz is not None, "Size could not be inferred, pass it in the init of your TensorPoint with `sz=...`" return self.sz if sz is None else sz def setup(self, dl): its = dl.do_item(0) for t in its: if isinstance(t, TensorPoint): self.c = t.numel() def encodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) def decodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) def encodes(self, x:TensorPoint): return _scale_pnts(x, self._get_sz(x), self.do_scale, self.y_first) def decodes(self, x:TensorPoint): return _unscale_pnts(x, self._get_sz(x)) ###Output _____no_output_____ ###Markdown To work with data augmentation, and in particular the `grid_sample` method, points need to be represented with coordinates going from -1 to 1 (-1 being top or left, 1 bottom or right), which will be done unless you pass `do_scale=False`. We also need to make sure they are following our convention of points being x,y coordinates, so pass along `y_first=True` if you have your data in an y,x format to add a flip.> Note: This transform automatically grabs the sizes of the images it sees before a `TensorPoint` object and embeds it in them. For this to work, those images need to be before any points in the order of your final tuple. If you don't have such images, you need to embed the size of the corresponding image when creating a `TensorPoint` by passing it with `sz=...`. ###Code def _pnt_lbl(x): return TensorPoint.create(pnts) def _pnt_open(fn): return PILImage(PILImage.create(fn).resize((28,35))) pnt_tds = DataSource([mnist_fn], [_pnt_open, [_pnt_lbl]]) pnt_tdl = TfmdDL(pnt_tds, bs=1, after_item=[PointScaler(), ToTensor()]) pnt_tdl.after_item.c x,y = pnt_tdl.one_batch() #Scaling and flipping properly done #NB: we added a point earlier at (9,17); formula below scales to (-1,1) coords test_close(y[0], tensor([[-1., -1.], [-1., 1.], [1., -1.], [1., 1.], [9/14-1, 17/17.5-1]])) a,b = pnt_tdl.decode_batch((x,y))[0] test_eq(b, tensor(pnts).float()) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorPoint) test_eq(type(a), TensorImage) test_eq(type(b), TensorPoint) pnt_tdl.show_batch(figsize=(2,2), cmap='Greys'); #export class BBoxLabeler(Transform): def setup(self, dl): self.vocab = dl.vocab def before_call(self): self.bbox,self.lbls = None,None def decode (self, x, **kwargs): self.bbox,self.lbls = None,None return self._call('decodes', x, **kwargs) def decodes(self, x:TensorMultiCategory): self.lbls = [self.vocab[a] for a in x] return x if self.bbox is None else LabeledBBox(self.bbox, self.lbls) def decodes(self, x:TensorBBox): self.bbox = x return self.bbox if self.lbls is None else LabeledBBox(self.bbox, self.lbls) #export #LabeledBBox can be sent in a tl with MultiCategorize (depending on the order of the tls) but it is already decoded. @MultiCategorize def decodes(self, x:LabeledBBox): return x #export @PointScaler def encodes(self, x:TensorBBox): pnts = self.encodes(TensorPoint(x.view(-1,2), sz=x._meta.get('sz', None))) return TensorBBox(pnts.view(-1, 4), sz=x._meta.get('sz', None)) @PointScaler def decodes(self, x:TensorBBox): pnts = self.decodes(TensorPoint(x.view(-1,2), sz=x._meta.get('sz', None))) return TensorBBox(pnts.view(-1, 4), sz=x._meta.get('sz', None)) def _coco_bb(x): return TensorBBox.create(bbox[0]) def _coco_lbl(x): return bbox[1] coco_tds = DataSource([coco_fn], [PILImage.create, [_coco_bb], [_coco_lbl, MultiCategorize(add_na=True)]], n_inp=1) coco_tdl = TfmdDL(coco_tds, bs=1, after_item=[BBoxLabeler(), PointScaler(), ToTensor()]) x,y,z = coco_tdl.one_batch() test_close(y[0], -1+tensor(bbox[0])/64) test_eq(z[0], tensor([1,1,1])) a,b,c = coco_tdl.decode_batch((x,y,z))[0] test_close(b, tensor(bbox[0]).float()) test_eq(c.bbox, b) test_eq(c.lbl, bbox[1]) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorBBox) test_eq(type(z), TensorMultiCategory) test_eq(type(a), TensorImage) test_eq(type(b), TensorBBox) test_eq(type(c), LabeledBBox) coco_tdl.show_batch(); #hide #test other direction works too coco_tds = DataSource([coco_fn], [PILImage.create, [_coco_lbl, MultiCategorize(add_na=True)], [_coco_bb]]) coco_tdl = TfmdDL(coco_tds, bs=1, after_item=[BBoxLabeler(), PointScaler(), ToTensor()]) x,y,z = coco_tdl.one_batch() test_close(z[0], -1+tensor(bbox[0])/64) test_eq(y[0], tensor([1,1,1])) a,b,c = coco_tdl.decode_batch((x,y,z))[0] test_eq(b, bbox[1]) test_close(c.bbox, tensor(bbox[0]).float()) test_eq(c.lbl, b) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorMultiCategory) test_eq(type(z), TensorBBox) test_eq(type(a), TensorImage) test_eq(type(b), MultiCategory) test_eq(type(c), LabeledBBox) ###Output _____no_output_____ ###Markdown Export - ###Code #hide from local.notebook.export import notebook2script notebook2script(all_fs=True) ###Output Converted 00_test.ipynb. Converted 01_core.ipynb. Converted 01a_utils.ipynb. Converted 01b_dispatch.ipynb. Converted 01c_transform.ipynb. Converted 02_script.ipynb. Converted 03_torch_core.ipynb. Converted 03a_layers.ipynb. Converted 04_dataloader.ipynb. Converted 05_data_core.ipynb. Converted 06_data_transforms.ipynb. Converted 07_data_block.ipynb. Converted 08_vision_core.ipynb. Converted 09_vision_augment.ipynb. Converted 09a_vision_data.ipynb. Converted 10_pets_tutorial.ipynb. Converted 11_vision_models_xresnet.ipynb. Converted 12_optimizer.ipynb. Converted 13_learner.ipynb. Converted 13a_metrics.ipynb. Converted 14_callback_schedule.ipynb. Converted 14a_callback_data.ipynb. Converted 15_callback_hook.ipynb. Converted 15a_vision_models_unet.ipynb. Converted 16_callback_progress.ipynb. Converted 17_callback_tracker.ipynb. Converted 18_callback_fp16.ipynb. Converted 19_callback_mixup.ipynb. Converted 20_interpret.ipynb. Converted 21_vision_learner.ipynb. Converted 22_tutorial_imagenette.ipynb. Converted 23_tutorial_transfer_learning.ipynb. Converted 30_text_core.ipynb. Converted 31_text_data.ipynb. Converted 32_text_models_awdlstm.ipynb. Converted 33_text_models_core.ipynb. Converted 34_callback_rnn.ipynb. Converted 35_tutorial_wikitext.ipynb. Converted 36_text_models_qrnn.ipynb. Converted 37_text_learner.ipynb. Converted 38_tutorial_ulmfit.ipynb. Converted 40_tabular_core.ipynb. Converted 41_tabular_model.ipynb. Converted 42_tabular_rapids.ipynb. Converted 50_data_block_examples.ipynb. Converted 60_medical_imaging.ipynb. Converted 65_medical_text.ipynb. Converted 90_notebook_core.ipynb. Converted 91_notebook_export.ipynb. Converted 92_notebook_showdoc.ipynb. Converted 93_notebook_export2html.ipynb. Converted 94_notebook_test.ipynb. Converted 95_index.ipynb. Converted 96_data_external.ipynb. Converted 97_utils_test.ipynb. Converted notebook2jekyll.ipynb. ###Markdown Core vision> Basic image opening/processing functionality Helpers ###Code #export imagenet_stats = ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) cifar_stats = ([0.491, 0.482, 0.447], [0.247, 0.243, 0.261]) mnist_stats = ([0.131], [0.308]) im = Image.open(TEST_IMAGE).resize((30,20)) #export if not hasattr(Image,'_patched'): _old_sz = Image.Image.size.fget @patch_property def size(x:Image.Image): return Tuple(_old_sz(x)) Image._patched = True #export @patch_property def n_px(x: Image.Image): return x.size[0] * x.size[1] ###Output _____no_output_____ ###Markdown `Image.n_px`> `Image.n_px` (property)Number of pixels in image ###Code test_eq(im.n_px, 30*20) #export @patch_property def shape(x: Image.Image): return x.size[1],x.size[0] ###Output _____no_output_____ ###Markdown `Image.shape`> `Image.shape` (property)Image (height,width) tuple (NB: opposite order of `Image.size()`, same order as numpy array and pytorch tensor) ###Code test_eq(im.shape, (20,30)) #export @patch_property def aspect(x: Image.Image): return x.size[0]/x.size[1] ###Output _____no_output_____ ###Markdown `Image.aspect`> `Image.aspect` (property)Aspect ratio of the image, i.e. `width/height` ###Code test_eq(im.aspect, 30/20) #export @patch def reshape(x: Image.Image, h, w, resample=0): "`resize` `x` to `(w,h)`" return x.resize((w,h), resample=resample) show_doc(Image.Image.reshape) test_eq(im.reshape(12,10).shape, (12,10)) a,b = L([5,2]).map(lambda x: math.floor(x * 4/3)) a,b #export @patch def resize_max(x: Image.Image, resample=0, max_px=None, max_h=None, max_w=None): "`resize` `x` to `max_px`, or `max_h`, or `max_w`" h,w = x.shape if max_px and x.n_px>max_px: h,w = Tuple(h,w).mul(math.sqrt(max_px/x.n_px)) if max_h and h>max_h: h,w = (max_h ,max_h*w/h) if max_w and w>max_w: h,w = (max_w*h/w,max_w ) return x.reshape(round(h), round(w), resample=resample) test_eq(im.resize_max(max_px=20*30).shape, (20,30)) test_eq(im.resize_max(max_px=300).n_px, 294) test_eq(im.resize_max(max_px=500, max_h=10, max_w=20).shape, (10,15)) test_eq(im.resize_max(max_px=500, max_h=14, max_w=15).shape, (10,15)) test_eq(im.resize_max(max_px=300, max_h=10, max_w=25).shape, (10,15)) show_doc(Image.Image.resize_max) #TODO function to resize_max all images in a path (optionally recursively) and save them somewhere (same relative dirs if recursive) ###Output _____no_output_____ ###Markdown Basic types This section regroups the basic types used in vision with the transform that create objects of those types. ###Code # TODO: docs #export def load_image(fn, mode=None, **kwargs): "Open and load a `PIL.Image` and convert to `mode`" im = Image.open(fn, **kwargs) im.load() im = im._new(im.im) return im.convert(mode) if mode else im #export class PILBase(Image.Image, metaclass=BypassNewMeta): _bypass_type=Image.Image default_batch_tfms = ByteToFloatTensor _show_args = {'cmap':'viridis'} _open_args = {'mode': 'RGB'} @classmethod def create(cls, fn, **kwargs)->None: "Open an `Image` from path `fn`" if isinstance(fn,Tensor): fn = fn.numpy() if isinstance(fn,ndarray): return cls(Image.fromarray(fn)) return cls(load_image(fn, **merge(cls._open_args, kwargs))) def show(self, ctx=None, **kwargs): "Show image using `merge(self._show_args, kwargs)`" return show_image(self, ctx=ctx, **merge(self._show_args, kwargs)) #export class PILImage(PILBase): pass #export class PILImageBW(PILImage): _show_args,_open_args = {'cmap':'Greys'},{'mode': 'L'} im = PILImage.create(TEST_IMAGE) test_eq(type(im), PILImage) test_eq(im.mode, 'RGB') im.resize((64,64)) ax = im.show(figsize=(1,1)) test_fig_exists(ax) #export class PILMask(PILBase): _open_args,_show_args = {'mode':'L'},{'alpha':0.5, 'cmap':'tab20'} im = PILMask.create(TEST_IMAGE) test_eq(type(im), PILMask) test_eq(im.mode, 'L') #export OpenMask = Transform(PILMask.create) OpenMask.loss_func = CrossEntropyLossFlat(axis=1) PILMask.create = OpenMask ###Output _____no_output_____ ###Markdown Images ###Code mnist = untar_data(URLs.MNIST_TINY) fns = get_image_files(mnist) mnist_fn = fns[0]; mnist_fn timg = Transform(PILImageBW.create) mnist_img = timg(mnist_fn) test_eq(mnist_img.size, (28,28)) assert isinstance(mnist_img, PILImageBW) mnist_img ###Output _____no_output_____ ###Markdown Segmentation masks ###Code camvid = untar_data(URLs.CAMVID_TINY) fns = get_image_files(camvid/'images') cam_fn = fns[0] mask_fn = camvid/'labels'/f'{cam_fn.stem}_P{cam_fn.suffix}' cam_img = PILImage.create(cam_fn) test_eq(cam_img.size, (128,96)) tmask = Transform(PILMask.create) mask = tmask(mask_fn) test_eq(type(mask), PILMask) test_eq(mask.size, (128,96)) _,axs = plt.subplots(1,3, figsize=(12,3)) cam_img.show(ctx=axs[0], title='image') mask.show(alpha=1, ctx=axs[1], vmin=1, vmax=30, title='mask') cam_img.show(ctx=axs[2], title='superimposed') mask.show(ctx=axs[2], vmin=1, vmax=30); ###Output _____no_output_____ ###Markdown Points ###Code # export class TensorPoint(TensorBase): "Basic type for points in an image" _show_args = dict(s=10, marker='.', c='r') @classmethod def create(cls, t, sz=None)->None: "Convert an array or a list of points `t` to a `Tensor`" return cls(tensor(t).view(-1, 2).float(), sz=sz) def show(self, ctx=None, **kwargs): if 'figsize' in kwargs: del kwargs['figsize'] x = self.view(-1,2) ctx.scatter(x[:, 0], x[:, 1], **{**self._show_args, **kwargs}) return ctx ###Output _____no_output_____ ###Markdown Points are expected to come as an array/tensor of shape `(n,2)` or as a list of lists with two elements. Unless you change the defaults in `PointScaler` (see later on), coordinates should go from 0 to width/height, with the first one being the column index (so from 0 to width) and the second one being the row index (so from 0 to height).> Note: This is differnt from the usual indeixing convention for arrays in numpy or in PyTorch, but it's the way points are expected by matplotlib or the internal functions in PyTorch like `F.grid_sample`. ###Code pnt_img = TensorImage(mnist_img.resize((28,35))) pnts = np.array([[0,0], [0,35], [28,0], [28,35], [9, 17]]) tfm = Transform(TensorPoint.create) tpnts = tfm(pnts) test_eq(tpnts.shape, [5,2]) test_eq(tpnts.dtype, torch.float32) ctx = pnt_img.show(figsize=(1,1), cmap='Greys') tpnts.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Bounding boxes ###Code # export def get_annotations(fname, prefix=None): "Open a COCO style json in `fname` and returns the lists of filenames (with maybe `prefix`) and labelled bboxes." annot_dict = json.load(open(fname)) id2images, id2bboxes, id2cats = {}, collections.defaultdict(list), collections.defaultdict(list) classes = {o['id']:o['name'] for o in annot_dict['categories']} for o in annot_dict['annotations']: bb = o['bbox'] id2bboxes[o['image_id']].append([bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]]) id2cats[o['image_id']].append(classes[o['category_id']]) id2images = {o['id']:ifnone(prefix, '') + o['file_name'] for o in annot_dict['images'] if o['id'] in id2bboxes} ids = list(id2images.keys()) return [id2images[k] for k in ids], [(id2bboxes[k], id2cats[k]) for k in ids] #hide #TODO explain and/or simplify this coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') annots = json.load(open(coco/'train.json')) test_eq(images, [k['file_name'] for k in annots['images']]) for _ in range(5): idx = random.randint(0, len(images)-1) fn = images[idx] i = 0 while annots['images'][i]['file_name'] != fn: i+=1 img_id = annots['images'][i]['id'] bbs = [ann for ann in annots['annotations'] if ann['image_id'] == img_id] i2o = {k['id']:k['name'] for k in annots['categories']} lbls = [i2o[bb['category_id']] for bb in bbs] bboxes = [bb['bbox'] for bb in bbs] bboxes = [[bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]] for bb in bboxes] test_eq(lbl_bbox[idx], [bboxes, lbls]) # export from matplotlib import patches, patheffects def _draw_outline(o, lw): o.set_path_effects([patheffects.Stroke(linewidth=lw, foreground='black'), patheffects.Normal()]) def _draw_rect(ax, b, color='white', text=None, text_size=14, hw=True, rev=False): lx,ly,w,h = b if rev: lx,ly,w,h = ly,lx,h,w if not hw: w,h = w-lx,h-ly patch = ax.add_patch(patches.Rectangle((lx,ly), w, h, fill=False, edgecolor=color, lw=2)) _draw_outline(patch, 4) if text is not None: patch = ax.text(lx,ly, text, verticalalignment='top', color=color, fontsize=text_size, weight='bold') _draw_outline(patch,1) # export class TensorBBox(TensorPoint): "Basic type for a tensor of bounding boxes in an image" @classmethod def create(cls, x, sz=None)->None: return cls(tensor(x).view(-1, 4).float(), sz=sz) def show(self, ctx=None, **kwargs): x = self.view(-1,4) for b in x: _draw_rect(ctx, b, hw=False, **kwargs) return ctx ###Output _____no_output_____ ###Markdown Bounding boxes are expected to come as tuple with an array/tensor of shape `(n,4)` or as a list of lists with four elements adn a list of corresponding labels. Unless you change the defaults in `BBoxScaler` (see later on), coordinates for each bounding box should go from 0 to height/width, with the following convetion: top, left, bottom, right.> Note: We use the same convention as for points with y axis being before x. ###Code # export class LabeledBBox(Tuple): "Basic type for a list of bounding boxes in an image" def show(self, ctx=None, **kwargs): for b,l in zip(self.bbox, self.lbl): if l != '#bg': ctx = retain_type(b, self.bbox).show(ctx=ctx, text=l) return ctx @classmethod def create(cls, x): return cls(x) bbox,lbl = add_props(lambda i,self: self[i]) coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') idx=2 coco_fn,bbox = coco/'train'/images[idx],lbl_bbox[idx] coco_img = timg(coco_fn) tbbox = LabeledBBox(TensorBBox(bbox[0]), bbox[1]) ctx = coco_img.show(figsize=(3,3), cmap='Greys') tbbox.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Basic Transforms Unless specifically metioned, all the following transforms can be used as single-item transforms (in one of the list in the `tfms` you pass to a `TfmdDS` or a `Datasource`) or tuple transform (in the `tuple_tfms` you pass to a `TfmdDS` or a `Datasource`). The safest way that will work accross applications is to always use them as `tuple_tfms`. For instance, if you have points or bounding boxes as targets and use `ImageResizer` as a single-item transform, when you get to `PointScaler` or `BBoxScaler` (which are tuple transforms) you won't have the correct size of the image to properly scale your points. ###Code class ImageResizer(Transform): order=10 "Resize image to `size` using `resample" def __init__(self, size, resample=Image.BILINEAR): if not is_listy(size): size=(size,size) self.size,self.resample = (size[1],size[0]),resample def encodes(self, o:PILImage): return o.resize(size=self.size, resample=self.resample) def encodes(self, o:PILMask): return o.resize(size=self.size, resample=Image.NEAREST) ###Output _____no_output_____ ###Markdown `size` can either be one integer (in which case images are resized to a square) or a tuple `height,width`.> Note: This is the usual convention for arrays or in PyTorch, but it's not the usual convention for PIL Image, which use the other way round. ###Code f = ImageResizer(14) test_eq(f(mnist_img).size, (14,14)) test_eq(f(mask).size, (14,14)) f = ImageResizer((32,28)) test_eq(f(mnist_img).size, (28,32))#PIL has width first test_eq(array(f(mnist_img)).shape, (32,28))#But numpy as height first and that is our convention # export def image2byte(img): "Transform image to byte tensor in `c*h*w` dim order." res = torch.ByteTensor(torch.ByteStorage.from_buffer(img.tobytes())) w,h = img.size return res.view(h,w,-1).permute(2,0,1) #export @ToTensor def encodes(self, o:PILImage): return TensorImage(image2byte(o)) @ToTensor def encodes(self, o:PILImageBW): return TensorImageBW(image2byte(o)) @ToTensor def encodes(self, o:PILMask): return TensorMask(image2byte(o)[0]) ###Output _____no_output_____ ###Markdown Any data augmentation transform that runs on PIL Images must be run before this transform. ###Code tfm = ToTensor() print(tfm) print(type(mnist_img)) print(type(tfm(mnist_img))) tfm = ToTensor() test_eq(tfm(mnist_img).shape, (1,28,28)) test_eq(type(tfm(mnist_img)), TensorImageBW) test_eq(tfm(mask).shape, (96,128)) test_eq(type(tfm(mask)), TensorMask) ###Output _____no_output_____ ###Markdown Let's confirm we can pipeline this with `PILImage.create`. ###Code pipe_img = Pipeline([PILImageBW.create, ToTensor()]) img = pipe_img(mnist_fn) pipe_img.show(img, figsize=(1,1)); def _cam_lbl(x): return mask_fn cam_tds = DataSource([cam_fn], [[PILImage.create, ToTensor()], [_cam_lbl, PILMask.create, ToTensor()]]) show_at(cam_tds, 0); PILMask.create.loss_func = CrossEntropyLossFlat(axis=1) ##export #def _scale_pnts(x, y, do_scale=True, y_first=False): # if y_first: y = y.flip(1) # sz = [x.shape[-1], x.shape[-2]] if isinstance(x, Tensor) else x.size # return y * 2/tensor(sz).float() - 1 if do_scale else y # #def _unscale_pnts(x, y): # sz = [x.shape[-1], x.shape[-2]] if isinstance(x, Tensor) else x.size # return (y+1) * tensor(sz).float()/2 ## export ##TODO: Transform on a whole tuple lose types, see if we can simplify that? #class PointScaler(ItemTransform): # "Scale a tensor representing points" # def __init__(self, do_scale=True, y_first=False): self.do_scale,self.y_first = do_scale,y_first # def encodes(self, o): return (o[0],TensorPoint(_scale_pnts(*o, self.do_scale, self.y_first))) # def decodes(self, o): return (o[0],TensorPoint(_unscale_pnts(*o))) # #TensorPoint.default_item_tfms = PointScaler ###Output _____no_output_____ ###Markdown To work with data augmentation, and in particular the `grid_sample` method, points need to be represented with coordinates going from -1 to 1 (-1 being top or left, 1 bottom or right), which will be done unless you pass `do_scale=False`. We also need to make sure they are following our convention of points being x,y coordinates, so pass along `y_first=True` if you have your data in an y,x format to add a flip.> Warning: This transform needs to run on the tuple level, before any transform that changes the image size. ###Code #export def _scale_pnts(y, sz, do_scale=True, y_first=False): if y_first: y = y.flip(1) res = y * 2/tensor(sz).float() - 1 if do_scale else y return TensorPoint(res, sz=sz) def _unscale_pnts(y, sz): return TensorPoint((y+1) * tensor(sz).float()/2, sz=sz) #export class PointScaler(Transform): "Scale a tensor representing points" loss_func = MSELossFlat() def __init__(self, do_scale=True, y_first=False): self.do_scale,self.y_first = do_scale,y_first def _grab_sz(self, x): self.sz = [x.shape[-1], x.shape[-2]] if isinstance(x, Tensor) else x.size return x def _get_sz(self, x): sz = getattr(x, '_meta', {}).get('sz', None) assert sz is not None or self.sz is not None, "Size could not be inferred, pass it in the init of your TensorPoint with `sz=...`" return self.sz if sz is None else sz def setup(self, dl): its = dl.do_item(0) for t in its: if isinstance(t, TensorPoint): self.c = t.numel() def encodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) def decodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) def encodes(self, x:TensorPoint): return _scale_pnts(x, self._get_sz(x), self.do_scale, self.y_first) def decodes(self, x:TensorPoint): return _unscale_pnts(x, self._get_sz(x)) TensorPoint.default_item_tfms = PointScaler ###Output _____no_output_____ ###Markdown To work with data augmentation, and in particular the `grid_sample` method, points need to be represented with coordinates going from -1 to 1 (-1 being top or left, 1 bottom or right), which will be done unless you pass `do_scale=False`. We also need to make sure they are following our convention of points being x,y coordinates, so pass along `y_first=True` if you have your data in an y,x format to add a flip.> Note: This transform automatically grabs the sizes of the images it sees before a `TensorPoint` object and embeds it in them. For this to work, those images need to be before any points in the order of your final tuple. If you don't have such images, you need to embed the size of the corresponding image when creating a `TensorPoint` by passing it with `sz=...`. ###Code def _pnt_lbl(x): return TensorPoint.create(pnts) def _pnt_open(fn): return PILImage(PILImage.create(fn).resize((28,35))) pnt_tds = DataSource([mnist_fn], [_pnt_open, [_pnt_lbl]]) pnt_tdl = TfmdDL(pnt_tds, bs=1, after_item=[PointScaler(), ToTensor()]) pnt_tdl.after_item.c x,y = pnt_tdl.one_batch() #Scaling and flipping properly done test_close(y[0], tensor([[-1., -1.], [-1., 1.], [1., -1.], [1., 1.], [9/14-1, 17/17.5-1]])) a,b = pnt_tdl.decode_batch((x,y))[0] test_eq(b, tensor(pnts).float()) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorPoint) test_eq(type(a), TensorImage) test_eq(type(b), TensorPoint) pnt_tdl.show_batch(figsize=(2,2), cmap='Greys'); # export #class BBoxScaler(PointScaler): # "Scale a tensor representing bounding boxes" # def encodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) # def decodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) # # def encodes(self, x:(BBox,TensorBBox)): # pnts = x.bbox.view(-1,2) # scaled_bb = _scale_pnts(pnts, self._get_sz(pnts), self.do_scale, self.y_first) # return TensorBBox((scaled_bb.view(-1,4),x.lbl)) # # def decodes(self, x:(BBox,TensorBBox)): # scaled_bb = _unscale_pnts(x.bbox.view(-1,2), self._get_sz(x.bbox.view(-1,2))) # return TensorBBox((scaled_bb.view(-1,4), x.lbl)) # export #class BBoxCategorize(Transform): # "Reversible transform of category string to `vocab` id" # order,state_args=1,'vocab' # def __init__(self, vocab=None): # self.vocab = vocab # self.o2i = None if vocab is None else {v:k for k,v in enumerate(vocab)} # # def setups(self, dsrc): # if not dsrc: return # vals = set() # for bb in dsrc: vals = vals.union(set(bb.lbl)) # self.vocab,self.otoi = uniqueify(list(vals), sort=True, bidir=True, start='#bg') # # def encodes(self, o:BBox): # return TensorBBox.create((o.bbox,tensor([self.otoi[o_] for o_ in o.lbl if o_ in self.otoi]))) # def decodes(self, o:TensorBBox): # return BBox((o.bbox,[self.vocab[i_] for i_ in o.lbl])) # #BBox.default_type_tfms,BBox.default_item_tfms = BBoxCategorize,BBoxScaler #export #TODO tests #def bb_pad(samples, pad_idx=0): # "Function that collect `samples` of labelled bboxes and adds padding with `pad_idx`." # max_len = max([len(s[1][1]) for s in samples]) # def _f(img,bbox,lbl): # bbox = torch.cat([bbox,bbox.new_zeros(max_len-bbox.shape[0], 4)]) # lbl = torch.cat([lbl, lbl .new_zeros(max_len-lbl .shape[0])+pad_idx]) # return img,TensorBBox((bbox,lbl)) # return [_f(x,*y) for x,y in samples] #export class BBoxLabeler(Transform): def setup(self, dl): self.vocab = dl.vocab def before_call(self): self.bbox,self.lbls = None,None def decode (self, x, **kwargs): self.bbox,self.lbls = None,None return self._call('decodes', x, **kwargs) def decodes(self, x:TensorMultiCategory): self.lbls = [self.vocab[a] for a in x] return x if self.bbox is None else LabeledBBox(self.bbox, self.lbls) def decodes(self, x:TensorBBox): self.bbox = x return self.bbox if self.lbls is None else LabeledBBox(self.bbox, self.lbls) #export #LabeledBBox can be sent in a tl with MultiCategorize (depending on the order of the tls) but it is already decoded. @MultiCategorize def decodes(self, x:LabeledBBox): return x #export @PointScaler def encodes(self, x:TensorBBox): pnts = self.encodes(TensorPoint(x.view(-1,2), sz=x._meta.get('sz', None))) return TensorBBox(pnts.view(-1, 4), sz=x._meta.get('sz', None)) @PointScaler def decodes(self, x:TensorBBox): pnts = self.decodes(TensorPoint(x.view(-1,2), sz=x._meta.get('sz', None))) return TensorBBox(pnts.view(-1, 4), sz=x._meta.get('sz', None)) # export #TODO: merge with padding def clip_remove_empty(bbox, label): "Clip bounding boxes with image border and label background the empty ones." bbox = torch.clamp(bbox, -1, 1) empty = ((bbox[...,2] - bbox[...,0])*(bbox[...,3] - bbox[...,1]) < 0.) return (bbox[~empty], label[~empty]) bb = tensor([[-2,-0.5,0.5,1.5], [-0.5,-0.5,0.5,0.5], [1,0.5,0.5,0.75], [-0.5,-0.5,0.5,0.5]]) bb,lbl = clip_remove_empty(bb, tensor([1,2,3,2])) test_eq(bb, tensor([[-1,-0.5,0.5,1.], [-0.5,-0.5,0.5,0.5], [-0.5,-0.5,0.5,0.5]])) test_eq(lbl, tensor([1,2,2])) #export #TODO tests def bb_pad(samples, pad_idx=0): "Function that collect `samples` of labelled bboxes and adds padding with `pad_idx`." samples = [(s[0], *clip_remove_empty(*s[1:])) for s in samples] max_len = max([len(s[2]) for s in samples]) def _f(img,bbox,lbl): bbox = torch.cat([bbox,bbox.new_zeros(max_len-bbox.shape[0], 4)]) lbl = torch.cat([lbl, lbl .new_zeros(max_len-lbl .shape[0])+pad_idx]) return img,bbox,lbl return [_f(*s) for s in samples] img1,img2 = TensorImage(torch.randn(16,16,3)),TensorImage(torch.randn(16,16,3)) bb1 = tensor([[-2,-0.5,0.5,1.5], [-0.5,-0.5,0.5,0.5], [1,0.5,0.5,0.75], [-0.5,-0.5,0.5,0.5]]) lbl1 = tensor([1, 2, 3, 2]) bb2 = tensor([[-0.5,-0.5,0.5,0.5], [-0.5,-0.5,0.5,0.5]]) lbl2 = tensor([2, 2]) samples = [(img1, bb1, lbl1), (img2, bb2, lbl2)] res = bb_pad(samples) non_empty = tensor([True,True,False,True]) test_eq(res[0][0], img1) test_eq(res[0][1], tensor([[-1,-0.5,0.5,1.], [-0.5,-0.5,0.5,0.5], [-0.5,-0.5,0.5,0.5]])) test_eq(res[0][2], tensor([1,2,2])) test_eq(res[1][0], img2) test_eq(res[1][1], tensor([[-0.5,-0.5,0.5,0.5], [-0.5,-0.5,0.5,0.5], [0,0,0,0]])) test_eq(res[1][2], tensor([2,2,0])) #export TensorBBox.dbunch_kwargs = {'before_batch': bb_pad} def _coco_bb(x): return TensorBBox.create(bbox[0]) def _coco_lbl(x): return bbox[1] coco_tds = DataSource([coco_fn], [PILImage.create, [_coco_bb], [_coco_lbl, MultiCategorize(add_na=True)]], n_inp=1) coco_tdl = TfmdDL(coco_tds, bs=1, after_item=[BBoxLabeler(), PointScaler(), ToTensor()]) x,y,z = coco_tdl.one_batch() test_close(y[0], -1+tensor(bbox[0])/64) test_eq(z[0], tensor([1,1,1])) a,b,c = coco_tdl.decode_batch((x,y,z))[0] test_close(b, tensor(bbox[0]).float()) test_eq(c.bbox, b) test_eq(c.lbl, bbox[1]) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorBBox) test_eq(type(z), TensorMultiCategory) test_eq(type(a), TensorImage) test_eq(type(b), TensorBBox) test_eq(type(c), LabeledBBox) coco_tdl.show_batch(); #hide #test other direction works too coco_tds = DataSource([coco_fn], [PILImage.create, [_coco_lbl, MultiCategorize(add_na=True)], [_coco_bb]]) coco_tdl = TfmdDL(coco_tds, bs=1, after_item=[BBoxLabeler(), PointScaler(), ToTensor()]) x,y,z = coco_tdl.one_batch() test_close(z[0], -1+tensor(bbox[0])/64) test_eq(y[0], tensor([1,1,1])) a,b,c = coco_tdl.decode_batch((x,y,z))[0] test_eq(b, bbox[1]) test_close(c.bbox, tensor(bbox[0]).float()) test_eq(c.lbl, b) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorMultiCategory) test_eq(type(z), TensorBBox) test_eq(type(a), TensorImage) test_eq(type(b), MultiCategory) test_eq(type(c), LabeledBBox) ###Output _____no_output_____ ###Markdown Show methods ###Code #export def _get_grid(n, rows=None, cols=None, add_vert=0, figsize=None, double=False): rows = rows or int(np.ceil(math.sqrt(n))) cols = cols or int(np.ceil(n/rows)) if double: cols*=2 ; n*=2 figsize = (cols*3, rows*3+add_vert) if figsize is None else figsize _,axs = subplots(rows, cols, figsize=figsize) axs = axs.flatten() for ax in axs[n:]: ax.set_axis_off() return axs #export @typedispatch def show_batch(x:TensorImage, y, its, ctxs=None, max_n=10, rows=None, cols=None, figsize=None, **kwargs): if ctxs is None: ctxs = _get_grid(min(len(its), max_n), rows=rows, cols=cols, figsize=figsize) ctxs = default_show_batch(x, y, its, ctxs=ctxs, max_n=max_n, **kwargs) return ctxs #export @typedispatch def show_results(x:TensorImage, y, its, ctxs=None, max_n=10, rows=None, cols=None, figsize=None, **kwargs): if ctxs is None: ctxs = _get_grid(min(len(its), max_n), rows=rows, cols=cols, add_vert=1, figsize=figsize) ctxs = default_show_results(x, y, its, ctxs=ctxs, max_n=max_n, **kwargs) return ctxs #export @typedispatch def show_results(x:TensorImage, y:TensorCategory, its, ctxs=None, max_n=10, rows=None, cols=None, figsize=None, **kwargs): if ctxs is None: ctxs = _get_grid(min(len(its), max_n), rows=rows, cols=cols, add_vert=1, figsize=figsize) for i in range(2): ctxs = [b.show(ctx=c, **kwargs) for b,c,_ in zip(its.itemgot(i),ctxs,range(max_n))] ctxs = [r.show(ctx=c, color='green' if b==r else 'red', **kwargs) for b,r,c,_ in zip(its.itemgot(1),its.itemgot(2),ctxs,range(max_n))] return ctxs #export @typedispatch def show_results(x:TensorImage, y:(TensorImageBase, TensorPoint, TensorBBox), its, ctxs=None, max_n=10, rows=None, cols=None, figsize=None, **kwargs): if ctxs is None: ctxs = _get_grid(min(len(its), max_n), rows=rows, cols=cols, add_vert=1, figsize=figsize, double=True) for i in range(2): ctxs[::2] = [b.show(ctx=c, **kwargs) for b,c,_ in zip(its.itemgot(i),ctxs[::2],range(max_n))] for i in [0,2]: ctxs[1::2] = [b.show(ctx=c, **kwargs) for b,c,_ in zip(its.itemgot(i),ctxs[1::2],range(max_n))] return ctxs ###Output _____no_output_____ ###Markdown Export - ###Code #hide from local.notebook.export import notebook2script notebook2script(all_fs=True) ###Output Converted 00_test.ipynb. Converted 01_core.ipynb. Converted 01a_utils.ipynb. Converted 01b_dispatch.ipynb. Converted 01c_transform.ipynb. Converted 02_script.ipynb. Converted 03_torch_core.ipynb. Converted 03a_layers.ipynb. Converted 04_dataloader.ipynb. Converted 05_data_core.ipynb. Converted 06_data_transforms.ipynb. Converted 07_data_block.ipynb. Converted 08_vision_core.ipynb. Converted 09_vision_augment.ipynb. Converted 10_pets_tutorial.ipynb. Converted 11_vision_models_xresnet.ipynb. Converted 12_optimizer.ipynb. Converted 13_learner.ipynb. Converted 13a_metrics.ipynb. Converted 14_callback_schedule.ipynb. Converted 14a_callback_data.ipynb. Converted 15_callback_hook.ipynb. Converted 15a_vision_models_unet.ipynb. Converted 16_callback_progress.ipynb. Converted 17_callback_tracker.ipynb. Converted 18_callback_fp16.ipynb. Converted 19_callback_mixup.ipynb. Converted 21_vision_learner.ipynb. Converted 22_tutorial_imagenette.ipynb. Converted 23_tutorial_transfer_learning.ipynb. Converted 30_text_core.ipynb. Converted 31_text_data.ipynb. Converted 32_text_models_awdlstm.ipynb. Converted 33_text_models_core.ipynb. Converted 34_callback_rnn.ipynb. Converted 35_tutorial_wikitext.ipynb. Converted 36_text_models_qrnn.ipynb. Converted 37_text_learner.ipynb. Converted 38_tutorial_ulmfit.ipynb. Converted 40_tabular_core.ipynb. Converted 41_tabular_model.ipynb. Converted 42_tabular_rapids.ipynb. Converted 50_data_block_examples.ipynb. Converted 60_medical_imaging.ipynb. Converted 65_medical_text.ipynb. Converted 90_notebook_core.ipynb. Converted 91_notebook_export.ipynb. Converted 92_notebook_showdoc.ipynb. Converted 93_notebook_export2html.ipynb. Converted 94_notebook_test.ipynb. Converted 95_index.ipynb. Converted 96_data_external.ipynb. Converted 97_utils_test.ipynb. Converted notebook2jekyll.ipynb. ###Markdown Core vision> Basic image opening/processing functionality Useful stats ###Code #export imagenet_stats = ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) cifar_stats = ([0.491, 0.482, 0.447], [0.247, 0.243, 0.261]) mnist_stats = ([0.131], [0.308]) ###Output _____no_output_____ ###Markdown Helpers ###Code im = Image.open(TEST_IMAGE).resize((30,20)) #export if not hasattr(Image,'_patched'): _old_sz = Image.Image.size.fget @patch_property def size(x:Image.Image): return Tuple(_old_sz(x)) Image._patched = True #export @patch_property def n_px(x: Image.Image): return x.size[0] * x.size[1] ###Output _____no_output_____ ###Markdown `Image.n_px`> `Image.n_px` (property)Number of pixels in image ###Code test_eq(im.n_px, 30*20) #export @patch_property def shape(x: Image.Image): return x.size[1],x.size[0] ###Output _____no_output_____ ###Markdown `Image.shape`> `Image.shape` (property)Image (height,width) tuple (NB: opposite order of `Image.size()`, same order as numpy array and pytorch tensor) ###Code test_eq(im.shape, (20,30)) #export @patch_property def aspect(x: Image.Image): return x.size[0]/x.size[1] ###Output _____no_output_____ ###Markdown `Image.aspect`> `Image.aspect` (property)Aspect ratio of the image, i.e. `width/height` ###Code test_eq(im.aspect, 30/20) #export @patch def reshape(x: Image.Image, h, w, resample=0): "`resize` `x` to `(w,h)`" return x.resize((w,h), resample=resample) show_doc(Image.Image.reshape) test_eq(im.reshape(12,10).shape, (12,10)) a,b = L([5,2]).map(lambda x: math.floor(x * 4/3)) a,b #export @patch def resize_max(x: Image.Image, resample=0, max_px=None, max_h=None, max_w=None): "`resize` `x` to `max_px`, or `max_h`, or `max_w`" h,w = x.shape if max_px and x.n_px>max_px: h,w = Tuple(h,w).mul(math.sqrt(max_px/x.n_px)) if max_h and h>max_h: h,w = (max_h ,max_h*w/h) if max_w and w>max_w: h,w = (max_w*h/w,max_w ) return x.reshape(round(h), round(w), resample=resample) test_eq(im.resize_max(max_px=20*30).shape, (20,30)) test_eq(im.resize_max(max_px=300).n_px, 294) test_eq(im.resize_max(max_px=500, max_h=10, max_w=20).shape, (10,15)) test_eq(im.resize_max(max_px=500, max_h=14, max_w=15).shape, (10,15)) test_eq(im.resize_max(max_px=300, max_h=10, max_w=25).shape, (10,15)) show_doc(Image.Image.resize_max) #TODO function to resize_max all images in a path (optionally recursively) and save them somewhere (same relative dirs if recursive) ###Output _____no_output_____ ###Markdown Basic types This section regroups the basic types used in vision with the transform that create objects of those types. ###Code # TODO: docs #export def load_image(fn, mode=None, **kwargs): "Open and load a `PIL.Image` and convert to `mode`" im = Image.open(fn, **kwargs) im.load() im = im._new(im.im) return im.convert(mode) if mode else im #export class PILBase(Image.Image, metaclass=BypassNewMeta): _bypass_type=Image.Image default_batch_tfms = ByteToFloatTensor _show_args = {'cmap':'viridis'} _open_args = {'mode': 'RGB'} @classmethod def create(cls, fn, **kwargs)->None: "Open an `Image` from path `fn`" if isinstance(fn,Tensor): fn = fn.numpy() if isinstance(fn,ndarray): return cls(Image.fromarray(fn)) return cls(load_image(fn, **merge(cls._open_args, kwargs))) def show(self, ctx=None, **kwargs): "Show image using `merge(self._show_args, kwargs)`" return show_image(self, ctx=ctx, **merge(self._show_args, kwargs)) #export class PILImage(PILBase): pass #export class PILImageBW(PILImage): _show_args,_open_args = {'cmap':'Greys'},{'mode': 'L'} im = PILImage.create(TEST_IMAGE) test_eq(type(im), PILImage) test_eq(im.mode, 'RGB') im.resize((64,64)) ax = im.show(figsize=(1,1)) test_fig_exists(ax) #export class PILMask(PILBase): _open_args,_show_args = {'mode':'L'},{'alpha':0.5, 'cmap':'tab20'} im = PILMask.create(TEST_IMAGE) test_eq(type(im), PILMask) test_eq(im.mode, 'L') ###Output _____no_output_____ ###Markdown Images ###Code mnist = untar_data(URLs.MNIST_TINY) fns = get_image_files(mnist) mnist_fn = fns[0]; mnist_fn timg = Transform(PILImageBW.create) mnist_img = timg(mnist_fn) test_eq(mnist_img.size, (28,28)) assert isinstance(mnist_img, PILImageBW) mnist_img ###Output _____no_output_____ ###Markdown Segmentation masks ###Code camvid = untar_data(URLs.CAMVID_TINY) fns = get_image_files(camvid/'images') cam_fn = fns[0] mask_fn = camvid/'labels'/f'{cam_fn.stem}_P{cam_fn.suffix}' cam_img = PILImage.create(cam_fn) test_eq(cam_img.size, (128,96)) tmask = Transform(PILMask.create) mask = tmask(mask_fn) test_eq(type(mask), PILMask) test_eq(mask.size, (128,96)) _,axs = plt.subplots(1,3, figsize=(12,3)) cam_img.show(ctx=axs[0], title='image') mask.show(alpha=1, ctx=axs[1], vmin=1, vmax=30, title='mask') cam_img.show(ctx=axs[2], title='superimposed') mask.show(ctx=axs[2], vmin=1, vmax=30); ###Output _____no_output_____ ###Markdown Points ###Code # export class TensorPoint(TensorBase): "Basic type for points in an image" _show_args = dict(s=10, marker='.', c='r') @classmethod def create(cls, t)->None: "Convert an array or a list of points `t` to a `Tensor`" return cls(tensor(t).view(-1, 2).float()) def show(self, ctx=None, **kwargs): if 'figsize' in kwargs: del kwargs['figsize'] ctx.scatter(self[:, 0], self[:, 1], **{**self._show_args, **kwargs}) return ctx ###Output _____no_output_____ ###Markdown Points are expected to come as an array/tensor of shape `(n,2)` or as a list of lists with two elements. Unless you change the defaults in `PointScaler` (see later on), coordinates should go from 0 to width/height, with the first one being the column index (so from 0 to width) and the second one being the row index (so from 0 to height).> Note: This is differnt from the usual indeixing convention for arrays in numpy or in PyTorch, but it's the way points are expected by matplotlib or the internal functions in PyTorch like `F.grid_sample`. ###Code pnt_img = TensorImage(mnist_img.resize((28,35))) pnts = np.array([[0,0], [0,35], [28,0], [28,35], [9, 17]]) tfm = Transform(TensorPoint.create) tpnts = tfm(pnts) test_eq(tpnts.shape, [5,2]) test_eq(tpnts.dtype, torch.float32) ctx = pnt_img.show(figsize=(1,1), cmap='Greys') tpnts.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Bounding boxes ###Code # export def get_annotations(fname, prefix=None): "Open a COCO style json in `fname` and returns the lists of filenames (with maybe `prefix`) and labelled bboxes." annot_dict = json.load(open(fname)) id2images, id2bboxes, id2cats = {}, collections.defaultdict(list), collections.defaultdict(list) classes = {o['id']:o['name'] for o in annot_dict['categories']} for o in annot_dict['annotations']: bb = o['bbox'] id2bboxes[o['image_id']].append([bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]]) id2cats[o['image_id']].append(classes[o['category_id']]) id2images = {o['id']:ifnone(prefix, '') + o['file_name'] for o in annot_dict['images'] if o['id'] in id2bboxes} ids = list(id2images.keys()) return [id2images[k] for k in ids], [(id2bboxes[k], id2cats[k]) for k in ids] #hide #TODO explain and/or simplify this coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') annots = json.load(open(coco/'train.json')) test_eq(images, [k['file_name'] for k in annots['images']]) for _ in range(5): idx = random.randint(0, len(images)-1) fn = images[idx] i = 0 while annots['images'][i]['file_name'] != fn: i+=1 img_id = annots['images'][i]['id'] bbs = [ann for ann in annots['annotations'] if ann['image_id'] == img_id] i2o = {k['id']:k['name'] for k in annots['categories']} lbls = [i2o[bb['category_id']] for bb in bbs] bboxes = [bb['bbox'] for bb in bbs] bboxes = [[bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]] for bb in bboxes] test_eq(lbl_bbox[idx], [bboxes, lbls]) # export from matplotlib import patches, patheffects def _draw_outline(o, lw): o.set_path_effects([patheffects.Stroke(linewidth=lw, foreground='black'), patheffects.Normal()]) def _draw_rect(ax, b, color='white', text=None, text_size=14, hw=True, rev=False): lx,ly,w,h = b if rev: lx,ly,w,h = ly,lx,h,w if not hw: w,h = w-lx,h-ly patch = ax.add_patch(patches.Rectangle((lx,ly), w, h, fill=False, edgecolor=color, lw=2)) _draw_outline(patch, 4) if text is not None: patch = ax.text(lx,ly, text, verticalalignment='top', color=color, fontsize=text_size, weight='bold') _draw_outline(patch,1) # export class BBox(Tuple): "Basic type for a list of bounding boxes in an image" def show(self, ctx=None, **kwargs): for b,l in zip(self.bbox, self.lbl): if l != '#bg': _draw_rect(ctx, b, hw=False, text=l) return ctx @classmethod def create(cls, x): return cls(x) bbox,lbl = add_props(lambda i,self: self[i]) # export class TensorBBox(Tuple): "Basic type for a tensor of bounding boxes in an image" @classmethod def create(cls, x): return cls(tensor(x[0]).view(-1, 4).float(), x[1]) bbox,lbl = add_props(lambda i,self: self[i]) ###Output _____no_output_____ ###Markdown Bounding boxes are expected to come as tuple with an array/tensor of shape `(n,4)` or as a list of lists with four elements adn a list of corresponding labels. Unless you change the defaults in `BBoxScaler` (see later on), coordinates for each bounding box should go from 0 to height/width, with the following convetion: top, left, bottom, right.> Note: We use the same convention as for points with y axis being before x. ###Code coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') idx=2 coco_fn,bbox = coco/'train'/images[idx],lbl_bbox[idx] coco_img = timg(coco_fn) tbbox = BBox(bbox) ctx = coco_img.show(figsize=(3,3), cmap='Greys') tbbox.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Basic Transforms Unless specifically metioned, all the following transforms can be used as single-item transforms (in one of the list in the `tfms` you pass to a `TfmdDS` or a `Datasource`) or tuple transform (in the `tuple_tfms` you pass to a `TfmdDS` or a `Datasource`). The safest way that will work accross applications is to always use them as `tuple_tfms`. For instance, if you have points or bounding boxes as targets and use `ImageResizer` as a single-item transform, when you get to `PointScaler` or `BBoxScaler` (which are tuple transforms) you won't have the correct size of the image to properly scale your points. ###Code class ImageResizer(Transform): order=10 "Resize image to `size` using `resample" def __init__(self, size, resample=Image.BILINEAR): if not is_listy(size): size=(size,size) self.size,self.resample = (size[1],size[0]),resample def encodes(self, o:PILImage): return o.resize(size=self.size, resample=self.resample) def encodes(self, o:PILMask): return o.resize(size=self.size, resample=Image.NEAREST) ###Output _____no_output_____ ###Markdown `size` can either be one integer (in which case images are resized to a square) or a tuple `height,width`.> Note: This is the usual convention for arrays or in PyTorch, but it's not the usual convention for PIL Image, which use the other way round. ###Code f = ImageResizer(14) test_eq(f(mnist_img).size, (14,14)) test_eq(f(mask).size, (14,14)) f = ImageResizer((32,28)) test_eq(f(mnist_img).size, (28,32))#PIL has width first test_eq(array(f(mnist_img)).shape, (32,28))#But numpy as height first and that is our convention # export def image2byte(img): "Transform image to byte tensor in `c*h*w` dim order." res = torch.ByteTensor(torch.ByteStorage.from_buffer(img.tobytes())) w,h = img.size return res.view(h,w,-1).permute(2,0,1) #export @ToTensor def encodes(self, o:PILImage): return TensorImage(image2byte(o)) @ToTensor def encodes(self, o:PILImageBW): return TensorImageBW(image2byte(o)) @ToTensor def encodes(self, o:PILMask): return TensorMask(image2byte(o)[0]) ###Output _____no_output_____ ###Markdown Any data augmentation transform that runs on PIL Images must be run before this transform. ###Code tfm = ToTensor() print(tfm) print(type(mnist_img)) print(type(tfm(mnist_img))) tfm = ToTensor() test_eq(tfm(mnist_img).shape, (1,28,28)) test_eq(type(tfm(mnist_img)), TensorImageBW) test_eq(tfm(mask).shape, (96,128)) test_eq(type(tfm(mask)), TensorMask) ###Output _____no_output_____ ###Markdown Let's confirm we can pipeline this with `PILImage.create`. ###Code pipe_img = Pipeline([PILImageBW.create, ToTensor()]) img = pipe_img(mnist_fn) pipe_img.show(img, figsize=(1,1)); def _cam_lbl(x): return mask_fn cam_tds = DataSource([cam_fn], [[PILImage.create, ToTensor()], [_cam_lbl, PILMask.create, ToTensor()]]) show_at(cam_tds, 0); #export def _scale_pnts(x, y, do_scale=True,y_first=False): if y_first: y = y.flip(1) sz = [x.shape[-1], x.shape[-2]] if isinstance(x, Tensor) else x.size return y * 2/tensor(sz).float() - 1 if do_scale else y def _unscale_pnts(x, y): sz = [x.shape[-1], x.shape[-2]] if isinstance(x, Tensor) else x.size return (y+1) * tensor(sz).float()/2 # export #TODO: Transform on a whole tuple lose types, see if we can simplify that? class PointScaler(ItemTransform): "Scale a tensor representing points" def __init__(self, do_scale=True, y_first=False): self.do_scale,self.y_first = do_scale,y_first def encodes(self, o): return (o[0],TensorPoint(_scale_pnts(*o, self.do_scale, self.y_first))) def decodes(self, o): return (o[0],TensorPoint(_unscale_pnts(*o))) TensorPoint.default_item_tfms = PointScaler ###Output _____no_output_____ ###Markdown To work with data augmentation, and in particular the `grid_sample` method, points need to be represented with coordinates going from -1 to 1 (-1 being top or left, 1 bottom or right), which will be done unless you pass `do_scale=False`. We also need to make sure they are following our convention of points being x,y coordinates, so pass along `y_first=True` if you have your data in an y,x format to add a flip.> Warning: This transform needs to run on the tuple level, before any transform that changes the image size. ###Code def _pnt_lbl(x): return TensorPoint.create(pnts) def _pnt_open(fn): return PILImage(PILImage.create(fn).resize((28,35))) pnt_tds = DataSource([mnist_fn], [_pnt_open, [_pnt_lbl]]) pnt_tdl = TfmdDL(pnt_tds, bs=1, after_item=[PointScaler(), ToTensor()]) x,y = pnt_tdl.one_batch() #Scaling and flipping properly done test_close(y[0], tensor([[-1., -1.], [-1., 1.], [1., -1.], [1., 1.], [9/14-1, 17/17.5-1]])) a,b = pnt_tdl.decode_batch((x,y))[0] test_eq(b, tensor(pnts).float()) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorPoint) test_eq(type(a), TensorImage) test_eq(type(b), TensorPoint) pnt_tdl.show_batch(figsize=(2,2), cmap='Greys'); # export class BBoxScaler(PointScaler): "Scale a tensor representing bounding boxes" def encodes(self, o): x,y = o scaled_bb = _scale_pnts(x, y.bbox.view(-1,2), self.do_scale, self.y_first) return (x,TensorBBox((scaled_bb.view(-1,4),y.lbl))) def decodes(self, o): x,y = o scaled_bb = _unscale_pnts(x, y.bbox.view(-1,2)) return (x, TensorBBox((scaled_bb.view(-1,4), y.lbl))) # export class BBoxCategorize(Transform): "Reversible transform of category string to `vocab` id" order,state_args=1,'vocab' def __init__(self, vocab=None): self.vocab = vocab self.o2i = None if vocab is None else {v:k for k,v in enumerate(vocab)} def setups(self, dsrc): if not dsrc: return vals = set() for bb in dsrc: vals = vals.union(set(bb.lbl)) self.vocab,self.otoi = uniqueify(list(vals), sort=True, bidir=True, start='#bg') def encodes(self, o:BBox): return TensorBBox.create((o.bbox,tensor([self.otoi[o_] for o_ in o.lbl if o_ in self.otoi]))) def decodes(self, o:TensorBBox): return BBox((o.bbox,[self.vocab[i_] for i_ in o.lbl])) BBox.default_type_tfms,BBox.default_item_tfms = BBoxCategorize,BBoxScaler #export #TODO tests def bb_pad(samples, pad_idx=0): "Function that collect `samples` of labelled bboxes and adds padding with `pad_idx`." max_len = max([len(s[1][1]) for s in samples]) def _f(img,bbox,lbl): bbox = torch.cat([bbox,bbox.new_zeros(max_len-bbox.shape[0], 4)]) lbl = torch.cat([lbl, lbl .new_zeros(max_len-lbl .shape[0])+pad_idx]) return img,TensorBBox((bbox,lbl)) return [_f(x,*y) for x,y in samples] def _coco_lbl(x): return BBox(bbox) tcat = BBoxCategorize() coco_tds = DataSource([coco_fn], [PILImage.create, [_coco_lbl, tcat]]) coco_tdl = TfmdDL(coco_tds, bs=1, after_item=[BBoxScaler(), ToTensor()]) x,y = coco_tdl.one_batch() y0 = y[0][0],y[1][0] #Scaling and flipping properly done test_close(y0[0], -1+tensor(bbox[0])/64) test_eq(y0[1], tensor([1,1,1])) a,b = coco_tdl.decode_batch((x,y))[0] test_close(b[0], tensor(bbox[0]).float()) test_eq(b[1], bbox[1]) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorBBox) test_eq(type(a), TensorImage) test_eq(type(b), BBox) coco_tdl.show_batch(); ###Output _____no_output_____ ###Markdown Show methods ###Code #export def _get_grid(n, rows=None, cols=None, add_vert=0, figsize=None, double=False): rows = rows or int(np.ceil(math.sqrt(n))) cols = cols or int(np.ceil(n/rows)) if double: cols*=2 ; n*=2 figsize = (cols*3, rows*3+add_vert) if figsize is None else figsize _,axs = subplots(rows, cols, figsize=figsize) axs = axs.flatten() for ax in axs[n:]: ax.set_axis_off() return axs #export @typedispatch def show_batch(x:TensorImage, y, its, ctxs=None, max_n=10, rows=None, cols=None, figsize=None, **kwargs): if ctxs is None: ctxs = _get_grid(min(len(its), max_n), rows=rows, cols=cols, figsize=figsize) ctxs = default_show_batch(x, y, its, ctxs=ctxs, max_n=max_n, **kwargs) return ctxs #export @typedispatch def show_results(x:TensorImage, y, its, ctxs=None, max_n=10, rows=None, cols=None, figsize=None, **kwargs): if ctxs is None: ctxs = _get_grid(min(len(its), max_n), rows=rows, cols=cols, add_vert=1, figsize=figsize) ctxs = default_show_results(x, y, its, ctxs=ctxs, max_n=max_n, **kwargs) return ctxs #export @typedispatch def show_results(x:TensorImage, y:TensorCategory, its, ctxs=None, max_n=10, rows=None, cols=None, figsize=None, **kwargs): if ctxs is None: ctxs = _get_grid(min(len(its), max_n), rows=rows, cols=cols, add_vert=1, figsize=figsize) for i in range(2): ctxs = [b.show(ctx=c, **kwargs) for b,c,_ in zip(its.itemgot(i),ctxs,range(max_n))] ctxs = [r.show(ctx=c, color='green' if b==r else 'red', **kwargs) for b,r,c,_ in zip(its.itemgot(1),its.itemgot(2),ctxs,range(max_n))] return ctxs #export @typedispatch def show_results(x:TensorImage, y:TensorImageBase, its, ctxs=None, max_n=10, rows=None, cols=None, figsize=None, **kwargs): if ctxs is None: ctxs = _get_grid(min(len(its), max_n), rows=rows, cols=cols, add_vert=1, figsize=figsize, double=True) for i in range(2): ctxs[::2] = [b.show(ctx=c, **kwargs) for b,c,_ in zip(its.itemgot(i),ctxs[::2],range(max_n))] for i in [0,2]: ctxs[1::2] = [b.show(ctx=c, **kwargs) for b,c,_ in zip(its.itemgot(i),ctxs[1::2],range(max_n))] return ctxs ###Output _____no_output_____ ###Markdown Export - ###Code #hide from local.notebook.export import notebook2script notebook2script(all_fs=True) ###Output Converted 00_test.ipynb. Converted 01_core.ipynb. Converted 01a_utils.ipynb. Converted 01b_dispatch.ipynb. Converted 01c_transform.ipynb. Converted 02_script.ipynb. Converted 03_torch_core.ipynb. Converted 04_dataloader.ipynb. Converted 05_data_core.ipynb. Converted 06_data_transforms.ipynb. Converted 07_data_block.ipynb. Converted 08_vision_core.ipynb. Converted 09_vision_augment.ipynb. Converted 10_pets_tutorial.ipynb. Converted 11_layers.ipynb. Converted 11a_vision_models_xresnet.ipynb. Converted 12_optimizer.ipynb. Converted 13_learner.ipynb. Converted 14_callback_schedule.ipynb. Converted 14a_callback_data.ipynb. Converted 15_callback_hook.ipynb. Converted 15a_vision_models_unet.ipynb. Converted 16_callback_progress.ipynb. Converted 17_callback_tracker.ipynb. Converted 18_callback_fp16.ipynb. Converted 19_callback_mixup.ipynb. Converted 20_metrics.ipynb. Converted 21_vision_learner.ipynb. Converted 22_tutorial_imagenette.ipynb. Converted 23_tutorial_transfer_learning.ipynb. Converted 30_text_core.ipynb. Converted 31_text_data.ipynb. Converted 32_text_models_awdlstm.ipynb. Converted 33_text_models_core.ipynb. Converted 34_callback_rnn.ipynb. Converted 35_tutorial_wikitext.ipynb. Converted 36_text_models_qrnn.ipynb. Converted 37_text_learner.ipynb. Converted 38_tutorial_ulmfit.ipynb. Converted 40_tabular_core.ipynb. Converted 41_tabular_model.ipynb. Converted 42_tabular_rapids.ipynb. Converted 50_data_block_examples.ipynb. Converted 60_medical_imaging.ipynb. Converted 65_medical_text.ipynb. Converted 90_notebook_core.ipynb. Converted 91_notebook_export.ipynb. Converted 92_notebook_showdoc.ipynb. Converted 93_notebook_export2html.ipynb. Converted 94_notebook_test.ipynb. Converted 95_index.ipynb. Converted 96_data_external.ipynb. Converted 97_utils_test.ipynb. Converted notebook2jekyll.ipynb. ###Markdown Core vision> Basic image opening/processing functionality Helpers ###Code #export imagenet_stats = ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) cifar_stats = ([0.491, 0.482, 0.447], [0.247, 0.243, 0.261]) mnist_stats = ([0.131], [0.308]) im = Image.open(TEST_IMAGE).resize((30,20)) #export if not hasattr(Image,'_patched'): _old_sz = Image.Image.size.fget @patch_property def size(x:Image.Image): return Tuple(_old_sz(x)) Image._patched = True #export @patch_property def n_px(x: Image.Image): return x.size[0] * x.size[1] ###Output _____no_output_____ ###Markdown `Image.n_px`> `Image.n_px` (property)Number of pixels in image ###Code test_eq(im.n_px, 30*20) #export @patch_property def shape(x: Image.Image): return x.size[1],x.size[0] ###Output _____no_output_____ ###Markdown `Image.shape`> `Image.shape` (property)Image (height,width) tuple (NB: opposite order of `Image.size()`, same order as numpy array and pytorch tensor) ###Code test_eq(im.shape, (20,30)) #export @patch_property def aspect(x: Image.Image): return x.size[0]/x.size[1] ###Output _____no_output_____ ###Markdown `Image.aspect`> `Image.aspect` (property)Aspect ratio of the image, i.e. `width/height` ###Code test_eq(im.aspect, 30/20) #export @patch def reshape(x: Image.Image, h, w, resample=0): "`resize` `x` to `(w,h)`" return x.resize((w,h), resample=resample) show_doc(Image.Image.reshape) test_eq(im.reshape(12,10).shape, (12,10)) #export @patch def resize_max(x: Image.Image, resample=0, max_px=None, max_h=None, max_w=None): "`resize` `x` to `max_px`, or `max_h`, or `max_w`" h,w = x.shape if max_px and x.n_px>max_px: h,w = Tuple(h,w).mul(math.sqrt(max_px/x.n_px)) if max_h and h>max_h: h,w = (max_h ,max_h*w/h) if max_w and w>max_w: h,w = (max_w*h/w,max_w ) return x.reshape(round(h), round(w), resample=resample) test_eq(im.resize_max(max_px=20*30).shape, (20,30)) test_eq(im.resize_max(max_px=300).n_px, 294) test_eq(im.resize_max(max_px=500, max_h=10, max_w=20).shape, (10,15)) test_eq(im.resize_max(max_px=500, max_h=14, max_w=15).shape, (10,15)) test_eq(im.resize_max(max_px=300, max_h=10, max_w=25).shape, (10,15)) show_doc(Image.Image.resize_max) #TODO function to resize_max all images in a path (optionally recursively) and save them somewhere (same relative dirs if recursive) ###Output _____no_output_____ ###Markdown Basic types This section regroups the basic types used in vision with the transform that create objects of those types. ###Code # TODO: docs #export def load_image(fn, mode=None, **kwargs): "Open and load a `PIL.Image` and convert to `mode`" im = Image.open(fn, **kwargs) im.load() im = im._new(im.im) return im.convert(mode) if mode else im #export class PILBase(Image.Image, metaclass=BypassNewMeta): _bypass_type=Image.Image default_batch_tfms = IntToFloatTensor _show_args = {'cmap':'viridis'} _open_args = {'mode': 'RGB'} @classmethod def create(cls, fn, **kwargs)->None: "Open an `Image` from path `fn`" if isinstance(fn,Tensor): fn = fn.numpy() if isinstance(fn,ndarray): return cls(Image.fromarray(fn)) return cls(load_image(fn, **merge(cls._open_args, kwargs))) def show(self, ctx=None, **kwargs): "Show image using `merge(self._show_args, kwargs)`" return show_image(self, ctx=ctx, **merge(self._show_args, kwargs)) #export class PILImage(PILBase): pass #export class PILImageBW(PILImage): _show_args,_open_args = {'cmap':'Greys'},{'mode': 'L'} im = PILImage.create(TEST_IMAGE) test_eq(type(im), PILImage) test_eq(im.mode, 'RGB') im.resize((64,64)) ax = im.show(figsize=(1,1)) test_fig_exists(ax) #export class PILMask(PILBase): _open_args,_show_args = {'mode':'L'},{'alpha':0.5, 'cmap':'tab20'} im = PILMask.create(TEST_IMAGE) test_eq(type(im), PILMask) test_eq(im.mode, 'L') #export OpenMask = Transform(PILMask.create) OpenMask.loss_func = CrossEntropyLossFlat(axis=1) PILMask.create = OpenMask ###Output _____no_output_____ ###Markdown Images ###Code mnist = untar_data(URLs.MNIST_TINY) fns = get_image_files(mnist) mnist_fn = TEST_IMAGE_BW timg = Transform(PILImageBW.create) mnist_img = timg(mnist_fn) test_eq(mnist_img.size, (28,28)) assert isinstance(mnist_img, PILImageBW) mnist_img ###Output _____no_output_____ ###Markdown Segmentation masks ###Code camvid = untar_data(URLs.CAMVID_TINY) fns = get_image_files(camvid/'images') cam_fn = fns[0] mask_fn = camvid/'labels'/f'{cam_fn.stem}_P{cam_fn.suffix}' cam_img = PILImage.create(cam_fn) test_eq(cam_img.size, (128,96)) tmask = Transform(PILMask.create) mask = tmask(mask_fn) test_eq(type(mask), PILMask) test_eq(mask.size, (128,96)) _,axs = plt.subplots(1,3, figsize=(12,3)) cam_img.show(ctx=axs[0], title='image') mask.show(alpha=1, ctx=axs[1], vmin=1, vmax=30, title='mask') cam_img.show(ctx=axs[2], title='superimposed') mask.show(ctx=axs[2], vmin=1, vmax=30); ###Output _____no_output_____ ###Markdown Points ###Code # export class TensorPoint(TensorBase): "Basic type for points in an image" _show_args = dict(s=10, marker='.', c='r') @classmethod def create(cls, t, sz=None)->None: "Convert an array or a list of points `t` to a `Tensor`" return cls(tensor(t).view(-1, 2).float(), sz=sz) def show(self, ctx=None, **kwargs): if 'figsize' in kwargs: del kwargs['figsize'] x = self.view(-1,2) ctx.scatter(x[:, 0], x[:, 1], **{**self._show_args, **kwargs}) return ctx ###Output _____no_output_____ ###Markdown Points are expected to come as an array/tensor of shape `(n,2)` or as a list of lists with two elements. Unless you change the defaults in `PointScaler` (see later on), coordinates should go from 0 to width/height, with the first one being the column index (so from 0 to width) and the second one being the row index (so from 0 to height).> Note: This is differnt from the usual indeixing convention for arrays in numpy or in PyTorch, but it's the way points are expected by matplotlib or the internal functions in PyTorch like `F.grid_sample`. ###Code pnt_img = TensorImage(mnist_img.resize((28,35))) pnts = np.array([[0,0], [0,35], [28,0], [28,35], [9, 17]]) tfm = Transform(TensorPoint.create) tpnts = tfm(pnts) test_eq(tpnts.shape, [5,2]) test_eq(tpnts.dtype, torch.float32) ctx = pnt_img.show(figsize=(1,1), cmap='Greys') tpnts.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Bounding boxes ###Code # export def get_annotations(fname, prefix=None): "Open a COCO style json in `fname` and returns the lists of filenames (with maybe `prefix`) and labelled bboxes." annot_dict = json.load(open(fname)) id2images, id2bboxes, id2cats = {}, collections.defaultdict(list), collections.defaultdict(list) classes = {o['id']:o['name'] for o in annot_dict['categories']} for o in annot_dict['annotations']: bb = o['bbox'] id2bboxes[o['image_id']].append([bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]]) id2cats[o['image_id']].append(classes[o['category_id']]) id2images = {o['id']:ifnone(prefix, '') + o['file_name'] for o in annot_dict['images'] if o['id'] in id2bboxes} ids = list(id2images.keys()) return [id2images[k] for k in ids], [(id2bboxes[k], id2cats[k]) for k in ids] #hide #TODO explain and/or simplify this coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') annots = json.load(open(coco/'train.json')) test_eq(images, [k['file_name'] for k in annots['images']]) for _ in range(5): idx = random.randint(0, len(images)-1) fn = images[idx] i = 0 while annots['images'][i]['file_name'] != fn: i+=1 img_id = annots['images'][i]['id'] bbs = [ann for ann in annots['annotations'] if ann['image_id'] == img_id] i2o = {k['id']:k['name'] for k in annots['categories']} lbls = [i2o[bb['category_id']] for bb in bbs] bboxes = [bb['bbox'] for bb in bbs] bboxes = [[bb[0],bb[1], bb[0]+bb[2], bb[1]+bb[3]] for bb in bboxes] test_eq(lbl_bbox[idx], [bboxes, lbls]) # export from matplotlib import patches, patheffects def _draw_outline(o, lw): o.set_path_effects([patheffects.Stroke(linewidth=lw, foreground='black'), patheffects.Normal()]) def _draw_rect(ax, b, color='white', text=None, text_size=14, hw=True, rev=False): lx,ly,w,h = b if rev: lx,ly,w,h = ly,lx,h,w if not hw: w,h = w-lx,h-ly patch = ax.add_patch(patches.Rectangle((lx,ly), w, h, fill=False, edgecolor=color, lw=2)) _draw_outline(patch, 4) if text is not None: patch = ax.text(lx,ly, text, verticalalignment='top', color=color, fontsize=text_size, weight='bold') _draw_outline(patch,1) # export class TensorBBox(TensorPoint): "Basic type for a tensor of bounding boxes in an image" @classmethod def create(cls, x, sz=None)->None: return cls(tensor(x).view(-1, 4).float(), sz=sz) def show(self, ctx=None, **kwargs): x = self.view(-1,4) for b in x: _draw_rect(ctx, b, hw=False, **kwargs) return ctx ###Output _____no_output_____ ###Markdown Bounding boxes are expected to come as tuple with an array/tensor of shape `(n,4)` or as a list of lists with four elements adn a list of corresponding labels. Unless you change the defaults in `BBoxScaler` (see later on), coordinates for each bounding box should go from 0 to height/width, with the following convetion: top, left, bottom, right.> Note: We use the same convention as for points with y axis being before x. ###Code # export class LabeledBBox(Tuple): "Basic type for a list of bounding boxes in an image" def show(self, ctx=None, **kwargs): for b,l in zip(self.bbox, self.lbl): if l != '#na#': ctx = retain_type(b, self.bbox).show(ctx=ctx, text=l) return ctx @classmethod def create(cls, x): return cls(x) bbox,lbl = add_props(lambda i,self: self[i]) coco = untar_data(URLs.COCO_TINY) images, lbl_bbox = get_annotations(coco/'train.json') idx=2 coco_fn,bbox = coco/'train'/images[idx],lbl_bbox[idx] coco_img = timg(coco_fn) tbbox = LabeledBBox(TensorBBox(bbox[0]), bbox[1]) ctx = coco_img.show(figsize=(3,3), cmap='Greys') tbbox.show(ctx=ctx); ###Output _____no_output_____ ###Markdown Basic Transforms Unless specifically metioned, all the following transforms can be used as single-item transforms (in one of the list in the `tfms` you pass to a `TfmdDS` or a `Datasource`) or tuple transform (in the `tuple_tfms` you pass to a `TfmdDS` or a `Datasource`). The safest way that will work accross applications is to always use them as `tuple_tfms`. For instance, if you have points or bounding boxes as targets and use `ImageResizer` as a single-item transform, when you get to `PointScaler` or `BBoxScaler` (which are tuple transforms) you won't have the correct size of the image to properly scale your points. ###Code class ImageResizer(Transform): order=10 "Resize image to `size` using `resample" def __init__(self, size, resample=Image.BILINEAR): if not is_listy(size): size=(size,size) self.size,self.resample = (size[1],size[0]),resample def encodes(self, o:PILImage): return o.resize(size=self.size, resample=self.resample) def encodes(self, o:PILMask): return o.resize(size=self.size, resample=Image.NEAREST) ###Output _____no_output_____ ###Markdown `size` can either be one integer (in which case images are resized to a square) or a tuple `height,width`.> Note: This is the usual convention for arrays or in PyTorch, but it's not the usual convention for PIL Image, which use the other way round. ###Code f = ImageResizer(14) test_eq(f(mnist_img).size, (14,14)) test_eq(f(mask).size, (14,14)) f = ImageResizer((32,28)) test_eq(f(mnist_img).size, (28,32))#PIL has width first test_eq(array(f(mnist_img)).shape, (32,28))#But numpy as height first and that is our convention # export def image2tensor(img): "Transform image to byte tensor in `c*h*w` dim order." res = tensor(img) if res.dim()==2: res = res.unsqueeze(-1) return res.permute(2,0,1) # export PILImage ._tensor_cls = TensorImage PILImageBW._tensor_cls = TensorImageBW PILMask ._tensor_cls = TensorMask #export @ToTensor def encodes(self, o:PILBase): return o._tensor_cls(image2tensor(o)) @ToTensor def encodes(self, o:PILMask): return o._tensor_cls(image2tensor(o)[0]) ###Output _____no_output_____ ###Markdown Any data augmentation transform that runs on PIL Images must be run before this transform. ###Code tfm = ToTensor() print(tfm) print(type(mnist_img)) print(type(tfm(mnist_img))) tfm = ToTensor() test_eq(tfm(mnist_img).shape, (1,28,28)) test_eq(type(tfm(mnist_img)), TensorImageBW) test_eq(tfm(mask).shape, (96,128)) test_eq(type(tfm(mask)), TensorMask) ###Output _____no_output_____ ###Markdown Let's confirm we can pipeline this with `PILImage.create`. ###Code pipe_img = Pipeline([PILImageBW.create, ToTensor()]) img = pipe_img(mnist_fn) test_eq(type(img), TensorImageBW) pipe_img.show(img, figsize=(1,1)); def _cam_lbl(x): return mask_fn cam_tds = DataSource([cam_fn], [[PILImage.create, ToTensor()], [_cam_lbl, PILMask.create, ToTensor()]]) show_at(cam_tds, 0); PILMask.create.loss_func = CrossEntropyLossFlat(axis=1) ##export #def _scale_pnts(x, y, do_scale=True, y_first=False): # if y_first: y = y.flip(1) # sz = [x.shape[-1], x.shape[-2]] if isinstance(x, Tensor) else x.size # return y * 2/tensor(sz).float() - 1 if do_scale else y # #def _unscale_pnts(x, y): # sz = [x.shape[-1], x.shape[-2]] if isinstance(x, Tensor) else x.size # return (y+1) * tensor(sz).float()/2 ## export ##TODO: Transform on a whole tuple lose types, see if we can simplify that? #class PointScaler(ItemTransform): # "Scale a tensor representing points" # def __init__(self, do_scale=True, y_first=False): self.do_scale,self.y_first = do_scale,y_first # def encodes(self, o): return (o[0],TensorPoint(_scale_pnts(*o, self.do_scale, self.y_first))) # def decodes(self, o): return (o[0],TensorPoint(_unscale_pnts(*o))) # #TensorPoint.default_item_tfms = PointScaler ###Output _____no_output_____ ###Markdown To work with data augmentation, and in particular the `grid_sample` method, points need to be represented with coordinates going from -1 to 1 (-1 being top or left, 1 bottom or right), which will be done unless you pass `do_scale=False`. We also need to make sure they are following our convention of points being x,y coordinates, so pass along `y_first=True` if you have your data in an y,x format to add a flip.> Warning: This transform needs to run on the tuple level, before any transform that changes the image size. ###Code #export def _scale_pnts(y, sz, do_scale=True, y_first=False): if y_first: y = y.flip(1) res = y * 2/tensor(sz).float() - 1 if do_scale else y return TensorPoint(res, sz=sz) def _unscale_pnts(y, sz): return TensorPoint((y+1) * tensor(sz).float()/2, sz=sz) #export class PointScaler(Transform): "Scale a tensor representing points" order,loss_func = 1,MSELossFlat() def __init__(self, do_scale=True, y_first=False): self.do_scale,self.y_first = do_scale,y_first def _grab_sz(self, x): self.sz = [x.shape[-1], x.shape[-2]] if isinstance(x, Tensor) else x.size return x def _get_sz(self, x): sz = getattr(x, '_meta', {}).get('sz', None) assert sz is not None or self.sz is not None, "Size could not be inferred, pass it in the init of your TensorPoint with `sz=...`" return self.sz if sz is None else sz def setup(self, dl): its = dl.do_item(0) for t in its: if isinstance(t, TensorPoint): self.c = t.numel() def encodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) def decodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) def encodes(self, x:TensorPoint): return _scale_pnts(x, self._get_sz(x), self.do_scale, self.y_first) def decodes(self, x:TensorPoint): return _unscale_pnts(x, self._get_sz(x)) TensorPoint.default_item_tfms = PointScaler ###Output _____no_output_____ ###Markdown To work with data augmentation, and in particular the `grid_sample` method, points need to be represented with coordinates going from -1 to 1 (-1 being top or left, 1 bottom or right), which will be done unless you pass `do_scale=False`. We also need to make sure they are following our convention of points being x,y coordinates, so pass along `y_first=True` if you have your data in an y,x format to add a flip.> Note: This transform automatically grabs the sizes of the images it sees before a `TensorPoint` object and embeds it in them. For this to work, those images need to be before any points in the order of your final tuple. If you don't have such images, you need to embed the size of the corresponding image when creating a `TensorPoint` by passing it with `sz=...`. ###Code def _pnt_lbl(x): return TensorPoint.create(pnts) def _pnt_open(fn): return PILImage(PILImage.create(fn).resize((28,35))) pnt_tds = DataSource([mnist_fn], [_pnt_open, [_pnt_lbl]]) pnt_tdl = TfmdDL(pnt_tds, bs=1, after_item=[PointScaler(), ToTensor()]) pnt_tdl.after_item.c x,y = pnt_tdl.one_batch() #Scaling and flipping properly done #NB: we added a point earlier at (9,17); formula below scales to (-1,1) coords test_close(y[0], tensor([[-1., -1.], [-1., 1.], [1., -1.], [1., 1.], [9/14-1, 17/17.5-1]])) a,b = pnt_tdl.decode_batch((x,y))[0] test_eq(b, tensor(pnts).float()) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorPoint) test_eq(type(a), TensorImage) test_eq(type(b), TensorPoint) pnt_tdl.show_batch(figsize=(2,2), cmap='Greys'); # export #class BBoxScaler(PointScaler): # "Scale a tensor representing bounding boxes" # def encodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) # def decodes(self, x:(PILBase,TensorImageBase)): return self._grab_sz(x) # # def encodes(self, x:(BBox,TensorBBox)): # pnts = x.bbox.view(-1,2) # scaled_bb = _scale_pnts(pnts, self._get_sz(pnts), self.do_scale, self.y_first) # return TensorBBox((scaled_bb.view(-1,4),x.lbl)) # # def decodes(self, x:(BBox,TensorBBox)): # scaled_bb = _unscale_pnts(x.bbox.view(-1,2), self._get_sz(x.bbox.view(-1,2))) # return TensorBBox((scaled_bb.view(-1,4), x.lbl)) # export #class BBoxCategorize(Transform): # "Reversible transform of category string to `vocab` id" # order,state_args=1,'vocab' # def __init__(self, vocab=None): # self.vocab = vocab # self.o2i = None if vocab is None else {v:k for k,v in enumerate(vocab)} # # def setups(self, dsrc): # if not dsrc: return # vals = set() # for bb in dsrc: vals = vals.union(set(bb.lbl)) # self.vocab,self.otoi = uniqueify(list(vals), sort=True, bidir=True, start='#bg') # # def encodes(self, o:BBox): # return TensorBBox.create((o.bbox,tensor([self.otoi[o_] for o_ in o.lbl if o_ in self.otoi]))) # def decodes(self, o:TensorBBox): # return BBox((o.bbox,[self.vocab[i_] for i_ in o.lbl])) # #BBox.default_type_tfms,BBox.default_item_tfms = BBoxCategorize,BBoxScaler #export #TODO tests #def bb_pad(samples, pad_idx=0): # "Function that collect `samples` of labelled bboxes and adds padding with `pad_idx`." # max_len = max([len(s[1][1]) for s in samples]) # def _f(img,bbox,lbl): # bbox = torch.cat([bbox,bbox.new_zeros(max_len-bbox.shape[0], 4)]) # lbl = torch.cat([lbl, lbl .new_zeros(max_len-lbl .shape[0])+pad_idx]) # return img,TensorBBox((bbox,lbl)) # return [_f(x,*y) for x,y in samples] #export class BBoxLabels(MultiCategory): create = MultiCategorize(add_na=True) default_type_tfms = None #export class BBoxLabeler(Transform): def setup(self, dl): self.vocab = dl.vocab def before_call(self): self.bbox,self.lbls = None,None def decode (self, x, **kwargs): self.bbox,self.lbls = None,None return self._call('decodes', x, **kwargs) def decodes(self, x:TensorMultiCategory): self.lbls = [self.vocab[a] for a in x] return x if self.bbox is None else LabeledBBox(self.bbox, self.lbls) def decodes(self, x:TensorBBox): self.bbox = x return self.bbox if self.lbls is None else LabeledBBox(self.bbox, self.lbls) #export BBoxLabels.default_item_tfms = BBoxLabeler #export #LabeledBBox can be sent in a tl with MultiCategorize (depending on the order of the tls) but it is already decoded. @MultiCategorize def decodes(self, x:LabeledBBox): return x #export @PointScaler def encodes(self, x:TensorBBox): pnts = self.encodes(TensorPoint(x.view(-1,2), sz=x._meta.get('sz', None))) return TensorBBox(pnts.view(-1, 4), sz=x._meta.get('sz', None)) @PointScaler def decodes(self, x:TensorBBox): pnts = self.decodes(TensorPoint(x.view(-1,2), sz=x._meta.get('sz', None))) return TensorBBox(pnts.view(-1, 4), sz=x._meta.get('sz', None)) # export def clip_remove_empty(bbox, label): "Clip bounding boxes with image border and label background the empty ones." bbox = torch.clamp(bbox, -1, 1) empty = ((bbox[...,2] - bbox[...,0])*(bbox[...,3] - bbox[...,1]) < 0.) return (bbox[~empty], label[~empty]) bb = tensor([[-2,-0.5,0.5,1.5], [-0.5,-0.5,0.5,0.5], [1,0.5,0.5,0.75], [-0.5,-0.5,0.5,0.5]]) bb,lbl = clip_remove_empty(bb, tensor([1,2,3,2])) test_eq(bb, tensor([[-1,-0.5,0.5,1.], [-0.5,-0.5,0.5,0.5], [-0.5,-0.5,0.5,0.5]])) test_eq(lbl, tensor([1,2,2])) #export def bb_pad(samples, pad_idx=0): "Function that collect `samples` of labelled bboxes and adds padding with `pad_idx`." samples = [(s[0], *clip_remove_empty(*s[1:])) for s in samples] max_len = max([len(s[2]) for s in samples]) def _f(img,bbox,lbl): bbox = torch.cat([bbox,bbox.new_zeros(max_len-bbox.shape[0], 4)]) lbl = torch.cat([lbl, lbl .new_zeros(max_len-lbl .shape[0])+pad_idx]) return img,bbox,lbl return [_f(*s) for s in samples] img1,img2 = TensorImage(torch.randn(16,16,3)),TensorImage(torch.randn(16,16,3)) bb1 = tensor([[-2,-0.5,0.5,1.5], [-0.5,-0.5,0.5,0.5], [1,0.5,0.5,0.75], [-0.5,-0.5,0.5,0.5]]) lbl1 = tensor([1, 2, 3, 2]) bb2 = tensor([[-0.5,-0.5,0.5,0.5], [-0.5,-0.5,0.5,0.5]]) lbl2 = tensor([2, 2]) samples = [(img1, bb1, lbl1), (img2, bb2, lbl2)] res = bb_pad(samples) non_empty = tensor([True,True,False,True]) test_eq(res[0][0], img1) test_eq(res[0][1], tensor([[-1,-0.5,0.5,1.], [-0.5,-0.5,0.5,0.5], [-0.5,-0.5,0.5,0.5]])) test_eq(res[0][2], tensor([1,2,2])) test_eq(res[1][0], img2) test_eq(res[1][1], tensor([[-0.5,-0.5,0.5,0.5], [-0.5,-0.5,0.5,0.5], [0,0,0,0]])) test_eq(res[1][2], tensor([2,2,0])) #export TensorBBox.dbunch_kwargs = {'before_batch': bb_pad} def _coco_bb(x): return TensorBBox.create(bbox[0]) def _coco_lbl(x): return bbox[1] coco_tds = DataSource([coco_fn], [PILImage.create, [_coco_bb], [_coco_lbl, MultiCategorize(add_na=True)]], n_inp=1) coco_tdl = TfmdDL(coco_tds, bs=1, after_item=[BBoxLabeler(), PointScaler(), ToTensor()]) x,y,z = coco_tdl.one_batch() test_close(y[0], -1+tensor(bbox[0])/64) test_eq(z[0], tensor([1,1,1])) a,b,c = coco_tdl.decode_batch((x,y,z))[0] test_close(b, tensor(bbox[0]).float()) test_eq(c.bbox, b) test_eq(c.lbl, bbox[1]) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorBBox) test_eq(type(z), TensorMultiCategory) test_eq(type(a), TensorImage) test_eq(type(b), TensorBBox) test_eq(type(c), LabeledBBox) coco_tdl.show_batch(); #hide #test other direction works too coco_tds = DataSource([coco_fn], [PILImage.create, [_coco_lbl, MultiCategorize(add_na=True)], [_coco_bb]]) coco_tdl = TfmdDL(coco_tds, bs=1, after_item=[BBoxLabeler(), PointScaler(), ToTensor()]) x,y,z = coco_tdl.one_batch() test_close(z[0], -1+tensor(bbox[0])/64) test_eq(y[0], tensor([1,1,1])) a,b,c = coco_tdl.decode_batch((x,y,z))[0] test_eq(b, bbox[1]) test_close(c.bbox, tensor(bbox[0]).float()) test_eq(c.lbl, b) #Check types test_eq(type(x), TensorImage) test_eq(type(y), TensorMultiCategory) test_eq(type(z), TensorBBox) test_eq(type(a), TensorImage) test_eq(type(b), MultiCategory) test_eq(type(c), LabeledBBox) ###Output _____no_output_____ ###Markdown Show methods ###Code #export def get_grid(n, rows=None, cols=None, add_vert=0, figsize=None, double=False, title=None): rows = rows or int(np.ceil(math.sqrt(n))) cols = cols or int(np.ceil(n/rows)) if double: cols*=2 ; n*=2 figsize = (cols*3, rows*3+add_vert) if figsize is None else figsize fig,axs = subplots(rows, cols, figsize=figsize) axs = axs.flatten() for ax in axs[n:]: ax.set_axis_off() if title is not None: fig.suptitle(title, weight='bold', size=14) return axs #export @typedispatch def show_batch(x:TensorImage, y, samples, ctxs=None, max_n=10, rows=None, cols=None, figsize=None, **kwargs): if ctxs is None: ctxs = get_grid(min(len(samples), max_n), rows=rows, cols=cols, figsize=figsize) ctxs = show_batch[object](x, y, samples, ctxs=ctxs, max_n=max_n, **kwargs) return ctxs #export @typedispatch def show_results(x:TensorImage, y, samples, outs, ctxs=None, max_n=10, rows=None, cols=None, figsize=None, **kwargs): if ctxs is None: ctxs = get_grid(min(len(samples), max_n), rows=rows, cols=cols, add_vert=1, figsize=figsize) ctxs = show_results[object](x, y, samples, outs, ctxs=ctxs, max_n=max_n, **kwargs) return ctxs #export @typedispatch def show_results(x:TensorImage, y:TensorCategory, samples, outs, ctxs=None, max_n=10, rows=None, cols=None, figsize=None, **kwargs): if ctxs is None: ctxs = get_grid(min(len(samples), max_n), rows=rows, cols=cols, add_vert=1, figsize=figsize) for i in range(2): ctxs = [b.show(ctx=c, **kwargs) for b,c,_ in zip(samples.itemgot(i),ctxs,range(max_n))] ctxs = [r.show(ctx=c, color='green' if b==r else 'red', **kwargs) for b,r,c,_ in zip(samples.itemgot(1),outs.itemgot(0),ctxs,range(max_n))] return ctxs #export @typedispatch def show_results(x:TensorImage, y:(TensorImageBase, TensorPoint, TensorBBox), samples, outs, ctxs=None, max_n=10, rows=None, cols=None, figsize=None, **kwargs): if ctxs is None: ctxs = get_grid(min(len(samples), max_n), rows=rows, cols=cols, add_vert=1, figsize=figsize, double=True) for i in range(2): ctxs[::2] = [b.show(ctx=c, **kwargs) for b,c,_ in zip(samples.itemgot(i),ctxs[::2],range(max_n))] for x in [samples,outs]: ctxs[1::2] = [b.show(ctx=c, **kwargs) for b,c,_ in zip(x.itemgot(0),ctxs[1::2],range(max_n))] return ctxs ###Output _____no_output_____ ###Markdown Export - ###Code #hide from local.notebook.export import notebook2script notebook2script(all_fs=True) ###Output Converted 00_test.ipynb. Converted 01_core.ipynb. Converted 01a_utils.ipynb. Converted 01b_dispatch.ipynb. Converted 01c_transform.ipynb. Converted 02_script.ipynb. Converted 03_torch_core.ipynb. Converted 03a_layers.ipynb. Converted 04_dataloader.ipynb. Converted 05_data_core.ipynb. Converted 06_data_transforms.ipynb. Converted 07_data_block.ipynb. Converted 08_vision_core.ipynb. Converted 09_vision_augment.ipynb. Converted 10_pets_tutorial.ipynb. Converted 11_vision_models_xresnet.ipynb. Converted 12_optimizer.ipynb. Converted 13_learner.ipynb. Converted 13a_metrics.ipynb. Converted 14_callback_schedule.ipynb. Converted 14a_callback_data.ipynb. Converted 15_callback_hook.ipynb. Converted 15a_vision_models_unet.ipynb. Converted 16_callback_progress.ipynb. Converted 17_callback_tracker.ipynb. Converted 18_callback_fp16.ipynb. Converted 19_callback_mixup.ipynb. Converted 20_interpret.ipynb. Converted 21_vision_learner.ipynb. Converted 22_tutorial_imagenette.ipynb. Converted 23_tutorial_transfer_learning.ipynb. Converted 30_text_core.ipynb. Converted 31_text_data.ipynb. Converted 32_text_models_awdlstm.ipynb. Converted 33_text_models_core.ipynb. Converted 34_callback_rnn.ipynb. Converted 35_tutorial_wikitext.ipynb. Converted 36_text_models_qrnn.ipynb. Converted 37_text_learner.ipynb. This cell doesn't have an export destination and was ignored: e Converted 38_tutorial_ulmfit.ipynb. Converted 40_tabular_core.ipynb. Converted 41_tabular_model.ipynb. Converted 42_tabular_rapids.ipynb. Converted 50_data_block_examples.ipynb. Converted 60_medical_imaging.ipynb. Converted 65_medical_text.ipynb. Converted 90_notebook_core.ipynb. Converted 91_notebook_export.ipynb. Converted 92_notebook_showdoc.ipynb. Converted 93_notebook_export2html.ipynb. Converted 94_notebook_test.ipynb. Converted 95_index.ipynb. Converted 96_data_external.ipynb. Converted 97_utils_test.ipynb. Converted notebook2jekyll.ipynb.
module02/lesson15/10_finding_pairs/pairs_candidates_solution.ipynb
###Markdown Checking if a pair of stocks is cointegrated Imports ###Code import numpy as np import pandas as pd from sklearn.linear_model import LinearRegression from statsmodels.tsa.stattools import adfuller import matplotlib.pyplot as plt import quiz_tests # Set plotting options %matplotlib inline plt.rc('figure', figsize=(16, 9)) # fix random generator so it's easier to reproduce results np.random.seed(2018) # use returns to create a price series drift = 100 r1 = np.random.normal(0, 1, 1000) s1 = pd.Series(np.cumsum(r1), name='s1') + drift #make second series offset = 10 noise = np.random.normal(0, 1, 1000) s2 = s1 + offset + noise s2.name = 's2' ## hedge ratio lr = LinearRegression() lr.fit(s1.values.reshape(-1,1),s2.values.reshape(-1,1)) hedge_ratio = lr.coef_[0][0] #spread spread = s2 - s1 * hedge_ratio ###Output _____no_output_____ ###Markdown Question Do you think we'll need the intercept when calculating the spread? Why or why not? Since the intercept is a constant, it's not necesary to include it in the spread, since it just shifts the spread up by a constant. We use the spread to check when it deviates from its historical average, so what matters going foward is how the spread differs from this average. Quiz Check if spread is stationary using Augmented Dickey Fuller TestThe [adfuller](http://www.statsmodels.org/dev/generated/statsmodels.tsa.stattools.adfuller.html) function is part of the statsmodel library.```adfuller(x, maxlag=None, regression='c', autolag='AIC', store=False, regresults=False)[source]adf (float) – Test statisticpvalue (float) – p-value...``` ###Code def is_spread_stationary(spread, p_level=0.05): """ spread: obtained from linear combination of two series with a hedge ratio p_level: level of significance required to reject null hypothesis of non-stationarity returns: True if spread can be considered stationary False otherwise """ adf_result = adfuller(spread) pvalue = adf_result[1] print(f"pvalue {pvalue:.4f}") if pvalue <= p_level: print(f"pvalue is <= {p_level}, assume spread is stationary") return True else: print(f"pvalue is > {p_level}, assume spread is not stationary") return False quiz_tests.test_is_spread_stationary(is_spread_stationary) # Try out your function print(f"Are the two series candidates for pairs trading? {is_spread_stationary(spread)}") ###Output pvalue 0.0000 pvalue is <= 0.05, assume spread is stationary Are the two series candidates for pairs trading? True
NoteBooks/Intro_MRI_Bloch_Solvers.ipynb
###Markdown Bloch Equation SolversThis notebook will investigate methods to solve the Bloch-Equations. Recall the Bloch Equations are:$$\frac{\partial M}{\partial t} = \gamma M \times B $$where $\gamma$ is the gyromagnetic ratio, $M$ is the magnetization, and $B$ is the magnetic field. It is most common to solve everything in the rotating frame. Where the entire corrdinate system rotates at the Larmor frequency ($\gamma B_0$), such that we only care about the frequency difference from that. There are two types of solvers:* Standard differential equation solvers. These solvers use Runge-Kutta or some othe solver to integrate the changes. These solvers are great for RF pulses or short blocks of time with large changes.* Nutation solvers. These solvers break the problem into periods of rotation and periods of relaxation. This enables much more computationally efficient solutions. This notebooks will show some ways to set this up, starting with the standard differential equation solver. ###Code # This is comment, Pyhton will ignore this line # Import libraries (load libraries which provide some functions) import numpy as np # array library import matplotlib.pyplot as plt # for plotting import math import cmath # Hit the play button to run this cell ###Output _____no_output_____ ###Markdown 1. Standard Solvers Bloch Equation SetupIn these solvers, we will use the matrix form of the Bloch equations\begin{align}\frac{\partial}{\partial t} \begin{bmatrix} M_x \\ M_y \\ M_z \end{bmatrix} = \begin{bmatrix} -1/T2 & \gamma B_z & -\gamma B_y \\ -\gamma B_z & -1/T2 & \gamma B_x \\ \gamma B_y & -\gamma B_x & -1/T1 \end{bmatrix} \begin{bmatrix}M_x \\ M_y \\ M_z \end{bmatrix} + \begin{bmatrix} 0 \\ 0 \\ M_0/T1 \end{bmatrix}\end{align}This is effectively a rotation about each axis at the rate $\gamma B$, a decay of $M_x$ and $M_y$ at a rate $M_{xy}/T2$, and a recovery of $M_z$ at a rate $(M_0-M_z)/T1$. $M_0$ is assumed to be 1 for simplicity. SolverThe solver will integrate the above equation. This could be done with a simple stepwise solver:\begin{align}M_x(t+\Delta t) = M_x(t) + \Delta \frac{\partial M_x(t)}{\partial t}\end{align}Where $\Delta$ is a step size and the temporal resolution of the discrete array. However, this method is not fully stable, especially since the derivative of $M_x$ depends on itself. You can very easily find odd conditions, such as the magnetization growing. Instead, more complex methods are needed, which allow for large steps sizes. This is a general issue with numerical solvers and not unique to MRI. The below code is a standard Runge-Kutta 4 (RK4) solver for a differential equation. It's essential numerical integration but is much more stable. It achives this stability by evaluating the derivative at the next steps. For more info see [Wikipedia](https://en.wikipedia.org/wiki/Runge%E2%80%93Kutta_methods) ###Code from scipy import interpolate def bloch_solver( B, time, freq, T1=2000, T2=2000, M0=[0,0,1], GAM=42.58e6*2*math.pi): # This is simple Rk4 solution to the Bloch Equations. # # Inputs: # B(array) -- Magentic Field [N x 3] (T) # time(array) -- Time of each point in waveforms (s) # T1 -- Longitudinal relaxation times (s) # T2 -- Transverse relaxation times (s) # Freq Offset -- Off center frequency in Hz # M0 -- Initial state of magnetization (not equilibrium magnetization) # Outputs: # MOutput -- Magnetization for each position in time # Convert frequency to rads/s act_freq = 2*math.pi*freq #Convert to roation rates (gamma*B) assert B.shape[1] == 3 Bx = GAM*B[:,0] By = GAM*B[:,1] Bz = GAM*B[:,2] + act_freq #Create a spline for interpolation spline_Bx = interpolate.splrep(time, Bx) spline_By = interpolate.splrep(time, By) spline_Bz = interpolate.splrep(time, Bz) #Initialize Mag = np.array(M0).reshape(3,1) # Output storage MOutput = np.zeros_like(B) #Runge-Kutta PDE Solution dt = time[2] - time[1] for count, t1 in enumerate(time): m1 = Mag bx = interpolate.splev(t1, spline_Bx) by = interpolate.splev(t1, spline_By) bz = interpolate.splev(t1, spline_Bz) k1 = np.array([[ -1/T2, bz, -by], [ -bz ,-1/T2, bx], [ by , -bx, -1/T1]])@ m1 + np.array([[0],[0],[1/T1]]) t2 = t1 + dt/2 bx = interpolate.splev(t2, spline_Bx) by = interpolate.splev(t2, spline_By) bz = interpolate.splev(t2, spline_Bz) m2 = Mag + k1*dt/2 k2 = np.array([[ -1/T2, bz, -by], [ -bz ,-1/T2, bx], [ by , -bx, -1/T1]])@ m2 + np.array([[0],[0],[1/T1]]) t3 = t1 + dt/2 bx = interpolate.splev(t3, spline_Bx) by = interpolate.splev(t3, spline_By) bz = interpolate.splev(t3, spline_Bz) m3 = Mag + k2*dt/2 k3 = np.array([[ -1/T2, bz, -by], [ -bz ,-1/T2, bx], [ by , -bx, -1/T1]])@m3 + np.array([[0],[0],[1/T1]]) t4 = t1 + dt bx = interpolate.splev(t4, spline_Bx) by = interpolate.splev(t4, spline_By) bz = interpolate.splev(t4, spline_Bz) m4 = Mag + k3*dt k4 = np.array([[ -1/T2, bz, -by], [ -bz ,-1/T2, bx], [ by , -bx, -1/T1]])@m4 + np.array([[0],[0],[1/T1]]) # Runge-Kutta averages the above terms Mag = Mag + dt/6*(k1 + 2*k2 + 2*k3 + k4); # Save to an array MOutput[count,0]= Mag[0] MOutput[count,1]= Mag[1] MOutput[count,2]= Mag[2] return MOutput ###Output _____no_output_____ ###Markdown Excitation ExperimentThis is the most simple experiment with a single excitation pulse followed by recovery. We assume magnetization starts at $M_0 \equiv 1$. We aim to apply an RF pulse to rotate the magnetization into the transverse plane. The amount we rotate the magnetization is the flip angle ($\alpha$). This is equal to:\begin{equation}\alpha = \int_{0}^{T_{RF}} \gamma B_1(t) dt\end{equation}This assumes the RF pulse is applied along a single axis and more complex relationships do exist for some pulse we won't cover (e.g. adiabatic pulses). The below code will just define a pulse envelope and a phase of the pulse. The phase sets the direction on the $B_1$ field (e.g. along x, y, or mixture). ###Code # Simulation settings dt = 2e-6 # systems run around 2 us resolution in the rotating frame Tmax = 30e-3 #total time to simulate gamma_bar = 42.58e6 gamma = math.pi*2.0*gamma_bar # RF Settings Trf = 2e-3 # period of the pulse flip = 90 #degrees excite_phase = 0 # excite phase in degrees # Define the time span time = np.arange(-1e-3,Tmax,dt) # define the RF pulse shape RF = np.zeros_like(time, dtype=np.complex) idx = (time<Trf) & (time>=0) RF[idx] = 1 # Rect function RF[idx] = np.exp( -16*(time[idx] - Trf/2)**2 / (Trf**2)) # Gaussian # Now lets scale the RF pulse amplitude to be correct rotation for on-resonance spins RF = RF*(flip/360) / np.sum(gamma_bar*RF*dt) #Rotate to the phase of excitation RF = RF*np.exp(2j*math.pi*excite_phase/360) # The RF is complex B = np.zeros( (len(time),3)) B[:,0] = np.real(RF) B[:,1] = np.imag(RF) # Plot plt.figure(figsize=(8,8)) plt.title('Applied magnetic fields vs. time') plt.plot(time,B[:,0],label='B_x') plt.plot(time,B[:,1],label='B_y') plt.plot(time,B[:,2],label='B_z') plt.xlabel('Time [s]') plt.ylabel('B1 [T]') plt.legend() plt.show() ###Output _____no_output_____ ###Markdown Simulate use the above define solversSome things to change:* Change T2 to some other values (typical range 1e-3 to 1), does the behavior match what you expect?* Change the frequency offset to other values. Why might the Larmor frequency and the RF frequency be slightly different?* Change T1 to some other values (typical range 100e-3 to 5), does the behavior match what you expect? ###Code # B is defined above # freq = offset frequency from assumed Larmor frequency # time is defined above # T1 is the longitudinal relaxation rate # T2 is the transverse longitudunal rate Mout = bloch_solver( B, time, freq=100, T1=2000, T2=20e-3) # Plot plt.figure(figsize=(8,8)) plt.plot(time,Mout[:,0],label=r'$M_x$') plt.plot(time,Mout[:,1],label=r'$M_y$') plt.plot(time,np.sqrt(Mout[:,1]**2+Mout[:,0]**2),'--',label=r'$M_{xy}$') plt.plot(time,Mout[:,2],label=r'$M_z$') plt.xlabel('Time [s]') plt.ylabel('M [a.u.]') plt.legend() plt.show() ###Output _____no_output_____ ###Markdown 2. Nutation Solvers Problem SetupNutation solvers break the Bloch equations into events. Events would include RF exitations, times of free recovery, and magnetic field changes from gradients (more on this last one later). The premise of this is that we know the solutions for the Bloch equations analytically. Such solvers can also include traditional solvers for events with unknown equations. Some we know include:* **Transverse magnetization decay wo/Frequency shift**\begin{equation}M_{xy}(t) = M_{xy}(t=0)e^{-t/T2}\end{equation}* **Longitudinal magnetization recovery**\begin{equation}M_{z}(t) = M_0 + ( M_{z}(t=0)-M_0)e^{-t/T1}\end{equation}* **RF pulse at exactly the Larmor frequency**\begin{equation}M_xy = R_z(-\theta)R_x(\alpha)R_z(\theta)M\end{equation}where $R_z$ is a rotation about $z$, $R_x$ is rotation about $x$, $\alpha$ is the flip angle, and $\theta$ is the phase of the RF pulse. ###Code class Event: def __init__(self, excite_flip=0, excite_phase=0, recovery_time=0, spoil=False): self.excite_flip = excite_flip self.excite_phase = excite_phase self.recovery_time = recovery_time self.spoil = spoil def bloch_nutation_solver( event_list, M0, T1, T2, freq ): # Inputs: # event_list -- Special structure with entries # .excite_flip flip angle of rotation # .excite_phase phase of excite degrees # .recovery_time time after excite to recover # .spoil (if 'true' this set the Mxy to zero at the recovery) # T1 -- Longitudinal relaxation times (s) # T2 -- Transverse relaxation times (s) # Freq Offset-- Off center frequency in Hz # M0 -- Initial state of magnetization (not equilibrium magnetization) # Outputs: # time -- Magnetization for each position in time # BOutput -- Magnetic field for each position in time (interpolated) # Initialize count = 0; time = [0,]; Mout = [M0,] M=M0; # Go through the event_list for event in event_list: theta = event.excite_phase * math.pi / 180 alpha = event.excite_flip * math.pi / 180 T = event.recovery_time spoil = event.spoil # Excite Rz = np.array([[math.cos(theta), math.sin(theta), 0], [-math.sin(theta), math.cos(theta), 0], [0, 0, 1]]) Rx = np.array([[1,0, 0], [0, math.cos(alpha), math.sin(alpha)], [0, -math.sin(alpha), math.cos(alpha)]]) M = np.linalg.inv(Rz)@Rx@Rz@M # Relaxation (Transverse) if spoil: Mxy = 0 else: Mxy = M[0] + 1j*M[1] Mxy = Mxy*cmath.exp( 2j*math.pi*freq*T)*math.exp(-T/T2) # Relaxation (Longitudinal) Mz = M[2] Mz = 1 + (Mz - 1)*math.exp(-T/T1); # Put back into [Mx; My; Mz] vector M = np.array([[Mxy[0].real], [Mxy[0].imag], [Mz[0]]]) # Store for output time.append(time[-1]+T) Mout.append(M) Mout = np.array(Mout) return time, Mout ###Output _____no_output_____ ###Markdown ExampleHere we simulate a situation in which we apply two RF pulses seperated by a gap:$90 ^{\circ} $ - Delay - $180 ^{\circ}$ - Free recoverySome experiments to consider:* Change the flip angles. * Change the delays (modify the range(100) to a desired delay in ms* Change T1 and T2 ###Code # Blank list of events event_list = [] # Excite 90 degree event_list.append( Event(excite_flip=90, excite_phase=0)) #Recover (100ms,could be one event but plotting possible here) for pos in range(100): event_list.append( Event(recovery_time=1e-3)) # Excite 180 degree event_list.append( Event(excite_flip=180, excite_phase=90)) #Recover (500ms,could be one event but plotting possible here) for pos in range(500): event_list.append( Event(recovery_time=1e-3)) # Simulate M0 = np.array([[0],[0],[1.0]]) T1 = 1 T2 = 1 freq = 3 time,Mout = bloch_nutation_solver( event_list, M0, T1, T2, freq ); plt.figure(figsize=(8,8)) plt.plot(time,Mout[:,0], label=r'$M_x$') plt.plot(time,Mout[:,1], label=r'$M_y$') plt.plot(time,np.sqrt(Mout[:,1]**2+Mout[:,0]**2), '--', label=r'$M_y$') plt.plot(time,Mout[:,2], label=r'$M_z$') plt.legend() plt.show() ###Output _____no_output_____
07_Horner_Plot.ipynb
###Markdown The Horner Plot methodHeat **always** flows from hotter to colder parts. During a drilling process, the relatively cool drilling fluid *cools* the surrounding in the vicinity of the borehole. If temperatures are measured shortly after the drilling ended, they therefore likely underestimate the true rock temperature at that depth. Over time, the measured temperatures will increase, as temperature in the borehole re-equilibrates with the temperature of the surrounding rockmass. As it is usually not possible to wait for the temperatures to re-equilibrate, there exist correction methods for those measured *Bottom Hole Temperatures* (often abbreviated as *BHT*).The Horner plot method is one correction method. It uses the behaviour of *in situ* temperature (and pressure) when disturbed by drilling. It plots the following linear equation: $$ T = T_\infty + \frac{Q}{4\pi\lambda} \times ln(1 + \frac{t_c}{\Delta t}) $$where $T$ is the bottom-hole temperature [°C], $T_\infty$ is the 'virgin rock temperature' or *in-situ* temperature [°C], $Q$ the heat flow per unit length [W/m], $\lambda$ the thermal conductivity of the rock [W/(m K)], $t_c$ the time between end of drilling and end of mud circulation and $\Delta t$ the time between end of mud circulation and measurement. This equation resembles a linear equation$$ y = b + m \times x $$where $ln(1 + \frac{t_c}{\Delta t})$ equals $x$ and $T$ equals $y$. Now if we have temperature measurements at different times $\Delta t$, we can assess the *in-situ* temperature by linear regression. Assume we have three temperatures measured at three different times at a depth of 1500 m: * $\Delta t_1$ = 10 h, T = 53 °C * $\Delta t_2$ = 15.5 h, T = 56.5 °C * $\Delta t_3$ = 20.5 h, T = 58.5 °C ###Code # import some libraries import numpy as np import matplotlib.pyplot as plt %matplotlib inline # Set up the variables tc = 3. Dt = np.array([10., 15.5, 24.5])#.reshape(3,1) T = np.array([53., 56.5, 59.8])#.reshape(3,1) # the y values in the linear equation # calculate x values for the linear equation x = np.log(1 + (tc/Dt)) ###Output _____no_output_____ ###Markdown Now that we know x ($ln(1 + \frac{t_c}{\Delta t})$) and y ($T$) values, we can do a linear regression to get $m$ and $b$ of the linear equation ###Code # linear regression m,b = np.polyfit(x,T,1) # set up a regression line x_reg = np.linspace(0,0.3,200) T_reg = m * x_reg + b # plot the results fig = plt.figure(figsize=(12,5)) dots, = plt.plot(x,T,'o', color='#660033', alpha=0.8) line, = plt.plot(x_reg,T_reg, '-', linewidth=2, color='#33ADAD') plt.xlabel('ln(1 + ($t_c/\Delta t$))', fontsize=16) plt.ylabel('Temperature [$^\circ$C]', fontsize=16) plt.tick_params(axis='both', which='major', labelsize=13) plt.legend([dots,line], ["Measured temperatures", "linear regression model"], loc=1, fontsize=18) ###Output _____no_output_____ ###Markdown We know that the slope $m$ equals $\frac{Q}{4\pi \lambda}$. If the mean thermal conductivity of the rocks equals 2.24 W m$^{-1}$ K$^{-1}$, we can calculate Q. ###Code tc = 2.24 Q = 4*np.pi*tc*m print("The heat flow per unit length Q is {} W/m (negative sign means flow into the borehole).".format(Q)) ###Output The heat flow per unit length Q is -1293.98727063 W/m (negative sign means flow into the borehole).
Week-1/Etivity1.2.ipynb
###Markdown **Artificial Intelligence - MSc**ET5003 - MACHINE LEARNING APPLICATIONS Instructor: Enrique Naredo ET5003_Etivity-1 Introduction16154541 -Darren White Explanation of the problemThe problem presented is to use Bayesian multinomial logistic regression to classify images from the MNIST database of handwritten digits. Dataset The MNIST database is a set of images of handwritten digits. The database contains 60,000 training images and 10,000 testing images. The database is based on the NIST database. The NIST database was compiled from American Census Bureau employees and high school students. The training set comes from the American Census Bureau employees and the testing set has been taken from American high school students. As a result of the difference between the groups it was posited that this database may not be efficient for machine learning.The MNIST database is compiled as follows;* 50% of the training data is taken from the NIST training set.* 50% of the training data is taken from the NIST test set.* 50% of the testing data is taken from the NIST training set.* 50% of the testing data is taken from the NIST test set.The MNIST database is maintained by Yann LeCun, (Courant Institute, NYU) Corinna Cortes, (Google Labs, New York) and Christopher J.C. Burges, (Microsoft Research, Redmond). Method Multinomial Logistic Regression (MLR) is used to classify the images in the MNIST database. MLR is an extension of Binary Logistic Regression (BLR) in which numerous binary models are deployed simultaneously. Multinomial logistic regression is used to classify categorial outcomes rather than continuous outcomes. Multinomial models do not assume normality, linearity or homoscedasticity. This makes it a strong modelling choice as real world data can often display these imperfections. Code Imports ###Code # Suppressing Warnings: import warnings warnings.filterwarnings("ignore") ! pip install opencv-python ! pip install scikit-image ! pip install arviz # Used to perform logistic regression, scoring, shuffle & split of dataset # to training & test from sklearn.linear_model import LogisticRegression from sklearn.metrics import accuracy_score from sklearn.model_selection import train_test_split from sklearn.utils import shuffle # Used to fetch MNIST dataset, works locally & on Google Colab from sklearn.datasets import fetch_openml # Used to generate probabilistic multinomial model import pymc3 as pm # Used to view plots of posterior import arviz as az # Used for numerical operations, generating tensors for PyMC3 usage import theano as tt # Used for numerical operations import numpy as np # Used to generate random numbers to draw samples from dataset import random # Used in plotting images & graphs import matplotlib.pyplot as plt from IPython.display import HTML %matplotlib inline ###Output _____no_output_____ ###Markdown Load Data ###Code mnist = fetch_openml('mnist_784', cache=False) X = mnist.data.astype('float32') y = mnist.target.astype('int64') X /= 255.0 X.min(), X.max() ###Output _____no_output_____ ###Markdown The use of `sklearn.datasets` through `fetch_openml()` to gather the MNIST dataset allows the notebook to run on both Google Colab and locally without change to the code. Preprocessing We split the MNIST data into a train and test set with 75% as training data and 25% as test data. ###Code # assigning features and labels X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=42) assert(X_train.shape[0] + X_test.shape[0] == mnist.data.shape[0]) X_train.shape, y_train.shape def plot_example(X: np.array, y: np.array, n: int=5, plot_title: str=None) -> None: """Plots the first 'n' images and their labels in a row. Args: X (numpy array): Image data with each row of array contining an image as a flat vector y (numpy array): Image labels n (int): Number of images to display plot_title (str): Title of the plot Returns: None """ fig, axs = plt.subplots(1, n) fig.suptitle(plot_title, fontsize=20) axs = axs.ravel() for i, (img, y) in enumerate(zip(X[:n].reshape(n, 28, 28), y[:n])): axs[i].axis('off') axs[i].imshow(img, cmap="Greys_r") axs[i].set_title(y) fig.tight_layout() fig.subplots_adjust(top=0.88) plt.show() plot_example(X=X_train, y=y_train, n=6, plot_title="Training Data") ###Output _____no_output_____ ###Markdown Building a Number Classifier from the MNIST Database A Bayesian Multinomial Logistic Regression (BMLR) shall be built to classify the handwritten numbers in the MNIST Database.Multinomial logistic regression is a classifiaction technique that is used to predict the category of an input or the probability of its membership to a category. This is calculated based on multiple independent variables that are either binary or continuous. Multinomial logistic regression allows for the dependent variable to be part of more than two categories (Czepiel, n.d.)(Carpita, et al., 2014).To build the classifier we must first understand its basic construction. The formula for BMLR is: $Pr(Y_{ik} = Pr(Y_i = k\mid x_i; \beta_1 , \beta_2 , ..., \beta_m) = \frac{\displaystyle\exp(\beta_{0k} + x_i \beta_k)}{\displaystyle\sum_{j=1}^{m}\exp(\beta_{0j} + x_i\beta_j)}$ with $k = 1,2,...$ where $\beta_k$ is a row vector of regression coefficients of $x$ for the $k$th category of $y$ Since multinomial logistic regression is an expansion of binary logistic regression we will first define a binary model. Logistic regression assumes that for a single data point $(x,y)$: $P(Y = 1 \mid X = x) = \sigma(z)$ where $z = \theta_0 + \displaystyle\sum_{i = 1}^{m} \theta_i x_i$ where $\theta$ is a vector of parameters of length $m$ the values of these parameters is found from $n$ training examples. This is equivalent to: $P(Y =1 \mid X = x) = \sigma(\theta^Tx)$ Maximum likelihood estimation (MLE) is used to choose the parameter values of the logistic regression. To do this we calculate the log-likelihood and find the values of $\theta$ that maximise it. Since the predictions being made are binary we can define each label as a Bernoulli random variable. The probability of one data point can thus be written as: $P(Y = y\mid X=x) = \sigma(\theta^Tx)^y \cdot [1 - \sigma(\theta^Tx)]^{(1-y)} $ The likelihood of all of the data is defined as follows: The likelihood of the independent training labels: $L(\theta) = \displaystyle\prod_{i =1}^n P(Y = y^{(i)} \mid X = x^{(i)})$ Using the likelihood of a Bernoulli we get $L(\theta) = \displaystyle\prod_{i =1}^n P(Y = y^{(i)} \mid X = x^{(i)}) = \displaystyle\prod_{i=1}^n\sigma(\theta^Tx^{(i)})^{y^{(i)}} \cdot [1-\sigma(\theta^Tx^{(i)})]^{(1-y^{(i)})}$ Therefore the log-likelihood of the logistic regression is: $LL(\theta) = \displaystyle\sum_{(i=1)}^ny^{(i)}\log[\sigma(\theta^Tx^{(i)}) + (1 - y^{(i)}) \log[1 - \sigma(\theta^Tx^{(i)})]$ By using a partial derivative of each parameter we can find the values of $\theta $ that maximise the log-likelihood. The partial derivative of $LL(\theta)$ is: $\frac{\partial LL(\theta)}{\partial\theta_j} = \displaystyle\sum_{i=1}^n [y^{(i)} - \sigma(\theta^Tx^{(i)})]x^{(i)}_j$ Using this various optimisation techniques can be deployed to identify the maximum likelihood. A typical binary logistic regression might use gradient descent. However multinomial classifiers will likely use more sophisticated techniques (Monroe, 2017). Classifier Dataset Summary ###Code # Number of training examples n_train = len(X_train) # Number of testing examples. n_test = len(X_test) # Shape of an MNIST image image_shape =X_train[0].shape # unique classes/labels in the dataset. alltotal = set(y_train ) # number of classes n_classes = len(alltotal ) # print information print("Number of training examples =", n_train) print("Number of testing examples =", n_test) print("Image data shape =", image_shape) print("Number of classes =", n_classes) ## plot histogram fig, ax = plt.subplots() # array with evenly spaced classes ind = np.arange(n_classes) # histogram n, bins, patches = ax.hist(y_train, n_classes, ec='black') # horizontal axis label ax.set_xlabel('classes') # vertical axis label ax.set_ylabel('counts') # plot title ax.set_title(r'Histogram of MNIST images') # show plot plt.figure(figsize=(10,8)) plt.show() ###Output _____no_output_____ ###Markdown We can see from the histogram that we have a relatively balanced dataset which should create good conditions for classification Data Preparation ###Code # Seed the run for repeatability np.random.seed(0) # Classes we will retain n_classes = 3 classes = [3, 7, 9] # The number of instances we'll keep for each of our 3 digits: N_per_class = 500 X = [] labels = [] for d in classes: imgs = X_train[np.where(y_train==d)[0],:] X.append(imgs[np.random.permutation(imgs.shape[0]),:][0:N_per_class,:]) labels.append(np.ones(N_per_class)*d) X_train2 = np.vstack(X).astype(np.float64) y_train2 = np.hstack(labels) ###Output _____no_output_____ ###Markdown We reduce the number of classes to 3 and rather than randomly select them for each notebook run we have explicitly selected them. This allows us to to discuss findings as a group based on the same data. We select all image indices of each desired class from `X_train`, randomly arrange them and append the first `inst_class` of them to the `inputs` array. ###Code print(X_train2.shape,y_train2.shape) # plot digits def plot_digits(instances, images_per_row=5, **options): """Plots images in rows Args: instances (numpy array): Numpy array of image data images_per_row (int): Number of images to print on each row Returns: None """ size = 28 images_per_row = min(len(instances), images_per_row) images = [instance.reshape(size,size) for instance in instances] n_rows = (len(instances) - 1) // images_per_row + 1 row_images = [] n_empty = n_rows * images_per_row - len(instances) images.append(np.zeros((size, size * n_empty))) for row in range(n_rows): rimages = images[row * images_per_row : (row + 1) * images_per_row] row_images.append(np.concatenate(rimages, axis=1)) image = np.concatenate(row_images, axis=0) plt.imshow(image, cmap='gist_yarg', **options) plt.axis("off") # Show random instances from each Digit: plt.figure(figsize=(8,8)) # Selecting a few label indices from each of the 3 classes to show: n_sample = 9 label_indices = [] for i in range(n_classes): label_indices += random.sample(range(i*N_per_class, (i+1)*N_per_class), n_sample) print(label_indices) # Plotting 'original' image plot_digits(X_train2[label_indices,:], images_per_row=9) plt.title("Original Image Samples", fontsize=14) ### we split the dataset in training and validation X_tr, X_val, y_tr, y_val = train_test_split(X_train2, y_train2, test_size=0.2, random_state=0) X_tr, y_tr = shuffle(X_tr, y_tr) print(X_tr.shape) print(X_val.shape) print(y_tr.shape) print(y_val.shape) # transform images into vectors X_trv = X_tr.flatten().reshape(X_tr.shape[0],X_tr.shape[1]) X_valv = X_val.flatten().reshape(X_val.shape[0],X_tr.shape[1]) print(X_trv.shape) print(X_valv.shape) print(y_tr.shape) print(y_val.shape) ###Output (1200, 784) (300, 784) (1200,) (300,) ###Markdown Given that the MNIST dataset is already in flat vector form, i.e. each image is already a one dimensional vector, the `flatten` step is not required. However we retain this step for future reference when using other datasets that may require flattening. Algorithm ###Code #General-recipe ML logistic regression clf = LogisticRegression(random_state=0, max_iter=2000, C=100, solver='lbfgs', multi_class='multinomial').fit(X_trv, y_tr) y_pred_logi = clf.predict(X_valv) y_pred_logi_prob = clf.predict_proba(X_valv) prob_classmax = np.max(y_pred_logi_prob,axis=1) print("Accuracy =", accuracy_score(y_pred_logi, y_val)) ###Output Accuracy = 0.92 ###Markdown Accuracy achieved of 0.92 is relatively good. We'll review the highest probablities for correctly and incorrectly classified images. ###Code # probability of general-recipe logistic regression in correct instances highest_prob_matches = np.sort(prob_classmax[y_val==y_pred_logi]) print(f"Probabilities of best scoring matches:\n{highest_prob_matches}") # probability of general-recipe logistic regression in wrong instances highest_prob_mismatches = np.sort(prob_classmax[y_val!=y_pred_logi]) print(f"Probabilities of best scoring mismatches:\n{highest_prob_mismatches}") mismatch_indices_gt_99 = np.intersect1d(np.where(y_val!=y_pred_logi), np.where(prob_classmax > 0.99)) print(f"Mismatch count above 99% probability : {len(mismatch_indices_gt_99)}") # Display mismatches above 99% probability display_cnt = len(mismatch_indices_gt_99) X_valv_mismatches = [] y_val_mismatches = [] y_pred_mismatches = [] compare = 'Comparison of actual vs predicted: \n' for idx in mismatch_indices_gt_99: X_valv_mismatches.append(X_valv[idx]) y_val_mismatches.append(y_val[idx]) y_pred_mismatches.append(y_pred_logi[idx]) compare += (f"y_pred:{y_pred_logi[idx]} y_val:{y_val[idx]}" +\ f", Pr({classes[0]}):{y_pred_logi_prob[idx][0]:.8f}" +\ f", Pr({classes[1]}):{y_pred_logi_prob[idx][1]:.8f}" +\ f", Pr({classes[2]}):{y_pred_logi_prob[idx][2]:.8f}\n") X_valv_mismatches = np.array(X_valv_mismatches) y_val_mismatches = np.array(y_val_mismatches) y_pred_mismatches = np.array(y_pred_mismatches) print(compare) plot_example(X=X_valv_mismatches, y=y_pred_mismatches, n=display_cnt, plot_title="Mismatches >99% probability (predictions labelled)") ###Output Comparison of actual vs predicted: y_pred:7.0 y_val:9.0, Pr(3):0.00000160, Pr(7):0.99999838, Pr(9):0.00000002 y_pred:7.0 y_val:9.0, Pr(3):0.00007013, Pr(7):0.99638036, Pr(9):0.00354951 y_pred:9.0 y_val:7.0, Pr(3):0.00051846, Pr(7):0.00205414, Pr(9):0.99742740 y_pred:9.0 y_val:7.0, Pr(3):0.00212730, Pr(7):0.00167353, Pr(9):0.99619916 y_pred:7.0 y_val:9.0, Pr(3):0.00000096, Pr(7):0.99945519, Pr(9):0.00054385 y_pred:9.0 y_val:7.0, Pr(3):0.00000448, Pr(7):0.00098430, Pr(9):0.99901122 y_pred:7.0 y_val:9.0, Pr(3):0.00000096, Pr(7):0.99999904, Pr(9):0.00000000 ###Markdown We observe seven of the wrong predictions by logistic regression having a confidence above 99%. On reviewing the images it is difficult to understand how these have been labelled incorrectly with such high confidence. the probability values for the correct labels are very low given our total probabiliy must sum to 1. Probabilistic ML ###Code X_trv.shape[1] import sklearn.preprocessing ## We use LabelBinarizer to transfor classes into counts # neg_label=0, pos_label=1 y_2_bin = sklearn.preprocessing.LabelBinarizer().fit_transform(y_tr.reshape(-1,1)) nf = X_trv.shape[1] # number of classes nc = len(classes) # floatX = float32 floatX = tt.config.floatX init_b = np.random.randn(nf, nc-1).astype(floatX) init_a = np.random.randn(nc-1).astype(floatX) with pm.Model() as multi_logistic: # Prior β = pm.Normal('beta', 0, sigma=100, shape=(nf, nc-1), testval=init_b) α = pm.Normal('alpha', 0, sigma=100, shape=(nc-1,), testval=init_a) # we need to consider nc-1 features because the model is not identifiable # the softmax turns a vector into a probability that sums up to one # therefore we add zeros to go back to dimension nc # so that softmax returns a vector of dimension nc β1 = tt.tensor.concatenate([np.zeros((nf,1)),β ],axis=1) α1 = tt.tensor.concatenate([[0],α ],) # Likelihood mu = pm.math.matrix_dot(X_trv,β1) + α1 # It doesn't work if the problem is binary p = tt.tensor.nnet.nnet.softmax(mu) observed = pm.Multinomial('likelihood', p=p, n=1, observed=y_2_bin) ###Output _____no_output_____ ###Markdown We set our priors as normal distributions with mean of 0 and $\sigma$ of 100. For $\alpha$ we specify a vector size of the class count minus one, i.e. $3-1=2$. For $\beta$ we specify a matrix size of the input pixel count times the class count minux one, i.e. $784x2$. ###Code y_2_bin with multi_logistic: #approx = pm.fit(300000, method='advi') # takes longer approx = pm.fit(3000, method='advi') plt.figure(figsize=(10,8)) plt.ylabel('Loss') plt.xlabel('Iteration') plt.plot(approx.hist) plt.title('Loss vs Iteration', fontsize=16) plt.show() ###Output _____no_output_____ ###Markdown The loss is seen to decrease as we iterate further on the model. ###Code # View graph of the posterior alpha & beta values dd = 300 posterior = approx.sample(draws=dd) az.plot_trace(posterior); # View summary table of the posterior with multi_logistic: display(az.summary(posterior, round_to=2)) ###Output arviz - WARNING - Shape validation failed: input_shape: (1, 300), minimum_shape: (chains=2, draws=4) ###Markdown The summary table and plots show our two alpha values for our multinomial three class problem. The 784 beta values correpond to the input feature set size of $28x28=784$ pixels per image. The right hand size of the plot shows the samples of the Markov chain plotted for beta and alpha values. ###Code ## The softmax function transforms each element of a collection by computing the exponential # of each element divided by the sum of the exponentials of all the elements. from scipy.special import softmax #select an image in the test set i = 10 #i = random.randint(0, dd) #select a sample in the posterior s = 100 #s = random.randint(0, dd) beta = np.hstack([np.zeros((nf,1)), posterior['beta'][s,:] ]) alpha = np.hstack([[0], posterior['alpha'][s,:] ]) image = X_valv[i,:].reshape(28,28) plt.figure(figsize=(2,2)) plt.imshow(image,cmap="Greys_r") np.set_printoptions(suppress=True) print("test image #" + str(i)) print("posterior sample #" + str(s)) print("true class=", y_val[i]) print("classes: " + str(classes)) print("estimated prob=",softmax((np.array([X_valv[i,:].dot(beta) + alpha])))[0,:]) # Bayesian prediction # return the class that has the highest posterior probability y_pred_Bayesian=[] for i in range(X_valv.shape[0]): val=np.zeros((1,len(classes))) for s in range(posterior['beta'].shape[0]): beta = np.hstack([np.zeros((nf,1)), posterior['beta'][s,:] ]) alpha = np.hstack([[0], posterior['alpha'][s,:] ]) val = val + softmax((np.array([X_valv[i,:].dot(beta) + alpha]))) mean_probability = val/posterior['beta'].shape[0] y_pred_Bayesian.append( np.argmax(mean_probability)) print(y_pred_Bayesian) # recall the classes we are using print(classes) # prediction array (using classes) nn = 10 # just an example np.array(classes)[y_pred_Bayesian[0:nn]] # using validation: y_val print("Accuracy=", accuracy_score(np.array(classes)[y_pred_Bayesian], y_val)) ###Output Accuracy= 0.9166666666666666 ###Markdown Selecting Differences ###Code y_predB=[] for i in range(X_valv.shape[0]): #print(i) val=[] for s in range(posterior['beta'].shape[0]): beta = np.hstack([np.zeros((nf,1)), posterior['beta'][s,:] ]) alpha = np.hstack([[0], posterior['alpha'][s,:] ]) val.append(softmax((np.array([X_valv[i,:].dot(beta) + alpha])))[0,:]) #mean probability valmean = np.mean(val,axis=0) #class with maximum mean probability classmax = np.argmax(valmean) #ranks ranks = np.array(val.copy()) ranks = ranks *0 #init colmax = np.argmax(np.array(val),axis=1) ranks[np.arange(0,len(colmax)),colmax]=1 y_predB.append( [classmax, valmean[classmax], np.std(ranks,axis=0)[classmax]]) y_predB= np.array(y_predB) # prediction array mm = 10 y_predB[0:mm,:] #sorting in descending order difficult = np.argsort(-y_predB[:,2]) y_predB[difficult[0:mm],:] #probability of general-recipe logistic regression in wrong instances prob_classmax[y_pred_logi != y_val] y_predB[y_pred_logi != y_val,:] ## Difficult & easy instances easy = np.argsort(y_predB[:,2]) print("Accuracy in easy instances =", accuracy_score(y_pred_logi[easy[0:100]], y_val[easy[0:100]])) difficult = np.argsort(-y_predB[:,2]) print("Accuracy in difficult instances =", accuracy_score(y_pred_logi[difficult[0:100]], y_val[difficult[0:100]])) # show 10 random 'easy' images fig, axs = plt.subplots(2,5, figsize=(15, 6)) fig.subplots_adjust(hspace = .2, wspace=.001) axs = axs.ravel() for i in range(10): index = easy[i] image = X_valv[index,:].reshape(28,28) axs[i].axis('off') axs[i].imshow(image,cmap="Greys_r") # show 10 random 'difficult' images fig, axs = plt.subplots(2,5, figsize=(15, 6)) fig.subplots_adjust(hspace = .2, wspace=.001) axs = axs.ravel() for i in range(10): index = difficult[i] image = X_valv[index,:].reshape(28,28) axs[i].axis('off') axs[i].imshow(image,cmap="Greys_r") # show 10 random 'easy' images fig, axs = plt.subplots(2,5, figsize=(15, 6)) fig.subplots_adjust(hspace = .2, wspace=.001) axs = axs.ravel() for i in range(10): index = easy[i] image = X_valv[index,:].reshape(28,28) axs[i].axis('off') axs[i].imshow(image,cmap="Greys_r") ###Output _____no_output_____ ###Markdown Predicted answers - easy ###Code plot_example(X=X_valv[easy], y=y_pred_logi[easy], n=6, plot_title="Predicted easy examples") ###Output _____no_output_____ ###Markdown Actual answers - easy ###Code plot_example(X=X_valv[easy], y=y_val[easy], n=6, plot_title="Actual easy examples") ###Output _____no_output_____ ###Markdown Predicted answers - difficult ###Code plot_example(X=X_valv[difficult], y=y_pred_logi[difficult], n=6, plot_title="Predicted Answers - difficult" ) ###Output _____no_output_____ ###Markdown Actual answers - difficult ###Code plot_example(X=X_valv[difficult], y=y_val[difficult], n=6, plot_title="Actual Answers - Difficult") ###Output _____no_output_____
tutorials/noise/7_accreditation.ipynb
###Markdown Accreditation protocol Accreditation Protocol (AP) is a protocol devised to characterize the reliability of noisy quantum devices.Given a noisy quantum device implementing a "target" quantum circuit, AP certifies an upper-bound on the variation distance between the probability distribution of the outputs returned by the device and the ideal probability distribution.This method is based on Ferracin et al, "Accrediting outputs of noisy intermediate-scale quantum devices", https://arxiv.org/abs/1811.09709.This notebook gives an example for how to use the ignis.characterization.accreditation module. This particular example shows how to accredit the outputs of a 4-qubit quantum circuit of depth 5. All the circuits are run using the noisy Aer simulator. ###Code #Import general libraries (needed for functions) import numpy as np from numpy import random import qiskit #Import Qiskit classes from qiskit import QuantumCircuit, QuantumRegister, ClassicalRegister, Aer, execute from qiskit.providers.aer.noise import NoiseModel from qiskit.providers.aer.noise.errors.standard_errors import depolarizing_error #Import the accreditation functions. from qiskit.ignis.verification.accreditation import AccreditationFitter,AccreditationCircuits ###Output _____no_output_____ ###Markdown Input to the protocol AP can accredit the outputs of a __target circuit__ that1) Takes as input $n$ qubits in the state $|{0}>$2) Ends with single-qubit measurements in the Pauli-$Z$ basis3) Is made of $m$ "bands", each band containing a round of single-qubit gates and a round of controlled-$Z$ gates.The accreditation is made by employing __trap circuits__, circuits that can be efficiently simulated on a classical computer and that whose outputs are used to witness the correct functionality of the device.Let's now draw a target quantum circuit!We start with a simple circuit to generate and measure 4-qubits GHZ states. ###Code # Create a Quantum Register with n_qb qubits. q_reg = QuantumRegister(4, 'q') # Create a Classical Register with n_qb bits. c_reg = ClassicalRegister(4, 's') # Create a Quantum Circuit acting on the q register target_circuit = QuantumCircuit(q_reg, c_reg) target_circuit.h(0) target_circuit.h(1) target_circuit.h(2) target_circuit.h(3) target_circuit.cz(0,1) target_circuit.cz(0,2) target_circuit.h(1) target_circuit.h(2) target_circuit.cz(0,3) target_circuit.cz(1,2) target_circuit.h(1) target_circuit.h(2) target_circuit.h(3) target_circuit.measure(q_reg, c_reg) target_circuit.draw(output = 'mpl') ###Output _____no_output_____ ###Markdown Generating accreditation circuits The function $accreditation\_circuits$ generates all the circuits required by AP, target and traps. It automatically appends random Pauli gates to the circuits (if the implementation is noisy, these random Pauli gates reduce the noise to Pauli errors ! ) It also returns the list $postp\_list$ of strings required to post-process the outputs, as well as the number $v\_zero$ indicating the circuit implementing the target.This is the target circuit with randomly chosen Pauli gates: ###Code accsys = AccreditationCircuits(target_circuit) v = 10 circ_list, postp_list, v_zero = accsys.generate_circuits(v) circ_list[(v_zero)%(v+1)].draw(output = 'mpl') ###Output _____no_output_____ ###Markdown This is how a trap looks like: ###Code circ_list[(v_zero+1)%(v+1)].draw(output = 'mpl') ###Output _____no_output_____ ###Markdown One can use the optional twoqubitgate argument to switch use cx instead of cz gates and can arbitrarily change the coupling map, in order to compile to the desired device topology (which in this case might lead to more layers than expected). ###Code accsys.target_circuit(target_circuit, two_qubit_gate='cx', coupling_map=[[0,1],[1,2],[2,3]] ) v = 10 circ_list, postp_list, v_zero = accsys.generate_circuits(v) circ_list[(v_zero)%(v+1)].draw(output = 'mpl') ###Output _____no_output_____ ###Markdown Simulate the ideal circuits Let's implement AP.We use $accreditation\_circuits$ to generate target and trap circuits.Then, we use the function $single\_protocol\_run$ to implement all these circuits, keeping the output of the target only if all of the traps return the correct output. ###Code simulator = qiskit.Aer.get_backend('qasm_simulator') test_1 = AccreditationFitter() # Create target and trap circuits with random Pauli gates accsys = AccreditationCircuits(target_circuit) circuit_list, postp_list, v_zero = accsys.generate_circuits(v) job = execute(circuit_list, simulator, shots=1) result = job.result() # Post-process the outputs and see if the protocol accepts test_1.single_protocol_run(result, postp_list, v_zero) print("Outputs of the target: ",test_1.outputs," , AP",test_1.flag,"these outputs!") ###Output Outputs of the target: ['0100'] , AP accepted these outputs! ###Markdown In the absence of noise, all traps return the expected output, therefore we always accept the output of the target.To obtain an upper-bound on the variation distance on the outputs of the target circuit, we need to implement AP $d$ times, each time with ___v___ different trap circuits. ###Code # Number of runs d = 20 test_2 = AccreditationFitter() for run in range(d): # Create target and trap circuits with random Pauli gates circuit_list, postp_list, v_zero = accsys.generate_circuits(v) # Implement all these circuits job = execute(circuit_list, simulator, shots=1) result = job.result() # Post-process the outputs and see if the protocol accepts test_2.single_protocol_run(result, postp_list, v_zero) print("Protocol run number",run+1,", outputs of the target",test_2.flag) print('\nAfter',test_2.num_runs,'runs, AP has accepted',test_2.N_acc,'outputs!') print('\nList of accepted outputs:\n', test_2.outputs) ###Output Protocol run number 1 , outputs of the target accepted Protocol run number 2 , outputs of the target accepted Protocol run number 3 , outputs of the target accepted Protocol run number 4 , outputs of the target accepted Protocol run number 5 , outputs of the target accepted Protocol run number 6 , outputs of the target accepted Protocol run number 7 , outputs of the target accepted Protocol run number 8 , outputs of the target accepted Protocol run number 9 , outputs of the target accepted Protocol run number 10 , outputs of the target accepted Protocol run number 11 , outputs of the target accepted Protocol run number 12 , outputs of the target accepted Protocol run number 13 , outputs of the target accepted Protocol run number 14 , outputs of the target accepted Protocol run number 15 , outputs of the target accepted Protocol run number 16 , outputs of the target accepted Protocol run number 17 , outputs of the target accepted Protocol run number 18 , outputs of the target accepted Protocol run number 19 , outputs of the target accepted Protocol run number 20 , outputs of the target accepted After 20 runs, AP has accepted 20 outputs! List of accepted outputs: ['0100', '0010', '1011', '1001', '0110', '1111', '1101', '0000', '1101', '0000', '0110', '1011', '0110', '1101', '0000', '0110', '1101', '0010', '1001', '1101', '0000'] ###Markdown The function $bound\_variation\_distance$ calculates the upper-bound on the variation distance (VD) using$$VD\leq \frac{\varepsilon}{N_{\textrm{acc}}/d-\theta}\textrm{ ,}$$where $\theta\in[0,1]$ is a positive number and$$\varepsilon= \frac{1.7}{v+1}$$is the maximum probability of accepting an incorrect state for the target.The function $bound\_variation\_distance$ also calculates the confidence in the bound as $$1-2\textrm{exp}\big(-2\theta d^2\big)$$ ###Code theta = 5/100 test_2.bound_variation_distance(theta) print("AP accepted",test_2.N_acc,"out of",test_2.num_runs,"times.") print("With confidence",test_2.confidence,"AP certifies that VD is upper-bounded by",test_2.bound) ###Output AP accepted 20 out of 20 times. With confidence 1.0 AP certifies that VD is upper-bounded by 0.16267942583732053 ###Markdown Defining the noise model We define a noise model for the simulator. We add depolarizing error probabilities to the controlled-$Z$ and single-qubit gates. ###Code noise_model = NoiseModel() p1q = 0.003 noise_model.add_all_qubit_quantum_error(depolarizing_error(p1q, 1), 'u1') noise_model.add_all_qubit_quantum_error(depolarizing_error(p1q, 1), 'u2') noise_model.add_all_qubit_quantum_error(depolarizing_error(p1q, 1), 'u3') p2q = 0.03 noise_model.add_all_qubit_quantum_error(depolarizing_error(p2q, 2), 'cx') basis_gates = ['u1','u2','u3','cx'] ###Output _____no_output_____ ###Markdown We then implement noisy circuits and pass their outputs to $single\_protocol\_run$. ###Code test_3 = AccreditationFitter() for run in range(d): # Create target and trap circuits with random Pauli gates circuit_list, postp_list, v_zero = accsys.generate_circuits(v) job = execute(circuit_list, simulator, noise_model=noise_model, basis_gates=basis_gates, shots=1, backend_options={'max_parallel_experiments': 0}) result = job.result() # Post-process the outputs and see if the protocol accepts test_3.single_protocol_run(result, postp_list, v_zero) print("Protocol run number",run+1,", outputs of the target",test_3.flag) print("\nAP accepted",test_3.N_acc,"out of",test_3.num_runs,"times.") print('\nList of accepted outputs:\n', test_3.outputs) theta = 5/100 test_3.bound_variation_distance(theta) print("\nWith confidence",test_3.confidence,"AP certifies that VD is upper-bounded by",test_3.bound) ###Output Protocol run number 1 , outputs of the target rejected Protocol run number 2 , outputs of the target rejected Protocol run number 3 , outputs of the target rejected Protocol run number 4 , outputs of the target rejected Protocol run number 5 , outputs of the target rejected Protocol run number 6 , outputs of the target rejected Protocol run number 7 , outputs of the target accepted Protocol run number 8 , outputs of the target rejected Protocol run number 9 , outputs of the target accepted Protocol run number 10 , outputs of the target rejected Protocol run number 11 , outputs of the target rejected Protocol run number 12 , outputs of the target rejected Protocol run number 13 , outputs of the target accepted Protocol run number 14 , outputs of the target rejected Protocol run number 15 , outputs of the target rejected Protocol run number 16 , outputs of the target rejected Protocol run number 17 , outputs of the target rejected Protocol run number 18 , outputs of the target accepted Protocol run number 19 , outputs of the target rejected Protocol run number 20 , outputs of the target rejected AP accepted 4 out of 20 times. List of accepted outputs: ['0100', '0010', '1011', '1001', '0110', '1111', '1101', '0000', '1101', '0000', '0110', '1011', '0110', '1101', '0000', '0110', '1101', '0010', '1001', '1101', '0000', '0100', '0110', '1111', '0100'] With confidence 1.0 AP certifies that VD is upper-bounded by 1 ###Markdown Changing the number of trap circuits per protocol run changes the upper-bound on the VD, but not the confidence.What number of trap circuits will ensure the minimal upper-bound for your target circuit? ###Code min_traps = 4 max_traps = 10 for num_trap_circs in range(min_traps,max_traps): test_4 = AccreditationFitter() for run in range(d): # Create target and trap circuits with random Pauli gates circuit_list, postp_list, v_zero = accsys.generate_circuits(num_trap_circs) job = execute(circuit_list, simulator, noise_model=noise_model, basis_gates=basis_gates, shots=1, backend_options={'max_parallel_experiments': 0}) result = job.result() # Post-process the outputs and see if the protocol accepts test_4.single_protocol_run(result, postp_list, v_zero) print("\nWith", num_trap_circs, "traps, AP accepted", test_4.N_acc, "out of", test_4.num_runs, "times.") test_4.bound_variation_distance(theta) print("With confidence", test_4.confidence, "AP with", num_trap_circs, "traps certifies that VD is upper-bounded by", test_4.bound) ###Output With 4 traps, AP accepted 16 out of 20 times. With confidence 1.0 AP with 4 traps certifies that VD is upper-bounded by 0.45333333333333337 With 5 traps, AP accepted 7 out of 20 times. With confidence 1.0 AP with 5 traps certifies that VD is upper-bounded by 0.9444444444444444 With 6 traps, AP accepted 11 out of 20 times. With confidence 1.0 AP with 6 traps certifies that VD is upper-bounded by 0.48571428571428577 With 7 traps, AP accepted 9 out of 20 times. With confidence 1.0 AP with 7 traps certifies that VD is upper-bounded by 0.53125 With 8 traps, AP accepted 7 out of 20 times. With confidence 1.0 AP with 8 traps certifies that VD is upper-bounded by 0.6296296296296298 With 9 traps, AP accepted 11 out of 20 times. With confidence 1.0 AP with 9 traps certifies that VD is upper-bounded by 0.33999999999999986 ###Markdown Accreditation protocol Accreditation Protocol (AP) is a protocol devised to characterize the reliability of noisy quantum devices.Given a noisy quantum device implementing a "target" quantum circuit, AP certifies an upper-bound on the variation distance between the probability distribution of the outputs returned by the device and the ideal probability distribution.This method is based on Ferracin et al, "Accrediting outputs of noisy intermediate-scale quantum devices", https://arxiv.org/abs/1811.09709.This notebook gives an example for how to use the ignis.characterization.accreditation module. This particular example shows how to accredit the outputs of a 4-qubit quantum circuit of depth 5. All the circuits are run using the noisy Aer simulator. ###Code #Import general libraries (needed for functions) import numpy as np from numpy import random import qiskit #Import Qiskit classes from qiskit import QuantumCircuit, QuantumRegister, ClassicalRegister, Aer, execute from qiskit.providers.aer.noise import NoiseModel from qiskit.providers.aer.noise.errors.standard_errors import depolarizing_error #Import the accreditation functions. from qiskit.ignis.verification.accreditation import AccreditationFitter,AccreditationCircuits ###Output _____no_output_____ ###Markdown Input to the protocol AP can accredit the outputs of a __target circuit__ that1) Takes as input $n$ qubits in the state $|{0}>$2) Ends with single-qubit measurements in the Pauli-$Z$ basis3) Is made of $m$ "bands", each band containing a round of single-qubit gates and a round of controlled-$Z$ gates.The accreditation is made by employing __trap circuits__, circuits that can be efficiently simulated on a classical computer and that whose outputs are used to witness the correct functionality of the device.Let's now draw a target quantum circuit!We start with a simple circuit to generate and measure 4-qubits GHZ states. ###Code # Create a Quantum Register with n_qb qubits. q_reg = QuantumRegister(4, 'q') # Create a Classical Register with n_qb bits. c_reg = ClassicalRegister(4, 's') # Create a Quantum Circuit acting on the q register target_circuit = QuantumCircuit(q_reg, c_reg) target_circuit.h(0) target_circuit.h(1) target_circuit.h(2) target_circuit.h(3) target_circuit.cz(0,1) target_circuit.cz(0,2) target_circuit.h(1) target_circuit.h(2) target_circuit.cz(0,3) target_circuit.cz(1,2) target_circuit.h(1) target_circuit.h(2) target_circuit.h(3) target_circuit.measure(q_reg, c_reg) target_circuit.draw(output = 'mpl') ###Output _____no_output_____ ###Markdown Generating accreditation circuits The function $accreditation\_circuits$ generates all the circuits required by AP, target and traps. It automatically appends random Pauli gates to the circuits (if the implementation is noisy, these random Pauli gates reduce the noise to Pauli errors ! ) It also returns the list $postp\_list$ of strings required to post-process the outputs, as well as the number $v\_zero$ indicating the circuit implementing the target.This is the target circuit with randomly chosen Pauli gates: ###Code accsys = AccreditationCircuits(target_circuit) v = 10 circ_list, postp_list, v_zero = accsys.generate_circuits(v) circ_list[(v_zero)%(v+1)].draw(output = 'mpl') ###Output _____no_output_____ ###Markdown This is how a trap looks like: ###Code circ_list[(v_zero+1)%(v+1)].draw(output = 'mpl') ###Output _____no_output_____ ###Markdown One can use the optional twoqubitgate arguement to switch use cx instead of cz gates and can arbitrarily change the coupling map, in order to compile to the desired device topology (which in this case might lead to more layers than expected). ###Code accsys.target_circuit(target_circuit, two_qubit_gate='cx', coupling_map=[[0,1],[1,2],[2,3]] ) v = 10 circ_list, postp_list, v_zero = accsys.generate_circuits(v) circ_list[(v_zero)%(v+1)].draw(output = 'mpl') ###Output _____no_output_____ ###Markdown Simulate the ideal circuits Let's implement AP.We use $accreditation\_circuits$ to generate target and trap circuits.Then, we use the function $single\_protocol\_run$ to implement all these circuits, keeping the output of the target only if all of the traps return the correct output. ###Code simulator = qiskit.Aer.get_backend('qasm_simulator') test_1 = AccreditationFitter() # Create target and trap circuits with random Pauli gates accsys = AccreditationCircuits(target_circuit) circuit_list, postp_list, v_zero = accsys.generate_circuits(v) job = execute(circuit_list, simulator, shots=1) result = job.result() # Post-process the outputs and see if the protocol accepts test_1.single_protocol_run(result, postp_list, v_zero) print("Outputs of the target: ",test_1.outputs," , AP",test_1.flag,"these outputs!") ###Output Outputs of the target: ['0100'] , AP accepted these outputs! ###Markdown In the absence of noise, all traps return the expected output, therefore we always accept the output of the target.To obtain an upper-bound on the variation distance on the outputs of the target circuit, we need to implement AP $d$ times, each time with ___v___ different trap circuits. ###Code # Number of runs d = 20 test_2 = AccreditationFitter() for run in range(d): # Create target and trap circuits with random Pauli gates circuit_list, postp_list, v_zero = accsys.generate_circuits(v) # Implement all these circuits job = execute(circuit_list, simulator, shots=1) result = job.result() # Post-process the outputs and see if the protocol accepts test_2.single_protocol_run(result, postp_list, v_zero) print("Protocol run number",run+1,", outputs of the target",test_2.flag) print('\nAfter',test_2.num_runs,'runs, AP has accepted',test_2.N_acc,'outputs!') print('\nList of accepted outputs:\n', test_2.outputs) ###Output Protocol run number 1 , outputs of the target accepted Protocol run number 2 , outputs of the target accepted Protocol run number 3 , outputs of the target accepted Protocol run number 4 , outputs of the target accepted Protocol run number 5 , outputs of the target accepted Protocol run number 6 , outputs of the target accepted Protocol run number 7 , outputs of the target accepted Protocol run number 8 , outputs of the target accepted Protocol run number 9 , outputs of the target accepted Protocol run number 10 , outputs of the target accepted Protocol run number 11 , outputs of the target accepted Protocol run number 12 , outputs of the target accepted Protocol run number 13 , outputs of the target accepted Protocol run number 14 , outputs of the target accepted Protocol run number 15 , outputs of the target accepted Protocol run number 16 , outputs of the target accepted Protocol run number 17 , outputs of the target accepted Protocol run number 18 , outputs of the target accepted Protocol run number 19 , outputs of the target accepted Protocol run number 20 , outputs of the target accepted After 20 runs, AP has accepted 20 outputs! List of accepted outputs: ['0100', '0010', '1011', '1001', '0110', '1111', '1101', '0000', '1101', '0000', '0110', '1011', '0110', '1101', '0000', '0110', '1101', '0010', '1001', '1101', '0000'] ###Markdown The function $bound\_variation\_distance$ calculates the upper-bound on the variation distance (VD) using$$VD\leq \frac{\varepsilon}{N_{\textrm{acc}}/d-\theta}\textrm{ ,}$$where $\theta\in[0,1]$ is a positive number and$$\varepsilon= \frac{1.7}{v+1}$$is the maximum probability of accepting an incorrect state for the target.The function $bound\_variation\_distance$ also calculates the confidence in the bound as $$1-2\textrm{exp}\big(-2\theta d^2\big)$$ ###Code theta = 5/100 test_2.bound_variation_distance(theta) print("AP accepted",test_2.N_acc,"out of",test_2.num_runs,"times.") print("With confidence",test_2.confidence,"AP certifies that VD is upper-bounded by",test_2.bound) ###Output AP accepted 20 out of 20 times. With confidence 1.0 AP certifies that VD is upper-bounded by 0.16267942583732053 ###Markdown Defining the noise model We define a noise model for the simulator. We add depolarizing error probabilities to the controlled-$Z$ and single-qubit gates. ###Code noise_model = NoiseModel() p1q = 0.003 noise_model.add_all_qubit_quantum_error(depolarizing_error(p1q, 1), 'u1') noise_model.add_all_qubit_quantum_error(depolarizing_error(p1q, 1), 'u2') noise_model.add_all_qubit_quantum_error(depolarizing_error(p1q, 1), 'u3') p2q = 0.03 noise_model.add_all_qubit_quantum_error(depolarizing_error(p2q, 2), 'cx') basis_gates = ['u1','u2','u3','cx'] ###Output _____no_output_____ ###Markdown We then implement noisy circuits and pass their outputs to $single\_protocol\_run$. ###Code test_3 = AccreditationFitter() for run in range(d): # Create target and trap circuits with random Pauli gates circuit_list, postp_list, v_zero = accsys.generate_circuits(v) job = execute(circuit_list, simulator, noise_model=noise_model, basis_gates=basis_gates, shots=1, backend_options={'max_parallel_experiments': 0}) result = job.result() # Post-process the outputs and see if the protocol accepts test_3.single_protocol_run(result, postp_list, v_zero) print("Protocol run number",run+1,", outputs of the target",test_3.flag) print("\nAP accepted",test_3.N_acc,"out of",test_3.num_runs,"times.") print('\nList of accepted outputs:\n', test_3.outputs) theta = 5/100 test_3.bound_variation_distance(theta) print("\nWith confidence",test_3.confidence,"AP certifies that VD is upper-bounded by",test_3.bound) ###Output Protocol run number 1 , outputs of the target rejected Protocol run number 2 , outputs of the target rejected Protocol run number 3 , outputs of the target rejected Protocol run number 4 , outputs of the target rejected Protocol run number 5 , outputs of the target rejected Protocol run number 6 , outputs of the target rejected Protocol run number 7 , outputs of the target accepted Protocol run number 8 , outputs of the target rejected Protocol run number 9 , outputs of the target accepted Protocol run number 10 , outputs of the target rejected Protocol run number 11 , outputs of the target rejected Protocol run number 12 , outputs of the target rejected Protocol run number 13 , outputs of the target accepted Protocol run number 14 , outputs of the target rejected Protocol run number 15 , outputs of the target rejected Protocol run number 16 , outputs of the target rejected Protocol run number 17 , outputs of the target rejected Protocol run number 18 , outputs of the target accepted Protocol run number 19 , outputs of the target rejected Protocol run number 20 , outputs of the target rejected AP accepted 4 out of 20 times. List of accepted outputs: ['0100', '0010', '1011', '1001', '0110', '1111', '1101', '0000', '1101', '0000', '0110', '1011', '0110', '1101', '0000', '0110', '1101', '0010', '1001', '1101', '0000', '0100', '0110', '1111', '0100'] With confidence 1.0 AP certifies that VD is upper-bounded by 1 ###Markdown Changing the number of trap circuits per protocol run changes the upper-bound on the VD, but not the confidence.What number of trap circuits will ensure the minimal upper-bound for your target circuit? ###Code min_traps = 4 max_traps = 10 for num_trap_circs in range(min_traps,max_traps): test_4 = AccreditationFitter() for run in range(d): # Create target and trap circuits with random Pauli gates circuit_list, postp_list, v_zero = accsys.generate_circuits(num_trap_circs) job = execute(circuit_list, simulator, noise_model=noise_model, basis_gates=basis_gates, shots=1, backend_options={'max_parallel_experiments': 0}) result = job.result() # Post-process the outputs and see if the protocol accepts test_4.single_protocol_run(result, postp_list, v_zero) print("\nWith", num_trap_circs, "traps, AP accepted", test_4.N_acc, "out of", test_4.num_runs, "times.") test_4.bound_variation_distance(theta) print("With confidence", test_4.confidence, "AP with", num_trap_circs, "traps certifies that VD is upper-bounded by", test_4.bound) ###Output With 4 traps, AP accepted 16 out of 20 times. With confidence 1.0 AP with 4 traps certifies that VD is upper-bounded by 0.45333333333333337 With 5 traps, AP accepted 7 out of 20 times. With confidence 1.0 AP with 5 traps certifies that VD is upper-bounded by 0.9444444444444444 With 6 traps, AP accepted 11 out of 20 times. With confidence 1.0 AP with 6 traps certifies that VD is upper-bounded by 0.48571428571428577 With 7 traps, AP accepted 9 out of 20 times. With confidence 1.0 AP with 7 traps certifies that VD is upper-bounded by 0.53125 With 8 traps, AP accepted 7 out of 20 times. With confidence 1.0 AP with 8 traps certifies that VD is upper-bounded by 0.6296296296296298 With 9 traps, AP accepted 11 out of 20 times. With confidence 1.0 AP with 9 traps certifies that VD is upper-bounded by 0.33999999999999986 ###Markdown Accreditation protocol **Accreditation Protocol (AP)** is a protocol devised to characterize the reliability of noisy quantum devices.Given a noisy quantum device implementing a "target" quantum circuit, AP certifies an upper-bound on the variation distance between the probability distribution of the outputs returned by the device and the ideal probability distribution.This method is based on Ferracin et al, "Accrediting outputs of noisy intermediate-scale quantum devices", https://arxiv.org/abs/1811.09709.This notebook gives an example for how to use the ``ignis.characterization.accreditation`` module. This particular example shows how to accredit the outputs of a 4-qubit quantum circuit of depth 5. All the circuits are run using the noisy Aer simulator. ###Code #Import general libraries (needed for functions) import numpy as np from numpy import random import qiskit #Import Qiskit classes from qiskit import QuantumCircuit, QuantumRegister, ClassicalRegister, Aer, execute from qiskit.providers.aer.noise import NoiseModel from qiskit.providers.aer.noise.errors.standard_errors import depolarizing_error #Import the accreditation functions. from qiskit.ignis.verification.accreditation import AccreditationFitter,AccreditationCircuits ###Output _____no_output_____ ###Markdown Input to the protocol AP can accredit the outputs of a __target circuit__ that1) Takes as input $n$ qubits in the state $|{0}>$2) Ends with single-qubit measurements in the Pauli-$Z$ basis3) Is made of $m$ "bands", each band containing a round of single-qubit gates and a round of controlled-$Z$ gates.The accreditation is made by employing __trap circuits__, circuits that can be efficiently simulated on a classical computer and that whose outputs are used to witness the correct functionality of the device.Let's now draw a target quantum circuit!We start with a simple circuit to generate and measure 4-qubits GHZ states. ###Code # Create a Quantum Register with n_qb qubits. q_reg = QuantumRegister(4, 'q') # Create a Classical Register with n_qb bits. c_reg = ClassicalRegister(4, 's') # Create a Quantum Circuit acting on the q register target_circuit = QuantumCircuit(q_reg, c_reg) target_circuit.h(0) target_circuit.h(1) target_circuit.h(2) target_circuit.h(3) target_circuit.cz(0,1) target_circuit.cz(0,2) target_circuit.h(1) target_circuit.h(2) target_circuit.cz(0,3) target_circuit.cz(1,2) target_circuit.h(1) target_circuit.h(2) target_circuit.h(3) target_circuit.measure(q_reg, c_reg) target_circuit.draw(output = 'mpl') ###Output _____no_output_____ ###Markdown Generating accreditation circuits The function $accreditation\_circuits$ generates all the circuits required by AP, target and traps. It automatically appends random Pauli gates to the circuits (if the implementation is noisy, these random Pauli gates reduce the noise to Pauli errors ! ) It also returns the list $postp\_list$ of strings required to post-process the outputs, as well as the number $v\_zero$ indicating the circuit implementing the target.This is the target circuit with randomly chosen Pauli gates: ###Code accsys = AccreditationCircuits(target_circuit) v = 10 circ_list, postp_list, v_zero = accsys.generate_circuits(v) circ_list[(v_zero)%(v+1)].draw(output = 'mpl') ###Output _____no_output_____ ###Markdown This is how a trap looks like: ###Code circ_list[(v_zero+1)%(v+1)].draw(output = 'mpl') ###Output _____no_output_____ ###Markdown One can use the optional twoqubitgate argument to switch use cx instead of cz gates and can arbitrarily change the coupling map, in order to compile to the desired device topology (which in this case might lead to more layers than expected). ###Code accsys.target_circuit(target_circuit, two_qubit_gate='cx', coupling_map=[[0,1],[1,2],[2,3]] ) v = 10 circ_list, postp_list, v_zero = accsys.generate_circuits(v) circ_list[(v_zero)%(v+1)].draw(output = 'mpl') ###Output _____no_output_____ ###Markdown Simulate the ideal circuits Let's implement AP.We use $accreditation\_circuits$ to generate target and trap circuits.Then, we use the function $single\_protocol\_run$ to implement all these circuits, keeping the output of the target only if all of the traps return the correct output. ###Code simulator = qiskit.Aer.get_backend('qasm_simulator') test_1 = AccreditationFitter() # Create target and trap circuits with random Pauli gates accsys = AccreditationCircuits(target_circuit) circuit_list, postp_list, v_zero = accsys.generate_circuits(v) job = execute(circuit_list, simulator, shots=1) result = job.result() # Post-process the outputs and see if the protocol accepts test_1.single_protocol_run(result, postp_list, v_zero) print("Outputs of the target: ",test_1.outputs," , AP",test_1.flag,"these outputs!") ###Output Outputs of the target: ['0100'] , AP accepted these outputs! ###Markdown In the absence of noise, all traps return the expected output, therefore we always accept the output of the target.To obtain an upper-bound on the variation distance on the outputs of the target circuit, we need to implement AP $d$ times, each time with ___v___ different trap circuits. ###Code # Number of runs d = 20 test_2 = AccreditationFitter() for run in range(d): # Create target and trap circuits with random Pauli gates circuit_list, postp_list, v_zero = accsys.generate_circuits(v) # Implement all these circuits job = execute(circuit_list, simulator, shots=1) result = job.result() # Post-process the outputs and see if the protocol accepts test_2.single_protocol_run(result, postp_list, v_zero) print("Protocol run number",run+1,", outputs of the target",test_2.flag) print('\nAfter',test_2.num_runs,'runs, AP has accepted',test_2.N_acc,'outputs!') print('\nList of accepted outputs:\n', test_2.outputs) ###Output Protocol run number 1 , outputs of the target accepted Protocol run number 2 , outputs of the target accepted Protocol run number 3 , outputs of the target accepted Protocol run number 4 , outputs of the target accepted Protocol run number 5 , outputs of the target accepted Protocol run number 6 , outputs of the target accepted Protocol run number 7 , outputs of the target accepted Protocol run number 8 , outputs of the target accepted Protocol run number 9 , outputs of the target accepted Protocol run number 10 , outputs of the target accepted Protocol run number 11 , outputs of the target accepted Protocol run number 12 , outputs of the target accepted Protocol run number 13 , outputs of the target accepted Protocol run number 14 , outputs of the target accepted Protocol run number 15 , outputs of the target accepted Protocol run number 16 , outputs of the target accepted Protocol run number 17 , outputs of the target accepted Protocol run number 18 , outputs of the target accepted Protocol run number 19 , outputs of the target accepted Protocol run number 20 , outputs of the target accepted After 20 runs, AP has accepted 20 outputs! List of accepted outputs: ['0100', '0010', '1011', '1001', '0110', '1111', '1101', '0000', '1101', '0000', '0110', '1011', '0110', '1101', '0000', '0110', '1101', '0010', '1001', '1101', '0000'] ###Markdown The function $bound\_variation\_distance$ calculates the upper-bound on the variation distance (VD) using$$VD\leq \frac{\varepsilon}{N_{\textrm{acc}}/d-\theta}\textrm{ ,}$$where $\theta\in[0,1]$ is a positive number and$$\varepsilon= \frac{1.7}{v+1}$$is the maximum probability of accepting an incorrect state for the target.The function $bound\_variation\_distance$ also calculates the confidence in the bound as $$1-2\textrm{exp}\big(-2\theta d^2\big)$$ ###Code theta = 5/100 test_2.bound_variation_distance(theta) print("AP accepted",test_2.N_acc,"out of",test_2.num_runs,"times.") print("With confidence",test_2.confidence,"AP certifies that VD is upper-bounded by",test_2.bound) ###Output AP accepted 20 out of 20 times. With confidence 1.0 AP certifies that VD is upper-bounded by 0.16267942583732053 ###Markdown Defining the noise model We define a noise model for the simulator. We add depolarizing error probabilities to the controlled-$Z$ and single-qubit gates. ###Code noise_model = NoiseModel() p1q = 0.003 noise_model.add_all_qubit_quantum_error(depolarizing_error(p1q, 1), 'u1') noise_model.add_all_qubit_quantum_error(depolarizing_error(p1q, 1), 'u2') noise_model.add_all_qubit_quantum_error(depolarizing_error(p1q, 1), 'u3') p2q = 0.03 noise_model.add_all_qubit_quantum_error(depolarizing_error(p2q, 2), 'cx') basis_gates = ['u1','u2','u3','cx'] ###Output _____no_output_____ ###Markdown We then implement noisy circuits and pass their outputs to $single\_protocol\_run$. ###Code test_3 = AccreditationFitter() for run in range(d): # Create target and trap circuits with random Pauli gates circuit_list, postp_list, v_zero = accsys.generate_circuits(v) job = execute(circuit_list, simulator, noise_model=noise_model, basis_gates=basis_gates, shots=1, backend_options={'max_parallel_experiments': 0}) result = job.result() # Post-process the outputs and see if the protocol accepts test_3.single_protocol_run(result, postp_list, v_zero) print("Protocol run number",run+1,", outputs of the target",test_3.flag) print("\nAP accepted",test_3.N_acc,"out of",test_3.num_runs,"times.") print('\nList of accepted outputs:\n', test_3.outputs) theta = 5/100 test_3.bound_variation_distance(theta) print("\nWith confidence",test_3.confidence,"AP certifies that VD is upper-bounded by",test_3.bound) ###Output Protocol run number 1 , outputs of the target rejected Protocol run number 2 , outputs of the target rejected Protocol run number 3 , outputs of the target rejected Protocol run number 4 , outputs of the target rejected Protocol run number 5 , outputs of the target rejected Protocol run number 6 , outputs of the target rejected Protocol run number 7 , outputs of the target accepted Protocol run number 8 , outputs of the target rejected Protocol run number 9 , outputs of the target accepted Protocol run number 10 , outputs of the target rejected Protocol run number 11 , outputs of the target rejected Protocol run number 12 , outputs of the target rejected Protocol run number 13 , outputs of the target accepted Protocol run number 14 , outputs of the target rejected Protocol run number 15 , outputs of the target rejected Protocol run number 16 , outputs of the target rejected Protocol run number 17 , outputs of the target rejected Protocol run number 18 , outputs of the target accepted Protocol run number 19 , outputs of the target rejected Protocol run number 20 , outputs of the target rejected AP accepted 4 out of 20 times. List of accepted outputs: ['0100', '0010', '1011', '1001', '0110', '1111', '1101', '0000', '1101', '0000', '0110', '1011', '0110', '1101', '0000', '0110', '1101', '0010', '1001', '1101', '0000', '0100', '0110', '1111', '0100'] With confidence 1.0 AP certifies that VD is upper-bounded by 1 ###Markdown Changing the number of trap circuits per protocol run changes the upper-bound on the VD, but not the confidence.What number of trap circuits will ensure the minimal upper-bound for your target circuit? ###Code min_traps = 4 max_traps = 10 for num_trap_circs in range(min_traps,max_traps): test_4 = AccreditationFitter() for run in range(d): # Create target and trap circuits with random Pauli gates circuit_list, postp_list, v_zero = accsys.generate_circuits(num_trap_circs) job = execute(circuit_list, simulator, noise_model=noise_model, basis_gates=basis_gates, shots=1, backend_options={'max_parallel_experiments': 0}) result = job.result() # Post-process the outputs and see if the protocol accepts test_4.single_protocol_run(result, postp_list, v_zero) print("\nWith", num_trap_circs, "traps, AP accepted", test_4.N_acc, "out of", test_4.num_runs, "times.") test_4.bound_variation_distance(theta) print("With confidence", test_4.confidence, "AP with", num_trap_circs, "traps certifies that VD is upper-bounded by", test_4.bound) ###Output With 4 traps, AP accepted 16 out of 20 times. With confidence 1.0 AP with 4 traps certifies that VD is upper-bounded by 0.45333333333333337 With 5 traps, AP accepted 7 out of 20 times. With confidence 1.0 AP with 5 traps certifies that VD is upper-bounded by 0.9444444444444444 With 6 traps, AP accepted 11 out of 20 times. With confidence 1.0 AP with 6 traps certifies that VD is upper-bounded by 0.48571428571428577 With 7 traps, AP accepted 9 out of 20 times. With confidence 1.0 AP with 7 traps certifies that VD is upper-bounded by 0.53125 With 8 traps, AP accepted 7 out of 20 times. With confidence 1.0 AP with 8 traps certifies that VD is upper-bounded by 0.6296296296296298 With 9 traps, AP accepted 11 out of 20 times. With confidence 1.0 AP with 9 traps certifies that VD is upper-bounded by 0.33999999999999986 ###Markdown Accreditation protocol Accreditation Protocol (AP) is a protocol devised to characterize the reliability of noisy quantum devices.Given a noisy quantum device implementing a "target" quantum circuit, AP certifies an upper-bound on the variation distance between the probability distribution of the outputs returned by the device and the ideal probability distribution.This method is based on Ferracin et al, "Accrediting outputs of noisy intermediate-scale quantum devices", https://arxiv.org/abs/1811.09709.This notebook gives an example for how to use the ignis.characterization.accreditation module. This particular example shows how to accredit the outputs of a 4-qubit quantum circuit of depth 5. All the circuits are run using the noisy Aer simulator. ###Code #Import general libraries (needed for functions) import numpy as np from numpy import random import qiskit #Import Qiskit classes from qiskit import QuantumCircuit, QuantumRegister, ClassicalRegister, Aer, execute from qiskit.providers.aer.noise import NoiseModel from qiskit.providers.aer.noise.errors.standard_errors import depolarizing_error #Import the accreditation functions. from qiskit.ignis.verification.accreditation import AccreditationFitter,AccreditationCircuits ###Output /opt/miniconda3/lib/python3.7/site-packages/qiskit_aqua-0.7.0-py3.7.egg/qiskit/aqua/operators/primitive_ops/pauli_op.py:25: DeprecationWarning: The module qiskit.extensions.standard is deprecated as of 0.14.0 and will be removed no earlier than 3 months after the release. You should import the standard gates from qiskit.circuit.library.standard_gates instead. from qiskit.extensions.standard import RZGate, RYGate, RXGate, XGate, YGate, ZGate, IGate ###Markdown Input to the protocol AP can accredit the outputs of a __target circuit__ that1) Takes as input $n$ qubits in the state $|{0}>$2) Ends with single-qubit measurements in the Pauli-$Z$ basis3) Is made of $m$ "bands", each band containing a round of single-qubit gates and a round of controlled-$Z$ gates.The accreditation is made by employing __trap circuits__, circuits that can be efficiently simulated on a classical computer and that whose outputs are used to witness the correct functionality of the device.Let's now draw a target quantum circuit!We start with a simple circuit to generate and measure 4-qubits GHZ states. ###Code # Create a Quantum Register with n_qb qubits. q_reg = QuantumRegister(4, 'q') # Create a Classical Register with n_qb bits. c_reg = ClassicalRegister(4, 's') # Create a Quantum Circuit acting on the q register target_circuit = QuantumCircuit(q_reg, c_reg) target_circuit.h(0) target_circuit.h(1) target_circuit.h(2) target_circuit.h(3) target_circuit.cz(0,1) target_circuit.cz(0,2) target_circuit.h(1) target_circuit.h(2) target_circuit.cz(0,3) target_circuit.cz(1,2) target_circuit.h(1) target_circuit.h(2) target_circuit.h(3) target_circuit.measure(q_reg, c_reg) target_circuit.draw(output = 'mpl') ###Output _____no_output_____ ###Markdown Generating accreditation circuits The function $accreditation\_circuits$ generates all the circuits required by AP, target and traps. It automatically appends random Pauli gates to the circuits (if the implementation is noisy, these random Pauli gates reduce the noise to Pauli errors ! ) It also returns the list $postp\_list$ of strings required to post-process the outputs, as well as the number $v\_zero$ indicating the circuit implementing the target.This is the target circuit with randomly chosen Pauli gates: ###Code accsys = AccreditationCircuits(target_circuit) v = 10 circ_list, postp_list, v_zero = accsys.generate_circuits(v) circ_list[(v_zero)%(v+1)].draw(output = 'mpl') ###Output _____no_output_____ ###Markdown This is how a trap looks like: ###Code circ_list[(v_zero+1)%(v+1)].draw(output = 'mpl') ###Output _____no_output_____ ###Markdown One can use the optional twoqubitgate arguement to switch use cx instead of cz gates and can arbitrarily change the coupling map, in order to compile to the desired device topology (which in this case might lead to more layers than expected). ###Code accsys.target_circuit(target_circuit, two_qubit_gate='cx', coupling_map=[[0,1],[1,2],[2,3]] ) v = 10 circ_list, postp_list, v_zero = accsys.generate_circuits(v) circ_list[(v_zero)%(v+1)].draw(output = 'mpl') ###Output _____no_output_____ ###Markdown Simulate the ideal circuits Let's implement AP.We use $accreditation\_circuits$ to generate target and trap circuits.Then, we use the function $single\_protocol\_run$ to implement all these circuits, keeping the output of the target only if all of the traps return the correct output. ###Code simulator = qiskit.Aer.get_backend('qasm_simulator') test_1 = AccreditationFitter() # Create target and trap circuits with random Pauli gates accsys = AccreditationCircuits(target_circuit) circuit_list, postp_list, v_zero = accsys.generate_circuits(v) job = execute(circuit_list, simulator, shots=1) result = job.result() # Post-process the outputs and see if the protocol accepts test_1.single_protocol_run(result, postp_list, v_zero) print("Outputs of the target: ",test_1.outputs," , AP",test_1.flag,"these outputs!") ###Output Outputs of the target: ['0100'] , AP accepted these outputs! ###Markdown In the absence of noise, all traps return the expected output, therefore we always accept the output of the target.To obtain an upper-bound on the variation distance on the outputs of the target circuit, we need to implement AP $d$ times, each time with ___v___ different trap circuits. ###Code # Number of runs d = 20 test_2 = AccreditationFitter() for run in range(d): # Create target and trap circuits with random Pauli gates circuit_list, postp_list, v_zero = accsys.generate_circuits(v) # Implement all these circuits job = execute(circuit_list, simulator, shots=1) result = job.result() # Post-process the outputs and see if the protocol accepts test_2.single_protocol_run(result, postp_list, v_zero) print("Protocol run number",run+1,", outputs of the target",test_2.flag) print('\nAfter',test_2.num_runs,'runs, AP has accepted',test_2.N_acc,'outputs!') print('\nList of accepted outputs:\n', test_2.outputs) ###Output Protocol run number 1 , outputs of the target accepted Protocol run number 2 , outputs of the target accepted Protocol run number 3 , outputs of the target accepted Protocol run number 4 , outputs of the target accepted Protocol run number 5 , outputs of the target accepted Protocol run number 6 , outputs of the target accepted Protocol run number 7 , outputs of the target accepted Protocol run number 8 , outputs of the target accepted Protocol run number 9 , outputs of the target accepted Protocol run number 10 , outputs of the target accepted Protocol run number 11 , outputs of the target accepted Protocol run number 12 , outputs of the target accepted Protocol run number 13 , outputs of the target accepted Protocol run number 14 , outputs of the target accepted Protocol run number 15 , outputs of the target accepted Protocol run number 16 , outputs of the target accepted Protocol run number 17 , outputs of the target accepted Protocol run number 18 , outputs of the target accepted Protocol run number 19 , outputs of the target accepted Protocol run number 20 , outputs of the target accepted After 20 runs, AP has accepted 20 outputs! List of accepted outputs: ['0100', '1011', '0000', '0000', '1011', '0010', '0100', '0000', '1111', '1001', '0010', '0110', '1111', '0110', '0110', '1001', '1101', '0100', '1001', '0010', '0100'] ###Markdown The function $bound\_variation\_distance$ calculates the upper-bound on the variation distance (VD) using$$VD\leq \frac{\varepsilon}{N_{\textrm{acc}}/d-\theta}\textrm{ ,}$$where $\theta\in[0,1]$ is a positive number and$$\varepsilon= \frac{1.7}{v+1}$$is the maximum probability of accepting an incorrect state for the target.The function $bound\_variation\_distance$ also calculates the confidence in the bound as $$1-2\textrm{exp}\big(-2\theta d^2\big)$$ ###Code theta = 5/100 test_2.bound_variation_distance(theta) print("AP accepted",test_2.N_acc,"out of",test_2.num_runs,"times.") print("With confidence",test_2.confidence,"AP certifies that VD is upper-bounded by",test_2.bound) ###Output AP accepted 20 out of 20 times. With confidence 1.0 AP certifies that VD is upper-bounded by 0.16267942583732053 ###Markdown Defining the noise model We define a noise model for the simulator. We add depolarizing error probabilities to the controlled-$Z$ and single-qubit gates. ###Code noise_model = NoiseModel() p1q = 0.003 noise_model.add_all_qubit_quantum_error(depolarizing_error(p1q, 1), 'u1') noise_model.add_all_qubit_quantum_error(depolarizing_error(p1q, 1), 'u2') noise_model.add_all_qubit_quantum_error(depolarizing_error(p1q, 1), 'u3') p2q = 0.03 noise_model.add_all_qubit_quantum_error(depolarizing_error(p2q, 2), 'cx') basis_gates = ['u1','u2','u3','cx'] ###Output _____no_output_____ ###Markdown We then implement noisy circuits and pass their outputs to $single\_protocol\_run$. ###Code test_3 = AccreditationFitter() for run in range(d): # Create target and trap circuits with random Pauli gates circuit_list, postp_list, v_zero = accsys.generate_circuits(v) job = execute(circuit_list, simulator, noise_model=noise_model, basis_gates=basis_gates, shots=1, backend_options={'max_parallel_experiments': 0}) result = job.result() # Post-process the outputs and see if the protocol accepts test_3.single_protocol_run(result, postp_list, v_zero) print("Protocol run number",run+1,", outputs of the target",test_3.flag) print("\nAP accepted",test_3.N_acc,"out of",test_3.num_runs,"times.") print('\nList of accepted outputs:\n', test_3.outputs) theta = 5/100 test_3.bound_variation_distance(theta) print("\nWith confidence",test_3.confidence,"AP certifies that VD is upper-bounded by",test_3.bound) ###Output Protocol run number 1 , outputs of the target accepted Protocol run number 2 , outputs of the target accepted Protocol run number 3 , outputs of the target rejected Protocol run number 4 , outputs of the target rejected Protocol run number 5 , outputs of the target rejected Protocol run number 6 , outputs of the target rejected Protocol run number 7 , outputs of the target accepted Protocol run number 8 , outputs of the target rejected Protocol run number 9 , outputs of the target accepted Protocol run number 10 , outputs of the target rejected Protocol run number 11 , outputs of the target accepted Protocol run number 12 , outputs of the target rejected Protocol run number 13 , outputs of the target rejected Protocol run number 14 , outputs of the target rejected Protocol run number 15 , outputs of the target rejected Protocol run number 16 , outputs of the target rejected Protocol run number 17 , outputs of the target rejected Protocol run number 18 , outputs of the target accepted Protocol run number 19 , outputs of the target accepted Protocol run number 20 , outputs of the target rejected AP accepted 7 out of 20 times. List of accepted outputs: ['0100', '1011', '0000', '0000', '1011', '0010', '0100', '0000', '1111', '1001', '0010', '0110', '1111', '0110', '0110', '1001', '1101', '0100', '1001', '0010', '0100', '1001', '1011', '0010', '0000', '0100', '0100', '0110'] With confidence 1.0 AP certifies that VD is upper-bounded by 0.5151515151515151 ###Markdown Changing the number of trap circuits per protocol run changes the upper-bound on the VD, but not the confidence.What number of trap circuits will ensure the minimal upper-bound for your target circuit? ###Code min_traps = 4 max_traps = 10 for num_trap_circs in range(min_traps,max_traps): test_4 = AccreditationFitter() for run in range(d): # Create target and trap circuits with random Pauli gates circuit_list, postp_list, v_zero = accsys.generate_circuits(num_trap_circs) job = execute(circuit_list, simulator, noise_model=noise_model, basis_gates=basis_gates, shots=1, backend_options={'max_parallel_experiments': 0}) result = job.result() # Post-process the outputs and see if the protocol accepts test_4.single_protocol_run(result, postp_list, v_zero) print("\nWith", num_trap_circs, "traps, AP accepted", test_4.N_acc, "out of", test_4.num_runs, "times.") test_4.bound_variation_distance(theta) print("With confidence", test_4.confidence, "AP with", num_trap_circs, "traps certifies that VD is upper-bounded by", test_4.bound) ###Output With 4 traps, AP accepted 15 out of 20 times. With confidence 1.0 AP with 4 traps certifies that VD is upper-bounded by 0.48571428571428577 With 5 traps, AP accepted 11 out of 20 times. With confidence 1.0 AP with 5 traps certifies that VD is upper-bounded by 0.5666666666666667
assignments/2019/assignment2/TensorFlow.ipynb
###Markdown What's this TensorFlow business?You've written a lot of code in this assignment to provide a whole host of neural network functionality. Dropout, Batch Norm, and 2D convolutions are some of the workhorses of deep learning in computer vision. You've also worked hard to make your code efficient and vectorized.For the last part of this assignment, though, we're going to leave behind your beautiful codebase and instead migrate to one of two popular deep learning frameworks: in this instance, TensorFlow (or PyTorch, if you choose to work with that notebook). What is it?TensorFlow is a system for executing computational graphs over Tensor objects, with native support for performing backpropogation for its Variables. In it, we work with Tensors which are n-dimensional arrays analogous to the numpy ndarray. Why?* Our code will now run on GPUs! Much faster training. Writing your own modules to run on GPUs is beyond the scope of this class, unfortunately.* We want you to be ready to use one of these frameworks for your project so you can experiment more efficiently than if you were writing every feature you want to use by hand. * We want you to stand on the shoulders of giants! TensorFlow and PyTorch are both excellent frameworks that will make your lives a lot easier, and now that you understand their guts, you are free to use them :) * We want you to be exposed to the sort of deep learning code you might run into in academia or industry. How will I learn TensorFlow?TensorFlow has many excellent tutorials available, including those from [Google themselves](https://www.tensorflow.org/get_started/get_started).Otherwise, this notebook will walk you through much of what you need to do to train models in TensorFlow. See the end of the notebook for some links to helpful tutorials if you want to learn more or need further clarification on topics that aren't fully explained here.**NOTE: This notebook is meant to teach you the latest version of Tensorflow 2.0. Most examples on the web today are still in 1.x, so be careful not to confuse the two when looking up documentation**. Install Tensorflow 2.0Tensorflow 2.0 is still not in a fully 100% stable release, but it's still usable and more intuitive than TF 1.x. Please make sure you have it installed before moving on in this notebook! Here are some steps to get started:1. Have the latest version of Anaconda installed on your machine.2. Create a new conda environment starting from Python 3.7. In this setup example, we'll call it `tf_20_env`.3. Run the command: `source activate tf_20_env`4. Then pip install TF 2.0 as described here: https://www.tensorflow.org/install/pip A guide on creating Anaconda enviornments: https://uoa-eresearch.github.io/eresearch-cookbook/recipe/2014/11/20/conda/This will give you an new enviornemnt to play in TF 2.0. Generally, if you plan to also use TensorFlow in your other projects, you might also want to keep a seperate Conda environment or virtualenv in Python 3.7 that has Tensorflow 1.9, so you can switch back and forth at will. Table of ContentsThis notebook has 5 parts. We will walk through TensorFlow at **three different levels of abstraction**, which should help you better understand it and prepare you for working on your project.1. Part I, Preparation: load the CIFAR-10 dataset.2. Part II, Barebone TensorFlow: **Abstraction Level 1**, we will work directly with low-level TensorFlow graphs. 3. Part III, Keras Model API: **Abstraction Level 2**, we will use `tf.keras.Model` to define arbitrary neural network architecture. 4. Part IV, Keras Sequential + Functional API: **Abstraction Level 3**, we will use `tf.keras.Sequential` to define a linear feed-forward network very conveniently, and then explore the functional libraries for building unique and uncommon models that require more flexibility.5. Part V, CIFAR-10 open-ended challenge: please implement your own network to get as high accuracy as possible on CIFAR-10. You can experiment with any layer, optimizer, hyperparameters or other advanced features. We will discuss Keras in more detail later in the notebook.Here is a table of comparison:| API | Flexibility | Convenience ||---------------|-------------|-------------|| Barebone | High | Low || `tf.keras.Model` | High | Medium || `tf.keras.Sequential` | Low | High | Part I: PreparationFirst, we load the CIFAR-10 dataset. This might take a few minutes to download the first time you run it, but after that the files should be cached on disk and loading should be faster.In previous parts of the assignment we used CS231N-specific code to download and read the CIFAR-10 dataset; however the `tf.keras.datasets` package in TensorFlow provides prebuilt utility functions for loading many common datasets.For the purposes of this assignment we will still write our own code to preprocess the data and iterate through it in minibatches. The `tf.data` package in TensorFlow provides tools for automating this process, but working with this package adds extra complication and is beyond the scope of this notebook. However using `tf.data` can be much more efficient than the simple approach used in this notebook, so you should consider using it for your project. ###Code import os import tensorflow as tf import numpy as np import math import timeit import matplotlib.pyplot as plt %matplotlib inline def load_cifar10(num_training=49000, num_validation=1000, num_test=10000): """ Fetch the CIFAR-10 dataset from the web and perform preprocessing to prepare it for the two-layer neural net classifier. These are the same steps as we used for the SVM, but condensed to a single function. """ # Load the raw CIFAR-10 dataset and use appropriate data types and shapes cifar10 = tf.keras.datasets.cifar10.load_data() (X_train, y_train), (X_test, y_test) = cifar10 X_train = np.asarray(X_train, dtype=np.float32) y_train = np.asarray(y_train, dtype=np.int32).flatten() X_test = np.asarray(X_test, dtype=np.float32) y_test = np.asarray(y_test, dtype=np.int32).flatten() # Subsample the data mask = range(num_training, num_training + num_validation) X_val = X_train[mask] y_val = y_train[mask] mask = range(num_training) X_train = X_train[mask] y_train = y_train[mask] mask = range(num_test) X_test = X_test[mask] y_test = y_test[mask] # Normalize the data: subtract the mean pixel and divide by std mean_pixel = X_train.mean(axis=(0, 1, 2), keepdims=True) std_pixel = X_train.std(axis=(0, 1, 2), keepdims=True) X_train = (X_train - mean_pixel) / std_pixel X_val = (X_val - mean_pixel) / std_pixel X_test = (X_test - mean_pixel) / std_pixel return X_train, y_train, X_val, y_val, X_test, y_test # If there are errors with SSL downloading involving self-signed certificates, # it may be that your Python version was recently installed on the current machine. # See: https://github.com/tensorflow/tensorflow/issues/10779 # To fix, run the command: /Applications/Python\ 3.7/Install\ Certificates.command # ...replacing paths as necessary. # Invoke the above function to get our data. NHW = (0, 1, 2) X_train, y_train, X_val, y_val, X_test, y_test = load_cifar10() print('Train data shape: ', X_train.shape) print('Train labels shape: ', y_train.shape, y_train.dtype) print('Validation data shape: ', X_val.shape) print('Validation labels shape: ', y_val.shape) print('Test data shape: ', X_test.shape) print('Test labels shape: ', y_test.shape) class Dataset(object): def __init__(self, X, y, batch_size, shuffle=False): """ Construct a Dataset object to iterate over data X and labels y Inputs: - X: Numpy array of data, of any shape - y: Numpy array of labels, of any shape but with y.shape[0] == X.shape[0] - batch_size: Integer giving number of elements per minibatch - shuffle: (optional) Boolean, whether to shuffle the data on each epoch """ assert X.shape[0] == y.shape[0], 'Got different numbers of data and labels' self.X, self.y = X, y self.batch_size, self.shuffle = batch_size, shuffle def __iter__(self): N, B = self.X.shape[0], self.batch_size idxs = np.arange(N) if self.shuffle: np.random.shuffle(idxs) return iter((self.X[i:i+B], self.y[i:i+B]) for i in range(0, N, B)) train_dset = Dataset(X_train, y_train, batch_size=64, shuffle=True) val_dset = Dataset(X_val, y_val, batch_size=64, shuffle=False) test_dset = Dataset(X_test, y_test, batch_size=64) # We can iterate through a dataset like this: for t, (x, y) in enumerate(train_dset): print(t, x.shape, y.shape) if t > 5: break ###Output 0 (64, 32, 32, 3) (64,) 1 (64, 32, 32, 3) (64,) 2 (64, 32, 32, 3) (64,) 3 (64, 32, 32, 3) (64,) 4 (64, 32, 32, 3) (64,) 5 (64, 32, 32, 3) (64,) 6 (64, 32, 32, 3) (64,) ###Markdown You can optionally **use GPU by setting the flag to True below**. It's not neccessary to use a GPU for this assignment; if you are working on Google Cloud then we recommend that you do not use a GPU, as it will be significantly more expensive. ###Code # Set up some global variables USE_GPU = False if USE_GPU: device = '/device:GPU:0' else: device = '/cpu:0' # Constant to control how often we print when training models print_every = 100 print('Using device: ', device) ###Output Using device: /cpu:0 ###Markdown Part II: Barebones TensorFlowTensorFlow ships with various high-level APIs which make it very convenient to define and train neural networks; we will cover some of these constructs in Part III and Part IV of this notebook. In this section we will start by building a model with basic TensorFlow constructs to help you better understand what's going on under the hood of the higher-level APIs.**"Barebones Tensorflow" is important to understanding the building blocks of TensorFlow, but much of it involves concepts from TensorFlow 1.x.** We will be working with legacy modules such as `tf.Variable`.Therefore, please read and understand the differences between legacy (1.x) TF and the new (2.0) TF. Historical background on TensorFlow 1.xTensorFlow 1.x is primarily a framework for working with **static computational graphs**. Nodes in the computational graph are Tensors which will hold n-dimensional arrays when the graph is run; edges in the graph represent functions that will operate on Tensors when the graph is run to actually perform useful computation.Before Tensorflow 2.0, we had to configure the graph into two phases. There are plenty of tutorials online that explain this two-step process. The process generally looks like the following for TF 1.x:1. **Build a computational graph that describes the computation that you want to perform**. This stage doesn't actually perform any computation; it just builds up a symbolic representation of your computation. This stage will typically define one or more `placeholder` objects that represent inputs to the computational graph.2. **Run the computational graph many times.** Each time the graph is run (e.g. for one gradient descent step) you will specify which parts of the graph you want to compute, and pass a `feed_dict` dictionary that will give concrete values to any `placeholder`s in the graph. The new paradigm in Tensorflow 2.0Now, with Tensorflow 2.0, we can simply adopt a functional form that is more Pythonic and similar in spirit to PyTorch and direct Numpy operation. Instead of the 2-step paradigm with computation graphs, making it (among other things) easier to debug TF code. You can read more details at https://www.tensorflow.org/guide/eager.The main difference between the TF 1.x and 2.0 approach is that the 2.0 approach doesn't make use of `tf.Session`, `tf.run`, `placeholder`, `feed_dict`. To get more details of what's different between the two version and how to convert between the two, check out the official migration guide: https://www.tensorflow.org/alpha/guide/migration_guideLater, in the rest of this notebook we'll focus on this new, simpler approach. TensorFlow warmup: Flatten FunctionWe can see this in action by defining a simple `flatten` function that will reshape image data for use in a fully-connected network.In TensorFlow, data for convolutional feature maps is typically stored in a Tensor of shape N x H x W x C where:- N is the number of datapoints (minibatch size)- H is the height of the feature map- W is the width of the feature map- C is the number of channels in the feature mapThis is the right way to represent the data when we are doing something like a 2D convolution, that needs spatial understanding of where the intermediate features are relative to each other. When we use fully connected affine layers to process the image, however, we want each datapoint to be represented by a single vector -- it's no longer useful to segregate the different channels, rows, and columns of the data. So, we use a "flatten" operation to collapse the `H x W x C` values per representation into a single long vector. Notice the `tf.reshape` call has the target shape as `(N, -1)`, meaning it will reshape/keep the first dimension to be N, and then infer as necessary what the second dimension is in the output, so we can collapse the remaining dimensions from the input properly.**NOTE**: TensorFlow and PyTorch differ on the default Tensor layout; TensorFlow uses N x H x W x C but PyTorch uses N x C x H x W. ###Code def flatten(x): """ Input: - TensorFlow Tensor of shape (N, D1, ..., DM) Output: - TensorFlow Tensor of shape (N, D1 * ... * DM) """ N = tf.shape(x)[0] return tf.reshape(x, (N, -1)) def test_flatten(): # Construct concrete values of the input data x using numpy x_np = np.arange(24).reshape((2, 3, 4)) print('x_np:\n', x_np, '\n') # Compute a concrete output value. x_flat_np = flatten(x_np) print('x_flat_np:\n', x_flat_np, '\n') test_flatten() ###Output x_np: [[[ 0 1 2 3] [ 4 5 6 7] [ 8 9 10 11]] [[12 13 14 15] [16 17 18 19] [20 21 22 23]]] x_flat_np: tf.Tensor( [[ 0 1 2 3 4 5 6 7 8 9 10 11] [12 13 14 15 16 17 18 19 20 21 22 23]], shape=(2, 12), dtype=int64) ###Markdown Barebones TensorFlow: Define a Two-Layer NetworkWe will now implement our first neural network with TensorFlow: a fully-connected ReLU network with two hidden layers and no biases on the CIFAR10 dataset. For now we will use only low-level TensorFlow operators to define the network; later we will see how to use the higher-level abstractions provided by `tf.keras` to simplify the process.We will define the forward pass of the network in the function `two_layer_fc`; this will accept TensorFlow Tensors for the inputs and weights of the network, and return a TensorFlow Tensor for the scores. After defining the network architecture in the `two_layer_fc` function, we will test the implementation by checking the shape of the output.**It's important that you read and understand this implementation.** ###Code def two_layer_fc(x, params): """ A fully-connected neural network; the architecture is: fully-connected layer -> ReLU -> fully connected layer. Note that we only need to define the forward pass here; TensorFlow will take care of computing the gradients for us. The input to the network will be a minibatch of data, of shape (N, d1, ..., dM) where d1 * ... * dM = D. The hidden layer will have H units, and the output layer will produce scores for C classes. Inputs: - x: A TensorFlow Tensor of shape (N, d1, ..., dM) giving a minibatch of input data. - params: A list [w1, w2] of TensorFlow Tensors giving weights for the network, where w1 has shape (D, H) and w2 has shape (H, C). Returns: - scores: A TensorFlow Tensor of shape (N, C) giving classification scores for the input data x. """ w1, w2 = params # Unpack the parameters x = flatten(x) # Flatten the input; now x has shape (N, D) h = tf.nn.relu(tf.matmul(x, w1)) # Hidden layer: h has shape (N, H) scores = tf.matmul(h, w2) # Compute scores of shape (N, C) return scores def two_layer_fc_test(): hidden_layer_size = 42 # Scoping our TF operations under a tf.device context manager # lets us tell TensorFlow where we want these Tensors to be # multiplied and/or operated on, e.g. on a CPU or a GPU. with tf.device(device): x = tf.zeros((64, 32, 32, 3)) w1 = tf.zeros((32 * 32 * 3, hidden_layer_size)) w2 = tf.zeros((hidden_layer_size, 10)) # Call our two_layer_fc function for the forward pass of the network. scores = two_layer_fc(x, [w1, w2]) print(scores.shape) two_layer_fc_test() ###Output (64, 10) ###Markdown Barebones TensorFlow: Three-Layer ConvNetHere you will complete the implementation of the function `three_layer_convnet` which will perform the forward pass of a three-layer convolutional network. The network should have the following architecture:1. A convolutional layer (with bias) with `channel_1` filters, each with shape `KW1 x KH1`, and zero-padding of two2. ReLU nonlinearity3. A convolutional layer (with bias) with `channel_2` filters, each with shape `KW2 x KH2`, and zero-padding of one4. ReLU nonlinearity5. Fully-connected layer with bias, producing scores for `C` classes.**HINT**: For convolutions: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/nn/conv2d; be careful with padding!**HINT**: For biases: https://www.tensorflow.org/performance/xla/broadcasting ###Code def three_layer_convnet(x, params): """ A three-layer convolutional network with the architecture described above. Inputs: - x: A TensorFlow Tensor of shape (N, H, W, 3) giving a minibatch of images - params: A list of TensorFlow Tensors giving the weights and biases for the network; should contain the following: - conv_w1: TensorFlow Tensor of shape (KH1, KW1, 3, channel_1) giving weights for the first convolutional layer. - conv_b1: TensorFlow Tensor of shape (channel_1,) giving biases for the first convolutional layer. - conv_w2: TensorFlow Tensor of shape (KH2, KW2, channel_1, channel_2) giving weights for the second convolutional layer - conv_b2: TensorFlow Tensor of shape (channel_2,) giving biases for the second convolutional layer. - fc_w: TensorFlow Tensor giving weights for the fully-connected layer. Can you figure out what the shape should be? - fc_b: TensorFlow Tensor giving biases for the fully-connected layer. Can you figure out what the shape should be? """ conv_w1, conv_b1, conv_w2, conv_b2, fc_w, fc_b = params scores = None ############################################################################ # TODO: Implement the forward pass for the three-layer ConvNet. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** # print('x.shape = ', x.shape) # print('conv_w1.shape = ', conv_w1.shape) # print('conv_b1.shape = ', conv_b1.shape) # print('conv_w2.shape = ', conv_w2.shape) # print('conv_b2.shape = ', conv_b2.shape) # print('fc_w.shape = ', fc_w.shape) # print('fc_b.shape = ', fc_b.shape) padding1 = tf.constant([[0,0],[2,2],[2,2],[0,0]]) padding2 = tf.constant([[0,0],[1,1],[1,1],[0,0]]) x = tf.pad(x, padding1, 'CONSTANT') l1 = tf.nn.conv2d(x, conv_w1, strides=[1,1,1,1], padding='VALID') + conv_b1 # print('l1 conv shape = ', l1.shape) l1 = tf.nn.relu(l1) # print('l1 relu shape = ', l1.shape) l1 = tf.pad(l1, padding2, 'CONSTANT') l2 = tf.nn.conv2d(l1, conv_w2, strides=[1,1,1,1], padding='VALID') + conv_b2 l2 = tf.nn.relu(l2) l2 = flatten(l2) scores = tf.matmul(l2, fc_w) + fc_b # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return scores ###Output _____no_output_____ ###Markdown After defing the forward pass of the three-layer ConvNet above, run the following cell to test your implementation. Like the two-layer network, we run the graph on a batch of zeros just to make sure the function doesn't crash, and produces outputs of the correct shape.When you run this function, `scores_np` should have shape `(64, 10)`. ###Code def three_layer_convnet_test(): with tf.device(device): x = tf.zeros((64, 32, 32, 3)) conv_w1 = tf.zeros((5, 5, 3, 6)) conv_b1 = tf.zeros((6,)) conv_w2 = tf.zeros((3, 3, 6, 9)) conv_b2 = tf.zeros((9,)) fc_w = tf.zeros((32 * 32 * 9, 10)) fc_b = tf.zeros((10,)) params = [conv_w1, conv_b1, conv_w2, conv_b2, fc_w, fc_b] scores = three_layer_convnet(x, params) # Inputs to convolutional layers are 4-dimensional arrays with shape # [batch_size, height, width, channels] print('scores_np has shape: ', scores.shape) three_layer_convnet_test() ###Output scores_np has shape: (64, 10) ###Markdown Barebones TensorFlow: Training StepWe now define the `training_step` function performs a single training step. This will take three basic steps:1. Compute the loss2. Compute the gradient of the loss with respect to all network weights3. Make a weight update step using (stochastic) gradient descent.We need to use a few new TensorFlow functions to do all of this:- For computing the cross-entropy loss we'll use `tf.nn.sparse_softmax_cross_entropy_with_logits`: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/nn/sparse_softmax_cross_entropy_with_logits- For averaging the loss across a minibatch of data we'll use `tf.reduce_mean`:https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/reduce_mean- For computing gradients of the loss with respect to the weights we'll use `tf.GradientTape` (useful for Eager execution): https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/GradientTape- We'll mutate the weight values stored in a TensorFlow Tensor using `tf.assign_sub` ("sub" is for subtraction): https://www.tensorflow.org/api_docs/python/tf/assign_sub ###Code def training_step(model_fn, x, y, params, learning_rate): with tf.GradientTape() as tape: scores = model_fn(x, params) # Forward pass of the model loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y, logits=scores) total_loss = tf.reduce_mean(loss) grad_params = tape.gradient(total_loss, params) # Make a vanilla gradient descent step on all of the model parameters # Manually update the weights using assign_sub() for w, grad_w in zip(params, grad_params): w.assign_sub(learning_rate * grad_w) return total_loss def train_part2(model_fn, init_fn, learning_rate): """ Train a model on CIFAR-10. Inputs: - model_fn: A Python function that performs the forward pass of the model using TensorFlow; it should have the following signature: scores = model_fn(x, params) where x is a TensorFlow Tensor giving a minibatch of image data, params is a list of TensorFlow Tensors holding the model weights, and scores is a TensorFlow Tensor of shape (N, C) giving scores for all elements of x. - init_fn: A Python function that initializes the parameters of the model. It should have the signature params = init_fn() where params is a list of TensorFlow Tensors holding the (randomly initialized) weights of the model. - learning_rate: Python float giving the learning rate to use for SGD. """ params = init_fn() # Initialize the model parameters for t, (x_np, y_np) in enumerate(train_dset): # Run the graph on a batch of training data. loss = training_step(model_fn, x_np, y_np, params, learning_rate) # Periodically print the loss and check accuracy on the val set. if t % print_every == 0: print('Iteration %d, loss = %.4f' % (t, loss)) check_accuracy(val_dset, x_np, model_fn, params) def check_accuracy(dset, x, model_fn, params): """ Check accuracy on a classification model, e.g. for validation. Inputs: - dset: A Dataset object against which to check accuracy - x: A TensorFlow placeholder Tensor where input images should be fed - model_fn: the Model we will be calling to make predictions on x - params: parameters for the model_fn to work with Returns: Nothing, but prints the accuracy of the model """ num_correct, num_samples = 0, 0 for x_batch, y_batch in dset: scores_np = model_fn(x_batch, params).numpy() y_pred = scores_np.argmax(axis=1) num_samples += x_batch.shape[0] num_correct += (y_pred == y_batch).sum() acc = float(num_correct) / num_samples print('Got %d / %d correct (%.2f%%)' % (num_correct, num_samples, 100 * acc)) ###Output _____no_output_____ ###Markdown Barebones TensorFlow: InitializationWe'll use the following utility method to initialize the weight matrices for our models using Kaiming's normalization method.[1] He et al, *Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification*, ICCV 2015, https://arxiv.org/abs/1502.01852 ###Code def create_matrix_with_kaiming_normal(shape): if len(shape) == 2: fan_in, fan_out = shape[0], shape[1] elif len(shape) == 4: fan_in, fan_out = np.prod(shape[:3]), shape[3] return tf.keras.backend.random_normal(shape) * np.sqrt(2.0 / fan_in) ###Output _____no_output_____ ###Markdown Barebones TensorFlow: Train a Two-Layer NetworkWe are finally ready to use all of the pieces defined above to train a two-layer fully-connected network on CIFAR-10.We just need to define a function to initialize the weights of the model, and call `train_part2`.Defining the weights of the network introduces another important piece of TensorFlow API: `tf.Variable`. A TensorFlow Variable is a Tensor whose value is stored in the graph and persists across runs of the computational graph; however unlike constants defined with `tf.zeros` or `tf.random_normal`, the values of a Variable can be mutated as the graph runs; these mutations will persist across graph runs. Learnable parameters of the network are usually stored in Variables.You don't need to tune any hyperparameters, but you should achieve validation accuracies above 40% after one epoch of training. ###Code def two_layer_fc_init(): """ Initialize the weights of a two-layer network, for use with the two_layer_network function defined above. You can use the `create_matrix_with_kaiming_normal` helper! Inputs: None Returns: A list of: - w1: TensorFlow tf.Variable giving the weights for the first layer - w2: TensorFlow tf.Variable giving the weights for the second layer """ hidden_layer_size = 4000 w1 = tf.Variable(create_matrix_with_kaiming_normal((3 * 32 * 32, 4000))) w2 = tf.Variable(create_matrix_with_kaiming_normal((4000, 10))) return [w1, w2] learning_rate = 1e-2 train_part2(two_layer_fc, two_layer_fc_init, learning_rate) ###Output Iteration 0, loss = 3.4696 Got 148 / 1000 correct (14.80%) Iteration 100, loss = 1.9044 Got 366 / 1000 correct (36.60%) Iteration 200, loss = 1.4232 Got 391 / 1000 correct (39.10%) Iteration 300, loss = 1.8703 Got 383 / 1000 correct (38.30%) Iteration 400, loss = 1.7725 Got 424 / 1000 correct (42.40%) Iteration 500, loss = 1.7954 Got 438 / 1000 correct (43.80%) Iteration 600, loss = 1.7827 Got 427 / 1000 correct (42.70%) Iteration 700, loss = 1.9754 Got 445 / 1000 correct (44.50%) ###Markdown Barebones TensorFlow: Train a three-layer ConvNetWe will now use TensorFlow to train a three-layer ConvNet on CIFAR-10.You need to implement the `three_layer_convnet_init` function. Recall that the architecture of the network is:1. Convolutional layer (with bias) with 32 5x5 filters, with zero-padding 22. ReLU3. Convolutional layer (with bias) with 16 3x3 filters, with zero-padding 14. ReLU5. Fully-connected layer (with bias) to compute scores for 10 classesYou don't need to do any hyperparameter tuning, but you should see validation accuracies above 43% after one epoch of training. ###Code x.shape = (64, 32, 32, 3) conv_w1.shape = (5, 5, 3, 6) conv_b1.shape = (6,) conv_w2.shape = (3, 3, 6, 9) conv_b2.shape = (9,) l1 conv shape = (64, 32, 32, 6) l1 relu shape = (64, 32, 32, 6) scores_np has shape: (64, 10) 9216 def three_layer_convnet_init(): """ Initialize the weights of a Three-Layer ConvNet, for use with the three_layer_convnet function defined above. You can use the `create_matrix_with_kaiming_normal` helper! Inputs: None Returns a list containing: - conv_w1: TensorFlow tf.Variable giving weights for the first conv layer - conv_b1: TensorFlow tf.Variable giving biases for the first conv layer - conv_w2: TensorFlow tf.Variable giving weights for the second conv layer - conv_b2: TensorFlow tf.Variable giving biases for the second conv layer - fc_w: TensorFlow tf.Variable giving weights for the fully-connected layer - fc_b: TensorFlow tf.Variable giving biases for the fully-connected layer """ params = None ############################################################################ # TODO: Initialize the parameters of the three-layer network. # #################§########################################################### # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** conv_w1 = tf.Variable(create_matrix_with_kaiming_normal((5,5,3,32))) conv_b1 = tf.Variable(np.zeros([32]), dtype=tf.float32 ) conv_w2 = tf.Variable(create_matrix_with_kaiming_normal((3,3, 32,16))) conv_b2 = tf.Variable(np.zeros([16]), dtype=tf.float32 ) fc_w = tf.Variable(create_matrix_with_kaiming_normal((32*32*16,10))) fc_b = tf.Variable(np.zeros([10]), dtype=tf.float32 ) params = [conv_w1, conv_b1, conv_w2, conv_b2, fc_w, fc_b] # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return params learning_rate = 3e-3 train_part2(three_layer_convnet, three_layer_convnet_init, learning_rate) ###Output _____no_output_____ ###Markdown Part III: Keras Model Subclassing APIImplementing a neural network using the low-level TensorFlow API is a good way to understand how TensorFlow works, but it's a little inconvenient - we had to manually keep track of all Tensors holding learnable parameters. This was fine for a small network, but could quickly become unweildy for a large complex model.Fortunately TensorFlow 2.0 provides higher-level APIs such as `tf.keras` which make it easy to build models out of modular, object-oriented layers. Further, TensorFlow 2.0 uses eager execution that evaluates operations immediately, without explicitly constructing any computational graphs. This makes it easy to write and debug models, and reduces the boilerplate code.In this part of the notebook we will define neural network models using the `tf.keras.Model` API. To implement your own model, you need to do the following:1. Define a new class which subclasses `tf.keras.Model`. Give your class an intuitive name that describes it, like `TwoLayerFC` or `ThreeLayerConvNet`.2. In the initializer `__init__()` for your new class, define all the layers you need as class attributes. The `tf.keras.layers` package provides many common neural-network layers, like `tf.keras.layers.Dense` for fully-connected layers and `tf.keras.layers.Conv2D` for convolutional layers. Under the hood, these layers will construct `Variable` Tensors for any learnable parameters. **Warning**: Don't forget to call `super(YourModelName, self).__init__()` as the first line in your initializer!3. Implement the `call()` method for your class; this implements the forward pass of your model, and defines the *connectivity* of your network. Layers defined in `__init__()` implement `__call__()` so they can be used as function objects that transform input Tensors into output Tensors. Don't define any new layers in `call()`; any layers you want to use in the forward pass should be defined in `__init__()`.After you define your `tf.keras.Model` subclass, you can instantiate it and use it like the model functions from Part II. Keras Model Subclassing API: Two-Layer NetworkHere is a concrete example of using the `tf.keras.Model` API to define a two-layer network. There are a few new bits of API to be aware of here:We use an `Initializer` object to set up the initial values of the learnable parameters of the layers; in particular `tf.initializers.VarianceScaling` gives behavior similar to the Kaiming initialization method we used in Part II. You can read more about it here: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/initializers/VarianceScalingWe construct `tf.keras.layers.Dense` objects to represent the two fully-connected layers of the model. In addition to multiplying their input by a weight matrix and adding a bias vector, these layer can also apply a nonlinearity for you. For the first layer we specify a ReLU activation function by passing `activation='relu'` to the constructor; the second layer uses softmax activation function. Finally, we use `tf.keras.layers.Flatten` to flatten the output from the previous fully-connected layer. ###Code class TwoLayerFC(tf.keras.Model): def __init__(self, hidden_size, num_classes): super(TwoLayerFC, self).__init__() initializer = tf.initializers.VarianceScaling(scale=2.0) self.fc1 = tf.keras.layers.Dense(hidden_size, activation='relu', kernel_initializer=initializer) self.fc2 = tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer) self.flatten = tf.keras.layers.Flatten() def call(self, x, training=False): x = self.flatten(x) x = self.fc1(x) x = self.fc2(x) return x def test_TwoLayerFC(): """ A small unit test to exercise the TwoLayerFC model above. """ input_size, hidden_size, num_classes = 50, 42, 10 x = tf.zeros((64, input_size)) model = TwoLayerFC(hidden_size, num_classes) with tf.device(device): scores = model(x) print(scores.shape) test_TwoLayerFC() ###Output _____no_output_____ ###Markdown Keras Model Subclassing API: Three-Layer ConvNetNow it's your turn to implement a three-layer ConvNet using the `tf.keras.Model` API. Your model should have the same architecture used in Part II:1. Convolutional layer with 5 x 5 kernels, with zero-padding of 22. ReLU nonlinearity3. Convolutional layer with 3 x 3 kernels, with zero-padding of 14. ReLU nonlinearity5. Fully-connected layer to give class scores6. Softmax nonlinearityYou should initialize the weights of your network using the same initialization method as was used in the two-layer network above.**Hint**: Refer to the documentation for `tf.keras.layers.Conv2D` and `tf.keras.layers.Dense`:https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/Conv2Dhttps://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/Dense ###Code class ThreeLayerConvNet(tf.keras.Model): def __init__(self, channel_1, channel_2, num_classes): super(ThreeLayerConvNet, self).__init__() ######################################################################## # TODO: Implement the __init__ method for a three-layer ConvNet. You # # should instantiate layer objects to be used in the forward pass. # ######################################################################## # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** initializer = tf.initializers.VarianceScaling(scale=2.0) self.conv1 = tf.keras.layers.Conv2D(filters=channel_1, kernel_size=(5,5), activation='relu', input_shape=(32,32,3), padding="valid") self.conv2 = tf.keras.layers.Conv2D(filters=channel_2, kernel_size=(3,3), activation='relu') self.fc = tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer) self.flatten = tf.keras.layers.Flatten() # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ######################################################################## # END OF YOUR CODE # ######################################################################## def call(self, x, training=False): scores = None ######################################################################## # TODO: Implement the forward pass for a three-layer ConvNet. You # # should use the layer objects defined in the __init__ method. # ######################################################################## # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** padding1 = tf.constant([[0,0],[2,2],[2,2],[0,0]]) padding2 = tf.constant([[0,0],[1,1],[1,1],[0,0]]) x = tf.pad(x, padding1, 'CONSTANT') x = self.conv1(x) x = tf.pad(x, padding2, 'CONSTANT') x = self.conv2(x) x = self.flatten(x) scores = self.fc(x) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ######################################################################## # END OF YOUR CODE # ######################################################################## return scores ###Output _____no_output_____ ###Markdown Once you complete the implementation of the `ThreeLayerConvNet` above you can run the following to ensure that your implementation does not crash and produces outputs of the expected shape. ###Code def test_ThreeLayerConvNet(): channel_1, channel_2, num_classes = 12, 8, 10 model = ThreeLayerConvNet(channel_1, channel_2, num_classes) with tf.device(device): x = tf.zeros((64, 3, 32, 32)) scores = model(x) print(scores.shape) test_ThreeLayerConvNet() ###Output _____no_output_____ ###Markdown Keras Model Subclassing API: Eager TrainingWhile keras models have a builtin training loop (using the `model.fit`), sometimes you need more customization. Here's an example, of a training loop implemented with eager execution.In particular, notice `tf.GradientTape`. Automatic differentiation is used in the backend for implementing backpropagation in frameworks like TensorFlow. During eager execution, `tf.GradientTape` is used to trace operations for computing gradients later. A particular `tf.GradientTape` can only compute one gradient; subsequent calls to tape will throw a runtime error. TensorFlow 2.0 ships with easy-to-use built-in metrics under `tf.keras.metrics` module. Each metric is an object, and we can use `update_state()` to add observations and `reset_state()` to clear all observations. We can get the current result of a metric by calling `result()` on the metric object. ###Code def train_part34(model_init_fn, optimizer_init_fn, num_epochs=1, is_training=False): """ Simple training loop for use with models defined using tf.keras. It trains a model for one epoch on the CIFAR-10 training set and periodically checks accuracy on the CIFAR-10 validation set. Inputs: - model_init_fn: A function that takes no parameters; when called it constructs the model we want to train: model = model_init_fn() - optimizer_init_fn: A function which takes no parameters; when called it constructs the Optimizer object we will use to optimize the model: optimizer = optimizer_init_fn() - num_epochs: The number of epochs to train for Returns: Nothing, but prints progress during trainingn """ with tf.device(device): # Compute the loss like we did in Part II loss_fn = tf.keras.losses.SparseCategoricalCrossentropy() model = model_init_fn() optimizer = optimizer_init_fn() train_loss = tf.keras.metrics.Mean(name='train_loss') train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy') val_loss = tf.keras.metrics.Mean(name='val_loss') val_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='val_accuracy') t = 0 for epoch in range(num_epochs): # Reset the metrics - https://www.tensorflow.org/alpha/guide/migration_guide#new-style_metrics train_loss.reset_states() train_accuracy.reset_states() for x_np, y_np in train_dset: with tf.GradientTape() as tape: # Use the model function to build the forward pass. scores = model(x_np, training=is_training) loss = loss_fn(y_np, scores) gradients = tape.gradient(loss, model.trainable_variables) optimizer.apply_gradients(zip(gradients, model.trainable_variables)) # Update the metrics train_loss.update_state(loss) train_accuracy.update_state(y_np, scores) if t % print_every == 0: val_loss.reset_states() val_accuracy.reset_states() for test_x, test_y in val_dset: # During validation at end of epoch, training set to False prediction = model(test_x, training=False) t_loss = loss_fn(test_y, prediction) val_loss.update_state(t_loss) val_accuracy.update_state(test_y, prediction) template = 'Iteration {}, Epoch {}, Loss: {}, Accuracy: {}, Val Loss: {}, Val Accuracy: {}' print (template.format(t, epoch+1, train_loss.result(), train_accuracy.result()*100, val_loss.result(), val_accuracy.result()*100)) t += 1 ###Output _____no_output_____ ###Markdown Keras Model Subclassing API: Train a Two-Layer NetworkWe can now use the tools defined above to train a two-layer network on CIFAR-10. We define the `model_init_fn` and `optimizer_init_fn` that construct the model and optimizer respectively when called. Here we want to train the model using stochastic gradient descent with no momentum, so we construct a `tf.keras.optimizers.SGD` function; you can [read about it here](https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/optimizers/SGD).You don't need to tune any hyperparameters here, but you should achieve validation accuracies above 40% after one epoch of training. ###Code hidden_size, num_classes = 4000, 10 learning_rate = 1e-2 def model_init_fn(): return TwoLayerFC(hidden_size, num_classes) def optimizer_init_fn(): return tf.keras.optimizers.SGD(learning_rate=learning_rate) train_part34(model_init_fn, optimizer_init_fn) ###Output _____no_output_____ ###Markdown Keras Model Subclassing API: Train a Three-Layer ConvNetHere you should use the tools we've defined above to train a three-layer ConvNet on CIFAR-10. Your ConvNet should use 32 filters in the first convolutional layer and 16 filters in the second layer.To train the model you should use gradient descent with Nesterov momentum 0.9. **HINT**: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/optimizers/SGDYou don't need to perform any hyperparameter tuning, but you should achieve validation accuracies above 50% after training for one epoch. ###Code learning_rate = 3e-3 channel_1, channel_2, num_classes = 32, 16, 10 def model_init_fn(): model = None ############################################################################ # TODO: Complete the implementation of model_fn. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** model = ThreeLayerConvNet(channel_1, channel_2, num_classes) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return model def optimizer_init_fn(): optimizer = None ############################################################################ # TODO: Complete the implementation of model_fn. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate, momentum=0.9) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return optimizer train_part34(model_init_fn, optimizer_init_fn) ###Output _____no_output_____ ###Markdown Part IV: Keras Sequential APIIn Part III we introduced the `tf.keras.Model` API, which allows you to define models with any number of learnable layers and with arbitrary connectivity between layers.However for many models you don't need such flexibility - a lot of models can be expressed as a sequential stack of layers, with the output of each layer fed to the next layer as input. If your model fits this pattern, then there is an even easier way to define your model: using `tf.keras.Sequential`. You don't need to write any custom classes; you simply call the `tf.keras.Sequential` constructor with a list containing a sequence of layer objects.One complication with `tf.keras.Sequential` is that you must define the shape of the input to the model by passing a value to the `input_shape` of the first layer in your model. Keras Sequential API: Two-Layer NetworkIn this subsection, we will rewrite the two-layer fully-connected network using `tf.keras.Sequential`, and train it using the training loop defined above.You don't need to perform any hyperparameter tuning here, but you should see validation accuracies above 40% after training for one epoch. ###Code learning_rate = 1e-2 def model_init_fn(): input_shape = (32, 32, 3) hidden_layer_size, num_classes = 4000, 10 initializer = tf.initializers.VarianceScaling(scale=2.0) layers = [ tf.keras.layers.Flatten(input_shape=input_shape), tf.keras.layers.Dense(hidden_layer_size, activation='relu', kernel_initializer=initializer), tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer), ] model = tf.keras.Sequential(layers) return model def optimizer_init_fn(): return tf.keras.optimizers.SGD(learning_rate=learning_rate) train_part34(model_init_fn, optimizer_init_fn) ###Output _____no_output_____ ###Markdown Abstracting Away the Training LoopIn the previous examples, we used a customised training loop to train models (e.g. `train_part34`). Writing your own training loop is only required if you need more flexibility and control during training your model. Alternately, you can also use built-in APIs like `tf.keras.Model.fit()` and `tf.keras.Model.evaluate` to train and evaluate a model. Also remember to configure your model for training by calling `tf.keras.Model.compile.You don't need to perform any hyperparameter tuning here, but you should see validation and test accuracies above 42% after training for one epoch. ###Code model = model_init_fn() model.compile(optimizer=tf.keras.optimizers.SGD(learning_rate=learning_rate), loss='sparse_categorical_crossentropy', metrics=[tf.keras.metrics.sparse_categorical_accuracy]) model.fit(X_train, y_train, batch_size=64, epochs=1, validation_data=(X_val, y_val)) model.evaluate(X_test, y_test) ###Output _____no_output_____ ###Markdown Keras Sequential API: Three-Layer ConvNetHere you should use `tf.keras.Sequential` to reimplement the same three-layer ConvNet architecture used in Part II and Part III. As a reminder, your model should have the following architecture:1. Convolutional layer with 32 5x5 kernels, using zero padding of 22. ReLU nonlinearity3. Convolutional layer with 16 3x3 kernels, using zero padding of 14. ReLU nonlinearity5. Fully-connected layer giving class scores6. Softmax nonlinearityYou should initialize the weights of the model using a `tf.initializers.VarianceScaling` as above.You should train the model using Nesterov momentum 0.9.You don't need to perform any hyperparameter search, but you should achieve accuracy above 45% after training for one epoch. ###Code def model_init_fn(): model = None ############################################################################ # TODO: Construct a three-layer ConvNet using tf.keras.Sequential. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** initializer = tf.initializers.VarianceScaling(scale=2.0) depth1, depth2 = 32, 16 ker_size1, ker_size2 = (5,5), (3,3) num_classes = 10 input_shape=(32,32,3) layers = [ tf.keras.layers.InputLayer(input_shape=input_shape), tf.keras.layers.Conv2D(filters=depth1, kernel_size=ker_size1, padding='same', activation = 'relu', kernel_initializer=initializer), # tf.keras.layers.ZeroPadding2D((1,1)), tf.keras.layers.Conv2D(filters=depth2, kernel_size=ker_size2, padding='same', activation = 'relu', kernel_initializer=initializer), tf.keras.layers.Flatten(), tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer), ] model = tf.keras.Sequential(layers) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return model learning_rate = 5e-4 def optimizer_init_fn(): optimizer = None ############################################################################ # TODO: Complete the implementation of model_fn. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate, momentum=0.9) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return optimizer train_part34(model_init_fn, optimizer_init_fn) ###Output _____no_output_____ ###Markdown We will also train this model with the built-in training loop APIs provided by TensorFlow. ###Code model = model_init_fn() model.compile(optimizer=optimizer_init_fn(), loss='sparse_categorical_crossentropy', metrics=[tf.keras.metrics.sparse_categorical_accuracy]) model.fit(X_train, y_train, batch_size=64, epochs=1, validation_data=(X_val, y_val)) model.evaluate(X_test, y_test) ###Output _____no_output_____ ###Markdown Part IV: Functional API Demonstration with a Two-Layer Network In the previous section, we saw how we can use `tf.keras.Sequential` to stack layers to quickly build simple models. But this comes at the cost of losing flexibility.Often we will have to write complex models that have non-sequential data flows: a layer can have **multiple inputs and/or outputs**, such as stacking the output of 2 previous layers together to feed as input to a third! (Some examples are residual connections and dense blocks.)In such cases, we can use Keras functional API to write models with complex topologies such as: 1. Multi-input models 2. Multi-output models 3. Models with shared layers (the same layer called several times) 4. Models with non-sequential data flows (e.g. residual connections)Writing a model with Functional API requires us to create a `tf.keras.Model` instance and explicitly write input tensors and output tensors for this model. ###Code def two_layer_fc_functional(input_shape, hidden_size, num_classes): initializer = tf.initializers.VarianceScaling(scale=2.0) inputs = tf.keras.Input(shape=input_shape) flattened_inputs = tf.keras.layers.Flatten()(inputs) fc1_output = tf.keras.layers.Dense(hidden_size, activation='relu', kernel_initializer=initializer)(flattened_inputs) scores = tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer)(fc1_output) # Instantiate the model given inputs and outputs. model = tf.keras.Model(inputs=inputs, outputs=scores) return model def test_two_layer_fc_functional(): """ A small unit test to exercise the TwoLayerFC model above. """ input_size, hidden_size, num_classes = 50, 42, 10 input_shape = (50,) x = tf.zeros((64, input_size)) model = two_layer_fc_functional(input_shape, hidden_size, num_classes) with tf.device(device): scores = model(x) print(scores.shape) test_two_layer_fc_functional() ###Output _____no_output_____ ###Markdown Keras Functional API: Train a Two-Layer NetworkYou can now train this two-layer network constructed using the functional API.You don't need to perform any hyperparameter tuning here, but you should see validation accuracies above 40% after training for one epoch. ###Code input_shape = (32, 32, 3) hidden_size, num_classes = 4000, 10 learning_rate = 1e-2 def model_init_fn(): return two_layer_fc_functional(input_shape, hidden_size, num_classes) def optimizer_init_fn(): return tf.keras.optimizers.SGD(learning_rate=learning_rate) train_part34(model_init_fn, optimizer_init_fn) ###Output _____no_output_____ ###Markdown Part V: CIFAR-10 open-ended challengeIn this section you can experiment with whatever ConvNet architecture you'd like on CIFAR-10.You should experiment with architectures, hyperparameters, loss functions, regularization, or anything else you can think of to train a model that achieves **at least 70%** accuracy on the **validation** set within 10 epochs. You can use the built-in train function, the `train_part34` function from above, or implement your own training loop.Describe what you did at the end of the notebook. Some things you can try:- **Filter size**: Above we used 5x5 and 3x3; is this optimal?- **Number of filters**: Above we used 16 and 32 filters. Would more or fewer do better?- **Pooling**: We didn't use any pooling above. Would this improve the model?- **Normalization**: Would your model be improved with batch normalization, layer normalization, group normalization, or some other normalization strategy?- **Network architecture**: The ConvNet above has only three layers of trainable parameters. Would a deeper model do better?- **Global average pooling**: Instead of flattening after the final convolutional layer, would global average pooling do better? This strategy is used for example in Google's Inception network and in Residual Networks.- **Regularization**: Would some kind of regularization improve performance? Maybe weight decay or dropout? NOTE: Batch Normalization / DropoutIf you are using Batch Normalization and Dropout, remember to pass `is_training=True` if you use the `train_part34()` function. BatchNorm and Dropout layers have different behaviors at training and inference time. `training` is a specific keyword argument reserved for this purpose in any `tf.keras.Model`'s `call()` function. Read more about this here : https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/BatchNormalizationmethodshttps://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/Dropoutmethods Tips for trainingFor each network architecture that you try, you should tune the learning rate and other hyperparameters. When doing this there are a couple important things to keep in mind: - If the parameters are working well, you should see improvement within a few hundred iterations- Remember the coarse-to-fine approach for hyperparameter tuning: start by testing a large range of hyperparameters for just a few training iterations to find the combinations of parameters that are working at all.- Once you have found some sets of parameters that seem to work, search more finely around these parameters. You may need to train for more epochs.- You should use the validation set for hyperparameter search, and save your test set for evaluating your architecture on the best parameters as selected by the validation set. Going above and beyondIf you are feeling adventurous there are many other features you can implement to try and improve your performance. You are **not required** to implement any of these, but don't miss the fun if you have time!- Alternative optimizers: you can try Adam, Adagrad, RMSprop, etc.- Alternative activation functions such as leaky ReLU, parametric ReLU, ELU, or MaxOut.- Model ensembles- Data augmentation- New Architectures - [ResNets](https://arxiv.org/abs/1512.03385) where the input from the previous layer is added to the output. - [DenseNets](https://arxiv.org/abs/1608.06993) where inputs into previous layers are concatenated together. - [This blog has an in-depth overview](https://chatbotslife.com/resnets-highwaynets-and-densenets-oh-my-9bb15918ee32) Have fun and happy training! ###Code class CustomConvNet(tf.keras.Model): def __init__(self): super(CustomConvNet, self).__init__() ############################################################################ # TODO: Construct a model that performs well on CIFAR-10 # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** initializer = tf.initializers.VarianceScaling(scale=2.0) self.conv1 = tf.keras.layers.Conv2D(filters=channel_1, kernel_size=(5,5), activation='relu', input_shape=(32,32,3), padding="valid") self.conv2 = tf.keras.layers.Conv2D(filters=channel_2, kernel_size=(3,3), activation='relu') self.fc = tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer) self.flatten = tf.keras.layers.Flatten() # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ######################################################################## # END OF YOUR CODE # ######################################################################## def call(self, x, training=False): scores = None ######################################################################## # TODO: Implement the forward pass for a three-layer ConvNet. You # # should use the layer objects defined in the __init__ method. # ######################################################################## # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** padding1 = tf.constant([[0,0],[2,2],[2,2],[0,0]]) padding2 = tf.constant([[0,0],[1,1],[1,1],[0,0]]) x = tf.pad(x, padding1, 'CONSTANT') x = self.conv1(x) x = tf.pad(x, padding2, 'CONSTANT') x = self.conv2(x) x = self.flatten(x) x = self.fc(x) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return x # device = '/device:GPU:0' # Change this to a CPU/GPU as you wish! device = '/cpu:0' # Change this to a CPU/GPU as you wish! print_every = 700 num_epochs = 10 model = CustomConvNet() def model_init_fn(): return CustomConvNet() def optimizer_init_fn(): learning_rate = 1e-3 return tf.keras.optimizers.Adam(learning_rate) train_part34(model_init_fn, optimizer_init_fn, num_epochs=num_epochs, is_training=True) ###Output Iteration 0, Epoch 1, Loss: 2.3400373458862305, Accuracy: 7.8125, Val Loss: 2.9617037773132324, Val Accuracy: 10.0 Iteration 700, Epoch 1, Loss: 1.413520336151123, Accuracy: 50.49705505371094, Val Loss: 1.1875035762786865, Val Accuracy: 58.20000076293945 Iteration 1400, Epoch 2, Loss: 1.0429129600524902, Accuracy: 63.917320251464844, Val Loss: 1.1137940883636475, Val Accuracy: 62.0 Iteration 2100, Epoch 3, Loss: 0.8839472532272339, Accuracy: 69.85116577148438, Val Loss: 1.1555709838867188, Val Accuracy: 63.0 Iteration 2800, Epoch 4, Loss: 0.7628107070922852, Accuracy: 74.03392028808594, Val Loss: 1.2096624374389648, Val Accuracy: 60.900001525878906 Iteration 3500, Epoch 5, Loss: 0.668714165687561, Accuracy: 77.58509826660156, Val Loss: 1.2593955993652344, Val Accuracy: 61.400001525878906 Iteration 4200, Epoch 6, Loss: 0.5919941067695618, Accuracy: 80.41610717773438, Val Loss: 1.4657975435256958, Val Accuracy: 57.79999923706055 Iteration 4900, Epoch 7, Loss: 0.5167434811592102, Accuracy: 82.61270141601562, Val Loss: 1.6434582471847534, Val Accuracy: 58.20000076293945 Iteration 5600, Epoch 8, Loss: 0.459024041891098, Accuracy: 84.54498291015625, Val Loss: 1.7989230155944824, Val Accuracy: 53.20000076293945 Iteration 6300, Epoch 9, Loss: 0.4310932159423828, Accuracy: 84.62789154052734, Val Loss: 2.1613223552703857, Val Accuracy: 54.5 Iteration 7000, Epoch 10, Loss: 0.3747079372406006, Accuracy: 87.16413116455078, Val Loss: 2.216815710067749, Val Accuracy: 57.599998474121094 ###Markdown What's this TensorFlow business?You've written a lot of code in this assignment to provide a whole host of neural network functionality. Dropout, Batch Norm, and 2D convolutions are some of the workhorses of deep learning in computer vision. You've also worked hard to make your code efficient and vectorized.For the last part of this assignment, though, we're going to leave behind your beautiful codebase and instead migrate to one of two popular deep learning frameworks: in this instance, TensorFlow (or PyTorch, if you choose to work with that notebook). What is it?TensorFlow is a system for executing computational graphs over Tensor objects, with native support for performing backpropogation for its Variables. In it, we work with Tensors which are n-dimensional arrays analogous to the numpy ndarray. Why?* Our code will now run on GPUs! Much faster training. Writing your own modules to run on GPUs is beyond the scope of this class, unfortunately.* We want you to be ready to use one of these frameworks for your project so you can experiment more efficiently than if you were writing every feature you want to use by hand. * We want you to stand on the shoulders of giants! TensorFlow and PyTorch are both excellent frameworks that will make your lives a lot easier, and now that you understand their guts, you are free to use them :) * We want you to be exposed to the sort of deep learning code you might run into in academia or industry. How will I learn TensorFlow?TensorFlow has many excellent tutorials available, including those from [Google themselves](https://www.tensorflow.org/get_started/get_started).Otherwise, this notebook will walk you through much of what you need to do to train models in TensorFlow. See the end of the notebook for some links to helpful tutorials if you want to learn more or need further clarification on topics that aren't fully explained here.**NOTE: This notebook is meant to teach you the latest version of Tensorflow 2.0. Most examples on the web today are still in 1.x, so be careful not to confuse the two when looking up documentation**. Install Tensorflow 2.0Tensorflow 2.0 is still not in a fully 100% stable release, but it's still usable and more intuitive than TF 1.x. Please make sure you have it installed before moving on in this notebook! Here are some steps to get started:1. Have the latest version of Anaconda installed on your machine.2. Create a new conda environment starting from Python 3.7. In this setup example, we'll call it `tf_20_env`.3. Run the command: `source activate tf_20_env`4. Then pip install TF 2.0 as described here: https://www.tensorflow.org/install/pip A guide on creating Anaconda enviornments: https://uoa-eresearch.github.io/eresearch-cookbook/recipe/2014/11/20/conda/This will give you an new enviornemnt to play in TF 2.0. Generally, if you plan to also use TensorFlow in your other projects, you might also want to keep a seperate Conda environment or virtualenv in Python 3.7 that has Tensorflow 1.9, so you can switch back and forth at will. Table of ContentsThis notebook has 5 parts. We will walk through TensorFlow at **three different levels of abstraction**, which should help you better understand it and prepare you for working on your project.1. Part I, Preparation: load the CIFAR-10 dataset.2. Part II, Barebone TensorFlow: **Abstraction Level 1**, we will work directly with low-level TensorFlow graphs. 3. Part III, Keras Model API: **Abstraction Level 2**, we will use `tf.keras.Model` to define arbitrary neural network architecture. 4. Part IV, Keras Sequential + Functional API: **Abstraction Level 3**, we will use `tf.keras.Sequential` to define a linear feed-forward network very conveniently, and then explore the functional libraries for building unique and uncommon models that require more flexibility.5. Part V, CIFAR-10 open-ended challenge: please implement your own network to get as high accuracy as possible on CIFAR-10. You can experiment with any layer, optimizer, hyperparameters or other advanced features. We will discuss Keras in more detail later in the notebook.Here is a table of comparison:| API | Flexibility | Convenience ||---------------|-------------|-------------|| Barebone | High | Low || `tf.keras.Model` | High | Medium || `tf.keras.Sequential` | Low | High | Part I: PreparationFirst, we load the CIFAR-10 dataset. This might take a few minutes to download the first time you run it, but after that the files should be cached on disk and loading should be faster.In previous parts of the assignment we used CS231N-specific code to download and read the CIFAR-10 dataset; however the `tf.keras.datasets` package in TensorFlow provides prebuilt utility functions for loading many common datasets.For the purposes of this assignment we will still write our own code to preprocess the data and iterate through it in minibatches. The `tf.data` package in TensorFlow provides tools for automating this process, but working with this package adds extra complication and is beyond the scope of this notebook. However using `tf.data` can be much more efficient than the simple approach used in this notebook, so you should consider using it for your project. ###Code import os import tensorflow as tf import numpy as np import math import timeit import matplotlib.pyplot as plt %matplotlib inline def load_cifar10(num_training=49000, num_validation=1000, num_test=10000): """ Fetch the CIFAR-10 dataset from the web and perform preprocessing to prepare it for the two-layer neural net classifier. These are the same steps as we used for the SVM, but condensed to a single function. """ # Load the raw CIFAR-10 dataset and use appropriate data types and shapes cifar10 = tf.keras.datasets.cifar10.load_data() (X_train, y_train), (X_test, y_test) = cifar10 X_train = np.asarray(X_train, dtype=np.float32) y_train = np.asarray(y_train, dtype=np.int32).flatten() X_test = np.asarray(X_test, dtype=np.float32) y_test = np.asarray(y_test, dtype=np.int32).flatten() # Subsample the data mask = range(num_training, num_training + num_validation) X_val = X_train[mask] y_val = y_train[mask] mask = range(num_training) X_train = X_train[mask] y_train = y_train[mask] mask = range(num_test) X_test = X_test[mask] y_test = y_test[mask] # Normalize the data: subtract the mean pixel and divide by std mean_pixel = X_train.mean(axis=(0, 1, 2), keepdims=True) std_pixel = X_train.std(axis=(0, 1, 2), keepdims=True) X_train = (X_train - mean_pixel) / std_pixel X_val = (X_val - mean_pixel) / std_pixel X_test = (X_test - mean_pixel) / std_pixel return X_train, y_train, X_val, y_val, X_test, y_test # If there are errors with SSL downloading involving self-signed certificates, # it may be that your Python version was recently installed on the current machine. # See: https://github.com/tensorflow/tensorflow/issues/10779 # To fix, run the command: /Applications/Python\ 3.7/Install\ Certificates.command # ...replacing paths as necessary. # Invoke the above function to get our data. NHW = (0, 1, 2) X_train, y_train, X_val, y_val, X_test, y_test = load_cifar10() print('Train data shape: ', X_train.shape) print('Train labels shape: ', y_train.shape, y_train.dtype) print('Validation data shape: ', X_val.shape) print('Validation labels shape: ', y_val.shape) print('Test data shape: ', X_test.shape) print('Test labels shape: ', y_test.shape) class Dataset(object): def __init__(self, X, y, batch_size, shuffle=False): """ Construct a Dataset object to iterate over data X and labels y Inputs: - X: Numpy array of data, of any shape - y: Numpy array of labels, of any shape but with y.shape[0] == X.shape[0] - batch_size: Integer giving number of elements per minibatch - shuffle: (optional) Boolean, whether to shuffle the data on each epoch """ assert X.shape[0] == y.shape[0], 'Got different numbers of data and labels' self.X, self.y = X, y self.batch_size, self.shuffle = batch_size, shuffle def __iter__(self): N, B = self.X.shape[0], self.batch_size idxs = np.arange(N) if self.shuffle: np.random.shuffle(idxs) return iter((self.X[i:i+B], self.y[i:i+B]) for i in range(0, N, B)) train_dset = Dataset(X_train, y_train, batch_size=64, shuffle=True) val_dset = Dataset(X_val, y_val, batch_size=64, shuffle=False) test_dset = Dataset(X_test, y_test, batch_size=64) # We can iterate through a dataset like this: for t, (x, y) in enumerate(train_dset): print(t, x.shape, y.shape) if t > 5: break ###Output _____no_output_____ ###Markdown You can optionally **use GPU by setting the flag to True below**. It's not neccessary to use a GPU for this assignment; if you are working on Google Cloud then we recommend that you do not use a GPU, as it will be significantly more expensive. ###Code # Set up some global variables USE_GPU = True if USE_GPU: device = '/device:GPU:0' else: device = '/cpu:0' # Constant to control how often we print when training models print_every = 100 print('Using device: ', device) ###Output _____no_output_____ ###Markdown Part II: Barebones TensorFlowTensorFlow ships with various high-level APIs which make it very convenient to define and train neural networks; we will cover some of these constructs in Part III and Part IV of this notebook. In this section we will start by building a model with basic TensorFlow constructs to help you better understand what's going on under the hood of the higher-level APIs.**"Barebones Tensorflow" is important to understanding the building blocks of TensorFlow, but much of it involves concepts from TensorFlow 1.x.** We will be working with legacy modules such as `tf.Variable`.Therefore, please read and understand the differences between legacy (1.x) TF and the new (2.0) TF. Historical background on TensorFlow 1.xTensorFlow 1.x is primarily a framework for working with **static computational graphs**. Nodes in the computational graph are Tensors which will hold n-dimensional arrays when the graph is run; edges in the graph represent functions that will operate on Tensors when the graph is run to actually perform useful computation.Before Tensorflow 2.0, we had to configure the graph into two phases. There are plenty of tutorials online that explain this two-step process. The process generally looks like the following for TF 1.x:1. **Build a computational graph that describes the computation that you want to perform**. This stage doesn't actually perform any computation; it just builds up a symbolic representation of your computation. This stage will typically define one or more `placeholder` objects that represent inputs to the computational graph.2. **Run the computational graph many times.** Each time the graph is run (e.g. for one gradient descent step) you will specify which parts of the graph you want to compute, and pass a `feed_dict` dictionary that will give concrete values to any `placeholder`s in the graph. The new paradigm in Tensorflow 2.0Now, with Tensorflow 2.0, we can simply adopt a functional form that is more Pythonic and similar in spirit to PyTorch and direct Numpy operation. Instead of the 2-step paradigm with computation graphs, making it (among other things) easier to debug TF code. You can read more details at https://www.tensorflow.org/guide/eager.The main difference between the TF 1.x and 2.0 approach is that the 2.0 approach doesn't make use of `tf.Session`, `tf.run`, `placeholder`, `feed_dict`. To get more details of what's different between the two version and how to convert between the two, check out the official migration guide: https://www.tensorflow.org/alpha/guide/migration_guideLater, in the rest of this notebook we'll focus on this new, simpler approach. TensorFlow warmup: Flatten FunctionWe can see this in action by defining a simple `flatten` function that will reshape image data for use in a fully-connected network.In TensorFlow, data for convolutional feature maps is typically stored in a Tensor of shape N x H x W x C where:- N is the number of datapoints (minibatch size)- H is the height of the feature map- W is the width of the feature map- C is the number of channels in the feature mapThis is the right way to represent the data when we are doing something like a 2D convolution, that needs spatial understanding of where the intermediate features are relative to each other. When we use fully connected affine layers to process the image, however, we want each datapoint to be represented by a single vector -- it's no longer useful to segregate the different channels, rows, and columns of the data. So, we use a "flatten" operation to collapse the `H x W x C` values per representation into a single long vector. Notice the `tf.reshape` call has the target shape as `(N, -1)`, meaning it will reshape/keep the first dimension to be N, and then infer as necessary what the second dimension is in the output, so we can collapse the remaining dimensions from the input properly.**NOTE**: TensorFlow and PyTorch differ on the default Tensor layout; TensorFlow uses N x H x W x C but PyTorch uses N x C x H x W. ###Code def flatten(x): """ Input: - TensorFlow Tensor of shape (N, D1, ..., DM) Output: - TensorFlow Tensor of shape (N, D1 * ... * DM) """ N = tf.shape(x)[0] return tf.reshape(x, (N, -1)) def test_flatten(): # Construct concrete values of the input data x using numpy x_np = np.arange(24).reshape((2, 3, 4)) print('x_np:\n', x_np, '\n') # Compute a concrete output value. x_flat_np = flatten(x_np) print('x_flat_np:\n', x_flat_np, '\n') test_flatten() ###Output _____no_output_____ ###Markdown Barebones TensorFlow: Define a Two-Layer NetworkWe will now implement our first neural network with TensorFlow: a fully-connected ReLU network with two hidden layers and no biases on the CIFAR10 dataset. For now we will use only low-level TensorFlow operators to define the network; later we will see how to use the higher-level abstractions provided by `tf.keras` to simplify the process.We will define the forward pass of the network in the function `two_layer_fc`; this will accept TensorFlow Tensors for the inputs and weights of the network, and return a TensorFlow Tensor for the scores. After defining the network architecture in the `two_layer_fc` function, we will test the implementation by checking the shape of the output.**It's important that you read and understand this implementation.** ###Code def two_layer_fc(x, params): """ A fully-connected neural network; the architecture is: fully-connected layer -> ReLU -> fully connected layer. Note that we only need to define the forward pass here; TensorFlow will take care of computing the gradients for us. The input to the network will be a minibatch of data, of shape (N, d1, ..., dM) where d1 * ... * dM = D. The hidden layer will have H units, and the output layer will produce scores for C classes. Inputs: - x: A TensorFlow Tensor of shape (N, d1, ..., dM) giving a minibatch of input data. - params: A list [w1, w2] of TensorFlow Tensors giving weights for the network, where w1 has shape (D, H) and w2 has shape (H, C). Returns: - scores: A TensorFlow Tensor of shape (N, C) giving classification scores for the input data x. """ w1, w2 = params # Unpack the parameters x = flatten(x) # Flatten the input; now x has shape (N, D) h = tf.nn.relu(tf.matmul(x, w1)) # Hidden layer: h has shape (N, H) scores = tf.matmul(h, w2) # Compute scores of shape (N, C) return scores def two_layer_fc_test(): hidden_layer_size = 42 # Scoping our TF operations under a tf.device context manager # lets us tell TensorFlow where we want these Tensors to be # multiplied and/or operated on, e.g. on a CPU or a GPU. with tf.device(device): x = tf.zeros((64, 32, 32, 3)) w1 = tf.zeros((32 * 32 * 3, hidden_layer_size)) w2 = tf.zeros((hidden_layer_size, 10)) # Call our two_layer_fc function for the forward pass of the network. scores = two_layer_fc(x, [w1, w2]) print(scores.shape) two_layer_fc_test() ###Output _____no_output_____ ###Markdown Barebones TensorFlow: Three-Layer ConvNetHere you will complete the implementation of the function `three_layer_convnet` which will perform the forward pass of a three-layer convolutional network. The network should have the following architecture:1. A convolutional layer (with bias) with `channel_1` filters, each with shape `KW1 x KH1`, and zero-padding of two2. ReLU nonlinearity3. A convolutional layer (with bias) with `channel_2` filters, each with shape `KW2 x KH2`, and zero-padding of one4. ReLU nonlinearity5. Fully-connected layer with bias, producing scores for `C` classes.**HINT**: For convolutions: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/nn/conv2d; be careful with padding!**HINT**: For biases: https://www.tensorflow.org/performance/xla/broadcasting ###Code def three_layer_convnet(x, params): """ A three-layer convolutional network with the architecture described above. Inputs: - x: A TensorFlow Tensor of shape (N, H, W, 3) giving a minibatch of images - params: A list of TensorFlow Tensors giving the weights and biases for the network; should contain the following: - conv_w1: TensorFlow Tensor of shape (KH1, KW1, 3, channel_1) giving weights for the first convolutional layer. - conv_b1: TensorFlow Tensor of shape (channel_1,) giving biases for the first convolutional layer. - conv_w2: TensorFlow Tensor of shape (KH2, KW2, channel_1, channel_2) giving weights for the second convolutional layer - conv_b2: TensorFlow Tensor of shape (channel_2,) giving biases for the second convolutional layer. - fc_w: TensorFlow Tensor giving weights for the fully-connected layer. Can you figure out what the shape should be? - fc_b: TensorFlow Tensor giving biases for the fully-connected layer. Can you figure out what the shape should be? """ conv_w1, conv_b1, conv_w2, conv_b2, fc_w, fc_b = params scores = None ############################################################################ # TODO: Implement the forward pass for the three-layer ConvNet. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** pass # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return scores ###Output _____no_output_____ ###Markdown After defing the forward pass of the three-layer ConvNet above, run the following cell to test your implementation. Like the two-layer network, we run the graph on a batch of zeros just to make sure the function doesn't crash, and produces outputs of the correct shape.When you run this function, `scores_np` should have shape `(64, 10)`. ###Code def three_layer_convnet_test(): with tf.device(device): x = tf.zeros((64, 32, 32, 3)) conv_w1 = tf.zeros((5, 5, 3, 6)) conv_b1 = tf.zeros((6,)) conv_w2 = tf.zeros((3, 3, 6, 9)) conv_b2 = tf.zeros((9,)) fc_w = tf.zeros((32 * 32 * 9, 10)) fc_b = tf.zeros((10,)) params = [conv_w1, conv_b1, conv_w2, conv_b2, fc_w, fc_b] scores = three_layer_convnet(x, params) # Inputs to convolutional layers are 4-dimensional arrays with shape # [batch_size, height, width, channels] print('scores_np has shape: ', scores.shape) three_layer_convnet_test() ###Output _____no_output_____ ###Markdown Barebones TensorFlow: Training StepWe now define the `training_step` function performs a single training step. This will take three basic steps:1. Compute the loss2. Compute the gradient of the loss with respect to all network weights3. Make a weight update step using (stochastic) gradient descent.We need to use a few new TensorFlow functions to do all of this:- For computing the cross-entropy loss we'll use `tf.nn.sparse_softmax_cross_entropy_with_logits`: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/nn/sparse_softmax_cross_entropy_with_logits- For averaging the loss across a minibatch of data we'll use `tf.reduce_mean`:https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/reduce_mean- For computing gradients of the loss with respect to the weights we'll use `tf.GradientTape` (useful for Eager execution): https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/GradientTape- We'll mutate the weight values stored in a TensorFlow Tensor using `tf.assign_sub` ("sub" is for subtraction): https://www.tensorflow.org/api_docs/python/tf/assign_sub ###Code def training_step(model_fn, x, y, params, learning_rate): with tf.GradientTape() as tape: scores = model_fn(x, params) # Forward pass of the model loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y, logits=scores) total_loss = tf.reduce_mean(loss) grad_params = tape.gradient(total_loss, params) # Make a vanilla gradient descent step on all of the model parameters # Manually update the weights using assign_sub() for w, grad_w in zip(params, grad_params): w.assign_sub(learning_rate * grad_w) return total_loss def train_part2(model_fn, init_fn, learning_rate): """ Train a model on CIFAR-10. Inputs: - model_fn: A Python function that performs the forward pass of the model using TensorFlow; it should have the following signature: scores = model_fn(x, params) where x is a TensorFlow Tensor giving a minibatch of image data, params is a list of TensorFlow Tensors holding the model weights, and scores is a TensorFlow Tensor of shape (N, C) giving scores for all elements of x. - init_fn: A Python function that initializes the parameters of the model. It should have the signature params = init_fn() where params is a list of TensorFlow Tensors holding the (randomly initialized) weights of the model. - learning_rate: Python float giving the learning rate to use for SGD. """ params = init_fn() # Initialize the model parameters for t, (x_np, y_np) in enumerate(train_dset): # Run the graph on a batch of training data. loss = training_step(model_fn, x_np, y_np, params, learning_rate) # Periodically print the loss and check accuracy on the val set. if t % print_every == 0: print('Iteration %d, loss = %.4f' % (t, loss)) check_accuracy(val_dset, x_np, model_fn, params) def check_accuracy(dset, x, model_fn, params): """ Check accuracy on a classification model, e.g. for validation. Inputs: - dset: A Dataset object against which to check accuracy - x: A TensorFlow placeholder Tensor where input images should be fed - model_fn: the Model we will be calling to make predictions on x - params: parameters for the model_fn to work with Returns: Nothing, but prints the accuracy of the model """ num_correct, num_samples = 0, 0 for x_batch, y_batch in dset: scores_np = model_fn(x_batch, params).numpy() y_pred = scores_np.argmax(axis=1) num_samples += x_batch.shape[0] num_correct += (y_pred == y_batch).sum() acc = float(num_correct) / num_samples print('Got %d / %d correct (%.2f%%)' % (num_correct, num_samples, 100 * acc)) ###Output _____no_output_____ ###Markdown Barebones TensorFlow: InitializationWe'll use the following utility method to initialize the weight matrices for our models using Kaiming's normalization method.[1] He et al, *Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification*, ICCV 2015, https://arxiv.org/abs/1502.01852 ###Code def create_matrix_with_kaiming_normal(shape): if len(shape) == 2: fan_in, fan_out = shape[0], shape[1] elif len(shape) == 4: fan_in, fan_out = np.prod(shape[:3]), shape[3] return tf.keras.backend.random_normal(shape) * np.sqrt(2.0 / fan_in) ###Output _____no_output_____ ###Markdown Barebones TensorFlow: Train a Two-Layer NetworkWe are finally ready to use all of the pieces defined above to train a two-layer fully-connected network on CIFAR-10.We just need to define a function to initialize the weights of the model, and call `train_part2`.Defining the weights of the network introduces another important piece of TensorFlow API: `tf.Variable`. A TensorFlow Variable is a Tensor whose value is stored in the graph and persists across runs of the computational graph; however unlike constants defined with `tf.zeros` or `tf.random_normal`, the values of a Variable can be mutated as the graph runs; these mutations will persist across graph runs. Learnable parameters of the network are usually stored in Variables.You don't need to tune any hyperparameters, but you should achieve validation accuracies above 40% after one epoch of training. ###Code def two_layer_fc_init(): """ Initialize the weights of a two-layer network, for use with the two_layer_network function defined above. You can use the `create_matrix_with_kaiming_normal` helper! Inputs: None Returns: A list of: - w1: TensorFlow tf.Variable giving the weights for the first layer - w2: TensorFlow tf.Variable giving the weights for the second layer """ hidden_layer_size = 4000 w1 = tf.Variable(create_matrix_with_kaiming_normal((3 * 32 * 32, 4000))) w2 = tf.Variable(create_matrix_with_kaiming_normal((4000, 10))) return [w1, w2] learning_rate = 1e-2 train_part2(two_layer_fc, two_layer_fc_init, learning_rate) ###Output _____no_output_____ ###Markdown Barebones TensorFlow: Train a three-layer ConvNetWe will now use TensorFlow to train a three-layer ConvNet on CIFAR-10.You need to implement the `three_layer_convnet_init` function. Recall that the architecture of the network is:1. Convolutional layer (with bias) with 32 5x5 filters, with zero-padding 22. ReLU3. Convolutional layer (with bias) with 16 3x3 filters, with zero-padding 14. ReLU5. Fully-connected layer (with bias) to compute scores for 10 classesYou don't need to do any hyperparameter tuning, but you should see validation accuracies above 43% after one epoch of training. ###Code def three_layer_convnet_init(): """ Initialize the weights of a Three-Layer ConvNet, for use with the three_layer_convnet function defined above. You can use the `create_matrix_with_kaiming_normal` helper! Inputs: None Returns a list containing: - conv_w1: TensorFlow tf.Variable giving weights for the first conv layer - conv_b1: TensorFlow tf.Variable giving biases for the first conv layer - conv_w2: TensorFlow tf.Variable giving weights for the second conv layer - conv_b2: TensorFlow tf.Variable giving biases for the second conv layer - fc_w: TensorFlow tf.Variable giving weights for the fully-connected layer - fc_b: TensorFlow tf.Variable giving biases for the fully-connected layer """ params = None ############################################################################ # TODO: Initialize the parameters of the three-layer network. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** pass # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return params learning_rate = 3e-3 train_part2(three_layer_convnet, three_layer_convnet_init, learning_rate) ###Output _____no_output_____ ###Markdown Part III: Keras Model Subclassing APIImplementing a neural network using the low-level TensorFlow API is a good way to understand how TensorFlow works, but it's a little inconvenient - we had to manually keep track of all Tensors holding learnable parameters. This was fine for a small network, but could quickly become unweildy for a large complex model.Fortunately TensorFlow 2.0 provides higher-level APIs such as `tf.keras` which make it easy to build models out of modular, object-oriented layers. Further, TensorFlow 2.0 uses eager execution that evaluates operations immediately, without explicitly constructing any computational graphs. This makes it easy to write and debug models, and reduces the boilerplate code.In this part of the notebook we will define neural network models using the `tf.keras.Model` API. To implement your own model, you need to do the following:1. Define a new class which subclasses `tf.keras.Model`. Give your class an intuitive name that describes it, like `TwoLayerFC` or `ThreeLayerConvNet`.2. In the initializer `__init__()` for your new class, define all the layers you need as class attributes. The `tf.keras.layers` package provides many common neural-network layers, like `tf.keras.layers.Dense` for fully-connected layers and `tf.keras.layers.Conv2D` for convolutional layers. Under the hood, these layers will construct `Variable` Tensors for any learnable parameters. **Warning**: Don't forget to call `super(YourModelName, self).__init__()` as the first line in your initializer!3. Implement the `call()` method for your class; this implements the forward pass of your model, and defines the *connectivity* of your network. Layers defined in `__init__()` implement `__call__()` so they can be used as function objects that transform input Tensors into output Tensors. Don't define any new layers in `call()`; any layers you want to use in the forward pass should be defined in `__init__()`.After you define your `tf.keras.Model` subclass, you can instantiate it and use it like the model functions from Part II. Keras Model Subclassing API: Two-Layer NetworkHere is a concrete example of using the `tf.keras.Model` API to define a two-layer network. There are a few new bits of API to be aware of here:We use an `Initializer` object to set up the initial values of the learnable parameters of the layers; in particular `tf.initializers.VarianceScaling` gives behavior similar to the Kaiming initialization method we used in Part II. You can read more about it here: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/initializers/VarianceScalingWe construct `tf.keras.layers.Dense` objects to represent the two fully-connected layers of the model. In addition to multiplying their input by a weight matrix and adding a bias vector, these layer can also apply a nonlinearity for you. For the first layer we specify a ReLU activation function by passing `activation='relu'` to the constructor; the second layer uses softmax activation function. Finally, we use `tf.keras.layers.Flatten` to flatten the output from the previous fully-connected layer. ###Code class TwoLayerFC(tf.keras.Model): def __init__(self, hidden_size, num_classes): super(TwoLayerFC, self).__init__() initializer = tf.initializers.VarianceScaling(scale=2.0) self.fc1 = tf.keras.layers.Dense(hidden_size, activation='relu', kernel_initializer=initializer) self.fc2 = tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer) self.flatten = tf.keras.layers.Flatten() def call(self, x, training=False): x = self.flatten(x) x = self.fc1(x) x = self.fc2(x) return x def test_TwoLayerFC(): """ A small unit test to exercise the TwoLayerFC model above. """ input_size, hidden_size, num_classes = 50, 42, 10 x = tf.zeros((64, input_size)) model = TwoLayerFC(hidden_size, num_classes) with tf.device(device): scores = model(x) print(scores.shape) test_TwoLayerFC() ###Output _____no_output_____ ###Markdown Keras Model Subclassing API: Three-Layer ConvNetNow it's your turn to implement a three-layer ConvNet using the `tf.keras.Model` API. Your model should have the same architecture used in Part II:1. Convolutional layer with 5 x 5 kernels, with zero-padding of 22. ReLU nonlinearity3. Convolutional layer with 3 x 3 kernels, with zero-padding of 14. ReLU nonlinearity5. Fully-connected layer to give class scores6. Softmax nonlinearityYou should initialize the weights of your network using the same initialization method as was used in the two-layer network above.**Hint**: Refer to the documentation for `tf.keras.layers.Conv2D` and `tf.keras.layers.Dense`:https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/Conv2Dhttps://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/Dense ###Code class ThreeLayerConvNet(tf.keras.Model): def __init__(self, channel_1, channel_2, num_classes): super(ThreeLayerConvNet, self).__init__() ######################################################################## # TODO: Implement the __init__ method for a three-layer ConvNet. You # # should instantiate layer objects to be used in the forward pass. # ######################################################################## # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** pass # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ######################################################################## # END OF YOUR CODE # ######################################################################## def call(self, x, training=False): scores = None ######################################################################## # TODO: Implement the forward pass for a three-layer ConvNet. You # # should use the layer objects defined in the __init__ method. # ######################################################################## # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** pass # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ######################################################################## # END OF YOUR CODE # ######################################################################## return scores ###Output _____no_output_____ ###Markdown Once you complete the implementation of the `ThreeLayerConvNet` above you can run the following to ensure that your implementation does not crash and produces outputs of the expected shape. ###Code def test_ThreeLayerConvNet(): channel_1, channel_2, num_classes = 12, 8, 10 model = ThreeLayerConvNet(channel_1, channel_2, num_classes) with tf.device(device): x = tf.zeros((64, 3, 32, 32)) scores = model(x) print(scores.shape) test_ThreeLayerConvNet() ###Output _____no_output_____ ###Markdown Keras Model Subclassing API: Eager TrainingWhile keras models have a builtin training loop (using the `model.fit`), sometimes you need more customization. Here's an example, of a training loop implemented with eager execution.In particular, notice `tf.GradientTape`. Automatic differentiation is used in the backend for implementing backpropagation in frameworks like TensorFlow. During eager execution, `tf.GradientTape` is used to trace operations for computing gradients later. A particular `tf.GradientTape` can only compute one gradient; subsequent calls to tape will throw a runtime error. TensorFlow 2.0 ships with easy-to-use built-in metrics under `tf.keras.metrics` module. Each metric is an object, and we can use `update_state()` to add observations and `reset_state()` to clear all observations. We can get the current result of a metric by calling `result()` on the metric object. ###Code def train_part34(model_init_fn, optimizer_init_fn, num_epochs=1, is_training=False): """ Simple training loop for use with models defined using tf.keras. It trains a model for one epoch on the CIFAR-10 training set and periodically checks accuracy on the CIFAR-10 validation set. Inputs: - model_init_fn: A function that takes no parameters; when called it constructs the model we want to train: model = model_init_fn() - optimizer_init_fn: A function which takes no parameters; when called it constructs the Optimizer object we will use to optimize the model: optimizer = optimizer_init_fn() - num_epochs: The number of epochs to train for Returns: Nothing, but prints progress during trainingn """ with tf.device(device): # Compute the loss like we did in Part II loss_fn = tf.keras.losses.SparseCategoricalCrossentropy() model = model_init_fn() optimizer = optimizer_init_fn() train_loss = tf.keras.metrics.Mean(name='train_loss') train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy') val_loss = tf.keras.metrics.Mean(name='val_loss') val_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='val_accuracy') t = 0 for epoch in range(num_epochs): # Reset the metrics - https://www.tensorflow.org/alpha/guide/migration_guide#new-style_metrics train_loss.reset_states() train_accuracy.reset_states() for x_np, y_np in train_dset: with tf.GradientTape() as tape: # Use the model function to build the forward pass. scores = model(x_np, training=is_training) loss = loss_fn(y_np, scores) gradients = tape.gradient(loss, model.trainable_variables) optimizer.apply_gradients(zip(gradients, model.trainable_variables)) # Update the metrics train_loss.update_state(loss) train_accuracy.update_state(y_np, scores) if t % print_every == 0: val_loss.reset_states() val_accuracy.reset_states() for test_x, test_y in val_dset: # During validation at end of epoch, training set to False prediction = model(test_x, training=False) t_loss = loss_fn(test_y, prediction) val_loss.update_state(t_loss) val_accuracy.update_state(test_y, prediction) template = 'Iteration {}, Epoch {}, Loss: {}, Accuracy: {}, Val Loss: {}, Val Accuracy: {}' print (template.format(t, epoch+1, train_loss.result(), train_accuracy.result()*100, val_loss.result(), val_accuracy.result()*100)) t += 1 ###Output _____no_output_____ ###Markdown Keras Model Subclassing API: Train a Two-Layer NetworkWe can now use the tools defined above to train a two-layer network on CIFAR-10. We define the `model_init_fn` and `optimizer_init_fn` that construct the model and optimizer respectively when called. Here we want to train the model using stochastic gradient descent with no momentum, so we construct a `tf.keras.optimizers.SGD` function; you can [read about it here](https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/optimizers/SGD).You don't need to tune any hyperparameters here, but you should achieve validation accuracies above 40% after one epoch of training. ###Code hidden_size, num_classes = 4000, 10 learning_rate = 1e-2 def model_init_fn(): return TwoLayerFC(hidden_size, num_classes) def optimizer_init_fn(): return tf.keras.optimizers.SGD(learning_rate=learning_rate) train_part34(model_init_fn, optimizer_init_fn) ###Output _____no_output_____ ###Markdown Keras Model Subclassing API: Train a Three-Layer ConvNetHere you should use the tools we've defined above to train a three-layer ConvNet on CIFAR-10. Your ConvNet should use 32 filters in the first convolutional layer and 16 filters in the second layer.To train the model you should use gradient descent with Nesterov momentum 0.9. **HINT**: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/optimizers/SGDYou don't need to perform any hyperparameter tuning, but you should achieve validation accuracies above 50% after training for one epoch. ###Code learning_rate = 3e-3 channel_1, channel_2, num_classes = 32, 16, 10 def model_init_fn(): model = None ############################################################################ # TODO: Complete the implementation of model_fn. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** pass # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return model def optimizer_init_fn(): optimizer = None ############################################################################ # TODO: Complete the implementation of model_fn. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** pass # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return optimizer train_part34(model_init_fn, optimizer_init_fn) ###Output _____no_output_____ ###Markdown Part IV: Keras Sequential APIIn Part III we introduced the `tf.keras.Model` API, which allows you to define models with any number of learnable layers and with arbitrary connectivity between layers.However for many models you don't need such flexibility - a lot of models can be expressed as a sequential stack of layers, with the output of each layer fed to the next layer as input. If your model fits this pattern, then there is an even easier way to define your model: using `tf.keras.Sequential`. You don't need to write any custom classes; you simply call the `tf.keras.Sequential` constructor with a list containing a sequence of layer objects.One complication with `tf.keras.Sequential` is that you must define the shape of the input to the model by passing a value to the `input_shape` of the first layer in your model. Keras Sequential API: Two-Layer NetworkIn this subsection, we will rewrite the two-layer fully-connected network using `tf.keras.Sequential`, and train it using the training loop defined above.You don't need to perform any hyperparameter tuning here, but you should see validation accuracies above 40% after training for one epoch. ###Code learning_rate = 1e-2 def model_init_fn(): input_shape = (32, 32, 3) hidden_layer_size, num_classes = 4000, 10 initializer = tf.initializers.VarianceScaling(scale=2.0) layers = [ tf.keras.layers.Flatten(input_shape=input_shape), tf.keras.layers.Dense(hidden_layer_size, activation='relu', kernel_initializer=initializer), tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer), ] model = tf.keras.Sequential(layers) return model def optimizer_init_fn(): return tf.keras.optimizers.SGD(learning_rate=learning_rate) train_part34(model_init_fn, optimizer_init_fn) ###Output _____no_output_____ ###Markdown Abstracting Away the Training LoopIn the previous examples, we used a customised training loop to train models (e.g. `train_part34`). Writing your own training loop is only required if you need more flexibility and control during training your model. Alternately, you can also use built-in APIs like `tf.keras.Model.fit()` and `tf.keras.Model.evaluate` to train and evaluate a model. Also remember to configure your model for training by calling `tf.keras.Model.compile.You don't need to perform any hyperparameter tuning here, but you should see validation and test accuracies above 42% after training for one epoch. ###Code model = model_init_fn() model.compile(optimizer=tf.keras.optimizers.SGD(learning_rate=learning_rate), loss='sparse_categorical_crossentropy', metrics=[tf.keras.metrics.sparse_categorical_accuracy]) model.fit(X_train, y_train, batch_size=64, epochs=1, validation_data=(X_val, y_val)) model.evaluate(X_test, y_test) ###Output _____no_output_____ ###Markdown Keras Sequential API: Three-Layer ConvNetHere you should use `tf.keras.Sequential` to reimplement the same three-layer ConvNet architecture used in Part II and Part III. As a reminder, your model should have the following architecture:1. Convolutional layer with 32 5x5 kernels, using zero padding of 22. ReLU nonlinearity3. Convolutional layer with 16 3x3 kernels, using zero padding of 14. ReLU nonlinearity5. Fully-connected layer giving class scores6. Softmax nonlinearityYou should initialize the weights of the model using a `tf.initializers.VarianceScaling` as above.You should train the model using Nesterov momentum 0.9.You don't need to perform any hyperparameter search, but you should achieve accuracy above 45% after training for one epoch. ###Code def model_init_fn(): model = None ############################################################################ # TODO: Construct a three-layer ConvNet using tf.keras.Sequential. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** pass # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return model learning_rate = 5e-4 def optimizer_init_fn(): optimizer = None ############################################################################ # TODO: Complete the implementation of model_fn. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** pass # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return optimizer train_part34(model_init_fn, optimizer_init_fn) ###Output _____no_output_____ ###Markdown We will also train this model with the built-in training loop APIs provided by TensorFlow. ###Code model = model_init_fn() model.compile(optimizer='sgd', loss='sparse_categorical_crossentropy', metrics=[tf.keras.metrics.sparse_categorical_accuracy]) model.fit(X_train, y_train, batch_size=64, epochs=1, validation_data=(X_val, y_val)) model.evaluate(X_test, y_test) ###Output _____no_output_____ ###Markdown Part IV: Functional API Demonstration with a Two-Layer Network In the previous section, we saw how we can use `tf.keras.Sequential` to stack layers to quickly build simple models. But this comes at the cost of losing flexibility.Often we will have to write complex models that have non-sequential data flows: a layer can have **multiple inputs and/or outputs**, such as stacking the output of 2 previous layers together to feed as input to a third! (Some examples are residual connections and dense blocks.)In such cases, we can use Keras functional API to write models with complex topologies such as: 1. Multi-input models 2. Multi-output models 3. Models with shared layers (the same layer called several times) 4. Models with non-sequential data flows (e.g. residual connections)Writing a model with Functional API requires us to create a `tf.keras.Model` instance and explicitly write input tensors and output tensors for this model. ###Code def two_layer_fc_functional(input_shape, hidden_size, num_classes): initializer = tf.initializers.VarianceScaling(scale=2.0) inputs = tf.keras.Input(shape=input_shape) flattened_inputs = tf.keras.layers.Flatten()(inputs) fc1_output = tf.keras.layers.Dense(hidden_size, activation='relu', kernel_initializer=initializer)(flattened_inputs) scores = tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer)(fc1_output) # Instantiate the model given inputs and outputs. model = tf.keras.Model(inputs=inputs, outputs=scores) return model def test_two_layer_fc_functional(): """ A small unit test to exercise the TwoLayerFC model above. """ input_size, hidden_size, num_classes = 50, 42, 10 input_shape = (50,) x = tf.zeros((64, input_size)) model = two_layer_fc_functional(input_shape, hidden_size, num_classes) with tf.device(device): scores = model(x) print(scores.shape) test_two_layer_fc_functional() ###Output _____no_output_____ ###Markdown Keras Functional API: Train a Two-Layer NetworkYou can now train this two-layer network constructed using the functional API.You don't need to perform any hyperparameter tuning here, but you should see validation accuracies above 40% after training for one epoch. ###Code input_shape = (32, 32, 3) hidden_size, num_classes = 4000, 10 learning_rate = 1e-2 def model_init_fn(): return two_layer_fc_functional(input_shape, hidden_size, num_classes) def optimizer_init_fn(): return tf.keras.optimizers.SGD(learning_rate=learning_rate) train_part34(model_init_fn, optimizer_init_fn) ###Output _____no_output_____ ###Markdown Part V: CIFAR-10 open-ended challengeIn this section you can experiment with whatever ConvNet architecture you'd like on CIFAR-10.You should experiment with architectures, hyperparameters, loss functions, regularization, or anything else you can think of to train a model that achieves **at least 70%** accuracy on the **validation** set within 10 epochs. You can use the built-in train function, the `train_part34` function from above, or implement your own training loop.Describe what you did at the end of the notebook. Some things you can try:- **Filter size**: Above we used 5x5 and 3x3; is this optimal?- **Number of filters**: Above we used 16 and 32 filters. Would more or fewer do better?- **Pooling**: We didn't use any pooling above. Would this improve the model?- **Normalization**: Would your model be improved with batch normalization, layer normalization, group normalization, or some other normalization strategy?- **Network architecture**: The ConvNet above has only three layers of trainable parameters. Would a deeper model do better?- **Global average pooling**: Instead of flattening after the final convolutional layer, would global average pooling do better? This strategy is used for example in Google's Inception network and in Residual Networks.- **Regularization**: Would some kind of regularization improve performance? Maybe weight decay or dropout? NOTE: Batch Normalization / DropoutIf you are using Batch Normalization and Dropout, remember to pass `is_training=True` if you use the `train_part34()` function. BatchNorm and Dropout layers have different behaviors at training and inference time. `training` is a specific keyword argument reserved for this purpose in any `tf.keras.Model`'s `call()` function. Read more about this here : https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/BatchNormalizationmethodshttps://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/Dropoutmethods Tips for trainingFor each network architecture that you try, you should tune the learning rate and other hyperparameters. When doing this there are a couple important things to keep in mind: - If the parameters are working well, you should see improvement within a few hundred iterations- Remember the coarse-to-fine approach for hyperparameter tuning: start by testing a large range of hyperparameters for just a few training iterations to find the combinations of parameters that are working at all.- Once you have found some sets of parameters that seem to work, search more finely around these parameters. You may need to train for more epochs.- You should use the validation set for hyperparameter search, and save your test set for evaluating your architecture on the best parameters as selected by the validation set. Going above and beyondIf you are feeling adventurous there are many other features you can implement to try and improve your performance. You are **not required** to implement any of these, but don't miss the fun if you have time!- Alternative optimizers: you can try Adam, Adagrad, RMSprop, etc.- Alternative activation functions such as leaky ReLU, parametric ReLU, ELU, or MaxOut.- Model ensembles- Data augmentation- New Architectures - [ResNets](https://arxiv.org/abs/1512.03385) where the input from the previous layer is added to the output. - [DenseNets](https://arxiv.org/abs/1608.06993) where inputs into previous layers are concatenated together. - [This blog has an in-depth overview](https://chatbotslife.com/resnets-highwaynets-and-densenets-oh-my-9bb15918ee32) Have fun and happy training! ###Code class CustomConvNet(tf.keras.Model): def __init__(self): super(CustomConvNet, self).__init__() ############################################################################ # TODO: Construct a model that performs well on CIFAR-10 # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** pass # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ def call(self, input_tensor, training=False): ############################################################################ # TODO: Construct a model that performs well on CIFAR-10 # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** pass # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return x device = '/device:GPU:0' # Change this to a CPU/GPU as you wish! # device = '/cpu:0' # Change this to a CPU/GPU as you wish! print_every = 700 num_epochs = 10 model = CustomConvNet() def model_init_fn(): return CustomConvNet() def optimizer_init_fn(): learning_rate = 1e-3 return tf.keras.optimizers.Adam(learning_rate) train_part34(model_init_fn, optimizer_init_fn, num_epochs=num_epochs, is_training=True) ###Output _____no_output_____ ###Markdown What's this TensorFlow business?You've written a lot of code in this assignment to provide a whole host of neural network functionality. Dropout, Batch Norm, and 2D convolutions are some of the workhorses of deep learning in computer vision. You've also worked hard to make your code efficient and vectorized.For the last part of this assignment, though, we're going to leave behind your beautiful codebase and instead migrate to one of two popular deep learning frameworks: in this instance, TensorFlow (or PyTorch, if you choose to work with that notebook). What is it?TensorFlow is a system for executing computational graphs over Tensor objects, with native support for performing backpropogation for its Variables. In it, we work with Tensors which are n-dimensional arrays analogous to the numpy ndarray. Why?* Our code will now run on GPUs! Much faster training. Writing your own modules to run on GPUs is beyond the scope of this class, unfortunately.* We want you to be ready to use one of these frameworks for your project so you can experiment more efficiently than if you were writing every feature you want to use by hand. * We want you to stand on the shoulders of giants! TensorFlow and PyTorch are both excellent frameworks that will make your lives a lot easier, and now that you understand their guts, you are free to use them :) * We want you to be exposed to the sort of deep learning code you might run into in academia or industry. How will I learn TensorFlow?TensorFlow has many excellent tutorials available, including those from [Google themselves](https://www.tensorflow.org/get_started/get_started).Otherwise, this notebook will walk you through much of what you need to do to train models in TensorFlow. See the end of the notebook for some links to helpful tutorials if you want to learn more or need further clarification on topics that aren't fully explained here.**NOTE: This notebook is meant to teach you the latest version of Tensorflow 2.0. Most examples on the web today are still in 1.x, so be careful not to confuse the two when looking up documentation**. Install Tensorflow 2.0Tensorflow 2.0 is still not in a fully 100% stable release, but it's still usable and more intuitive than TF 1.x. Please make sure you have it installed before moving on in this notebook! Here are some steps to get started:1. Have the latest version of Anaconda installed on your machine.2. Create a new conda environment starting from Python 3.7. In this setup example, we'll call it `tf_20_env`.3. Run the command: `source activate tf_20_env`4. Then pip install TF 2.0 as described here: https://www.tensorflow.org/install/pip A guide on creating Anaconda enviornments: https://uoa-eresearch.github.io/eresearch-cookbook/recipe/2014/11/20/conda/This will give you an new enviornemnt to play in TF 2.0. Generally, if you plan to also use TensorFlow in your other projects, you might also want to keep a seperate Conda environment or virtualenv in Python 3.7 that has Tensorflow 1.9, so you can switch back and forth at will. Table of ContentsThis notebook has 5 parts. We will walk through TensorFlow at **three different levels of abstraction**, which should help you better understand it and prepare you for working on your project.1. Part I, Preparation: load the CIFAR-10 dataset.2. Part II, Barebone TensorFlow: **Abstraction Level 1**, we will work directly with low-level TensorFlow graphs. 3. Part III, Keras Model API: **Abstraction Level 2**, we will use `tf.keras.Model` to define arbitrary neural network architecture. 4. Part IV, Keras Sequential + Functional API: **Abstraction Level 3**, we will use `tf.keras.Sequential` to define a linear feed-forward network very conveniently, and then explore the functional libraries for building unique and uncommon models that require more flexibility.5. Part V, CIFAR-10 open-ended challenge: please implement your own network to get as high accuracy as possible on CIFAR-10. You can experiment with any layer, optimizer, hyperparameters or other advanced features. We will discuss Keras in more detail later in the notebook.Here is a table of comparison:| API | Flexibility | Convenience ||---------------|-------------|-------------|| Barebone | High | Low || `tf.keras.Model` | High | Medium || `tf.keras.Sequential` | Low | High | Part I: PreparationFirst, we load the CIFAR-10 dataset. This might take a few minutes to download the first time you run it, but after that the files should be cached on disk and loading should be faster.In previous parts of the assignment we used CS231N-specific code to download and read the CIFAR-10 dataset; however the `tf.keras.datasets` package in TensorFlow provides prebuilt utility functions for loading many common datasets.For the purposes of this assignment we will still write our own code to preprocess the data and iterate through it in minibatches. The `tf.data` package in TensorFlow provides tools for automating this process, but working with this package adds extra complication and is beyond the scope of this notebook. However using `tf.data` can be much more efficient than the simple approach used in this notebook, so you should consider using it for your project. ###Code import os import tensorflow as tf import numpy as np import math import timeit import matplotlib.pyplot as plt %matplotlib inline print(tf.__version__) # need tf 2.0 def load_cifar10(num_training=49000, num_validation=1000, num_test=10000): """ Fetch the CIFAR-10 dataset from the web and perform preprocessing to prepare it for the two-layer neural net classifier. These are the same steps as we used for the SVM, but condensed to a single function. """ # Load the raw CIFAR-10 dataset and use appropriate data types and shapes cifar10 = tf.keras.datasets.cifar10.load_data() (X_train, y_train), (X_test, y_test) = cifar10 X_train = np.asarray(X_train, dtype=np.float32) y_train = np.asarray(y_train, dtype=np.int32).flatten() X_test = np.asarray(X_test, dtype=np.float32) y_test = np.asarray(y_test, dtype=np.int32).flatten() # Subsample the data mask = range(num_training, num_training + num_validation) X_val = X_train[mask] y_val = y_train[mask] mask = range(num_training) X_train = X_train[mask] y_train = y_train[mask] mask = range(num_test) X_test = X_test[mask] y_test = y_test[mask] # Normalize the data: subtract the mean pixel and divide by std mean_pixel = X_train.mean(axis=(0, 1, 2), keepdims=True) std_pixel = X_train.std(axis=(0, 1, 2), keepdims=True) X_train = (X_train - mean_pixel) / std_pixel X_val = (X_val - mean_pixel) / std_pixel X_test = (X_test - mean_pixel) / std_pixel return X_train, y_train, X_val, y_val, X_test, y_test # If there are errors with SSL downloading involving self-signed certificates, # it may be that your Python version was recently installed on the current machine. # See: https://github.com/tensorflow/tensorflow/issues/10779 # To fix, run the command: /Applications/Python\ 3.7/Install\ Certificates.command # ...replacing paths as necessary. # Invoke the above function to get our data. NHW = (0, 1, 2) X_train, y_train, X_val, y_val, X_test, y_test = load_cifar10() print('Train data shape: ', X_train.shape) print('Train labels shape: ', y_train.shape, y_train.dtype) print('Validation data shape: ', X_val.shape) print('Validation labels shape: ', y_val.shape) print('Test data shape: ', X_test.shape) print('Test labels shape: ', y_test.shape) class Dataset(object): def __init__(self, X, y, batch_size, shuffle=False): """ Construct a Dataset object to iterate over data X and labels y Inputs: - X: Numpy array of data, of any shape - y: Numpy array of labels, of any shape but with y.shape[0] == X.shape[0] - batch_size: Integer giving number of elements per minibatch - shuffle: (optional) Boolean, whether to shuffle the data on each epoch """ assert X.shape[0] == y.shape[0], 'Got different numbers of data and labels' self.X, self.y = X, y self.batch_size, self.shuffle = batch_size, shuffle def __iter__(self): N, B = self.X.shape[0], self.batch_size idxs = np.arange(N) if self.shuffle: np.random.shuffle(idxs) return iter((self.X[i:i+B], self.y[i:i+B]) for i in range(0, N, B)) train_dset = Dataset(X_train, y_train, batch_size=64, shuffle=True) val_dset = Dataset(X_val, y_val, batch_size=64, shuffle=False) test_dset = Dataset(X_test, y_test, batch_size=64) # We can iterate through a dataset like this: for t, (x, y) in enumerate(train_dset): print(t, x.shape, y.shape) if t > 5: break ###Output 0 (64, 32, 32, 3) (64,) 1 (64, 32, 32, 3) (64,) 2 (64, 32, 32, 3) (64,) 3 (64, 32, 32, 3) (64,) 4 (64, 32, 32, 3) (64,) 5 (64, 32, 32, 3) (64,) 6 (64, 32, 32, 3) (64,) ###Markdown You can optionally **use GPU by setting the flag to True below**. It's not neccessary to use a GPU for this assignment; if you are working on Google Cloud then we recommend that you do not use a GPU, as it will be significantly more expensive. ###Code # Set up some global variables USE_GPU = True if USE_GPU: device = '/device:GPU:0' else: device = '/cpu:0' # Constant to control how often we print when training models print_every = 100 print('Using device: ', device) ###Output Using device: /device:GPU:0 ###Markdown Part II: Barebones TensorFlowTensorFlow ships with various high-level APIs which make it very convenient to define and train neural networks; we will cover some of these constructs in Part III and Part IV of this notebook. In this section we will start by building a model with basic TensorFlow constructs to help you better understand what's going on under the hood of the higher-level APIs.**"Barebones Tensorflow" is important to understanding the building blocks of TensorFlow, but much of it involves concepts from TensorFlow 1.x.** We will be working with legacy modules such as `tf.Variable`.Therefore, please read and understand the differences between legacy (1.x) TF and the new (2.0) TF. Historical background on TensorFlow 1.xTensorFlow 1.x is primarily a framework for working with **static computational graphs**. Nodes in the computational graph are Tensors which will hold n-dimensional arrays when the graph is run; edges in the graph represent functions that will operate on Tensors when the graph is run to actually perform useful computation.Before Tensorflow 2.0, we had to configure the graph into two phases. There are plenty of tutorials online that explain this two-step process. The process generally looks like the following for TF 1.x:1. **Build a computational graph that describes the computation that you want to perform**. This stage doesn't actually perform any computation; it just builds up a symbolic representation of your computation. This stage will typically define one or more `placeholder` objects that represent inputs to the computational graph.2. **Run the computational graph many times.** Each time the graph is run (e.g. for one gradient descent step) you will specify which parts of the graph you want to compute, and pass a `feed_dict` dictionary that will give concrete values to any `placeholder`s in the graph. The new paradigm in Tensorflow 2.0Now, with Tensorflow 2.0, we can simply adopt a functional form that is more Pythonic and similar in spirit to PyTorch and direct Numpy operation. Instead of the 2-step paradigm with computation graphs, making it (among other things) easier to debug TF code. You can read more details at https://www.tensorflow.org/guide/eager.The main difference between the TF 1.x and 2.0 approach is that the 2.0 approach doesn't make use of `tf.Session`, `tf.run`, `placeholder`, `feed_dict`. To get more details of what's different between the two version and how to convert between the two, check out the official migration guide: https://www.tensorflow.org/alpha/guide/migration_guideLater, in the rest of this notebook we'll focus on this new, simpler approach. TensorFlow warmup: Flatten FunctionWe can see this in action by defining a simple `flatten` function that will reshape image data for use in a fully-connected network.In TensorFlow, data for convolutional feature maps is typically stored in a Tensor of shape N x H x W x C where:- N is the number of datapoints (minibatch size)- H is the height of the feature map- W is the width of the feature map- C is the number of channels in the feature mapThis is the right way to represent the data when we are doing something like a 2D convolution, that needs spatial understanding of where the intermediate features are relative to each other. When we use fully connected affine layers to process the image, however, we want each datapoint to be represented by a single vector -- it's no longer useful to segregate the different channels, rows, and columns of the data. So, we use a "flatten" operation to collapse the `H x W x C` values per representation into a single long vector. Notice the `tf.reshape` call has the target shape as `(N, -1)`, meaning it will reshape/keep the first dimension to be N, and then infer as necessary what the second dimension is in the output, so we can collapse the remaining dimensions from the input properly.**NOTE**: TensorFlow and PyTorch differ on the default Tensor layout; TensorFlow uses N x H x W x C but PyTorch uses N x C x H x W. ###Code def flatten(x): """ Input: - TensorFlow Tensor of shape (N, D1, ..., DM) Output: - TensorFlow Tensor of shape (N, D1 * ... * DM) """ N = tf.shape(x)[0] return tf.reshape(x, (N, -1)) def test_flatten(): # Construct concrete values of the input data x using numpy x_np = np.arange(24).reshape((2, 3, 4)) print('x_np:\n', x_np, '\n') # Compute a concrete output value. x_flat_np = flatten(x_np) print('x_flat_np:\n', x_flat_np, '\n') test_flatten() ###Output x_np: [[[ 0 1 2 3] [ 4 5 6 7] [ 8 9 10 11]] [[12 13 14 15] [16 17 18 19] [20 21 22 23]]] x_flat_np: tf.Tensor( [[ 0 1 2 3 4 5 6 7 8 9 10 11] [12 13 14 15 16 17 18 19 20 21 22 23]], shape=(2, 12), dtype=int32) ###Markdown Barebones TensorFlow: Define a Two-Layer NetworkWe will now implement our first neural network with TensorFlow: a fully-connected ReLU network with two hidden layers and no biases on the CIFAR10 dataset. For now we will use only low-level TensorFlow operators to define the network; later we will see how to use the higher-level abstractions provided by `tf.keras` to simplify the process.We will define the forward pass of the network in the function `two_layer_fc`; this will accept TensorFlow Tensors for the inputs and weights of the network, and return a TensorFlow Tensor for the scores. After defining the network architecture in the `two_layer_fc` function, we will test the implementation by checking the shape of the output.**It's important that you read and understand this implementation.** ###Code def two_layer_fc(x, params): """ A fully-connected neural network; the architecture is: fully-connected layer -> ReLU -> fully connected layer. Note that we only need to define the forward pass here; TensorFlow will take care of computing the gradients for us. The input to the network will be a minibatch of data, of shape (N, d1, ..., dM) where d1 * ... * dM = D. The hidden layer will have H units, and the output layer will produce scores for C classes. Inputs: - x: A TensorFlow Tensor of shape (N, d1, ..., dM) giving a minibatch of input data. - params: A list [w1, w2] of TensorFlow Tensors giving weights for the network, where w1 has shape (D, H) and w2 has shape (H, C). Returns: - scores: A TensorFlow Tensor of shape (N, C) giving classification scores for the input data x. """ w1, w2 = params # Unpack the parameters x = flatten(x) # Flatten the input; now x has shape (N, D) h = tf.nn.relu(tf.matmul(x, w1)) # Hidden layer: h has shape (N, H) scores = tf.matmul(h, w2) # Compute scores of shape (N, C) return scores def two_layer_fc_test(): hidden_layer_size = 42 # Scoping our TF operations under a tf.device context manager # lets us tell TensorFlow where we want these Tensors to be # multiplied and/or operated on, e.g. on a CPU or a GPU. with tf.device(device): x = tf.zeros((64, 32, 32, 3)) w1 = tf.zeros((32 * 32 * 3, hidden_layer_size)) w2 = tf.zeros((hidden_layer_size, 10)) # Call our two_layer_fc function for the forward pass of the network. scores = two_layer_fc(x, [w1, w2]) print(scores.shape) two_layer_fc_test() ###Output (64, 10) ###Markdown Barebones TensorFlow: Three-Layer ConvNetHere you will complete the implementation of the function `three_layer_convnet` which will perform the forward pass of a three-layer convolutional network. The network should have the following architecture:1. A convolutional layer (with bias) with `channel_1` filters, each with shape `KW1 x KH1`, and zero-padding of two2. ReLU nonlinearity3. A convolutional layer (with bias) with `channel_2` filters, each with shape `KW2 x KH2`, and zero-padding of one4. ReLU nonlinearity5. Fully-connected layer with bias, producing scores for `C` classes.**HINT**: For convolutions: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/nn/conv2d; be careful with padding!**HINT**: For biases: https://www.tensorflow.org/performance/xla/broadcasting ###Code def three_layer_convnet(x, params): """ A three-layer convolutional network with the architecture described above. Inputs: - x: A TensorFlow Tensor of shape (N, H, W, 3) giving a minibatch of images - params: A list of TensorFlow Tensors giving the weights and biases for the network; should contain the following: - conv_w1: TensorFlow Tensor of shape (KH1, KW1, 3, channel_1) giving weights for the first convolutional layer. - conv_b1: TensorFlow Tensor of shape (channel_1,) giving biases for the first convolutional layer. - conv_w2: TensorFlow Tensor of shape (KH2, KW2, channel_1, channel_2) giving weights for the second convolutional layer - conv_b2: TensorFlow Tensor of shape (channel_2,) giving biases for the second convolutional layer. - fc_w: TensorFlow Tensor giving weights for the fully-connected layer. Can you figure out what the shape should be? - fc_b: TensorFlow Tensor giving biases for the fully-connected layer. Can you figure out what the shape should be? """ conv_w1, conv_b1, conv_w2, conv_b2, fc_w, fc_b = params scores = None ############################################################################ # TODO: Implement the forward pass for the three-layer ConvNet. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** conv1 = tf.nn.conv2d(x, conv_w1, 1, [[0, 0], [2, 2], [2, 2], [0, 0]]) + conv_b1 relu1 = tf.nn.relu(conv1) conv2 = tf.nn.conv2d(relu1, conv_w2, 1, [[0, 0], [1, 1], [1, 1], [0, 0]]) + conv_b2 relu2 = tf.nn.relu(conv2) relu2_flat = flatten(relu2) scores = tf.matmul(relu2_flat, fc_w) + fc_b # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return scores ###Output _____no_output_____ ###Markdown After defing the forward pass of the three-layer ConvNet above, run the following cell to test your implementation. Like the two-layer network, we run the graph on a batch of zeros just to make sure the function doesn't crash, and produces outputs of the correct shape.When you run this function, `scores_np` should have shape `(64, 10)`. ###Code def three_layer_convnet_test(): with tf.device(device): x = tf.zeros((64, 32, 32, 3)) conv_w1 = tf.zeros((5, 5, 3, 6)) conv_b1 = tf.zeros((6,)) conv_w2 = tf.zeros((3, 3, 6, 9)) conv_b2 = tf.zeros((9,)) fc_w = tf.zeros((32 * 32 * 9, 10)) fc_b = tf.zeros((10,)) params = [conv_w1, conv_b1, conv_w2, conv_b2, fc_w, fc_b] scores = three_layer_convnet(x, params) # Inputs to convolutional layers are 4-dimensional arrays with shape # [batch_size, height, width, channels] print('scores_np has shape: ', scores.shape) three_layer_convnet_test() ###Output scores_np has shape: (64, 10) ###Markdown Barebones TensorFlow: Training StepWe now define the `training_step` function performs a single training step. This will take three basic steps:1. Compute the loss2. Compute the gradient of the loss with respect to all network weights3. Make a weight update step using (stochastic) gradient descent.We need to use a few new TensorFlow functions to do all of this:- For computing the cross-entropy loss we'll use `tf.nn.sparse_softmax_cross_entropy_with_logits`: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/nn/sparse_softmax_cross_entropy_with_logits- For averaging the loss across a minibatch of data we'll use `tf.reduce_mean`:https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/reduce_mean- For computing gradients of the loss with respect to the weights we'll use `tf.GradientTape` (useful for Eager execution): https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/GradientTape- We'll mutate the weight values stored in a TensorFlow Tensor using `tf.assign_sub` ("sub" is for subtraction): https://www.tensorflow.org/api_docs/python/tf/assign_sub ###Code def training_step(model_fn, x, y, params, learning_rate): with tf.GradientTape() as tape: scores = model_fn(x, params) # Forward pass of the model loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y, logits=scores) total_loss = tf.reduce_mean(loss) grad_params = tape.gradient(total_loss, params) # Make a vanilla gradient descent step on all of the model parameters # Manually update the weights using assign_sub() for w, grad_w in zip(params, grad_params): w.assign_sub(learning_rate * grad_w) return total_loss def train_part2(model_fn, init_fn, learning_rate): """ Train a model on CIFAR-10. Inputs: - model_fn: A Python function that performs the forward pass of the model using TensorFlow; it should have the following signature: scores = model_fn(x, params) where x is a TensorFlow Tensor giving a minibatch of image data, params is a list of TensorFlow Tensors holding the model weights, and scores is a TensorFlow Tensor of shape (N, C) giving scores for all elements of x. - init_fn: A Python function that initializes the parameters of the model. It should have the signature params = init_fn() where params is a list of TensorFlow Tensors holding the (randomly initialized) weights of the model. - learning_rate: Python float giving the learning rate to use for SGD. """ params = init_fn() # Initialize the model parameters for t, (x_np, y_np) in enumerate(train_dset): # Run the graph on a batch of training data. loss = training_step(model_fn, x_np, y_np, params, learning_rate) # Periodically print the loss and check accuracy on the val set. if t % print_every == 0: print('Iteration %d, loss = %.4f' % (t, loss)) check_accuracy(val_dset, x_np, model_fn, params) def check_accuracy(dset, x, model_fn, params): """ Check accuracy on a classification model, e.g. for validation. Inputs: - dset: A Dataset object against which to check accuracy - x: A TensorFlow placeholder Tensor where input images should be fed - model_fn: the Model we will be calling to make predictions on x - params: parameters for the model_fn to work with Returns: Nothing, but prints the accuracy of the model """ num_correct, num_samples = 0, 0 for x_batch, y_batch in dset: scores_np = model_fn(x_batch, params).numpy() y_pred = scores_np.argmax(axis=1) num_samples += x_batch.shape[0] num_correct += (y_pred == y_batch).sum() acc = float(num_correct) / num_samples print('Got %d / %d correct (%.2f%%)' % (num_correct, num_samples, 100 * acc)) ###Output _____no_output_____ ###Markdown Barebones TensorFlow: InitializationWe'll use the following utility method to initialize the weight matrices for our models using Kaiming's normalization method.[1] He et al, *Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification*, ICCV 2015, https://arxiv.org/abs/1502.01852 ###Code def create_matrix_with_kaiming_normal(shape): if len(shape) == 2: fan_in, fan_out = shape[0], shape[1] elif len(shape) == 4: fan_in, fan_out = np.prod(shape[:3]), shape[3] return tf.keras.backend.random_normal(shape) * np.sqrt(2.0 / fan_in) ###Output _____no_output_____ ###Markdown Barebones TensorFlow: Train a Two-Layer NetworkWe are finally ready to use all of the pieces defined above to train a two-layer fully-connected network on CIFAR-10.We just need to define a function to initialize the weights of the model, and call `train_part2`.Defining the weights of the network introduces another important piece of TensorFlow API: `tf.Variable`. A TensorFlow Variable is a Tensor whose value is stored in the graph and persists across runs of the computational graph; however unlike constants defined with `tf.zeros` or `tf.random_normal`, the values of a Variable can be mutated as the graph runs; these mutations will persist across graph runs. Learnable parameters of the network are usually stored in Variables.You don't need to tune any hyperparameters, but you should achieve validation accuracies above 40% after one epoch of training. ###Code def two_layer_fc_init(): """ Initialize the weights of a two-layer network, for use with the two_layer_network function defined above. You can use the `create_matrix_with_kaiming_normal` helper! Inputs: None Returns: A list of: - w1: TensorFlow tf.Variable giving the weights for the first layer - w2: TensorFlow tf.Variable giving the weights for the second layer """ hidden_layer_size = 4000 w1 = tf.Variable(create_matrix_with_kaiming_normal((3 * 32 * 32, 4000))) w2 = tf.Variable(create_matrix_with_kaiming_normal((4000, 10))) return [w1, w2] learning_rate = 1e-2 train_part2(two_layer_fc, two_layer_fc_init, learning_rate) ###Output Iteration 0, loss = 3.1796 Got 138 / 1000 correct (13.80%) Iteration 100, loss = 1.8606 Got 371 / 1000 correct (37.10%) Iteration 200, loss = 1.4913 Got 389 / 1000 correct (38.90%) Iteration 300, loss = 1.8782 Got 363 / 1000 correct (36.30%) Iteration 400, loss = 1.8509 Got 426 / 1000 correct (42.60%) Iteration 500, loss = 1.8515 Got 419 / 1000 correct (41.90%) Iteration 600, loss = 1.9118 Got 428 / 1000 correct (42.80%) Iteration 700, loss = 1.8966 Got 450 / 1000 correct (45.00%) ###Markdown Barebones TensorFlow: Train a three-layer ConvNetWe will now use TensorFlow to train a three-layer ConvNet on CIFAR-10.You need to implement the `three_layer_convnet_init` function. Recall that the architecture of the network is:1. Convolutional layer (with bias) with 32 5x5 filters, with zero-padding 22. ReLU3. Convolutional layer (with bias) with 16 3x3 filters, with zero-padding 14. ReLU5. Fully-connected layer (with bias) to compute scores for 10 classesYou don't need to do any hyperparameter tuning, but you should see validation accuracies above 43% after one epoch of training. ###Code def three_layer_convnet_init(): """ Initialize the weights of a Three-Layer ConvNet, for use with the three_layer_convnet function defined above. You can use the `create_matrix_with_kaiming_normal` helper! Inputs: None Returns a list containing: - conv_w1: TensorFlow tf.Variable giving weights for the first conv layer - conv_b1: TensorFlow tf.Variable giving biases for the first conv layer - conv_w2: TensorFlow tf.Variable giving weights for the second conv layer - conv_b2: TensorFlow tf.Variable giving biases for the second conv layer - fc_w: TensorFlow tf.Variable giving weights for the fully-connected layer - fc_b: TensorFlow tf.Variable giving biases for the fully-connected layer """ params = None ############################################################################ # TODO: Initialize the parameters of the three-layer network. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** # a sample input is 32 x 32 x 3 conv_w1 = tf.Variable(create_matrix_with_kaiming_normal((5, 5, 3, 32))) conv_b1 = tf.Variable(create_matrix_with_kaiming_normal((1, 32))) conv_w2 = tf.Variable(create_matrix_with_kaiming_normal((3, 3, 32, 16))) conv_b2 = tf.Variable(create_matrix_with_kaiming_normal((1, 16))) fc_w = tf.Variable(create_matrix_with_kaiming_normal((32 * 32 * 16, 10))) # the input size after two convs is 32 x 32 x 16. fc_b = tf.Variable(create_matrix_with_kaiming_normal((1, 10))) params = [conv_w1, conv_b1, conv_w2, conv_b2, fc_w, fc_b] # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return params learning_rate = 3e-3 train_part2(three_layer_convnet, three_layer_convnet_init, learning_rate) ###Output Iteration 0, loss = 4.8878 Got 79 / 1000 correct (7.90%) Iteration 100, loss = 1.9799 Got 348 / 1000 correct (34.80%) Iteration 200, loss = 1.7410 Got 391 / 1000 correct (39.10%) Iteration 300, loss = 1.7193 Got 390 / 1000 correct (39.00%) Iteration 400, loss = 1.7154 Got 427 / 1000 correct (42.70%) Iteration 500, loss = 1.6703 Got 438 / 1000 correct (43.80%) Iteration 600, loss = 1.6941 Got 446 / 1000 correct (44.60%) Iteration 700, loss = 1.6794 Got 460 / 1000 correct (46.00%) ###Markdown Part III: Keras Model Subclassing APIImplementing a neural network using the low-level TensorFlow API is a good way to understand how TensorFlow works, but it's a little inconvenient - we had to manually keep track of all Tensors holding learnable parameters. This was fine for a small network, but could quickly become unweildy for a large complex model.Fortunately TensorFlow 2.0 provides higher-level APIs such as `tf.keras` which make it easy to build models out of modular, object-oriented layers. Further, TensorFlow 2.0 uses eager execution that evaluates operations immediately, without explicitly constructing any computational graphs. This makes it easy to write and debug models, and reduces the boilerplate code.In this part of the notebook we will define neural network models using the `tf.keras.Model` API. To implement your own model, you need to do the following:1. Define a new class which subclasses `tf.keras.Model`. Give your class an intuitive name that describes it, like `TwoLayerFC` or `ThreeLayerConvNet`.2. In the initializer `__init__()` for your new class, define all the layers you need as class attributes. The `tf.keras.layers` package provides many common neural-network layers, like `tf.keras.layers.Dense` for fully-connected layers and `tf.keras.layers.Conv2D` for convolutional layers. Under the hood, these layers will construct `Variable` Tensors for any learnable parameters. **Warning**: Don't forget to call `super(YourModelName, self).__init__()` as the first line in your initializer!3. Implement the `call()` method for your class; this implements the forward pass of your model, and defines the *connectivity* of your network. Layers defined in `__init__()` implement `__call__()` so they can be used as function objects that transform input Tensors into output Tensors. Don't define any new layers in `call()`; any layers you want to use in the forward pass should be defined in `__init__()`.After you define your `tf.keras.Model` subclass, you can instantiate it and use it like the model functions from Part II. Keras Model Subclassing API: Two-Layer NetworkHere is a concrete example of using the `tf.keras.Model` API to define a two-layer network. There are a few new bits of API to be aware of here:We use an `Initializer` object to set up the initial values of the learnable parameters of the layers; in particular `tf.initializers.VarianceScaling` gives behavior similar to the Kaiming initialization method we used in Part II. You can read more about it here: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/initializers/VarianceScalingWe construct `tf.keras.layers.Dense` objects to represent the two fully-connected layers of the model. In addition to multiplying their input by a weight matrix and adding a bias vector, these layer can also apply a nonlinearity for you. For the first layer we specify a ReLU activation function by passing `activation='relu'` to the constructor; the second layer uses softmax activation function. Finally, we use `tf.keras.layers.Flatten` to flatten the output from the previous fully-connected layer. ###Code class TwoLayerFC(tf.keras.Model): def __init__(self, hidden_size, num_classes): super().__init__() #super(TwoLayerFC, self).__init__() initializer = tf.initializers.VarianceScaling(scale=2.0) self.flatten = tf.keras.layers.Flatten() self.fc1 = tf.keras.layers.Dense(hidden_size, activation='relu', kernel_initializer=initializer) self.fc2 = tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer) def call(self, x, training=False): x = self.flatten(x) x = self.fc1(x) x = self.fc2(x) return x def test_TwoLayerFC(): """ A small unit test to exercise the TwoLayerFC model above. """ input_size, hidden_size, num_classes = 50, 42, 10 x = tf.zeros((64, input_size)) model = TwoLayerFC(hidden_size, num_classes) with tf.device(device): scores = model(x) print(scores.shape) test_TwoLayerFC() ###Output (64, 10) ###Markdown Keras Model Subclassing API: Three-Layer ConvNetNow it's your turn to implement a three-layer ConvNet using the `tf.keras.Model` API. Your model should have the same architecture used in Part II:1. Convolutional layer with 5 x 5 kernels, with zero-padding of 22. ReLU nonlinearity3. Convolutional layer with 3 x 3 kernels, with zero-padding of 14. ReLU nonlinearity5. Fully-connected layer to give class scores6. Softmax nonlinearityYou should initialize the weights of your network using the same initialization method as was used in the two-layer network above.**Hint**: Refer to the documentation for `tf.keras.layers.Conv2D` and `tf.keras.layers.Dense`:https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/Conv2Dhttps://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/Dense ###Code class ThreeLayerConvNet(tf.keras.Model): def __init__(self, channel_1, channel_2, num_classes): super().__init__() ######################################################################## # TODO: Implement the __init__ method for a three-layer ConvNet. You # # should instantiate layer objects to be used in the forward pass. # ######################################################################## # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** initializer = tf.initializers.VarianceScaling(scale=2.0) self.conv1 = tf.keras.layers.Conv2D(channel_1, (5, 5), padding='same', activation='relu', kernel_initializer=initializer) self.conv2 = tf.keras.layers.Conv2D(channel_2, (3, 3), padding='same', activation='relu', kernel_initializer=initializer) self.flatten = tf.keras.layers.Flatten() self.fc = tf.keras.layers.Dense(num_classes, activation='softmax') # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ######################################################################## # END OF YOUR CODE # ######################################################################## def call(self, x, training=False): scores = None ######################################################################## # TODO: Implement the forward pass for a three-layer ConvNet. You # # should use the layer objects defined in the __init__ method. # ######################################################################## # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** x = self.conv1(x) x = self.conv2(x) x = self.flatten(x) scores = self.fc(x) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ######################################################################## # END OF YOUR CODE # ######################################################################## return scores ###Output _____no_output_____ ###Markdown Once you complete the implementation of the `ThreeLayerConvNet` above you can run the following to ensure that your implementation does not crash and produces outputs of the expected shape. ###Code def test_ThreeLayerConvNet(): channel_1, channel_2, num_classes = 12, 8, 10 model = ThreeLayerConvNet(channel_1, channel_2, num_classes) with tf.device(device): x = tf.zeros((64, 3, 32, 32)) scores = model(x) print(scores.shape) test_ThreeLayerConvNet() ###Output (64, 10) ###Markdown Keras Model Subclassing API: Eager TrainingWhile keras models have a builtin training loop (using the `model.fit`), sometimes you need more customization. Here's an example, of a training loop implemented with eager execution.In particular, notice `tf.GradientTape`. Automatic differentiation is used in the backend for implementing backpropagation in frameworks like TensorFlow. During eager execution, `tf.GradientTape` is used to trace operations for computing gradients later. A particular `tf.GradientTape` can only compute one gradient; subsequent calls to tape will throw a runtime error. TensorFlow 2.0 ships with easy-to-use built-in metrics under `tf.keras.metrics` module. Each metric is an object, and we can use `update_state()` to add observations and `reset_state()` to clear all observations. We can get the current result of a metric by calling `result()` on the metric object. ###Code def train_part34(model_init_fn, optimizer_init_fn, num_epochs=1, is_training=False): """ Simple training loop for use with models defined using tf.keras. It trains a model for one epoch on the CIFAR-10 training set and periodically checks accuracy on the CIFAR-10 validation set. Inputs: - model_init_fn: A function that takes no parameters; when called it constructs the model we want to train: model = model_init_fn() - optimizer_init_fn: A function which takes no parameters; when called it constructs the Optimizer object we will use to optimize the model: optimizer = optimizer_init_fn() - num_epochs: The number of epochs to train for Returns: Nothing, but prints progress during trainingn """ with tf.device(device): # Compute the loss like we did in Part II loss_fn = tf.keras.losses.SparseCategoricalCrossentropy() model = model_init_fn() optimizer = optimizer_init_fn() train_loss = tf.keras.metrics.Mean(name='train_loss') train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy') val_loss = tf.keras.metrics.Mean(name='val_loss') val_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='val_accuracy') t = 0 for epoch in range(num_epochs): # Reset the metrics - https://www.tensorflow.org/alpha/guide/migration_guide#new-style_metrics train_loss.reset_states() train_accuracy.reset_states() for x_np, y_np in train_dset: with tf.GradientTape() as tape: # Use the model function to build the forward pass. scores = model(x_np, training=is_training) loss = loss_fn(y_np, scores) gradients = tape.gradient(loss, model.trainable_variables) optimizer.apply_gradients(zip(gradients, model.trainable_variables)) # Update the metrics train_loss.update_state(loss) train_accuracy.update_state(y_np, scores) if t % print_every == 0: val_loss.reset_states() val_accuracy.reset_states() for test_x, test_y in val_dset: # During validation at end of epoch, training set to False prediction = model(test_x, training=False) t_loss = loss_fn(test_y, prediction) val_loss.update_state(t_loss) val_accuracy.update_state(test_y, prediction) template = 'Iteration {}, Epoch {}, Loss: {}, Accuracy: {}, Val Loss: {}, Val Accuracy: {}' print (template.format(t, epoch+1, train_loss.result(), train_accuracy.result()*100, val_loss.result(), val_accuracy.result()*100)) t += 1 ###Output _____no_output_____ ###Markdown Keras Model Subclassing API: Train a Two-Layer NetworkWe can now use the tools defined above to train a two-layer network on CIFAR-10. We define the `model_init_fn` and `optimizer_init_fn` that construct the model and optimizer respectively when called. Here we want to train the model using stochastic gradient descent with no momentum, so we construct a `tf.keras.optimizers.SGD` function; you can [read about it here](https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/optimizers/SGD).You don't need to tune any hyperparameters here, but you should achieve validation accuracies above 40% after one epoch of training. ###Code hidden_size, num_classes = 4000, 10 learning_rate = 1e-2 def model_init_fn(): return TwoLayerFC(hidden_size, num_classes) def optimizer_init_fn(): return tf.keras.optimizers.SGD(learning_rate=learning_rate) train_part34(model_init_fn, optimizer_init_fn, num_epochs=1) ###Output Iteration 0, Epoch 1, Loss: 2.8405842781066895, Accuracy: 10.9375, Val Loss: 2.9787204265594482, Val Accuracy: 11.800000190734863 Iteration 100, Epoch 1, Loss: 2.2377357482910156, Accuracy: 27.691831588745117, Val Loss: 1.9073878526687622, Val Accuracy: 38.29999923706055 Iteration 200, Epoch 1, Loss: 2.0827560424804688, Accuracy: 31.778606414794922, Val Loss: 1.834181785583496, Val Accuracy: 39.89999771118164 Iteration 300, Epoch 1, Loss: 2.000749349594116, Accuracy: 33.80917739868164, Val Loss: 1.8754873275756836, Val Accuracy: 36.099998474121094 Iteration 400, Epoch 1, Loss: 1.9329524040222168, Accuracy: 35.582916259765625, Val Loss: 1.7133322954177856, Val Accuracy: 41.20000076293945 Iteration 500, Epoch 1, Loss: 1.8873575925827026, Accuracy: 36.76708984375, Val Loss: 1.6488999128341675, Val Accuracy: 41.60000228881836 Iteration 600, Epoch 1, Loss: 1.8568066358566284, Accuracy: 37.72878646850586, Val Loss: 1.6973183155059814, Val Accuracy: 41.60000228881836 Iteration 700, Epoch 1, Loss: 1.830464243888855, Accuracy: 38.46290969848633, Val Loss: 1.6309181451797485, Val Accuracy: 42.5 ###Markdown Keras Model Subclassing API: Train a Three-Layer ConvNetHere you should use the tools we've defined above to train a three-layer ConvNet on CIFAR-10. Your ConvNet should use 32 filters in the first convolutional layer and 16 filters in the second layer.To train the model you should use gradient descent with Nesterov momentum 0.9. **HINT**: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/optimizers/SGDYou don't need to perform any hyperparameter tuning, but you should achieve validation accuracies above 50% after training for one epoch. ###Code learning_rate = 3e-3 channel_1, channel_2, num_classes = 32, 16, 10 def model_init_fn(): model = None ############################################################################ # TODO: Complete the implementation of model_fn. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** model = ThreeLayerConvNet(channel_1, channel_2, num_classes) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return model def optimizer_init_fn(): optimizer = None ############################################################################ # TODO: Complete the implementation of model_fn. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate, nesterov=True, momentum=0.9) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return optimizer train_part34(model_init_fn, optimizer_init_fn, num_epochs=1) ###Output Iteration 0, Epoch 1, Loss: 3.154758930206299, Accuracy: 10.9375, Val Loss: 5.975311279296875, Val Accuracy: 8.399999618530273 Iteration 100, Epoch 1, Loss: 2.041959047317505, Accuracy: 31.07982635498047, Val Loss: 1.68193519115448, Val Accuracy: 40.70000076293945 Iteration 200, Epoch 1, Loss: 1.8226760625839233, Accuracy: 37.31343078613281, Val Loss: 1.5124176740646362, Val Accuracy: 47.400001525878906 Iteration 300, Epoch 1, Loss: 1.714808464050293, Accuracy: 40.785919189453125, Val Loss: 1.4400432109832764, Val Accuracy: 49.0 Iteration 400, Epoch 1, Loss: 1.6360254287719727, Accuracy: 42.94731903076172, Val Loss: 1.3696250915527344, Val Accuracy: 52.29999923706055 Iteration 500, Epoch 1, Loss: 1.5823934078216553, Accuracy: 44.61701583862305, Val Loss: 1.3425835371017456, Val Accuracy: 51.5 Iteration 600, Epoch 1, Loss: 1.5470515489578247, Accuracy: 45.705074310302734, Val Loss: 1.3165476322174072, Val Accuracy: 52.79999923706055 Iteration 700, Epoch 1, Loss: 1.5153937339782715, Accuracy: 46.83710479736328, Val Loss: 1.3428152799606323, Val Accuracy: 52.79999923706055 ###Markdown Part IV: Keras Sequential APIIn Part III we introduced the `tf.keras.Model` API, which allows you to define models with any number of learnable layers and with arbitrary connectivity between layers.However for many models you don't need such flexibility - a lot of models can be expressed as a sequential stack of layers, with the output of each layer fed to the next layer as input. If your model fits this pattern, then there is an even easier way to define your model: using `tf.keras.Sequential`. You don't need to write any custom classes; you simply call the `tf.keras.Sequential` constructor with a list containing a sequence of layer objects.One complication with `tf.keras.Sequential` is that you must define the shape of the input to the model by passing a value to the `input_shape` of the first layer in your model. Keras Sequential API: Two-Layer NetworkIn this subsection, we will rewrite the two-layer fully-connected network using `tf.keras.Sequential`, and train it using the training loop defined above.You don't need to perform any hyperparameter tuning here, but you should see validation accuracies above 40% after training for one epoch. ###Code learning_rate = 1e-2 def model_init_fn(): input_shape = (32, 32, 3) hidden_layer_size, num_classes = 4000, 10 initializer = tf.initializers.VarianceScaling(scale=2.0) layers = [ tf.keras.layers.Flatten(input_shape=input_shape), tf.keras.layers.Dense(hidden_layer_size, activation='relu', kernel_initializer=initializer), tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer), ] model = tf.keras.Sequential(layers) return model def optimizer_init_fn(): return tf.keras.optimizers.SGD(learning_rate=learning_rate) train_part34(model_init_fn, optimizer_init_fn) ###Output Iteration 0, Epoch 1, Loss: 3.049715995788574, Accuracy: 9.375, Val Loss: 2.892936944961548, Val Accuracy: 11.699999809265137 Iteration 100, Epoch 1, Loss: 2.2131028175354004, Accuracy: 29.068687438964844, Val Loss: 1.8789817094802856, Val Accuracy: 39.89999771118164 Iteration 200, Epoch 1, Loss: 2.063352584838867, Accuracy: 32.75808334350586, Val Loss: 1.8269221782684326, Val Accuracy: 39.79999923706055 Iteration 300, Epoch 1, Loss: 1.9906622171401978, Accuracy: 34.55668640136719, Val Loss: 1.851961374282837, Val Accuracy: 39.20000076293945 Iteration 400, Epoch 1, Loss: 1.9242477416992188, Accuracy: 36.264808654785156, Val Loss: 1.7083684206008911, Val Accuracy: 43.70000076293945 Iteration 500, Epoch 1, Loss: 1.8799132108688354, Accuracy: 37.34406280517578, Val Loss: 1.6594358682632446, Val Accuracy: 44.29999923706055 Iteration 600, Epoch 1, Loss: 1.848968744277954, Accuracy: 38.2149543762207, Val Loss: 1.6731466054916382, Val Accuracy: 43.29999923706055 Iteration 700, Epoch 1, Loss: 1.8228310346603394, Accuracy: 38.906471252441406, Val Loss: 1.6165885925292969, Val Accuracy: 45.69999694824219 ###Markdown Abstracting Away the Training LoopIn the previous examples, we used a customised training loop to train models (e.g. `train_part34`). Writing your own training loop is only required if you need more flexibility and control during training your model. Alternately, you can also use built-in APIs like `tf.keras.Model.fit()` and `tf.keras.Model.evaluate` to train and evaluate a model. Also remember to configure your model for training by calling `tf.keras.Model.compile.You don't need to perform any hyperparameter tuning here, but you should see validation and test accuracies above 42% after training for one epoch. ###Code model = model_init_fn() model.compile(optimizer=tf.keras.optimizers.SGD(learning_rate=learning_rate), loss='sparse_categorical_crossentropy', metrics=[tf.keras.metrics.sparse_categorical_accuracy]) model.fit(X_train, y_train, batch_size=64, epochs=1, validation_data=(X_val, y_val)) model.evaluate(X_test, y_test) ###Output Train on 49000 samples, validate on 1000 samples 49000/49000 [==============================] - 3s 57us/sample - loss: 1.8204 - sparse_categorical_accuracy: 0.3874 - val_loss: 1.6748 - val_sparse_categorical_accuracy: 0.4180 10000/1 [================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================] - 1s 77us/sample - loss: 1.6879 - sparse_categorical_accuracy: 0.4223 ###Markdown Keras Sequential API: Three-Layer ConvNetHere you should use `tf.keras.Sequential` to reimplement the same three-layer ConvNet architecture used in Part II and Part III. As a reminder, your model should have the following architecture:1. Convolutional layer with 32 5x5 kernels, using zero padding of 22. ReLU nonlinearity3. Convolutional layer with 16 3x3 kernels, using zero padding of 14. ReLU nonlinearity5. Fully-connected layer giving class scores6. Softmax nonlinearityYou should initialize the weights of the model using a `tf.initializers.VarianceScaling` as above.You should train the model using Nesterov momentum 0.9.You don't need to perform any hyperparameter search, but you should achieve accuracy above 45% after training for one epoch. ###Code def model_init_fn(): model = None ############################################################################ # TODO: Construct a three-layer ConvNet using tf.keras.Sequential. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** input_shape = (32, 32, 3) num_classes = 10 initializer = tf.initializers.VarianceScaling(scale=2.0) layers = [ tf.keras.layers.Conv2D(32, (5, 5), padding='same', activation='relu', kernel_initializer=initializer, input_shape=input_shape), tf.keras.layers.Conv2D(16, (3, 3), padding='same', activation='relu', kernel_initializer=initializer), tf.keras.layers.Flatten(), tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer), ] model = tf.keras.Sequential(layers) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return model learning_rate = 5e-4 def optimizer_init_fn(): optimizer = None ############################################################################ # TODO: Complete the implementation of model_fn. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate, nesterov=True, momentum=0.9) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return optimizer train_part34(model_init_fn, optimizer_init_fn) ###Output Iteration 0, Epoch 1, Loss: 2.9712705612182617, Accuracy: 9.375, Val Loss: 3.2196428775787354, Val Accuracy: 12.700000762939453 Iteration 100, Epoch 1, Loss: 2.0532257556915283, Accuracy: 29.981433868408203, Val Loss: 1.8071397542953491, Val Accuracy: 36.29999923706055 Iteration 200, Epoch 1, Loss: 1.899694800376892, Accuracy: 34.16511154174805, Val Loss: 1.6743762493133545, Val Accuracy: 41.0 Iteration 300, Epoch 1, Loss: 1.8233054876327515, Accuracy: 36.59675979614258, Val Loss: 1.6346663236618042, Val Accuracy: 45.0 Iteration 400, Epoch 1, Loss: 1.7578901052474976, Accuracy: 38.88325881958008, Val Loss: 1.5955396890640259, Val Accuracy: 45.20000076293945 Iteration 500, Epoch 1, Loss: 1.7127082347869873, Accuracy: 40.272579193115234, Val Loss: 1.5494569540023804, Val Accuracy: 46.29999923706055 Iteration 600, Epoch 1, Loss: 1.6833456754684448, Accuracy: 41.28015899658203, Val Loss: 1.5043326616287231, Val Accuracy: 48.79999923706055 Iteration 700, Epoch 1, Loss: 1.6563769578933716, Accuracy: 42.189727783203125, Val Loss: 1.480652928352356, Val Accuracy: 48.0 ###Markdown We will also train this model with the built-in training loop APIs provided by TensorFlow. ###Code model = model_init_fn() model.compile(optimizer='sgd', loss='sparse_categorical_crossentropy', metrics=[tf.keras.metrics.sparse_categorical_accuracy]) model.fit(X_train, y_train, batch_size=64, epochs=1, validation_data=(X_val, y_val)) model.evaluate(X_test, y_test) ###Output Train on 49000 samples, validate on 1000 samples 49000/49000 [==============================] - 4s 89us/sample - loss: 1.5548 - sparse_categorical_accuracy: 0.4531 - val_loss: 1.4325 - val_sparse_categorical_accuracy: 0.4920 10000/1 [================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================] - 1s 93us/sample - loss: 1.3228 - sparse_categorical_accuracy: 0.4854 ###Markdown Part IV: Functional API Demonstration with a Two-Layer Network In the previous section, we saw how we can use `tf.keras.Sequential` to stack layers to quickly build simple models. But this comes at the cost of losing flexibility.Often we will have to write complex models that have non-sequential data flows: a layer can have **multiple inputs and/or outputs**, such as stacking the output of 2 previous layers together to feed as input to a third! (Some examples are residual connections and dense blocks.)In such cases, we can use Keras functional API to write models with complex topologies such as: 1. Multi-input models 2. Multi-output models 3. Models with shared layers (the same layer called several times) 4. Models with non-sequential data flows (e.g. residual connections)Writing a model with Functional API requires us to create a `tf.keras.Model` instance and explicitly write input tensors and output tensors for this model. ###Code def two_layer_fc_functional(input_shape, hidden_size, num_classes): initializer = tf.initializers.VarianceScaling(scale=2.0) inputs = tf.keras.Input(shape=input_shape) flattened_inputs = tf.keras.layers.Flatten()(inputs) fc1_output = tf.keras.layers.Dense(hidden_size, activation='relu', kernel_initializer=initializer)(flattened_inputs) scores = tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer)(fc1_output) # Instantiate the model given inputs and outputs. model = tf.keras.Model(inputs=inputs, outputs=scores) return model def test_two_layer_fc_functional(): """ A small unit test to exercise the TwoLayerFC model above. """ input_size, hidden_size, num_classes = 50, 42, 10 input_shape = (50,) x = tf.zeros((64, input_size)) model = two_layer_fc_functional(input_shape, hidden_size, num_classes) with tf.device(device): scores = model(x) print(scores.shape) test_two_layer_fc_functional() ###Output (64, 10) ###Markdown Keras Functional API: Train a Two-Layer NetworkYou can now train this two-layer network constructed using the functional API.You don't need to perform any hyperparameter tuning here, but you should see validation accuracies above 40% after training for one epoch. ###Code input_shape = (32, 32, 3) hidden_size, num_classes = 4000, 10 learning_rate = 1e-2 def model_init_fn(): return two_layer_fc_functional(input_shape, hidden_size, num_classes) def optimizer_init_fn(): return tf.keras.optimizers.SGD(learning_rate=learning_rate) train_part34(model_init_fn, optimizer_init_fn) ###Output Iteration 0, Epoch 1, Loss: 2.8402795791625977, Accuracy: 17.1875, Val Loss: 2.810950517654419, Val Accuracy: 15.09999942779541 Iteration 100, Epoch 1, Loss: 2.226226568222046, Accuracy: 28.496286392211914, Val Loss: 1.9141902923583984, Val Accuracy: 38.60000228881836 Iteration 200, Epoch 1, Loss: 2.080077886581421, Accuracy: 32.06623077392578, Val Loss: 1.8725242614746094, Val Accuracy: 38.29999923706055 Iteration 300, Epoch 1, Loss: 2.0010735988616943, Accuracy: 34.29194259643555, Val Loss: 1.8864822387695312, Val Accuracy: 37.400001525878906 Iteration 400, Epoch 1, Loss: 1.9321510791778564, Accuracy: 36.06218719482422, Val Loss: 1.7192223072052002, Val Accuracy: 41.70000076293945 Iteration 500, Epoch 1, Loss: 1.886867642402649, Accuracy: 37.12263107299805, Val Loss: 1.6607885360717773, Val Accuracy: 44.0 Iteration 600, Epoch 1, Loss: 1.8582278490066528, Accuracy: 37.86917495727539, Val Loss: 1.698154330253601, Val Accuracy: 42.39999771118164 Iteration 700, Epoch 1, Loss: 1.8322076797485352, Accuracy: 38.547611236572266, Val Loss: 1.6500277519226074, Val Accuracy: 42.39999771118164 ###Markdown Part V: CIFAR-10 open-ended challengeIn this section you can experiment with whatever ConvNet architecture you'd like on CIFAR-10.You should experiment with architectures, hyperparameters, loss functions, regularization, or anything else you can think of to train a model that achieves **at least 70%** accuracy on the **validation** set within 10 epochs. You can use the built-in train function, the `train_part34` function from above, or implement your own training loop.Describe what you did at the end of the notebook. Some things you can try:- **Filter size**: Above we used 5x5 and 3x3; is this optimal?- **Number of filters**: Above we used 16 and 32 filters. Would more or fewer do better?- **Pooling**: We didn't use any pooling above. Would this improve the model?- **Normalization**: Would your model be improved with batch normalization, layer normalization, group normalization, or some other normalization strategy?- **Network architecture**: The ConvNet above has only three layers of trainable parameters. Would a deeper model do better?- **Global average pooling**: Instead of flattening after the final convolutional layer, would global average pooling do better? This strategy is used for example in Google's Inception network and in Residual Networks.- **Regularization**: Would some kind of regularization improve performance? Maybe weight decay or dropout? NOTE: Batch Normalization / DropoutIf you are using Batch Normalization and Dropout, remember to pass `is_training=True` if you use the `train_part34()` function. BatchNorm and Dropout layers have different behaviors at training and inference time. `training` is a specific keyword argument reserved for this purpose in any `tf.keras.Model`'s `call()` function. Read more about this here : https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/BatchNormalizationmethodshttps://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/Dropoutmethods Tips for trainingFor each network architecture that you try, you should tune the learning rate and other hyperparameters. When doing this there are a couple important things to keep in mind: - If the parameters are working well, you should see improvement within a few hundred iterations- Remember the coarse-to-fine approach for hyperparameter tuning: start by testing a large range of hyperparameters for just a few training iterations to find the combinations of parameters that are working at all.- Once you have found some sets of parameters that seem to work, search more finely around these parameters. You may need to train for more epochs.- You should use the validation set for hyperparameter search, and save your test set for evaluating your architecture on the best parameters as selected by the validation set. Going above and beyondIf you are feeling adventurous there are many other features you can implement to try and improve your performance. You are **not required** to implement any of these, but don't miss the fun if you have time!- Alternative optimizers: you can try Adam, Adagrad, RMSprop, etc.- Alternative activation functions such as leaky ReLU, parametric ReLU, ELU, or MaxOut.- Model ensembles- Data augmentation- New Architectures - [ResNets](https://arxiv.org/abs/1512.03385) where the input from the previous layer is added to the output. - [DenseNets](https://arxiv.org/abs/1608.06993) where inputs into previous layers are concatenated together. - [This blog has an in-depth overview](https://chatbotslife.com/resnets-highwaynets-and-densenets-oh-my-9bb15918ee32) Have fun and happy training! ###Code class CustomConvNet(tf.keras.Model): def __init__(self): super().__init__() ############################################################################ # TODO: Construct a model that performs well on CIFAR-10 # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** # input is [N, 32, 32, 3] initializer = tf.initializers.VarianceScaling(scale=2.0) self.conv11 = tf.keras.layers.Conv2D(512, (3, 3), padding='same', kernel_initializer=initializer) # 32, 32, 128 self.prelu11 = tf.keras.layers.PReLU(alpha_initializer=initializer) self.bn11 = tf.keras.layers.BatchNormalization() self.conv12 = tf.keras.layers.Conv2D(256, (3, 3), padding='same', kernel_initializer=initializer) # 32, 32, 128 self.prelu12 = tf.keras.layers.PReLU(alpha_initializer=initializer) self.bn12 = tf.keras.layers.BatchNormalization() self.conv13 = tf.keras.layers.Conv2D(128, (3, 3), padding='same', kernel_initializer=initializer) # 32, 32, 128 self.prelu13 = tf.keras.layers.PReLU(alpha_initializer=initializer) self.bn13 = tf.keras.layers.BatchNormalization() self.conv2 = tf.keras.layers.Conv2D(64, (3, 3), padding='same', kernel_initializer=initializer) # 32, 32, 64 self.prelu2 = tf.keras.layers.PReLU(alpha_initializer=initializer) self.bn2 = tf.keras.layers.BatchNormalization() self.maxpool2 = tf.keras.layers.MaxPool2D((2, 2), padding='same') # 16, 16, 64 self.conv3 = tf.keras.layers.Conv2D(32, (3, 3), padding='same', kernel_initializer=initializer) # 16, 16, 32 self.prelu3 = tf.keras.layers.PReLU(alpha_initializer=initializer) self.bn3 = tf.keras.layers.BatchNormalization() self.maxpool3 = tf.keras.layers.MaxPool2D((2, 2), padding='same') # 8, 8, 32 self.flatten = tf.keras.layers.Flatten() self.fc = tf.keras.layers.Dense(10, activation='softmax', kernel_initializer=initializer) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ def call(self, input_tensor, training=False): ############################################################################ # TODO: Construct a model that performs well on CIFAR-10 # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** x = input_tensor x = self.conv11(x) x = self.prelu11(x) x = self.bn11(x, training) x = self.conv12(x) x = self.prelu12(x) x = self.bn12(x, training) x = self.conv13(x) x = self.prelu13(x) x = self.bn13(x, training) x = self.conv2(x) x = self.prelu2(x) x = self.bn2(x, training) x = self.maxpool2(x) x = self.conv3(x) x = self.prelu3(x) x = self.bn3(x, training) x = self.maxpool3(x) x = self.flatten(x) x = self.fc(x) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return x device = '/device:GPU:0' # Change this to a CPU/GPU as you wish! # device = '/cpu:0' # Change this to a CPU/GPU as you wish! print_every = 300 num_epochs = 10 # model = CustomConvNet() # model = CustomResNet() def model_init_fn(): return CustomConvNet() def optimizer_init_fn(): learning_rate = 1e-3 return tf.keras.optimizers.Adam(learning_rate) train_part34(model_init_fn, optimizer_init_fn, num_epochs=num_epochs, is_training=True) ###Output Iteration 0, Epoch 1, Loss: 4.086295127868652, Accuracy: 12.5, Val Loss: 4.507420539855957, Val Accuracy: 13.199999809265137 Iteration 300, Epoch 1, Loss: 1.6533818244934082, Accuracy: 43.55793380737305, Val Loss: 1.471709132194519, Val Accuracy: 50.19999694824219 Iteration 600, Epoch 1, Loss: 1.4560245275497437, Accuracy: 49.78681564331055, Val Loss: 1.2510604858398438, Val Accuracy: 57.20000457763672 Iteration 900, Epoch 2, Loss: 1.007521152496338, Accuracy: 64.97685241699219, Val Loss: 1.0304434299468994, Val Accuracy: 63.900001525878906 Iteration 1200, Epoch 2, Loss: 0.9641342163085938, Accuracy: 66.3936767578125, Val Loss: 0.9729200601577759, Val Accuracy: 65.80000305175781 Iteration 1500, Epoch 2, Loss: 0.926266610622406, Accuracy: 67.551025390625, Val Loss: 0.9301822781562805, Val Accuracy: 67.19999694824219 Iteration 1800, Epoch 3, Loss: 0.7934252619743347, Accuracy: 72.49070739746094, Val Loss: 0.8698236346244812, Val Accuracy: 70.30000305175781 Iteration 2100, Epoch 3, Loss: 0.763357937335968, Accuracy: 73.67091369628906, Val Loss: 0.83391273021698, Val Accuracy: 71.5 Iteration 2400, Epoch 4, Loss: 0.6582642793655396, Accuracy: 77.12378692626953, Val Loss: 0.7901626825332642, Val Accuracy: 73.69999694824219 Iteration 2700, Epoch 4, Loss: 0.6490508913993835, Accuracy: 77.60545349121094, Val Loss: 0.8368052840232849, Val Accuracy: 71.80000305175781 Iteration 3000, Epoch 4, Loss: 0.6276915669441223, Accuracy: 78.28724670410156, Val Loss: 0.8429279923439026, Val Accuracy: 72.79999542236328 Iteration 3300, Epoch 5, Loss: 0.5404477119445801, Accuracy: 81.48734283447266, Val Loss: 0.8269108533859253, Val Accuracy: 74.5 Iteration 3600, Epoch 5, Loss: 0.5194686055183411, Accuracy: 82.16072845458984, Val Loss: 0.9136738181114197, Val Accuracy: 71.0 Iteration 3900, Epoch 6, Loss: 0.42960742115974426, Accuracy: 86.35562896728516, Val Loss: 0.8334789276123047, Val Accuracy: 73.69999694824219 Iteration 4200, Epoch 6, Loss: 0.409514844417572, Accuracy: 86.14386749267578, Val Loss: 0.859889030456543, Val Accuracy: 72.0 Iteration 4500, Epoch 6, Loss: 0.3866899609565735, Accuracy: 86.96441650390625, Val Loss: 0.9733158349990845, Val Accuracy: 71.80000305175781 Iteration 4800, Epoch 7, Loss: 0.31196486949920654, Accuracy: 89.12347412109375, Val Loss: 0.9835796356201172, Val Accuracy: 71.9000015258789 Iteration 5100, Epoch 7, Loss: 0.29267334938049316, Accuracy: 90.00928497314453, Val Loss: 1.1022939682006836, Val Accuracy: 72.79999542236328 Iteration 5400, Epoch 8, Loss: 0.24182185530662537, Accuracy: 91.34615325927734, Val Loss: 1.1322561502456665, Val Accuracy: 71.69999694824219 ###Markdown What's this TensorFlow business?You've written a lot of code in this assignment to provide a whole host of neural network functionality. Dropout, Batch Norm, and 2D convolutions are some of the workhorses of deep learning in computer vision. You've also worked hard to make your code efficient and vectorized.For the last part of this assignment, though, we're going to leave behind your beautiful codebase and instead migrate to one of two popular deep learning frameworks: in this instance, TensorFlow (or PyTorch, if you choose to work with that notebook). What is it?TensorFlow is a system for executing computational graphs over Tensor objects, with native support for performing backpropogation for its Variables. In it, we work with Tensors which are n-dimensional arrays analogous to the numpy ndarray. Why?* Our code will now run on GPUs! Much faster training. Writing your own modules to run on GPUs is beyond the scope of this class, unfortunately.* We want you to be ready to use one of these frameworks for your project so you can experiment more efficiently than if you were writing every feature you want to use by hand. * We want you to stand on the shoulders of giants! TensorFlow and PyTorch are both excellent frameworks that will make your lives a lot easier, and now that you understand their guts, you are free to use them :) * We want you to be exposed to the sort of deep learning code you might run into in academia or industry. How will I learn TensorFlow?TensorFlow has many excellent tutorials available, including those from [Google themselves](https://www.tensorflow.org/get_started/get_started).Otherwise, this notebook will walk you through much of what you need to do to train models in TensorFlow. See the end of the notebook for some links to helpful tutorials if you want to learn more or need further clarification on topics that aren't fully explained here.**NOTE: This notebook is meant to teach you the latest version of Tensorflow 2.0. Most examples on the web today are still in 1.x, so be careful not to confuse the two when looking up documentation**. Install Tensorflow 2.0Tensorflow 2.0 is still not in a fully 100% stable release, but it's still usable and more intuitive than TF 1.x. Please make sure you have it installed before moving on in this notebook! Here are some steps to get started:1. Have the latest version of Anaconda installed on your machine.2. Create a new conda environment starting from Python 3.7. In this setup example, we'll call it `tf_20_env`.3. Run the command: `source activate tf_20_env`4. Then pip install TF 2.0 as described here: https://www.tensorflow.org/install/pip A guide on creating Anaconda enviornments: https://uoa-eresearch.github.io/eresearch-cookbook/recipe/2014/11/20/conda/This will give you an new enviornemnt to play in TF 2.0. Generally, if you plan to also use TensorFlow in your other projects, you might also want to keep a seperate Conda environment or virtualenv in Python 3.7 that has Tensorflow 1.9, so you can switch back and forth at will. Table of ContentsThis notebook has 5 parts. We will walk through TensorFlow at **three different levels of abstraction**, which should help you better understand it and prepare you for working on your project.1. Part I, Preparation: load the CIFAR-10 dataset.2. Part II, Barebone TensorFlow: **Abstraction Level 1**, we will work directly with low-level TensorFlow graphs. 3. Part III, Keras Model API: **Abstraction Level 2**, we will use `tf.keras.Model` to define arbitrary neural network architecture. 4. Part IV, Keras Sequential + Functional API: **Abstraction Level 3**, we will use `tf.keras.Sequential` to define a linear feed-forward network very conveniently, and then explore the functional libraries for building unique and uncommon models that require more flexibility.5. Part V, CIFAR-10 open-ended challenge: please implement your own network to get as high accuracy as possible on CIFAR-10. You can experiment with any layer, optimizer, hyperparameters or other advanced features. We will discuss Keras in more detail later in the notebook.Here is a table of comparison:| API | Flexibility | Convenience ||---------------|-------------|-------------|| Barebone | High | Low || `tf.keras.Model` | High | Medium || `tf.keras.Sequential` | Low | High | Part I: PreparationFirst, we load the CIFAR-10 dataset. This might take a few minutes to download the first time you run it, but after that the files should be cached on disk and loading should be faster.In previous parts of the assignment we used CS231N-specific code to download and read the CIFAR-10 dataset; however the `tf.keras.datasets` package in TensorFlow provides prebuilt utility functions for loading many common datasets.For the purposes of this assignment we will still write our own code to preprocess the data and iterate through it in minibatches. The `tf.data` package in TensorFlow provides tools for automating this process, but working with this package adds extra complication and is beyond the scope of this notebook. However using `tf.data` can be much more efficient than the simple approach used in this notebook, so you should consider using it for your project. ###Code import os import tensorflow as tf import numpy as np import math import timeit import matplotlib.pyplot as plt %matplotlib inline def load_cifar10(num_training=49000, num_validation=1000, num_test=10000): """ Fetch the CIFAR-10 dataset from the web and perform preprocessing to prepare it for the two-layer neural net classifier. These are the same steps as we used for the SVM, but condensed to a single function. """ # Load the raw CIFAR-10 dataset and use appropriate data types and shapes cifar10 = tf.keras.datasets.cifar10.load_data() (X_train, y_train), (X_test, y_test) = cifar10 X_train = np.asarray(X_train, dtype=np.float32) y_train = np.asarray(y_train, dtype=np.int32).flatten() X_test = np.asarray(X_test, dtype=np.float32) y_test = np.asarray(y_test, dtype=np.int32).flatten() # Subsample the data mask = range(num_training, num_training + num_validation) X_val = X_train[mask] y_val = y_train[mask] mask = range(num_training) X_train = X_train[mask] y_train = y_train[mask] mask = range(num_test) X_test = X_test[mask] y_test = y_test[mask] # Normalize the data: subtract the mean pixel and divide by std mean_pixel = X_train.mean(axis=(0, 1, 2), keepdims=True) std_pixel = X_train.std(axis=(0, 1, 2), keepdims=True) X_train = (X_train - mean_pixel) / std_pixel X_val = (X_val - mean_pixel) / std_pixel X_test = (X_test - mean_pixel) / std_pixel return X_train, y_train, X_val, y_val, X_test, y_test # If there are errors with SSL downloading involving self-signed certificates, # it may be that your Python version was recently installed on the current machine. # See: https://github.com/tensorflow/tensorflow/issues/10779 # To fix, run the command: /Applications/Python\ 3.7/Install\ Certificates.command # ...replacing paths as necessary. # Invoke the above function to get our data. NHW = (0, 1, 2) X_train, y_train, X_val, y_val, X_test, y_test = load_cifar10() print('Train data shape: ', X_train.shape) print('Train labels shape: ', y_train.shape, y_train.dtype) print('Validation data shape: ', X_val.shape) print('Validation labels shape: ', y_val.shape) print('Test data shape: ', X_test.shape) print('Test labels shape: ', y_test.shape) class Dataset(object): def __init__(self, X, y, batch_size, shuffle=False): """ Construct a Dataset object to iterate over data X and labels y Inputs: - X: Numpy array of data, of any shape - y: Numpy array of labels, of any shape but with y.shape[0] == X.shape[0] - batch_size: Integer giving number of elements per minibatch - shuffle: (optional) Boolean, whether to shuffle the data on each epoch """ assert X.shape[0] == y.shape[0], 'Got different numbers of data and labels' self.X, self.y = X, y self.batch_size, self.shuffle = batch_size, shuffle def __iter__(self): N, B = self.X.shape[0], self.batch_size idxs = np.arange(N) if self.shuffle: np.random.shuffle(idxs) return iter((self.X[i:i+B], self.y[i:i+B]) for i in range(0, N, B)) train_dset = Dataset(X_train, y_train, batch_size=64, shuffle=True) val_dset = Dataset(X_val, y_val, batch_size=64, shuffle=False) test_dset = Dataset(X_test, y_test, batch_size=64) # We can iterate through a dataset like this: for t, (x, y) in enumerate(train_dset): print(t, x.shape, y.shape) if t > 5: break ###Output 0 (64, 32, 32, 3) (64,) 1 (64, 32, 32, 3) (64,) 2 (64, 32, 32, 3) (64,) 3 (64, 32, 32, 3) (64,) 4 (64, 32, 32, 3) (64,) 5 (64, 32, 32, 3) (64,) 6 (64, 32, 32, 3) (64,) ###Markdown You can optionally **use GPU by setting the flag to True below**. It's not neccessary to use a GPU for this assignment; if you are working on Google Cloud then we recommend that you do not use a GPU, as it will be significantly more expensive. ###Code # Set up some global variables USE_GPU = True if USE_GPU: device = '/device:GPU:0' else: device = '/cpu:0' # Constant to control how often we print when training models print_every = 100 print('Using device: ', device) ###Output Using device: /device:GPU:0 ###Markdown Part II: Barebones TensorFlowTensorFlow ships with various high-level APIs which make it very convenient to define and train neural networks; we will cover some of these constructs in Part III and Part IV of this notebook. In this section we will start by building a model with basic TensorFlow constructs to help you better understand what's going on under the hood of the higher-level APIs.**"Barebones Tensorflow" is important to understanding the building blocks of TensorFlow, but much of it involves concepts from TensorFlow 1.x.** We will be working with legacy modules such as `tf.Variable`.Therefore, please read and understand the differences between legacy (1.x) TF and the new (2.0) TF. Historical background on TensorFlow 1.xTensorFlow 1.x is primarily a framework for working with **static computational graphs**. Nodes in the computational graph are Tensors which will hold n-dimensional arrays when the graph is run; edges in the graph represent functions that will operate on Tensors when the graph is run to actually perform useful computation.Before Tensorflow 2.0, we had to configure the graph into two phases. There are plenty of tutorials online that explain this two-step process. The process generally looks like the following for TF 1.x:1. **Build a computational graph that describes the computation that you want to perform**. This stage doesn't actually perform any computation; it just builds up a symbolic representation of your computation. This stage will typically define one or more `placeholder` objects that represent inputs to the computational graph.2. **Run the computational graph many times.** Each time the graph is run (e.g. for one gradient descent step) you will specify which parts of the graph you want to compute, and pass a `feed_dict` dictionary that will give concrete values to any `placeholder`s in the graph. The new paradigm in Tensorflow 2.0Now, with Tensorflow 2.0, we can simply adopt a functional form that is more Pythonic and similar in spirit to PyTorch and direct Numpy operation. Instead of the 2-step paradigm with computation graphs, making it (among other things) easier to debug TF code. You can read more details at https://www.tensorflow.org/guide/eager.The main difference between the TF 1.x and 2.0 approach is that the 2.0 approach doesn't make use of `tf.Session`, `tf.run`, `placeholder`, `feed_dict`. To get more details of what's different between the two version and how to convert between the two, check out the official migration guide: https://www.tensorflow.org/alpha/guide/migration_guideLater, in the rest of this notebook we'll focus on this new, simpler approach. TensorFlow warmup: Flatten FunctionWe can see this in action by defining a simple `flatten` function that will reshape image data for use in a fully-connected network.In TensorFlow, data for convolutional feature maps is typically stored in a Tensor of shape N x H x W x C where:- N is the number of datapoints (minibatch size)- H is the height of the feature map- W is the width of the feature map- C is the number of channels in the feature mapThis is the right way to represent the data when we are doing something like a 2D convolution, that needs spatial understanding of where the intermediate features are relative to each other. When we use fully connected affine layers to process the image, however, we want each datapoint to be represented by a single vector -- it's no longer useful to segregate the different channels, rows, and columns of the data. So, we use a "flatten" operation to collapse the `H x W x C` values per representation into a single long vector. Notice the `tf.reshape` call has the target shape as `(N, -1)`, meaning it will reshape/keep the first dimension to be N, and then infer as necessary what the second dimension is in the output, so we can collapse the remaining dimensions from the input properly.**NOTE**: TensorFlow and PyTorch differ on the default Tensor layout; TensorFlow uses N x H x W x C but PyTorch uses N x C x H x W. ###Code def flatten(x): """ Input: - TensorFlow Tensor of shape (N, D1, ..., DM) Output: - TensorFlow Tensor of shape (N, D1 * ... * DM) """ N = tf.shape(x)[0] return tf.reshape(x, (N, -1)) def test_flatten(): # Construct concrete values of the input data x using numpy x_np = np.arange(24).reshape((2, 3, 4)) print('x_np:\n', x_np, '\n') # Compute a concrete output value. x_flat_np = flatten(x_np) print('x_flat_np:\n', x_flat_np, '\n') test_flatten() ###Output x_np: [[[ 0 1 2 3] [ 4 5 6 7] [ 8 9 10 11]] [[12 13 14 15] [16 17 18 19] [20 21 22 23]]] x_flat_np: tf.Tensor( [[ 0 1 2 3 4 5 6 7 8 9 10 11] [12 13 14 15 16 17 18 19 20 21 22 23]], shape=(2, 12), dtype=int64) ###Markdown Barebones TensorFlow: Define a Two-Layer NetworkWe will now implement our first neural network with TensorFlow: a fully-connected ReLU network with two hidden layers and no biases on the CIFAR10 dataset. For now we will use only low-level TensorFlow operators to define the network; later we will see how to use the higher-level abstractions provided by `tf.keras` to simplify the process.We will define the forward pass of the network in the function `two_layer_fc`; this will accept TensorFlow Tensors for the inputs and weights of the network, and return a TensorFlow Tensor for the scores. After defining the network architecture in the `two_layer_fc` function, we will test the implementation by checking the shape of the output.**It's important that you read and understand this implementation.** ###Code def two_layer_fc(x, params): """ A fully-connected neural network; the architecture is: fully-connected layer -> ReLU -> fully connected layer. Note that we only need to define the forward pass here; TensorFlow will take care of computing the gradients for us. The input to the network will be a minibatch of data, of shape (N, d1, ..., dM) where d1 * ... * dM = D. The hidden layer will have H units, and the output layer will produce scores for C classes. Inputs: - x: A TensorFlow Tensor of shape (N, d1, ..., dM) giving a minibatch of input data. - params: A list [w1, w2] of TensorFlow Tensors giving weights for the network, where w1 has shape (D, H) and w2 has shape (H, C). Returns: - scores: A TensorFlow Tensor of shape (N, C) giving classification scores for the input data x. """ w1, w2 = params # Unpack the parameters x = flatten(x) # Flatten the input; now x has shape (N, D) h = tf.nn.relu(tf.matmul(x, w1)) # Hidden layer: h has shape (N, H) scores = tf.matmul(h, w2) # Compute scores of shape (N, C) return scores def two_layer_fc_test(): hidden_layer_size = 42 # Scoping our TF operations under a tf.device context manager # lets us tell TensorFlow where we want these Tensors to be # multiplied and/or operated on, e.g. on a CPU or a GPU. with tf.device(device): x = tf.zeros((64, 32, 32, 3)) w1 = tf.zeros((32 * 32 * 3, hidden_layer_size)) w2 = tf.zeros((hidden_layer_size, 10)) # Call our two_layer_fc function for the forward pass of the network. scores = two_layer_fc(x, [w1, w2]) print(scores.shape) two_layer_fc_test() ###Output (64, 10) ###Markdown Barebones TensorFlow: Three-Layer ConvNetHere you will complete the implementation of the function `three_layer_convnet` which will perform the forward pass of a three-layer convolutional network. The network should have the following architecture:1. A convolutional layer (with bias) with `channel_1` filters, each with shape `KW1 x KH1`, and zero-padding of two2. ReLU nonlinearity3. A convolutional layer (with bias) with `channel_2` filters, each with shape `KW2 x KH2`, and zero-padding of one4. ReLU nonlinearity5. Fully-connected layer with bias, producing scores for `C` classes.**HINT**: For convolutions: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/nn/conv2d; be careful with padding!**HINT**: For biases: https://www.tensorflow.org/performance/xla/broadcasting ###Code def three_layer_convnet(x, params): """ A three-layer convolutional network with the architecture described above. Inputs: - x: A TensorFlow Tensor of shape (N, H, W, 3) giving a minibatch of images - params: A list of TensorFlow Tensors giving the weights and biases for the network; should contain the following: - conv_w1: TensorFlow Tensor of shape (KH1, KW1, 3, channel_1) giving weights for the first convolutional layer. - conv_b1: TensorFlow Tensor of shape (channel_1,) giving biases for the first convolutional layer. - conv_w2: TensorFlow Tensor of shape (KH2, KW2, channel_1, channel_2) giving weights for the second convolutional layer - conv_b2: TensorFlow Tensor of shape (channel_2,) giving biases for the second convolutional layer. - fc_w: TensorFlow Tensor giving weights for the fully-connected layer. Can you figure out what the shape should be? - fc_b: TensorFlow Tensor giving biases for the fully-connected layer. Can you figure out what the shape should be? """ conv_w1, conv_b1, conv_w2, conv_b2, fc_w, fc_b = params scores = None ############################################################################ # TODO: Implement the forward pass for the three-layer ConvNet. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** with tf.device(device): c1 = tf.nn.conv2d(x, conv_w1, [1], "SAME") c1b = tf.add(c1, conv_b1) c1 = tf.nn.conv2d(x, conv_w1, [1], "SAME") c1b = tf.add(c1, conv_b1) c1r = tf.nn.relu(c1b) c2 = tf.nn.conv2d(c1r, conv_w2, [1], "SAME") c2b = tf.add(c2, conv_b2) c2r = tf.nn.relu(c2b) c2f = flatten(c2r) f = tf.matmul(c2f, fc_w) scores = tf.add(f, fc_b) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return scores ###Output _____no_output_____ ###Markdown After defing the forward pass of the three-layer ConvNet above, run the following cell to test your implementation. Like the two-layer network, we run the graph on a batch of zeros just to make sure the function doesn't crash, and produces outputs of the correct shape.When you run this function, `scores_np` should have shape `(64, 10)`. ###Code def three_layer_convnet_test(): with tf.device(device): x = tf.zeros((64, 32, 32, 3)) conv_w1 = tf.zeros((5, 5, 3, 6)) conv_b1 = tf.zeros((6,)) conv_w2 = tf.zeros((3, 3, 6, 9)) conv_b2 = tf.zeros((9,)) fc_w = tf.zeros((32 * 32 * 9, 10)) fc_b = tf.zeros((10,)) params = [conv_w1, conv_b1, conv_w2, conv_b2, fc_w, fc_b] scores = three_layer_convnet(x, params) # Inputs to convolutional layers are 4-dimensional arrays with shape # [batch_size, height, width, channels] print('scores_np has shape: ', scores.shape) three_layer_convnet_test() ###Output scores_np has shape: (64, 10) ###Markdown Barebones TensorFlow: Training StepWe now define the `training_step` function performs a single training step. This will take three basic steps:1. Compute the loss2. Compute the gradient of the loss with respect to all network weights3. Make a weight update step using (stochastic) gradient descent.We need to use a few new TensorFlow functions to do all of this:- For computing the cross-entropy loss we'll use `tf.nn.sparse_softmax_cross_entropy_with_logits`: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/nn/sparse_softmax_cross_entropy_with_logits- For averaging the loss across a minibatch of data we'll use `tf.reduce_mean`:https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/reduce_mean- For computing gradients of the loss with respect to the weights we'll use `tf.GradientTape` (useful for Eager execution): https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/GradientTape- We'll mutate the weight values stored in a TensorFlow Tensor using `tf.assign_sub` ("sub" is for subtraction): https://www.tensorflow.org/api_docs/python/tf/assign_sub ###Code def training_step(model_fn, x, y, params, learning_rate): with tf.GradientTape() as tape: scores = model_fn(x, params) # Forward pass of the model loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y, logits=scores) total_loss = tf.reduce_mean(loss) grad_params = tape.gradient(total_loss, params) # Make a vanilla gradient descent step on all of the model parameters # Manually update the weights using assign_sub() for w, grad_w in zip(params, grad_params): w.assign_sub(learning_rate * grad_w) return total_loss def train_part2(model_fn, init_fn, learning_rate): """ Train a model on CIFAR-10. Inputs: - model_fn: A Python function that performs the forward pass of the model using TensorFlow; it should have the following signature: scores = model_fn(x, params) where x is a TensorFlow Tensor giving a minibatch of image data, params is a list of TensorFlow Tensors holding the model weights, and scores is a TensorFlow Tensor of shape (N, C) giving scores for all elements of x. - init_fn: A Python function that initializes the parameters of the model. It should have the signature params = init_fn() where params is a list of TensorFlow Tensors holding the (randomly initialized) weights of the model. - learning_rate: Python float giving the learning rate to use for SGD. """ params = init_fn() # Initialize the model parameters for t, (x_np, y_np) in enumerate(train_dset): # Run the graph on a batch of training data. loss = training_step(model_fn, x_np, y_np, params, learning_rate) # Periodically print the loss and check accuracy on the val set. if t % print_every == 0: print('Iteration %d, loss = %.4f' % (t, loss)) check_accuracy(val_dset, x_np, model_fn, params) def check_accuracy(dset, x, model_fn, params): """ Check accuracy on a classification model, e.g. for validation. Inputs: - dset: A Dataset object against which to check accuracy - x: A TensorFlow placeholder Tensor where input images should be fed - model_fn: the Model we will be calling to make predictions on x - params: parameters for the model_fn to work with Returns: Nothing, but prints the accuracy of the model """ num_correct, num_samples = 0, 0 for x_batch, y_batch in dset: scores_np = model_fn(x_batch, params).numpy() y_pred = scores_np.argmax(axis=1) num_samples += x_batch.shape[0] num_correct += (y_pred == y_batch).sum() acc = float(num_correct) / num_samples print('Got %d / %d correct (%.2f%%)' % (num_correct, num_samples, 100 * acc)) ###Output _____no_output_____ ###Markdown Barebones TensorFlow: InitializationWe'll use the following utility method to initialize the weight matrices for our models using Kaiming's normalization method.[1] He et al, *Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification*, ICCV 2015, https://arxiv.org/abs/1502.01852 ###Code def create_matrix_with_kaiming_normal(shape): if len(shape) == 2: fan_in, fan_out = shape[0], shape[1] elif len(shape) == 4: fan_in, fan_out = np.prod(shape[:3]), shape[3] return tf.keras.backend.random_normal(shape) * np.sqrt(2.0 / fan_in) ###Output _____no_output_____ ###Markdown Barebones TensorFlow: Train a Two-Layer NetworkWe are finally ready to use all of the pieces defined above to train a two-layer fully-connected network on CIFAR-10.We just need to define a function to initialize the weights of the model, and call `train_part2`.Defining the weights of the network introduces another important piece of TensorFlow API: `tf.Variable`. A TensorFlow Variable is a Tensor whose value is stored in the graph and persists across runs of the computational graph; however unlike constants defined with `tf.zeros` or `tf.random_normal`, the values of a Variable can be mutated as the graph runs; these mutations will persist across graph runs. Learnable parameters of the network are usually stored in Variables.You don't need to tune any hyperparameters, but you should achieve validation accuracies above 40% after one epoch of training. ###Code def two_layer_fc_init(): """ Initialize the weights of a two-layer network, for use with the two_layer_network function defined above. You can use the `create_matrix_with_kaiming_normal` helper! Inputs: None Returns: A list of: - w1: TensorFlow tf.Variable giving the weights for the first layer - w2: TensorFlow tf.Variable giving the weights for the second layer """ hidden_layer_size = 4000 w1 = tf.Variable(create_matrix_with_kaiming_normal((3 * 32 * 32, 4000))) w2 = tf.Variable(create_matrix_with_kaiming_normal((4000, 10))) return [w1, w2] learning_rate = 1e-2 train_part2(two_layer_fc, two_layer_fc_init, learning_rate) ###Output Iteration 0, loss = 2.7718 Got 118 / 1000 correct (11.80%) Iteration 100, loss = 1.8125 Got 372 / 1000 correct (37.20%) Iteration 200, loss = 1.4477 Got 396 / 1000 correct (39.60%) Iteration 300, loss = 1.7683 Got 366 / 1000 correct (36.60%) Iteration 400, loss = 1.7126 Got 426 / 1000 correct (42.60%) Iteration 500, loss = 1.8523 Got 445 / 1000 correct (44.50%) Iteration 600, loss = 1.8816 Got 434 / 1000 correct (43.40%) Iteration 700, loss = 1.9445 Got 442 / 1000 correct (44.20%) ###Markdown Barebones TensorFlow: Train a three-layer ConvNetWe will now use TensorFlow to train a three-layer ConvNet on CIFAR-10.You need to implement the `three_layer_convnet_init` function. Recall that the architecture of the network is:1. Convolutional layer (with bias) with 32 5x5 filters, with zero-padding 22. ReLU3. Convolutional layer (with bias) with 16 3x3 filters, with zero-padding 14. ReLU5. Fully-connected layer (with bias) to compute scores for 10 classesYou don't need to do any hyperparameter tuning, but you should see validation accuracies above 43% after one epoch of training. ###Code def three_layer_convnet_init(): """ Initialize the weights of a Three-Layer ConvNet, for use with the three_layer_convnet function defined above. You can use the `create_matrix_with_kaiming_normal` helper! Inputs: None Returns a list containing: - conv_w1: TensorFlow tf.Variable giving weights for the first conv layer - conv_b1: TensorFlow tf.Variable giving biases for the first conv layer - conv_w2: TensorFlow tf.Variable giving weights for the second conv layer - conv_b2: TensorFlow tf.Variable giving biases for the second conv layer - fc_w: TensorFlow tf.Variable giving weights for the fully-connected layer - fc_b: TensorFlow tf.Variable giving biases for the fully-connected layer """ params = None ############################################################################ # TODO: Initialize the parameters of the three-layer network. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** conv_w1 = tf.Variable(create_matrix_with_kaiming_normal((5, 5, 3, 32))) conv_b1 = tf.Variable(tf.zeros((32))) conv_w2 = tf.Variable(create_matrix_with_kaiming_normal((3, 3, 32, 16))) conv_b2 = tf.Variable(tf.zeros((16))) fc_w = tf.Variable(create_matrix_with_kaiming_normal((32 * 32 * 16, 10))) fc_b = tf.Variable(tf.zeros((10))) params = (conv_w1, conv_b1, conv_w2, conv_b2, fc_w, fc_b) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return params learning_rate = 3e-3 train_part2(three_layer_convnet, three_layer_convnet_init, learning_rate) ###Output Iteration 0, loss = 2.7355 Got 107 / 1000 correct (10.70%) Iteration 100, loss = 1.8522 Got 361 / 1000 correct (36.10%) Iteration 200, loss = 1.4544 Got 407 / 1000 correct (40.70%) Iteration 300, loss = 1.7427 Got 416 / 1000 correct (41.60%) Iteration 400, loss = 1.6566 Got 450 / 1000 correct (45.00%) Iteration 500, loss = 1.5731 Got 458 / 1000 correct (45.80%) Iteration 600, loss = 1.5616 Got 467 / 1000 correct (46.70%) Iteration 700, loss = 1.6404 Got 503 / 1000 correct (50.30%) ###Markdown Part III: Keras Model Subclassing APIImplementing a neural network using the low-level TensorFlow API is a good way to understand how TensorFlow works, but it's a little inconvenient - we had to manually keep track of all Tensors holding learnable parameters. This was fine for a small network, but could quickly become unweildy for a large complex model.Fortunately TensorFlow 2.0 provides higher-level APIs such as `tf.keras` which make it easy to build models out of modular, object-oriented layers. Further, TensorFlow 2.0 uses eager execution that evaluates operations immediately, without explicitly constructing any computational graphs. This makes it easy to write and debug models, and reduces the boilerplate code.In this part of the notebook we will define neural network models using the `tf.keras.Model` API. To implement your own model, you need to do the following:1. Define a new class which subclasses `tf.keras.Model`. Give your class an intuitive name that describes it, like `TwoLayerFC` or `ThreeLayerConvNet`.2. In the initializer `__init__()` for your new class, define all the layers you need as class attributes. The `tf.keras.layers` package provides many common neural-network layers, like `tf.keras.layers.Dense` for fully-connected layers and `tf.keras.layers.Conv2D` for convolutional layers. Under the hood, these layers will construct `Variable` Tensors for any learnable parameters. **Warning**: Don't forget to call `super(YourModelName, self).__init__()` as the first line in your initializer!3. Implement the `call()` method for your class; this implements the forward pass of your model, and defines the *connectivity* of your network. Layers defined in `__init__()` implement `__call__()` so they can be used as function objects that transform input Tensors into output Tensors. Don't define any new layers in `call()`; any layers you want to use in the forward pass should be defined in `__init__()`.After you define your `tf.keras.Model` subclass, you can instantiate it and use it like the model functions from Part II. Keras Model Subclassing API: Two-Layer NetworkHere is a concrete example of using the `tf.keras.Model` API to define a two-layer network. There are a few new bits of API to be aware of here:We use an `Initializer` object to set up the initial values of the learnable parameters of the layers; in particular `tf.initializers.VarianceScaling` gives behavior similar to the Kaiming initialization method we used in Part II. You can read more about it here: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/initializers/VarianceScalingWe construct `tf.keras.layers.Dense` objects to represent the two fully-connected layers of the model. In addition to multiplying their input by a weight matrix and adding a bias vector, these layer can also apply a nonlinearity for you. For the first layer we specify a ReLU activation function by passing `activation='relu'` to the constructor; the second layer uses softmax activation function. Finally, we use `tf.keras.layers.Flatten` to flatten the output from the previous fully-connected layer. ###Code class TwoLayerFC(tf.keras.Model): def __init__(self, hidden_size, num_classes): super(TwoLayerFC, self).__init__() initializer = tf.initializers.VarianceScaling(scale=2.0) self.fc1 = tf.keras.layers.Dense(hidden_size, activation='relu', kernel_initializer=initializer) self.fc2 = tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer) self.flatten = tf.keras.layers.Flatten() def call(self, x, training=False): x = self.flatten(x) x = self.fc1(x) x = self.fc2(x) return x def test_TwoLayerFC(): """ A small unit test to exercise the TwoLayerFC model above. """ input_size, hidden_size, num_classes = 50, 42, 10 x = tf.zeros((64, input_size)) model = TwoLayerFC(hidden_size, num_classes) with tf.device(device): scores = model(x) print(scores.shape) test_TwoLayerFC() ###Output (64, 10) ###Markdown Keras Model Subclassing API: Three-Layer ConvNetNow it's your turn to implement a three-layer ConvNet using the `tf.keras.Model` API. Your model should have the same architecture used in Part II:1. Convolutional layer with 5 x 5 kernels, with zero-padding of 22. ReLU nonlinearity3. Convolutional layer with 3 x 3 kernels, with zero-padding of 14. ReLU nonlinearity5. Fully-connected layer to give class scores6. Softmax nonlinearityYou should initialize the weights of your network using the same initialization method as was used in the two-layer network above.**Hint**: Refer to the documentation for `tf.keras.layers.Conv2D` and `tf.keras.layers.Dense`:https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/Conv2Dhttps://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/Dense ###Code class ThreeLayerConvNet(tf.keras.Model): def __init__(self, channel_1, channel_2, num_classes): super(ThreeLayerConvNet, self).__init__() ######################################################################## # TODO: Implement the __init__ method for a three-layer ConvNet. You # # should instantiate layer objects to be used in the forward pass. # ######################################################################## # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** super(ThreeLayerConvNet, self).__init__() initializer = tf.initializers.VarianceScaling(scale=2.0) self.conv1 = tf.keras.layers.Conv2D(channel_1, (5, 5), (1,1), padding = 'same', activation='relu', kernel_initializer=initializer) self.conv2 = tf.keras.layers.Conv2D(channel_2, (3, 3), (1,1), padding = 'same', activation='relu', kernel_initializer=initializer) self.fc1 = tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer) self.flatten = tf.keras.layers.Flatten() # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ######################################################################## # END OF YOUR CODE # ######################################################################## def call(self, x, training=False): scores = None ######################################################################## # TODO: Implement the forward pass for a three-layer ConvNet. You # # should use the layer objects defined in the __init__ method. # ######################################################################## # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** x = self.conv1(x) x = self.conv2(x) x = self.flatten(x) x = self.fc1(x) return x # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ######################################################################## # END OF YOUR CODE # ######################################################################## return scores ###Output _____no_output_____ ###Markdown Once you complete the implementation of the `ThreeLayerConvNet` above you can run the following to ensure that your implementation does not crash and produces outputs of the expected shape. ###Code def test_ThreeLayerConvNet(): channel_1, channel_2, num_classes = 12, 8, 10 model = ThreeLayerConvNet(channel_1, channel_2, num_classes) with tf.device(device): x = tf.zeros((64, 3, 32, 32)) scores = model(x) print(scores.shape) test_ThreeLayerConvNet() ###Output (64, 10) ###Markdown Keras Model Subclassing API: Eager TrainingWhile keras models have a builtin training loop (using the `model.fit`), sometimes you need more customization. Here's an example, of a training loop implemented with eager execution.In particular, notice `tf.GradientTape`. Automatic differentiation is used in the backend for implementing backpropagation in frameworks like TensorFlow. During eager execution, `tf.GradientTape` is used to trace operations for computing gradients later. A particular `tf.GradientTape` can only compute one gradient; subsequent calls to tape will throw a runtime error. TensorFlow 2.0 ships with easy-to-use built-in metrics under `tf.keras.metrics` module. Each metric is an object, and we can use `update_state()` to add observations and `reset_state()` to clear all observations. We can get the current result of a metric by calling `result()` on the metric object. ###Code def train_part34(model_init_fn, optimizer_init_fn, num_epochs=1, is_training=False): """ Simple training loop for use with models defined using tf.keras. It trains a model for one epoch on the CIFAR-10 training set and periodically checks accuracy on the CIFAR-10 validation set. Inputs: - model_init_fn: A function that takes no parameters; when called it constructs the model we want to train: model = model_init_fn() - optimizer_init_fn: A function which takes no parameters; when called it constructs the Optimizer object we will use to optimize the model: optimizer = optimizer_init_fn() - num_epochs: The number of epochs to train for Returns: Nothing, but prints progress during trainingn """ with tf.device(device): # Compute the loss like we did in Part II loss_fn = tf.keras.losses.SparseCategoricalCrossentropy() model = model_init_fn() optimizer = optimizer_init_fn() train_loss = tf.keras.metrics.Mean(name='train_loss') train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy') val_loss = tf.keras.metrics.Mean(name='val_loss') val_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='val_accuracy') t = 0 for epoch in range(num_epochs): # Reset the metrics - https://www.tensorflow.org/alpha/guide/migration_guide#new-style_metrics train_loss.reset_states() train_accuracy.reset_states() for x_np, y_np in train_dset: with tf.GradientTape() as tape: # Use the model function to build the forward pass. scores = model(x_np, training=is_training) loss = loss_fn(y_np, scores) gradients = tape.gradient(loss, model.trainable_variables) optimizer.apply_gradients(zip(gradients, model.trainable_variables)) # Update the metrics train_loss.update_state(loss) train_accuracy.update_state(y_np, scores) if t % print_every == 0: val_loss.reset_states() val_accuracy.reset_states() for test_x, test_y in val_dset: # During validation at end of epoch, training set to False prediction = model(test_x, training=False) t_loss = loss_fn(test_y, prediction) val_loss.update_state(t_loss) val_accuracy.update_state(test_y, prediction) template = 'Iteration {}, Epoch {}, Loss: {}, Accuracy: {}, Val Loss: {}, Val Accuracy: {}' print (template.format(t, epoch+1, train_loss.result(), train_accuracy.result()*100, val_loss.result(), val_accuracy.result()*100)) t += 1 return model ###Output _____no_output_____ ###Markdown Keras Model Subclassing API: Train a Two-Layer NetworkWe can now use the tools defined above to train a two-layer network on CIFAR-10. We define the `model_init_fn` and `optimizer_init_fn` that construct the model and optimizer respectively when called. Here we want to train the model using stochastic gradient descent with no momentum, so we construct a `tf.keras.optimizers.SGD` function; you can [read about it here](https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/optimizers/SGD).You don't need to tune any hyperparameters here, but you should achieve validation accuracies above 40% after one epoch of training. ###Code hidden_size, num_classes = 4000, 10 learning_rate = 1e-2 def model_init_fn(): return TwoLayerFC(hidden_size, num_classes) def optimizer_init_fn(): return tf.keras.optimizers.SGD(learning_rate=learning_rate) train_part34(model_init_fn, optimizer_init_fn) ###Output Iteration 0, Epoch 1, Loss: 3.2483625411987305, Accuracy: 7.8125, Val Loss: 3.1048901081085205, Val Accuracy: 9.800000190734863 Iteration 100, Epoch 1, Loss: 2.2455251216888428, Accuracy: 29.161510467529297, Val Loss: 1.883112907409668, Val Accuracy: 37.900001525878906 Iteration 200, Epoch 1, Loss: 2.0907742977142334, Accuracy: 32.22947692871094, Val Loss: 1.8729472160339355, Val Accuracy: 38.80000305175781 Iteration 300, Epoch 1, Loss: 2.0080342292785645, Accuracy: 34.05315399169922, Val Loss: 1.9259060621261597, Val Accuracy: 37.20000076293945 Iteration 400, Epoch 1, Loss: 1.9369709491729736, Accuracy: 35.94529342651367, Val Loss: 1.7247862815856934, Val Accuracy: 42.099998474121094 Iteration 500, Epoch 1, Loss: 1.8901138305664062, Accuracy: 37.019710540771484, Val Loss: 1.6591670513153076, Val Accuracy: 42.20000076293945 Iteration 600, Epoch 1, Loss: 1.858939528465271, Accuracy: 37.879573822021484, Val Loss: 1.6988977193832397, Val Accuracy: 42.0 Iteration 700, Epoch 1, Loss: 1.8310751914978027, Accuracy: 38.61670684814453, Val Loss: 1.6381797790527344, Val Accuracy: 46.20000076293945 ###Markdown Keras Model Subclassing API: Train a Three-Layer ConvNetHere you should use the tools we've defined above to train a three-layer ConvNet on CIFAR-10. Your ConvNet should use 32 filters in the first convolutional layer and 16 filters in the second layer.To train the model you should use gradient descent with Nesterov momentum 0.9. **HINT**: https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/optimizers/SGDYou don't need to perform any hyperparameter tuning, but you should achieve validation accuracies above 50% after training for one epoch. ###Code learning_rate = 3e-3 channel_1, channel_2, num_classes = 32, 16, 10 def model_init_fn(): model = None ############################################################################ # TODO: Complete the implementation of model_fn. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** model = ThreeLayerConvNet(channel_1, channel_2, num_classes) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return model def optimizer_init_fn(): optimizer = None ############################################################################ # TODO: Complete the implementation of model_fn. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate, momentum = .9, nesterov = True) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return optimizer model34 = train_part34(model_init_fn, optimizer_init_fn) ###Output Iteration 0, Epoch 1, Loss: 2.8665101528167725, Accuracy: 7.8125, Val Loss: 3.5604350566864014, Val Accuracy: 9.300000190734863 Iteration 100, Epoch 1, Loss: 1.880402684211731, Accuracy: 34.26670837402344, Val Loss: 1.6414457559585571, Val Accuracy: 44.70000076293945 Iteration 200, Epoch 1, Loss: 1.7286632061004639, Accuracy: 39.42008590698242, Val Loss: 1.4922603368759155, Val Accuracy: 47.60000228881836 Iteration 300, Epoch 1, Loss: 1.6467211246490479, Accuracy: 42.08367919921875, Val Loss: 1.4390604496002197, Val Accuracy: 49.099998474121094 Iteration 400, Epoch 1, Loss: 1.579654574394226, Accuracy: 44.45137023925781, Val Loss: 1.3616528511047363, Val Accuracy: 52.20000076293945 Iteration 500, Epoch 1, Loss: 1.5306543111801147, Accuracy: 46.154563903808594, Val Loss: 1.3258088827133179, Val Accuracy: 54.79999923706055 Iteration 600, Epoch 1, Loss: 1.5000053644180298, Accuracy: 47.18177795410156, Val Loss: 1.281928539276123, Val Accuracy: 55.19999694824219 Iteration 700, Epoch 1, Loss: 1.472947597503662, Accuracy: 48.165565490722656, Val Loss: 1.2728898525238037, Val Accuracy: 57.099998474121094 ###Markdown Part IV: Keras Sequential APIIn Part III we introduced the `tf.keras.Model` API, which allows you to define models with any number of learnable layers and with arbitrary connectivity between layers.However for many models you don't need such flexibility - a lot of models can be expressed as a sequential stack of layers, with the output of each layer fed to the next layer as input. If your model fits this pattern, then there is an even easier way to define your model: using `tf.keras.Sequential`. You don't need to write any custom classes; you simply call the `tf.keras.Sequential` constructor with a list containing a sequence of layer objects.One complication with `tf.keras.Sequential` is that you must define the shape of the input to the model by passing a value to the `input_shape` of the first layer in your model. Keras Sequential API: Two-Layer NetworkIn this subsection, we will rewrite the two-layer fully-connected network using `tf.keras.Sequential`, and train it using the training loop defined above.You don't need to perform any hyperparameter tuning here, but you should see validation accuracies above 40% after training for one epoch. ###Code learning_rate = 1e-2 def model_init_fn(): input_shape = (32, 32, 3) hidden_layer_size, num_classes = 4000, 10 initializer = tf.initializers.VarianceScaling(scale=2.0) layers = [ tf.keras.layers.Flatten(input_shape=input_shape), tf.keras.layers.Dense(hidden_layer_size, activation='relu', kernel_initializer=initializer), tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer), ] model = tf.keras.Sequential(layers) return model def optimizer_init_fn(): return tf.keras.optimizers.SGD(learning_rate=learning_rate) train_part34(model_init_fn, optimizer_init_fn) ###Output Iteration 0, Epoch 1, Loss: 3.1283693313598633, Accuracy: 12.5, Val Loss: 2.907970428466797, Val Accuracy: 13.40000057220459 Iteration 100, Epoch 1, Loss: 2.2596824169158936, Accuracy: 27.61448097229004, Val Loss: 1.8809293508529663, Val Accuracy: 39.0 Iteration 200, Epoch 1, Loss: 2.090951681137085, Accuracy: 31.895212173461914, Val Loss: 1.8405581712722778, Val Accuracy: 38.80000305175781 Iteration 300, Epoch 1, Loss: 2.0104644298553467, Accuracy: 33.85589599609375, Val Loss: 1.901405930519104, Val Accuracy: 36.599998474121094 Iteration 400, Epoch 1, Loss: 1.9390699863433838, Accuracy: 35.83618927001953, Val Loss: 1.7299988269805908, Val Accuracy: 41.900001525878906 Iteration 500, Epoch 1, Loss: 1.8952276706695557, Accuracy: 36.860652923583984, Val Loss: 1.6673927307128906, Val Accuracy: 42.39999771118164 Iteration 600, Epoch 1, Loss: 1.8639533519744873, Accuracy: 37.689788818359375, Val Loss: 1.682585597038269, Val Accuracy: 42.79999923706055 Iteration 700, Epoch 1, Loss: 1.837199330329895, Accuracy: 38.35145950317383, Val Loss: 1.6344739198684692, Val Accuracy: 42.599998474121094 ###Markdown Abstracting Away the Training LoopIn the previous examples, we used a customised training loop to train models (e.g. `train_part34`). Writing your own training loop is only required if you need more flexibility and control during training your model. Alternately, you can also use built-in APIs like `tf.keras.Model.fit()` and `tf.keras.Model.evaluate` to train and evaluate a model. Also remember to configure your model for training by calling `tf.keras.Model.compile.You don't need to perform any hyperparameter tuning here, but you should see validation and test accuracies above 42% after training for one epoch. ###Code model = model_init_fn() #model.compile(optimizer=tf.keras.optimizers.SGD(learning_rate=learning_rate), model.compile(optimizer=tf.keras.optimizers.SGD(learning_rate=learning_rate, momentum = .9, nesterov = True), loss='sparse_categorical_crossentropy', metrics=[tf.keras.metrics.sparse_categorical_accuracy]) model.fit(X_train, y_train, batch_size=64, epochs=2, validation_data=(X_val, y_val)) model.evaluate(X_test, y_test) ###Output Train on 49000 samples, validate on 1000 samples Epoch 1/2 49000/49000 [==============================] - 7s 149us/sample - loss: 2.2677 - sparse_categorical_accuracy: 0.3791 - val_loss: 2.3114 - val_sparse_categorical_accuracy: 0.4140 Epoch 2/2 49000/49000 [==============================] - 7s 139us/sample - loss: 1.8612 - sparse_categorical_accuracy: 0.4630 - val_loss: 2.1996 - val_sparse_categorical_accuracy: 0.4240 10000/10000 [==============================] - 1s 118us/sample - loss: 2.1615 - sparse_categorical_accuracy: 0.4194 ###Markdown Keras Sequential API: Three-Layer ConvNetHere you should use `tf.keras.Sequential` to reimplement the same three-layer ConvNet architecture used in Part II and Part III. As a reminder, your model should have the following architecture:1. Convolutional layer with 32 5x5 kernels, using zero padding of 22. ReLU nonlinearity3. Convolutional layer with 16 3x3 kernels, using zero padding of 14. ReLU nonlinearity5. Fully-connected layer giving class scores6. Softmax nonlinearityYou should initialize the weights of the model using a `tf.initializers.VarianceScaling` as above.You should train the model using Nesterov momentum 0.9.You don't need to perform any hyperparameter search, but you should achieve accuracy above 45% after training for one epoch. ###Code def model_init_fn(): model = None ############################################################################ # TODO: Construct a three-layer ConvNet using tf.keras.Sequential. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** input_shape = (32, 32, 3) num_classes = 10 initializer = tf.initializers.VarianceScaling(scale=2.0) layers = [ tf.keras.layers.Conv2D(channel_1, (5, 5), (1,1), input_shape=input_shape, padding = 'same', activation='relu', kernel_initializer=initializer), tf.keras.layers.Conv2D(channel_2, (3, 3), (1,1), padding = 'same', activation='relu', kernel_initializer=initializer), tf.keras.layers.Flatten(), tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer) ] model = tf.keras.Sequential(layers) return model # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return model learning_rate = 5e-4 def optimizer_init_fn(): optimizer = None ############################################################################ # TODO: Complete the implementation of model_fn. # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate, momentum = .9, nesterov = True) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return optimizer model34 = train_part34(model_init_fn, optimizer_init_fn) ###Output Iteration 0, Epoch 1, Loss: 3.1202919483184814, Accuracy: 10.9375, Val Loss: 2.727996349334717, Val Accuracy: 10.300000190734863 Iteration 100, Epoch 1, Loss: 2.000793933868408, Accuracy: 30.136138916015625, Val Loss: 1.7997872829437256, Val Accuracy: 37.70000076293945 Iteration 200, Epoch 1, Loss: 1.8710850477218628, Accuracy: 34.95802307128906, Val Loss: 1.6573798656463623, Val Accuracy: 41.900001525878906 Iteration 300, Epoch 1, Loss: 1.7957477569580078, Accuracy: 37.2923583984375, Val Loss: 1.6201965808868408, Val Accuracy: 46.39999771118164 Iteration 400, Epoch 1, Loss: 1.7333133220672607, Accuracy: 39.4092903137207, Val Loss: 1.5384198427200317, Val Accuracy: 47.5 Iteration 500, Epoch 1, Loss: 1.6867010593414307, Accuracy: 40.90256881713867, Val Loss: 1.4892431497573853, Val Accuracy: 49.20000076293945 Iteration 600, Epoch 1, Loss: 1.6554102897644043, Accuracy: 41.997711181640625, Val Loss: 1.4652884006500244, Val Accuracy: 50.599998474121094 Iteration 700, Epoch 1, Loss: 1.6274172067642212, Accuracy: 43.056793212890625, Val Loss: 1.4231661558151245, Val Accuracy: 51.79999923706055 ###Markdown We will also train this model with the built-in training loop APIs provided by TensorFlow. ###Code model = model_init_fn() #model.compile(optimizer='sgd', model.compile(optimizer=tf.keras.optimizers.SGD(learning_rate=learning_rate), loss='sparse_categorical_crossentropy', metrics=[tf.keras.metrics.sparse_categorical_accuracy]) model.fit(X_train, y_train, batch_size=64, epochs=1, validation_data=(X_val, y_val)) model.evaluate(X_test, y_test) ###Output Train on 49000 samples, validate on 1000 samples 49000/49000 [==============================] - 7s 148us/sample - loss: 1.9793 - sparse_categorical_accuracy: 0.3033 - val_loss: 1.7637 - val_sparse_categorical_accuracy: 0.3940 10000/10000 [==============================] - 1s 108us/sample - loss: 1.7820 - sparse_categorical_accuracy: 0.3748 ###Markdown Part IV: Functional API Demonstration with a Two-Layer Network In the previous section, we saw how we can use `tf.keras.Sequential` to stack layers to quickly build simple models. But this comes at the cost of losing flexibility.Often we will have to write complex models that have non-sequential data flows: a layer can have **multiple inputs and/or outputs**, such as stacking the output of 2 previous layers together to feed as input to a third! (Some examples are residual connections and dense blocks.)In such cases, we can use Keras functional API to write models with complex topologies such as: 1. Multi-input models 2. Multi-output models 3. Models with shared layers (the same layer called several times) 4. Models with non-sequential data flows (e.g. residual connections)Writing a model with Functional API requires us to create a `tf.keras.Model` instance and explicitly write input tensors and output tensors for this model. ###Code def two_layer_fc_functional(input_shape, hidden_size, num_classes): initializer = tf.initializers.VarianceScaling(scale=2.0) inputs = tf.keras.Input(shape=input_shape) flattened_inputs = tf.keras.layers.Flatten()(inputs) fc1_output = tf.keras.layers.Dense(hidden_size, activation='relu', kernel_initializer=initializer)(flattened_inputs) scores = tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer)(fc1_output) # Instantiate the model given inputs and outputs. model = tf.keras.Model(inputs=inputs, outputs=scores) return model def test_two_layer_fc_functional(): """ A small unit test to exercise the TwoLayerFC model above. """ input_size, hidden_size, num_classes = 50, 42, 10 input_shape = (50,) x = tf.zeros((64, input_size)) model = two_layer_fc_functional(input_shape, hidden_size, num_classes) with tf.device(device): scores = model(x) print(scores.shape) test_two_layer_fc_functional() ###Output (64, 10) ###Markdown Keras Functional API: Train a Two-Layer NetworkYou can now train this two-layer network constructed using the functional API.You don't need to perform any hyperparameter tuning here, but you should see validation accuracies above 40% after training for one epoch. ###Code input_shape = (32, 32, 3) hidden_size, num_classes = 4000, 10 learning_rate = 1e-2 def model_init_fn(): return two_layer_fc_functional(input_shape, hidden_size, num_classes) def optimizer_init_fn(): return tf.keras.optimizers.SGD(learning_rate=learning_rate) train_part34(model_init_fn, optimizer_init_fn) ###Output Iteration 0, Epoch 1, Loss: 2.975888967514038, Accuracy: 9.375, Val Loss: 3.124917507171631, Val Accuracy: 11.0 Iteration 100, Epoch 1, Loss: 2.26572322845459, Accuracy: 28.078588485717773, Val Loss: 1.9556703567504883, Val Accuracy: 36.5 Iteration 200, Epoch 1, Loss: 2.0940442085266113, Accuracy: 32.03513717651367, Val Loss: 1.818089246749878, Val Accuracy: 39.70000076293945 Iteration 300, Epoch 1, Loss: 2.0151140689849854, Accuracy: 33.99086380004883, Val Loss: 1.9358595609664917, Val Accuracy: 36.70000076293945 Iteration 400, Epoch 1, Loss: 1.9442771673202515, Accuracy: 35.81281280517578, Val Loss: 1.7535817623138428, Val Accuracy: 41.10000228881836 Iteration 500, Epoch 1, Loss: 1.8974992036819458, Accuracy: 36.86376953125, Val Loss: 1.7121632099151611, Val Accuracy: 41.400001525878906 Iteration 600, Epoch 1, Loss: 1.8658170700073242, Accuracy: 37.80418014526367, Val Loss: 1.7267868518829346, Val Accuracy: 40.900001525878906 Iteration 700, Epoch 1, Loss: 1.8383498191833496, Accuracy: 38.61001968383789, Val Loss: 1.6557785272598267, Val Accuracy: 42.0 ###Markdown Part V: CIFAR-10 open-ended challengeIn this section you can experiment with whatever ConvNet architecture you'd like on CIFAR-10.You should experiment with architectures, hyperparameters, loss functions, regularization, or anything else you can think of to train a model that achieves **at least 70%** accuracy on the **validation** set within 10 epochs. You can use the built-in train function, the `train_part34` function from above, or implement your own training loop.Describe what you did at the end of the notebook. Some things you can try:- **Filter size**: Above we used 5x5 and 3x3; is this optimal?- **Number of filters**: Above we used 16 and 32 filters. Would more or fewer do better?- **Pooling**: We didn't use any pooling above. Would this improve the model?- **Normalization**: Would your model be improved with batch normalization, layer normalization, group normalization, or some other normalization strategy?- **Network architecture**: The ConvNet above has only three layers of trainable parameters. Would a deeper model do better?- **Global average pooling**: Instead of flattening after the final convolutional layer, would global average pooling do better? This strategy is used for example in Google's Inception network and in Residual Networks.- **Regularization**: Would some kind of regularization improve performance? Maybe weight decay or dropout? NOTE: Batch Normalization / DropoutIf you are using Batch Normalization and Dropout, remember to pass `is_training=True` if you use the `train_part34()` function. BatchNorm and Dropout layers have different behaviors at training and inference time. `training` is a specific keyword argument reserved for this purpose in any `tf.keras.Model`'s `call()` function. Read more about this here : https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/BatchNormalizationmethodshttps://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers/Dropoutmethods Tips for trainingFor each network architecture that you try, you should tune the learning rate and other hyperparameters. When doing this there are a couple important things to keep in mind: - If the parameters are working well, you should see improvement within a few hundred iterations- Remember the coarse-to-fine approach for hyperparameter tuning: start by testing a large range of hyperparameters for just a few training iterations to find the combinations of parameters that are working at all.- Once you have found some sets of parameters that seem to work, search more finely around these parameters. You may need to train for more epochs.- You should use the validation set for hyperparameter search, and save your test set for evaluating your architecture on the best parameters as selected by the validation set. Going above and beyondIf you are feeling adventurous there are many other features you can implement to try and improve your performance. You are **not required** to implement any of these, but don't miss the fun if you have time!- Alternative optimizers: you can try Adam, Adagrad, RMSprop, etc.- Alternative activation functions such as leaky ReLU, parametric ReLU, ELU, or MaxOut.- Model ensembles- Data augmentation- New Architectures - [ResNets](https://arxiv.org/abs/1512.03385) where the input from the previous layer is added to the output. - [DenseNets](https://arxiv.org/abs/1608.06993) where inputs into previous layers are concatenated together. - [This blog has an in-depth overview](https://chatbotslife.com/resnets-highwaynets-and-densenets-oh-my-9bb15918ee32) Have fun and happy training! ###Code class CustomConvNet(tf.keras.Model): def __init__(self): super(CustomConvNet, self).__init__() ############################################################################ # TODO: Construct a model that performs well on CIFAR-10 # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** reg_strength = 0.01 # Use scaling of 2 because of RELU initializer = tf.initializers.VarianceScaling(scale=2.0) self.dropout0 = tf.keras.layers.Dropout(.5) self.conv1 = tf.keras.layers.Conv2D(channel_1, (5, 5), (1,1), padding = 'same', #activation='none', kernel_initializer=initializer, kernel_regularizer= tf.keras.regularizers.l2(reg_strength), bias_regularizer=tf.keras.regularizers.l2(reg_strength)) self.norm1 = tf.keras.layers.BatchNormalization(axis=-1) self.active1 = tf.keras.layers.Activation("relu") self.pool1 = tf.keras.layers.MaxPool2D() self.conv2 = tf.keras.layers.Conv2D(channel_2, (3, 3), (1,1), padding = 'same', #activation='relu', kernel_initializer=initializer, kernel_regularizer= tf.keras.regularizers.l2(reg_strength), bias_regularizer=tf.keras.regularizers.l2(reg_strength)) self.norm2 = tf.keras.layers.BatchNormalization(axis=-1) self.active2 = tf.keras.layers.Activation("relu") self.pool2 = tf.keras.layers.MaxPool2D() self.conv3 = tf.keras.layers.Conv2D(channel_2, (3, 3), (1,1), padding = 'same', #activation='relu', kernel_initializer=initializer, kernel_regularizer= tf.keras.regularizers.l2(reg_strength), bias_regularizer=tf.keras.regularizers.l2(reg_strength)) self.norm3 = tf.keras.layers.BatchNormalization(axis=-1) self.active3 = tf.keras.layers.Activation("relu") self.pool3 = tf.keras.layers.MaxPool2D() self.flatten0 = tf.keras.layers.Flatten() self.fc0 = tf.keras.layers.Dense(100, activation='relu', kernel_initializer=initializer, kernel_regularizer= tf.keras.regularizers.l2(reg_strength), bias_regularizer=tf.keras.regularizers.l2(reg_strength)) self.fc1 = tf.keras.layers.Dense(num_classes, activation='softmax', kernel_initializer=initializer, kernel_regularizer= tf.keras.regularizers.l2(reg_strength), bias_regularizer=tf.keras.regularizers.l2(reg_strength)) # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ def call(self, input_tensor, training=False): ############################################################################ # TODO: Construct a model that performs well on CIFAR-10 # ############################################################################ # *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** x = input_tensor x = self.conv1(x) x = self.norm1(x, training = training) x = self.active1(x) x = self.pool1(x) x = self.conv2(x) x = self.norm2(x, training = training) x = self.active2(x) x = self.pool2(x) x = self.conv3(x) x = self.norm3(x, training = training) x = self.active3(x) x = self.pool3(x) x = self.flatten0(x) x = self.fc0(x) x = self.fc1(x) return x # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)***** ############################################################################ # END OF YOUR CODE # ############################################################################ return x device = '/device:GPU:0' # Change this to a CPU/GPU as you wish! # device = '/cpu:0' # Change this to a CPU/GPU as you wish! print_every = 700 num_epochs = 10 model = CustomConvNet() def model_init_fn(): return CustomConvNet() def optimizer_init_fn(): learning_rate = 1e-3 #return tf.keras.optimizers.SGD(learning_rate=learning_rate, momentum = .9, nesterov = True) return tf.keras.optimizers.Adam(learning_rate) train_part34(model_init_fn, optimizer_init_fn, num_epochs=num_epochs, is_training=True) ###Output Iteration 0, Epoch 1, Loss: 3.0353410243988037, Accuracy: 14.0625, Val Loss: 4.090989112854004, Val Accuracy: 12.800000190734863 Iteration 700, Epoch 1, Loss: 1.503588318824768, Accuracy: 45.62678527832031, Val Loss: 1.31296706199646, Val Accuracy: 52.499996185302734 Iteration 1400, Epoch 2, Loss: 1.1449296474456787, Accuracy: 59.17322540283203, Val Loss: 1.1885653734207153, Val Accuracy: 58.499996185302734 Iteration 2100, Epoch 3, Loss: 1.012212872505188, Accuracy: 64.03229522705078, Val Loss: 1.0230681896209717, Val Accuracy: 63.900001525878906 Iteration 2800, Epoch 4, Loss: 0.9305073022842407, Accuracy: 67.0290756225586, Val Loss: 0.9436802268028259, Val Accuracy: 66.5 Iteration 3500, Epoch 5, Loss: 0.8777076005935669, Accuracy: 68.9716796875, Val Loss: 0.9642516374588013, Val Accuracy: 66.19999694824219
1A.ipynb
###Markdown Face RecognitionIn this assignment, you will build a face recognition system. Many of the ideas presented here are from [FaceNet](https://arxiv.org/pdf/1503.03832.pdf). In lecture, we also talked about [DeepFace](https://research.fb.com/wp-content/uploads/2016/11/deepface-closing-the-gap-to-human-level-performance-in-face-verification.pdf). Face recognition problems commonly fall into two categories: - **Face Verification** - "is this the claimed person?". For example, at some airports, you can pass through customs by letting a system scan your passport and then verifying that you (the person carrying the passport) are the correct person. A mobile phone that unlocks using your face is also using face verification. This is a 1:1 matching problem. - **Face Recognition** - "who is this person?". For example, the video lecture showed a [face recognition video](https://www.youtube.com/watch?v=wr4rx0Spihs) of Baidu employees entering the office without needing to otherwise identify themselves. This is a 1:K matching problem. FaceNet learns a neural network that encodes a face image into a vector of 128 numbers. By comparing two such vectors, you can then determine if two pictures are of the same person. **In this assignment, you will:**- Implement the triplet loss function- Use a pretrained model to map face images into 128-dimensional encodings- Use these encodings to perform face verification and face recognition Channels-first notation* In this exercise, we will be using a pre-trained model which represents ConvNet activations using a **"channels first"** convention, as opposed to the "channels last" convention used in lecture and previous programming assignments. * In other words, a batch of images will be of shape $(m, n_C, n_H, n_W)$ instead of $(m, n_H, n_W, n_C)$. * Both of these conventions have a reasonable amount of traction among open-source implementations; there isn't a uniform standard yet within the deep learning community. Updates If you were working on the notebook before this update...* The current notebook is version "3a".* You can find your original work saved in the notebook with the previous version name ("v3") * To view the file directory, go to the menu "File->Open", and this will open a new tab that shows the file directory. List of updates* `triplet_loss`: Additional Hints added.* `verify`: Hints added.* `who_is_it`: corrected hints given in the comments.* Spelling and formatting updates for easier reading. Load packagesLet's load the required packages. ###Code from keras.models import Sequential from keras.layers import Conv2D, ZeroPadding2D, Activation, Input, concatenate from keras.models import Model from keras.layers.normalization import BatchNormalization from keras.layers.pooling import MaxPooling2D, AveragePooling2D from keras.layers.merge import Concatenate from keras.layers.core import Lambda, Flatten, Dense from keras.initializers import glorot_uniform from keras.engine.topology import Layer from keras import backend as K K.set_image_data_format('channels_first') import cv2 import os import numpy as np from numpy import genfromtxt import pandas as pd import tensorflow as tf from fr_utils import * from inception_blocks_v2 import * %matplotlib inline %load_ext autoreload %autoreload 2 np.set_printoptions(threshold=np.nan) ###Output Using TensorFlow backend. ###Markdown 0 - Naive Face VerificationIn Face Verification, you're given two images and you have to determine if they are of the same person. The simplest way to do this is to compare the two images pixel-by-pixel. If the distance between the raw images are less than a chosen threshold, it may be the same person! **Figure 1** * Of course, this algorithm performs really poorly, since the pixel values change dramatically due to variations in lighting, orientation of the person's face, even minor changes in head position, and so on. * You'll see that rather than using the raw image, you can learn an encoding, $f(img)$. * By using an encoding for each image, an element-wise comparison produces a more accurate judgement as to whether two pictures are of the same person. 1 - Encoding face images into a 128-dimensional vector 1.1 - Using a ConvNet to compute encodingsThe FaceNet model takes a lot of data and a long time to train. So following common practice in applied deep learning, let's load weights that someone else has already trained. The network architecture follows the Inception model from [Szegedy *et al.*](https://arxiv.org/abs/1409.4842). We have provided an inception network implementation. You can look in the file `inception_blocks_v2.py` to see how it is implemented (do so by going to "File->Open..." at the top of the Jupyter notebook. This opens the file directory that contains the '.py' file). The key things you need to know are:- This network uses 96x96 dimensional RGB images as its input. Specifically, inputs a face image (or batch of $m$ face images) as a tensor of shape $(m, n_C, n_H, n_W) = (m, 3, 96, 96)$ - It outputs a matrix of shape $(m, 128)$ that encodes each input face image into a 128-dimensional vectorRun the cell below to create the model for face images. ###Code FRmodel = faceRecoModel(input_shape=(3, 96, 96)) print("Total Params:", FRmodel.count_params()) ###Output Total Params: 3743280 ###Markdown ** Expected Output **Total Params: 3743280 By using a 128-neuron fully connected layer as its last layer, the model ensures that the output is an encoding vector of size 128. You then use the encodings to compare two face images as follows: **Figure 2**: By computing the distance between two encodings and thresholding, you can determine if the two pictures represent the same personSo, an encoding is a good one if: - The encodings of two images of the same person are quite similar to each other. - The encodings of two images of different persons are very different.The triplet loss function formalizes this, and tries to "push" the encodings of two images of the same person (Anchor and Positive) closer together, while "pulling" the encodings of two images of different persons (Anchor, Negative) further apart. **Figure 3**: In the next part, we will call the pictures from left to right: Anchor (A), Positive (P), Negative (N) 1.2 - The Triplet LossFor an image $x$, we denote its encoding $f(x)$, where $f$ is the function computed by the neural network.<!--We will also add a normalization step at the end of our model so that $\mid \mid f(x) \mid \mid_2 = 1$ (means the vector of encoding should be of norm 1).!-->Training will use triplets of images $(A, P, N)$: - A is an "Anchor" image--a picture of a person. - P is a "Positive" image--a picture of the same person as the Anchor image.- N is a "Negative" image--a picture of a different person than the Anchor image.These triplets are picked from our training dataset. We will write $(A^{(i)}, P^{(i)}, N^{(i)})$ to denote the $i$-th training example. You'd like to make sure that an image $A^{(i)}$ of an individual is closer to the Positive $P^{(i)}$ than to the Negative image $N^{(i)}$) by at least a margin $\alpha$:$$\mid \mid f(A^{(i)}) - f(P^{(i)}) \mid \mid_2^2 + \alpha < \mid \mid f(A^{(i)}) - f(N^{(i)}) \mid \mid_2^2$$You would thus like to minimize the following "triplet cost":$$\mathcal{J} = \sum^{m}_{i=1} \large[ \small \underbrace{\mid \mid f(A^{(i)}) - f(P^{(i)}) \mid \mid_2^2}_\text{(1)} - \underbrace{\mid \mid f(A^{(i)}) - f(N^{(i)}) \mid \mid_2^2}_\text{(2)} + \alpha \large ] \small_+ \tag{3}$$Here, we are using the notation "$[z]_+$" to denote $max(z,0)$. Notes:- The term (1) is the squared distance between the anchor "A" and the positive "P" for a given triplet; you want this to be small. - The term (2) is the squared distance between the anchor "A" and the negative "N" for a given triplet, you want this to be relatively large. It has a minus sign preceding it because minimizing the negative of the term is the same as maximizing that term.- $\alpha$ is called the margin. It is a hyperparameter that you pick manually. We will use $\alpha = 0.2$. Most implementations also rescale the encoding vectors to haven L2 norm equal to one (i.e., $\mid \mid f(img)\mid \mid_2$=1); you won't have to worry about that in this assignment.**Exercise**: Implement the triplet loss as defined by formula (3). Here are the 4 steps:1. Compute the distance between the encodings of "anchor" and "positive": $\mid \mid f(A^{(i)}) - f(P^{(i)}) \mid \mid_2^2$2. Compute the distance between the encodings of "anchor" and "negative": $\mid \mid f(A^{(i)}) - f(N^{(i)}) \mid \mid_2^2$3. Compute the formula per training example: $ \mid \mid f(A^{(i)}) - f(P^{(i)}) \mid \mid_2^2 - \mid \mid f(A^{(i)}) - f(N^{(i)}) \mid \mid_2^2 + \alpha$3. Compute the full formula by taking the max with zero and summing over the training examples:$$\mathcal{J} = \sum^{m}_{i=1} \large[ \small \mid \mid f(A^{(i)}) - f(P^{(i)}) \mid \mid_2^2 - \mid \mid f(A^{(i)}) - f(N^{(i)}) \mid \mid_2^2+ \alpha \large ] \small_+ \tag{3}$$ Hints* Useful functions: `tf.reduce_sum()`, `tf.square()`, `tf.subtract()`, `tf.add()`, `tf.maximum()`.* For steps 1 and 2, you will sum over the entries of $\mid \mid f(A^{(i)}) - f(P^{(i)}) \mid \mid_2^2$ and $\mid \mid f(A^{(i)}) - f(N^{(i)}) \mid \mid_2^2$. * For step 4 you will sum over the training examples. Additional Hints* Recall that the square of the L2 norm is the sum of the squared differences: $||x - y||_{2}^{2} = \sum_{i=1}^{N}(x_{i} - y_{i})^{2}$* Note that the `anchor`, `positive` and `negative` encodings are of shape `(m,128)`, where m is the number of training examples and 128 is the number of elements used to encode a single example.* For steps 1 and 2, you will maintain the number of `m` training examples and sum along the 128 values of each encoding. [tf.reduce_sum](https://www.tensorflow.org/api_docs/python/tf/math/reduce_sum) has an `axis` parameter. This chooses along which axis the sums are applied. * Note that one way to choose the last axis in a tensor is to use negative indexing (`axis=-1`).* In step 4, when summing over training examples, the result will be a single scalar value.* For `tf.reduce_sum` to sum across all axes, keep the default value `axis=None`. ###Code # GRADED FUNCTION: triplet_loss def triplet_loss(y_true, y_pred, alpha = 0.2): """ Implementation of the triplet loss as defined by formula (3) Arguments: y_true -- true labels, required when you define a loss in Keras, you don't need it in this function. y_pred -- python list containing three objects: anchor -- the encodings for the anchor images, of shape (None, 128) positive -- the encodings for the positive images, of shape (None, 128) negative -- the encodings for the negative images, of shape (None, 128) Returns: loss -- real number, value of the loss """ anchor, positive, negative = y_pred[0], y_pred[1], y_pred[2] ### START CODE HERE ### (≈ 4 lines) # Step 1: Compute the (encoding) distance between the anchor and the positive pos_dist = tf.reduce_sum((anchor-positive)**2,axis=-1) # Step 2: Compute the (encoding) distance between the anchor and the negative neg_dist = tf.reduce_sum((anchor-negative)**2,axis=-1) # Step 3: subtract the two previous distances and add alpha. basic_loss = pos_dist+alpha-neg_dist # Step 4: Take the maximum of basic_loss and 0.0. Sum over the training examples. loss = tf.reduce_sum(tf.maximum(basic_loss,0)) ### END CODE HERE ### return loss with tf.Session() as test: tf.set_random_seed(1) y_true = (None, None, None) y_pred = (tf.random_normal([3, 128], mean=6, stddev=0.1, seed = 1), tf.random_normal([3, 128], mean=1, stddev=1, seed = 1), tf.random_normal([3, 128], mean=3, stddev=4, seed = 1)) loss = triplet_loss(y_true, y_pred) print("loss = " + str(loss.eval())) ###Output loss = 528.143 ###Markdown **Expected Output**: **loss** 528.143 2 - Loading the pre-trained modelFaceNet is trained by minimizing the triplet loss. But since training requires a lot of data and a lot of computation, we won't train it from scratch here. Instead, we load a previously trained model. Load a model using the following cell; this might take a couple of minutes to run. ###Code FRmodel.compile(optimizer = 'adam', loss = triplet_loss, metrics = ['accuracy']) load_weights_from_FaceNet(FRmodel) ###Output _____no_output_____ ###Markdown Here are some examples of distances between the encodings between three individuals: **Figure 4**: Example of distance outputs between three individuals' encodingsLet's now use this model to perform face verification and face recognition! 3 - Applying the model You are building a system for an office building where the building manager would like to offer facial recognition to allow the employees to enter the building.You'd like to build a **Face verification** system that gives access to the list of people who live or work there. To get admitted, each person has to swipe an ID card (identification card) to identify themselves at the entrance. The face recognition system then checks that they are who they claim to be. 3.1 - Face VerificationLet's build a database containing one encoding vector for each person who is allowed to enter the office. To generate the encoding we use `img_to_encoding(image_path, model)`, which runs the forward propagation of the model on the specified image. Run the following code to build the database (represented as a python dictionary). This database maps each person's name to a 128-dimensional encoding of their face. ###Code database = {} database["danielle"] = img_to_encoding("images/danielle.png", FRmodel) database["younes"] = img_to_encoding("images/younes.jpg", FRmodel) database["tian"] = img_to_encoding("images/tian.jpg", FRmodel) database["andrew"] = img_to_encoding("images/andrew.jpg", FRmodel) database["kian"] = img_to_encoding("images/kian.jpg", FRmodel) database["dan"] = img_to_encoding("images/dan.jpg", FRmodel) database["sebastiano"] = img_to_encoding("images/sebastiano.jpg", FRmodel) database["bertrand"] = img_to_encoding("images/bertrand.jpg", FRmodel) database["kevin"] = img_to_encoding("images/kevin.jpg", FRmodel) database["felix"] = img_to_encoding("images/felix.jpg", FRmodel) database["benoit"] = img_to_encoding("images/benoit.jpg", FRmodel) database["arnaud"] = img_to_encoding("images/arnaud.jpg", FRmodel) ###Output _____no_output_____ ###Markdown Now, when someone shows up at your front door and swipes their ID card (thus giving you their name), you can look up their encoding in the database, and use it to check if the person standing at the front door matches the name on the ID.**Exercise**: Implement the verify() function which checks if the front-door camera picture (`image_path`) is actually the person called "identity". You will have to go through the following steps:1. Compute the encoding of the image from `image_path`.2. Compute the distance between this encoding and the encoding of the identity image stored in the database.3. Open the door if the distance is less than 0.7, else do not open it.* As presented above, you should use the L2 distance [np.linalg.norm](https://docs.scipy.org/doc/numpy/reference/generated/numpy.linalg.norm.html). * (Note: In this implementation, compare the L2 distance, not the square of the L2 distance, to the threshold 0.7.) Hints* `identity` is a string that is also a key in the `database` dictionary.* `img_to_encoding` has two parameters: the `image_path` and `model`. ###Code # GRADED FUNCTION: verify def verify(image_path, identity, database, model): """ Function that verifies if the person on the "image_path" image is "identity". Arguments: image_path -- path to an image identity -- string, name of the person you'd like to verify the identity. Has to be an employee who works in the office. database -- python dictionary mapping names of allowed people's names (strings) to their encodings (vectors). model -- your Inception model instance in Keras Returns: dist -- distance between the image_path and the image of "identity" in the database. door_open -- True, if the door should open. False otherwise. """ ### START CODE HERE ### # Step 1: Compute the encoding for the image. Use img_to_encoding() see example above. (≈ 1 line) encoding = img_to_encoding(image_path,model) # Step 2: Compute distance with identity's image (≈ 1 line) dist = np.linalg.norm(encoding-database[identity]) # Step 3: Open the door if dist < 0.7, else don't open (≈ 3 lines) if dist<0.7: print("It's " + str(identity) + ", welcome in!") door_open = True else: print("It's not " + str(identity) + ", please go away") door_open = False ### END CODE HERE ### return dist, door_open ###Output _____no_output_____ ###Markdown Younes is trying to enter the office and the camera takes a picture of him ("images/camera_0.jpg"). Let's run your verification algorithm on this picture: ###Code verify("images/camera_0.jpg", "younes", database, FRmodel) ###Output It's younes, welcome in! ###Markdown **Expected Output**: **It's younes, welcome in!** (0.65939283, True) Benoit, who does not work in the office, stole Kian's ID card and tried to enter the office. The camera took a picture of Benoit ("images/camera_2.jpg). Let's run the verification algorithm to check if benoit can enter. ###Code verify("images/camera_2.jpg", "kian", database, FRmodel) ###Output It's not kian, please go away ###Markdown **Expected Output**: **It's not kian, please go away** (0.86224014, False) 3.2 - Face RecognitionYour face verification system is mostly working well. But since Kian got his ID card stolen, when he came back to the office the next day and couldn't get in! To solve this, you'd like to change your face verification system to a face recognition system. This way, no one has to carry an ID card anymore. An authorized person can just walk up to the building, and the door will unlock for them! You'll implement a face recognition system that takes as input an image, and figures out if it is one of the authorized persons (and if so, who). Unlike the previous face verification system, we will no longer get a person's name as one of the inputs. **Exercise**: Implement `who_is_it()`. You will have to go through the following steps:1. Compute the target encoding of the image from image_path2. Find the encoding from the database that has smallest distance with the target encoding. - Initialize the `min_dist` variable to a large enough number (100). It will help you keep track of what is the closest encoding to the input's encoding. - Loop over the database dictionary's names and encodings. To loop use `for (name, db_enc) in database.items()`. - Compute the L2 distance between the target "encoding" and the current "encoding" from the database. - If this distance is less than the min_dist, then set `min_dist` to `dist`, and `identity` to `name`. ###Code # GRADED FUNCTION: who_is_it def who_is_it(image_path, database, model): """ Implements face recognition for the office by finding who is the person on the image_path image. Arguments: image_path -- path to an image database -- database containing image encodings along with the name of the person on the image model -- your Inception model instance in Keras Returns: min_dist -- the minimum distance between image_path encoding and the encodings from the database identity -- string, the name prediction for the person on image_path """ ### START CODE HERE ### ## Step 1: Compute the target "encoding" for the image. Use img_to_encoding() see example above. ## (≈ 1 line) encoding = img_to_encoding(image_path,model) ## Step 2: Find the closest encoding ## # Initialize "min_dist" to a large value, say 100 (≈1 line) min_dist = 100 # Loop over the database dictionary's names and encodings. for (name, db_enc) in database.items(): # Compute L2 distance between the target "encoding" and the current db_enc from the database. (≈ 1 line) dist = np.linalg.norm(encoding-db_enc) # If this distance is less than the min_dist, then set min_dist to dist, and identity to name. (≈ 3 lines) if dist<min_dist: min_dist = dist identity = name ### END CODE HERE ### if min_dist > 0.7: print("Not in the database.") else: print ("it's " + str(identity) + ", the distance is " + str(min_dist)) return min_dist, identity ###Output _____no_output_____ ###Markdown Younes is at the front-door and the camera takes a picture of him ("images/camera_0.jpg"). Let's see if your who_it_is() algorithm identifies Younes. ###Code who_is_it("images/camera_0.jpg", database, FRmodel) ###Output it's younes, the distance is 0.659393 ###Markdown Autonomous driving - Car detectionWelcome to your week 3 programming assignment. You will learn about object detection using the very powerful YOLO model. Many of the ideas in this notebook are described in the two YOLO papers: [Redmon et al., 2016](https://arxiv.org/abs/1506.02640) and [Redmon and Farhadi, 2016](https://arxiv.org/abs/1612.08242). **You will learn to**:- Use object detection on a car detection dataset- Deal with bounding boxes Updates If you were working on the notebook before this update...* The current notebook is version "3a".* You can find your original work saved in the notebook with the previous version name ("v3") * To view the file directory, go to the menu "File->Open", and this will open a new tab that shows the file directory. List of updates* Clarified "YOLO" instructions preceding the code. * Added details about anchor boxes.* Added explanation of how score is calculated.* `yolo_filter_boxes`: added additional hints. Clarify syntax for argmax and max.* `iou`: clarify instructions for finding the intersection.* `iou`: give variable names for all 8 box vertices, for clarity. Adds `width` and `height` variables for clarity.* `iou`: add test cases to check handling of non-intersecting boxes, intersection at vertices, or intersection at edges.* `yolo_non_max_suppression`: clarify syntax for tf.image.non_max_suppression and keras.gather.* "convert output of the model to usable bounding box tensors": Provides a link to the definition of `yolo_head`.* `predict`: hint on calling sess.run.* Spelling, grammar, wording and formatting updates to improve clarity. Import librariesRun the following cell to load the packages and dependencies that you will find useful as you build the object detector! ###Code import argparse import os import matplotlib.pyplot as plt from matplotlib.pyplot import imshow import scipy.io import scipy.misc import numpy as np import pandas as pd import PIL import tensorflow as tf from keras import backend as K from keras.layers import Input, Lambda, Conv2D from keras.models import load_model, Model from yolo_utils import read_classes, read_anchors, generate_colors, preprocess_image, draw_boxes, scale_boxes from yad2k.models.keras_yolo import yolo_head, yolo_boxes_to_corners, preprocess_true_boxes, yolo_loss, yolo_body %matplotlib inline ###Output Using TensorFlow backend. ###Markdown **Important Note**: As you can see, we import Keras's backend as K. This means that to use a Keras function in this notebook, you will need to write: `K.function(...)`. 1 - Problem StatementYou are working on a self-driving car. As a critical component of this project, you'd like to first build a car detection system. To collect data, you've mounted a camera to the hood (meaning the front) of the car, which takes pictures of the road ahead every few seconds while you drive around. Pictures taken from a car-mounted camera while driving around Silicon Valley. We thank [drive.ai](htps://www.drive.ai/) for providing this dataset.You've gathered all these images into a folder and have labelled them by drawing bounding boxes around every car you found. Here's an example of what your bounding boxes look like. **Figure 1** : **Definition of a box** If you have 80 classes that you want the object detector to recognize, you can represent the class label $c$ either as an integer from 1 to 80, or as an 80-dimensional vector (with 80 numbers) one component of which is 1 and the rest of which are 0. The video lectures had used the latter representation; in this notebook, we will use both representations, depending on which is more convenient for a particular step. In this exercise, you will learn how "You Only Look Once" (YOLO) performs object detection, and then apply it to car detection. Because the YOLO model is very computationally expensive to train, we will load pre-trained weights for you to use. 2 - YOLO "You Only Look Once" (YOLO) is a popular algorithm because it achieves high accuracy while also being able to run in real-time. This algorithm "only looks once" at the image in the sense that it requires only one forward propagation pass through the network to make predictions. After non-max suppression, it then outputs recognized objects together with the bounding boxes. 2.1 - Model details Inputs and outputs- The **input** is a batch of images, and each image has the shape (m, 608, 608, 3)- The **output** is a list of bounding boxes along with the recognized classes. Each bounding box is represented by 6 numbers $(p_c, b_x, b_y, b_h, b_w, c)$ as explained above. If you expand $c$ into an 80-dimensional vector, each bounding box is then represented by 85 numbers. Anchor Boxes* Anchor boxes are chosen by exploring the training data to choose reasonable height/width ratios that represent the different classes. For this assignment, 5 anchor boxes were chosen for you (to cover the 80 classes), and stored in the file './model_data/yolo_anchors.txt'* The dimension for anchor boxes is the second to last dimension in the encoding: $(m, n_H,n_W,anchors,classes)$.* The YOLO architecture is: IMAGE (m, 608, 608, 3) -> DEEP CNN -> ENCODING (m, 19, 19, 5, 85). EncodingLet's look in greater detail at what this encoding represents. **Figure 2** : **Encoding architecture for YOLO** If the center/midpoint of an object falls into a grid cell, that grid cell is responsible for detecting that object. Since we are using 5 anchor boxes, each of the 19 x19 cells thus encodes information about 5 boxes. Anchor boxes are defined only by their width and height.For simplicity, we will flatten the last two last dimensions of the shape (19, 19, 5, 85) encoding. So the output of the Deep CNN is (19, 19, 425). **Figure 3** : **Flattening the last two last dimensions** Class scoreNow, for each box (of each cell) we will compute the following element-wise product and extract a probability that the box contains a certain class. The class score is $score_{c,i} = p_{c} \times c_{i}$: the probability that there is an object $p_{c}$ times the probability that the object is a certain class $c_{i}$. **Figure 4** : **Find the class detected by each box** Example of figure 4* In figure 4, let's say for box 1 (cell 1), the probability that an object exists is $p_{1}=0.60$. So there's a 60% chance that an object exists in box 1 (cell 1). * The probability that the object is the class "category 3 (a car)" is $c_{3}=0.73$. * The score for box 1 and for category "3" is $score_{1,3}=0.60 \times 0.73 = 0.44$. * Let's say we calculate the score for all 80 classes in box 1, and find that the score for the car class (class 3) is the maximum. So we'll assign the score 0.44 and class "3" to this box "1". Visualizing classesHere's one way to visualize what YOLO is predicting on an image:- For each of the 19x19 grid cells, find the maximum of the probability scores (taking a max across the 80 classes, one maximum for each of the 5 anchor boxes).- Color that grid cell according to what object that grid cell considers the most likely.Doing this results in this picture: **Figure 5** : Each one of the 19x19 grid cells is colored according to which class has the largest predicted probability in that cell. Note that this visualization isn't a core part of the YOLO algorithm itself for making predictions; it's just a nice way of visualizing an intermediate result of the algorithm. Visualizing bounding boxesAnother way to visualize YOLO's output is to plot the bounding boxes that it outputs. Doing that results in a visualization like this: **Figure 6** : Each cell gives you 5 boxes. In total, the model predicts: 19x19x5 = 1805 boxes just by looking once at the image (one forward pass through the network)! Different colors denote different classes. Non-Max suppressionIn the figure above, we plotted only boxes for which the model had assigned a high probability, but this is still too many boxes. You'd like to reduce the algorithm's output to a much smaller number of detected objects. To do so, you'll use **non-max suppression**. Specifically, you'll carry out these steps: - Get rid of boxes with a low score (meaning, the box is not very confident about detecting a class; either due to the low probability of any object, or low probability of this particular class).- Select only one box when several boxes overlap with each other and detect the same object. 2.2 - Filtering with a threshold on class scoresYou are going to first apply a filter by thresholding. You would like to get rid of any box for which the class "score" is less than a chosen threshold. The model gives you a total of 19x19x5x85 numbers, with each box described by 85 numbers. It is convenient to rearrange the (19,19,5,85) (or (19,19,425)) dimensional tensor into the following variables: - `box_confidence`: tensor of shape $(19 \times 19, 5, 1)$ containing $p_c$ (confidence probability that there's some object) for each of the 5 boxes predicted in each of the 19x19 cells.- `boxes`: tensor of shape $(19 \times 19, 5, 4)$ containing the midpoint and dimensions $(b_x, b_y, b_h, b_w)$ for each of the 5 boxes in each cell.- `box_class_probs`: tensor of shape $(19 \times 19, 5, 80)$ containing the "class probabilities" $(c_1, c_2, ... c_{80})$ for each of the 80 classes for each of the 5 boxes per cell. **Exercise**: Implement `yolo_filter_boxes()`.1. Compute box scores by doing the elementwise product as described in Figure 4 ($p \times c$). The following code may help you choose the right operator: ```pythona = np.random.randn(19*19, 5, 1)b = np.random.randn(19*19, 5, 80)c = a * b shape of c will be (19*19, 5, 80)```This is an example of **broadcasting** (multiplying vectors of different sizes).2. For each box, find: - the index of the class with the maximum box score - the corresponding box score **Useful references** * [Keras argmax](https://keras.io/backend/argmax) * [Keras max](https://keras.io/backend/max) **Additional Hints** * For the `axis` parameter of `argmax` and `max`, if you want to select the **last** axis, one way to do so is to set `axis=-1`. This is similar to Python array indexing, where you can select the last position of an array using `arrayname[-1]`. * Applying `max` normally collapses the axis for which the maximum is applied. `keepdims=False` is the default option, and allows that dimension to be removed. We don't need to keep the last dimension after applying the maximum here. * Even though the documentation shows `keras.backend.argmax`, use `keras.argmax`. Similarly, use `keras.max`.3. Create a mask by using a threshold. As a reminder: `([0.9, 0.3, 0.4, 0.5, 0.1] < 0.4)` returns: `[False, True, False, False, True]`. The mask should be True for the boxes you want to keep. 4. Use TensorFlow to apply the mask to `box_class_scores`, `boxes` and `box_classes` to filter out the boxes we don't want. You should be left with just the subset of boxes you want to keep. **Useful reference**: * [boolean mask](https://www.tensorflow.org/api_docs/python/tf/boolean_mask) **Additional Hints**: * For the `tf.boolean_mask`, we can keep the default `axis=None`.**Reminder**: to call a Keras function, you should use `K.function(...)`. ###Code # GRADED FUNCTION: yolo_filter_boxes def yolo_filter_boxes(box_confidence, boxes, box_class_probs, threshold = .6): """Filters YOLO boxes by thresholding on object and class confidence. Arguments: box_confidence -- tensor of shape (19, 19, 5, 1) boxes -- tensor of shape (19, 19, 5, 4) box_class_probs -- tensor of shape (19, 19, 5, 80) threshold -- real value, if [ highest class probability score < threshold], then get rid of the corresponding box Returns: scores -- tensor of shape (None,), containing the class probability score for selected boxes boxes -- tensor of shape (None, 4), containing (b_x, b_y, b_h, b_w) coordinates of selected boxes classes -- tensor of shape (None,), containing the index of the class detected by the selected boxes Note: "None" is here because you don't know the exact number of selected boxes, as it depends on the threshold. For example, the actual output size of scores would be (10,) if there are 10 boxes. """ # Step 1: Compute box scores ### START CODE HERE ### (≈ 1 line) box_scores = box_class_probs*box_confidence ### END CODE HERE ### # Step 2: Find the box_classes using the max box_scores, keep track of the corresponding score ### START CODE HERE ### (≈ 2 lines) box_classes = K.argmax(box_scores,axis=-1) box_class_scores = K.max(box_scores,axis=-1) ### END CODE HERE ### # Step 3: Create a filtering mask based on "box_class_scores" by using "threshold". The mask should have the # same dimension as box_class_scores, and be True for the boxes you want to keep (with probability >= threshold) ### START CODE HERE ### (≈ 1 line) filtering_mask = box_class_scores>=threshold ### END CODE HERE ### # Step 4: Apply the mask to box_class_scores, boxes and box_classes ### START CODE HERE ### (≈ 3 lines) scores = tf.boolean_mask(box_class_scores,filtering_mask) boxes = tf.boolean_mask(boxes,filtering_mask) classes = tf.boolean_mask(box_classes,filtering_mask) ### END CODE HERE ### return scores, boxes, classes with tf.Session() as test_a: box_confidence = tf.random_normal([19, 19, 5, 1], mean=1, stddev=4, seed = 1) boxes = tf.random_normal([19, 19, 5, 4], mean=1, stddev=4, seed = 1) box_class_probs = tf.random_normal([19, 19, 5, 80], mean=1, stddev=4, seed = 1) scores, boxes, classes = yolo_filter_boxes(box_confidence, boxes, box_class_probs, threshold = 0.5) print("scores[2] = " + str(scores[2].eval())) print("boxes[2] = " + str(boxes[2].eval())) print("classes[2] = " + str(classes[2].eval())) print("scores.shape = " + str(scores.shape)) print("boxes.shape = " + str(boxes.shape)) print("classes.shape = " + str(classes.shape)) ###Output scores[2] = 10.7506 boxes[2] = [ 8.42653275 3.27136683 -0.5313437 -4.94137383] classes[2] = 7 scores.shape = (?,) boxes.shape = (?, 4) classes.shape = (?,) ###Markdown **Expected Output**: **scores[2]** 10.7506 **boxes[2]** [ 8.42653275 3.27136683 -0.5313437 -4.94137383] **classes[2]** 7 **scores.shape** (?,) **boxes.shape** (?, 4) **classes.shape** (?,) **Note** In the test for `yolo_filter_boxes`, we're using random numbers to test the function. In real data, the `box_class_probs` would contain non-zero values between 0 and 1 for the probabilities. The box coordinates in `boxes` would also be chosen so that lengths and heights are non-negative. 2.3 - Non-max suppression Even after filtering by thresholding over the class scores, you still end up with a lot of overlapping boxes. A second filter for selecting the right boxes is called non-maximum suppression (NMS). **Figure 7** : In this example, the model has predicted 3 cars, but it's actually 3 predictions of the same car. Running non-max suppression (NMS) will select only the most accurate (highest probability) of the 3 boxes. Non-max suppression uses the very important function called **"Intersection over Union"**, or IoU. **Figure 8** : Definition of "Intersection over Union". **Exercise**: Implement iou(). Some hints:- In this code, we use the convention that (0,0) is the top-left corner of an image, (1,0) is the upper-right corner, and (1,1) is the lower-right corner. In other words, the (0,0) origin starts at the top left corner of the image. As x increases, we move to the right. As y increases, we move down.- For this exercise, we define a box using its two corners: upper left $(x_1, y_1)$ and lower right $(x_2,y_2)$, instead of using the midpoint, height and width. (This makes it a bit easier to calculate the intersection).- To calculate the area of a rectangle, multiply its height $(y_2 - y_1)$ by its width $(x_2 - x_1)$. (Since $(x_1,y_1)$ is the top left and $x_2,y_2$ are the bottom right, these differences should be non-negative.- To find the **intersection** of the two boxes $(xi_{1}, yi_{1}, xi_{2}, yi_{2})$: - Feel free to draw some examples on paper to clarify this conceptually. - The top left corner of the intersection $(xi_{1}, yi_{1})$ is found by comparing the top left corners $(x_1, y_1)$ of the two boxes and finding a vertex that has an x-coordinate that is closer to the right, and y-coordinate that is closer to the bottom. - The bottom right corner of the intersection $(xi_{2}, yi_{2})$ is found by comparing the bottom right corners $(x_2,y_2)$ of the two boxes and finding a vertex whose x-coordinate is closer to the left, and the y-coordinate that is closer to the top. - The two boxes **may have no intersection**. You can detect this if the intersection coordinates you calculate end up being the top right and/or bottom left corners of an intersection box. Another way to think of this is if you calculate the height $(y_2 - y_1)$ or width $(x_2 - x_1)$ and find that at least one of these lengths is negative, then there is no intersection (intersection area is zero). - The two boxes may intersect at the **edges or vertices**, in which case the intersection area is still zero. This happens when either the height or width (or both) of the calculated intersection is zero.**Additional Hints**- `xi1` = **max**imum of the x1 coordinates of the two boxes- `yi1` = **max**imum of the y1 coordinates of the two boxes- `xi2` = **min**imum of the x2 coordinates of the two boxes- `yi2` = **min**imum of the y2 coordinates of the two boxes- `inter_area` = You can use `max(height, 0)` and `max(width, 0)` ###Code # GRADED FUNCTION: iou def iou(box1, box2): """Implement the intersection over union (IoU) between box1 and box2      Arguments: box1 -- first box, list object with coordinates (box1_x1, box1_y1, box1_x2, box_1_y2)     box2 -- second box, list object with coordinates (box2_x1, box2_y1, box2_x2, box2_y2)     """ # Assign variable names to coordinates for clarity (box1_x1, box1_y1, box1_x2, box1_y2) = box1 (box2_x1, box2_y1, box2_x2, box2_y2) = box2 # Calculate the (yi1, xi1, yi2, xi2) coordinates of the intersection of box1 and box2. Calculate its Area. ### START CODE HERE ### (≈ 7 lines) xi1 = max(box1[0],box2[0]) yi1 = max(box1[1],box2[1]) xi2 = min(box1[2],box2[2]) yi2 = min(box1[3],box2[3]) inter_width = max(xi2-xi1,0) inter_height = max(yi2-yi1,0) inter_area = inter_width*inter_height ### END CODE HERE ###     # Calculate the Union area by using Formula: Union(A,B) = A + B - Inter(A,B) ### START CODE HERE ### (≈ 3 lines) box1_area = (box1[2]-box1[0])*(box1[3]-box1[1]) box2_area = (box2[2]-box2[0])*(box2[3]-box2[1]) union_area = box1_area+box2_area-inter_area ### END CODE HERE ### # compute the IoU ### START CODE HERE ### (≈ 1 line) iou = inter_area/union_area ### END CODE HERE ### return iou ## Test case 1: boxes intersect box1 = (2, 1, 4, 3) box2 = (1, 2, 3, 4) print("iou for intersecting boxes = " + str(iou(box1, box2))) ## Test case 2: boxes do not intersect box1 = (1,2,3,4) box2 = (5,6,7,8) print("iou for non-intersecting boxes = " + str(iou(box1,box2))) ## Test case 3: boxes intersect at vertices only box1 = (1,1,2,2) box2 = (2,2,3,3) print("iou for boxes that only touch at vertices = " + str(iou(box1,box2))) ## Test case 4: boxes intersect at edge only box1 = (1,1,3,3) box2 = (2,3,3,4) print("iou for boxes that only touch at edges = " + str(iou(box1,box2))) ###Output iou for intersecting boxes = 0.14285714285714285 iou for non-intersecting boxes = 0.0 iou for boxes that only touch at vertices = 0.0 iou for boxes that only touch at edges = 0.0 ###Markdown **Expected Output**:```iou for intersecting boxes = 0.14285714285714285iou for non-intersecting boxes = 0.0iou for boxes that only touch at vertices = 0.0iou for boxes that only touch at edges = 0.0``` YOLO non-max suppressionYou are now ready to implement non-max suppression. The key steps are: 1. Select the box that has the highest score.2. Compute the overlap of this box with all other boxes, and remove boxes that overlap significantly (iou >= `iou_threshold`).3. Go back to step 1 and iterate until there are no more boxes with a lower score than the currently selected box.This will remove all boxes that have a large overlap with the selected boxes. Only the "best" boxes remain.**Exercise**: Implement yolo_non_max_suppression() using TensorFlow. TensorFlow has two built-in functions that are used to implement non-max suppression (so you don't actually need to use your `iou()` implementation):** Reference documentation ** - [tf.image.non_max_suppression()](https://www.tensorflow.org/api_docs/python/tf/image/non_max_suppression)```tf.image.non_max_suppression( boxes, scores, max_output_size, iou_threshold=0.5, name=None)```Note that in the version of tensorflow used here, there is no parameter `score_threshold` (it's shown in the documentation for the latest version) so trying to set this value will result in an error message: *got an unexpected keyword argument 'score_threshold.*- [K.gather()](https://www.tensorflow.org/api_docs/python/tf/keras/backend/gather) Even though the documentation shows `tf.keras.backend.gather()`, you can use `keras.gather()`. ```keras.gather( reference, indices)``` ###Code # GRADED FUNCTION: yolo_non_max_suppression def yolo_non_max_suppression(scores, boxes, classes, max_boxes = 10, iou_threshold = 0.5): """ Applies Non-max suppression (NMS) to set of boxes Arguments: scores -- tensor of shape (None,), output of yolo_filter_boxes() boxes -- tensor of shape (None, 4), output of yolo_filter_boxes() that have been scaled to the image size (see later) classes -- tensor of shape (None,), output of yolo_filter_boxes() max_boxes -- integer, maximum number of predicted boxes you'd like iou_threshold -- real value, "intersection over union" threshold used for NMS filtering Returns: scores -- tensor of shape (, None), predicted score for each box boxes -- tensor of shape (4, None), predicted box coordinates classes -- tensor of shape (, None), predicted class for each box Note: The "None" dimension of the output tensors has obviously to be less than max_boxes. Note also that this function will transpose the shapes of scores, boxes, classes. This is made for convenience. """ max_boxes_tensor = K.variable(max_boxes, dtype='int32') # tensor to be used in tf.image.non_max_suppression() K.get_session().run(tf.variables_initializer([max_boxes_tensor])) # initialize variable max_boxes_tensor # Use tf.image.non_max_suppression() to get the list of indices corresponding to boxes you keep ### START CODE HERE ### (≈ 1 line) nms_indices = tf.image.non_max_suppression(boxes,scores,max_boxes,iou_threshold) ### END CODE HERE ### # Use K.gather() to select only nms_indices from scores, boxes and classes ### START CODE HERE ### (≈ 3 lines) scores = K.gather(scores,nms_indices) boxes = K.gather(boxes,nms_indices) classes = K.gather(classes,nms_indices) ### END CODE HERE ### return scores, boxes, classes with tf.Session() as test_b: scores = tf.random_normal([54,], mean=1, stddev=4, seed = 1) boxes = tf.random_normal([54, 4], mean=1, stddev=4, seed = 1) classes = tf.random_normal([54,], mean=1, stddev=4, seed = 1) scores, boxes, classes = yolo_non_max_suppression(scores, boxes, classes) print("scores[2] = " + str(scores[2].eval())) print("boxes[2] = " + str(boxes[2].eval())) print("classes[2] = " + str(classes[2].eval())) print("scores.shape = " + str(scores.eval().shape)) print("boxes.shape = " + str(boxes.eval().shape)) print("classes.shape = " + str(classes.eval().shape)) ###Output scores[2] = 6.9384 boxes[2] = [-5.299932 3.13798141 4.45036697 0.95942086] classes[2] = -2.24527 scores.shape = (10,) boxes.shape = (10, 4) classes.shape = (10,) ###Markdown **Expected Output**: **scores[2]** 6.9384 **boxes[2]** [-5.299932 3.13798141 4.45036697 0.95942086] **classes[2]** -2.24527 **scores.shape** (10,) **boxes.shape** (10, 4) **classes.shape** (10,) 2.4 Wrapping up the filteringIt's time to implement a function taking the output of the deep CNN (the 19x19x5x85 dimensional encoding) and filtering through all the boxes using the functions you've just implemented. **Exercise**: Implement `yolo_eval()` which takes the output of the YOLO encoding and filters the boxes using score threshold and NMS. There's just one last implementational detail you have to know. There're a few ways of representing boxes, such as via their corners or via their midpoint and height/width. YOLO converts between a few such formats at different times, using the following functions (which we have provided): ```pythonboxes = yolo_boxes_to_corners(box_xy, box_wh) ```which converts the yolo box coordinates (x,y,w,h) to box corners' coordinates (x1, y1, x2, y2) to fit the input of `yolo_filter_boxes````pythonboxes = scale_boxes(boxes, image_shape)```YOLO's network was trained to run on 608x608 images. If you are testing this data on a different size image--for example, the car detection dataset had 720x1280 images--this step rescales the boxes so that they can be plotted on top of the original 720x1280 image. Don't worry about these two functions; we'll show you where they need to be called. ###Code # GRADED FUNCTION: yolo_eval def yolo_eval(yolo_outputs, image_shape = (720., 1280.), max_boxes=10, score_threshold=.6, iou_threshold=.5): """ Converts the output of YOLO encoding (a lot of boxes) to your predicted boxes along with their scores, box coordinates and classes. Arguments: yolo_outputs -- output of the encoding model (for image_shape of (608, 608, 3)), contains 4 tensors: box_confidence: tensor of shape (None, 19, 19, 5, 1) box_xy: tensor of shape (None, 19, 19, 5, 2) box_wh: tensor of shape (None, 19, 19, 5, 2) box_class_probs: tensor of shape (None, 19, 19, 5, 80) image_shape -- tensor of shape (2,) containing the input shape, in this notebook we use (608., 608.) (has to be float32 dtype) max_boxes -- integer, maximum number of predicted boxes you'd like score_threshold -- real value, if [ highest class probability score < threshold], then get rid of the corresponding box iou_threshold -- real value, "intersection over union" threshold used for NMS filtering Returns: scores -- tensor of shape (None, ), predicted score for each box boxes -- tensor of shape (None, 4), predicted box coordinates classes -- tensor of shape (None,), predicted class for each box """ ### START CODE HERE ### # Retrieve outputs of the YOLO model (≈1 line) box_confidence, box_xy, box_wh, box_class_probs = yolo_outputs # Convert boxes to be ready for filtering functions (convert boxes box_xy and box_wh to corner coordinates) boxes = yolo_boxes_to_corners(box_xy, box_wh) # Use one of the functions you've implemented to perform Score-filtering with a threshold of score_threshold (≈1 line) scores, boxes, classes = yolo_filter_boxes(box_confidence,boxes,box_class_probs,score_threshold) # Scale boxes back to original image shape. boxes = scale_boxes(boxes, image_shape) # Use one of the functions you've implemented to perform Non-max suppression with # maximum number of boxes set to max_boxes and a threshold of iou_threshold (≈1 line) scores, boxes, classes = yolo_non_max_suppression(scores,boxes,classes,max_boxes,iou_threshold) ### END CODE HERE ### return scores, boxes, classes with tf.Session() as test_b: yolo_outputs = (tf.random_normal([19, 19, 5, 1], mean=1, stddev=4, seed = 1), tf.random_normal([19, 19, 5, 2], mean=1, stddev=4, seed = 1), tf.random_normal([19, 19, 5, 2], mean=1, stddev=4, seed = 1), tf.random_normal([19, 19, 5, 80], mean=1, stddev=4, seed = 1)) scores, boxes, classes = yolo_eval(yolo_outputs) print("scores[2] = " + str(scores[2].eval())) print("boxes[2] = " + str(boxes[2].eval())) print("classes[2] = " + str(classes[2].eval())) print("scores.shape = " + str(scores.eval().shape)) print("boxes.shape = " + str(boxes.eval().shape)) print("classes.shape = " + str(classes.eval().shape)) ###Output scores[2] = 138.791 boxes[2] = [ 1292.32971191 -278.52166748 3876.98925781 -835.56494141] classes[2] = 54 scores.shape = (10,) boxes.shape = (10, 4) classes.shape = (10,) ###Markdown **Expected Output**: **scores[2]** 138.791 **boxes[2]** [ 1292.32971191 -278.52166748 3876.98925781 -835.56494141] **classes[2]** 54 **scores.shape** (10,) **boxes.shape** (10, 4) **classes.shape** (10,) Summary for YOLO:- Input image (608, 608, 3)- The input image goes through a CNN, resulting in a (19,19,5,85) dimensional output. - After flattening the last two dimensions, the output is a volume of shape (19, 19, 425): - Each cell in a 19x19 grid over the input image gives 425 numbers. - 425 = 5 x 85 because each cell contains predictions for 5 boxes, corresponding to 5 anchor boxes, as seen in lecture. - 85 = 5 + 80 where 5 is because $(p_c, b_x, b_y, b_h, b_w)$ has 5 numbers, and 80 is the number of classes we'd like to detect- You then select only few boxes based on: - Score-thresholding: throw away boxes that have detected a class with a score less than the threshold - Non-max suppression: Compute the Intersection over Union and avoid selecting overlapping boxes- This gives you YOLO's final output. 3 - Test YOLO pre-trained model on images In this part, you are going to use a pre-trained model and test it on the car detection dataset. We'll need a session to execute the computation graph and evaluate the tensors. ###Code sess = K.get_session() ###Output _____no_output_____ ###Markdown 3.1 - Defining classes, anchors and image shape.* Recall that we are trying to detect 80 classes, and are using 5 anchor boxes. * We have gathered the information on the 80 classes and 5 boxes in two files "coco_classes.txt" and "yolo_anchors.txt". * We'll read class names and anchors from text files.* The car detection dataset has 720x1280 images, which we've pre-processed into 608x608 images. ###Code class_names = read_classes("model_data/coco_classes.txt") anchors = read_anchors("model_data/yolo_anchors.txt") image_shape = (720., 1280.) ###Output _____no_output_____ ###Markdown 3.2 - Loading a pre-trained model* Training a YOLO model takes a very long time and requires a fairly large dataset of labelled bounding boxes for a large range of target classes. * You are going to load an existing pre-trained Keras YOLO model stored in "yolo.h5". * These weights come from the official YOLO website, and were converted using a function written by Allan Zelener. References are at the end of this notebook. Technically, these are the parameters from the "YOLOv2" model, but we will simply refer to it as "YOLO" in this notebook.Run the cell below to load the model from this file. ###Code yolo_model = load_model("model_data/yolo.h5") ###Output /opt/conda/lib/python3.6/site-packages/keras/models.py:251: UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually. warnings.warn('No training configuration found in save file: ' ###Markdown This loads the weights of a trained YOLO model. Here's a summary of the layers your model contains. ###Code yolo_model.summary() ###Output ____________________________________________________________________________________________________ Layer (type) Output Shape Param # Connected to ==================================================================================================== input_1 (InputLayer) (None, 608, 608, 3) 0 ____________________________________________________________________________________________________ conv2d_1 (Conv2D) (None, 608, 608, 32) 864 input_1[0][0] ____________________________________________________________________________________________________ batch_normalization_1 (BatchNorm (None, 608, 608, 32) 128 conv2d_1[0][0] ____________________________________________________________________________________________________ leaky_re_lu_1 (LeakyReLU) (None, 608, 608, 32) 0 batch_normalization_1[0][0] ____________________________________________________________________________________________________ max_pooling2d_1 (MaxPooling2D) (None, 304, 304, 32) 0 leaky_re_lu_1[0][0] ____________________________________________________________________________________________________ conv2d_2 (Conv2D) (None, 304, 304, 64) 18432 max_pooling2d_1[0][0] ____________________________________________________________________________________________________ batch_normalization_2 (BatchNorm (None, 304, 304, 64) 256 conv2d_2[0][0] ____________________________________________________________________________________________________ leaky_re_lu_2 (LeakyReLU) (None, 304, 304, 64) 0 batch_normalization_2[0][0] ____________________________________________________________________________________________________ max_pooling2d_2 (MaxPooling2D) (None, 152, 152, 64) 0 leaky_re_lu_2[0][0] ____________________________________________________________________________________________________ conv2d_3 (Conv2D) (None, 152, 152, 128) 73728 max_pooling2d_2[0][0] ____________________________________________________________________________________________________ batch_normalization_3 (BatchNorm (None, 152, 152, 128) 512 conv2d_3[0][0] ____________________________________________________________________________________________________ leaky_re_lu_3 (LeakyReLU) (None, 152, 152, 128) 0 batch_normalization_3[0][0] ____________________________________________________________________________________________________ conv2d_4 (Conv2D) (None, 152, 152, 64) 8192 leaky_re_lu_3[0][0] ____________________________________________________________________________________________________ batch_normalization_4 (BatchNorm (None, 152, 152, 64) 256 conv2d_4[0][0] ____________________________________________________________________________________________________ leaky_re_lu_4 (LeakyReLU) (None, 152, 152, 64) 0 batch_normalization_4[0][0] ____________________________________________________________________________________________________ conv2d_5 (Conv2D) (None, 152, 152, 128) 73728 leaky_re_lu_4[0][0] ____________________________________________________________________________________________________ batch_normalization_5 (BatchNorm (None, 152, 152, 128) 512 conv2d_5[0][0] ____________________________________________________________________________________________________ leaky_re_lu_5 (LeakyReLU) (None, 152, 152, 128) 0 batch_normalization_5[0][0] ____________________________________________________________________________________________________ max_pooling2d_3 (MaxPooling2D) (None, 76, 76, 128) 0 leaky_re_lu_5[0][0] ____________________________________________________________________________________________________ conv2d_6 (Conv2D) (None, 76, 76, 256) 294912 max_pooling2d_3[0][0] ____________________________________________________________________________________________________ batch_normalization_6 (BatchNorm (None, 76, 76, 256) 1024 conv2d_6[0][0] ____________________________________________________________________________________________________ leaky_re_lu_6 (LeakyReLU) (None, 76, 76, 256) 0 batch_normalization_6[0][0] ____________________________________________________________________________________________________ conv2d_7 (Conv2D) (None, 76, 76, 128) 32768 leaky_re_lu_6[0][0] ____________________________________________________________________________________________________ batch_normalization_7 (BatchNorm (None, 76, 76, 128) 512 conv2d_7[0][0] ____________________________________________________________________________________________________ leaky_re_lu_7 (LeakyReLU) (None, 76, 76, 128) 0 batch_normalization_7[0][0] ____________________________________________________________________________________________________ conv2d_8 (Conv2D) (None, 76, 76, 256) 294912 leaky_re_lu_7[0][0] ____________________________________________________________________________________________________ batch_normalization_8 (BatchNorm (None, 76, 76, 256) 1024 conv2d_8[0][0] ____________________________________________________________________________________________________ leaky_re_lu_8 (LeakyReLU) (None, 76, 76, 256) 0 batch_normalization_8[0][0] ____________________________________________________________________________________________________ max_pooling2d_4 (MaxPooling2D) (None, 38, 38, 256) 0 leaky_re_lu_8[0][0] ____________________________________________________________________________________________________ conv2d_9 (Conv2D) (None, 38, 38, 512) 1179648 max_pooling2d_4[0][0] ____________________________________________________________________________________________________ batch_normalization_9 (BatchNorm (None, 38, 38, 512) 2048 conv2d_9[0][0] ____________________________________________________________________________________________________ leaky_re_lu_9 (LeakyReLU) (None, 38, 38, 512) 0 batch_normalization_9[0][0] ____________________________________________________________________________________________________ conv2d_10 (Conv2D) (None, 38, 38, 256) 131072 leaky_re_lu_9[0][0] ____________________________________________________________________________________________________ batch_normalization_10 (BatchNor (None, 38, 38, 256) 1024 conv2d_10[0][0] ____________________________________________________________________________________________________ leaky_re_lu_10 (LeakyReLU) (None, 38, 38, 256) 0 batch_normalization_10[0][0] ____________________________________________________________________________________________________ conv2d_11 (Conv2D) (None, 38, 38, 512) 1179648 leaky_re_lu_10[0][0] ____________________________________________________________________________________________________ batch_normalization_11 (BatchNor (None, 38, 38, 512) 2048 conv2d_11[0][0] ____________________________________________________________________________________________________ leaky_re_lu_11 (LeakyReLU) (None, 38, 38, 512) 0 batch_normalization_11[0][0] ____________________________________________________________________________________________________ conv2d_12 (Conv2D) (None, 38, 38, 256) 131072 leaky_re_lu_11[0][0] ____________________________________________________________________________________________________ batch_normalization_12 (BatchNor (None, 38, 38, 256) 1024 conv2d_12[0][0] ____________________________________________________________________________________________________ leaky_re_lu_12 (LeakyReLU) (None, 38, 38, 256) 0 batch_normalization_12[0][0] ____________________________________________________________________________________________________ conv2d_13 (Conv2D) (None, 38, 38, 512) 1179648 leaky_re_lu_12[0][0] ____________________________________________________________________________________________________ batch_normalization_13 (BatchNor (None, 38, 38, 512) 2048 conv2d_13[0][0] ____________________________________________________________________________________________________ leaky_re_lu_13 (LeakyReLU) (None, 38, 38, 512) 0 batch_normalization_13[0][0] ____________________________________________________________________________________________________ max_pooling2d_5 (MaxPooling2D) (None, 19, 19, 512) 0 leaky_re_lu_13[0][0] ____________________________________________________________________________________________________ conv2d_14 (Conv2D) (None, 19, 19, 1024) 4718592 max_pooling2d_5[0][0] ____________________________________________________________________________________________________ batch_normalization_14 (BatchNor (None, 19, 19, 1024) 4096 conv2d_14[0][0] ____________________________________________________________________________________________________ leaky_re_lu_14 (LeakyReLU) (None, 19, 19, 1024) 0 batch_normalization_14[0][0] ____________________________________________________________________________________________________ conv2d_15 (Conv2D) (None, 19, 19, 512) 524288 leaky_re_lu_14[0][0] ____________________________________________________________________________________________________ batch_normalization_15 (BatchNor (None, 19, 19, 512) 2048 conv2d_15[0][0] ____________________________________________________________________________________________________ leaky_re_lu_15 (LeakyReLU) (None, 19, 19, 512) 0 batch_normalization_15[0][0] ____________________________________________________________________________________________________ conv2d_16 (Conv2D) (None, 19, 19, 1024) 4718592 leaky_re_lu_15[0][0] ____________________________________________________________________________________________________ batch_normalization_16 (BatchNor (None, 19, 19, 1024) 4096 conv2d_16[0][0] ____________________________________________________________________________________________________ leaky_re_lu_16 (LeakyReLU) (None, 19, 19, 1024) 0 batch_normalization_16[0][0] ____________________________________________________________________________________________________ conv2d_17 (Conv2D) (None, 19, 19, 512) 524288 leaky_re_lu_16[0][0] ____________________________________________________________________________________________________ batch_normalization_17 (BatchNor (None, 19, 19, 512) 2048 conv2d_17[0][0] ____________________________________________________________________________________________________ leaky_re_lu_17 (LeakyReLU) (None, 19, 19, 512) 0 batch_normalization_17[0][0] ____________________________________________________________________________________________________ conv2d_18 (Conv2D) (None, 19, 19, 1024) 4718592 leaky_re_lu_17[0][0] ____________________________________________________________________________________________________ batch_normalization_18 (BatchNor (None, 19, 19, 1024) 4096 conv2d_18[0][0] ____________________________________________________________________________________________________ leaky_re_lu_18 (LeakyReLU) (None, 19, 19, 1024) 0 batch_normalization_18[0][0] ____________________________________________________________________________________________________ conv2d_19 (Conv2D) (None, 19, 19, 1024) 9437184 leaky_re_lu_18[0][0] ____________________________________________________________________________________________________ batch_normalization_19 (BatchNor (None, 19, 19, 1024) 4096 conv2d_19[0][0] ____________________________________________________________________________________________________ conv2d_21 (Conv2D) (None, 38, 38, 64) 32768 leaky_re_lu_13[0][0] ____________________________________________________________________________________________________ leaky_re_lu_19 (LeakyReLU) (None, 19, 19, 1024) 0 batch_normalization_19[0][0] ____________________________________________________________________________________________________ batch_normalization_21 (BatchNor (None, 38, 38, 64) 256 conv2d_21[0][0] ____________________________________________________________________________________________________ conv2d_20 (Conv2D) (None, 19, 19, 1024) 9437184 leaky_re_lu_19[0][0] ____________________________________________________________________________________________________ leaky_re_lu_21 (LeakyReLU) (None, 38, 38, 64) 0 batch_normalization_21[0][0] ____________________________________________________________________________________________________ batch_normalization_20 (BatchNor (None, 19, 19, 1024) 4096 conv2d_20[0][0] ____________________________________________________________________________________________________ space_to_depth_x2 (Lambda) (None, 19, 19, 256) 0 leaky_re_lu_21[0][0] ____________________________________________________________________________________________________ leaky_re_lu_20 (LeakyReLU) (None, 19, 19, 1024) 0 batch_normalization_20[0][0] ____________________________________________________________________________________________________ concatenate_1 (Concatenate) (None, 19, 19, 1280) 0 space_to_depth_x2[0][0] leaky_re_lu_20[0][0] ____________________________________________________________________________________________________ conv2d_22 (Conv2D) (None, 19, 19, 1024) 11796480 concatenate_1[0][0] ____________________________________________________________________________________________________ batch_normalization_22 (BatchNor (None, 19, 19, 1024) 4096 conv2d_22[0][0] ____________________________________________________________________________________________________ leaky_re_lu_22 (LeakyReLU) (None, 19, 19, 1024) 0 batch_normalization_22[0][0] ____________________________________________________________________________________________________ conv2d_23 (Conv2D) (None, 19, 19, 425) 435625 leaky_re_lu_22[0][0] ==================================================================================================== Total params: 50,983,561 Trainable params: 50,962,889 Non-trainable params: 20,672 ____________________________________________________________________________________________________ ###Markdown **Note**: On some computers, you may see a warning message from Keras. Don't worry about it if you do--it is fine.**Reminder**: this model converts a preprocessed batch of input images (shape: (m, 608, 608, 3)) into a tensor of shape (m, 19, 19, 5, 85) as explained in Figure (2). 3.3 - Convert output of the model to usable bounding box tensorsThe output of `yolo_model` is a (m, 19, 19, 5, 85) tensor that needs to pass through non-trivial processing and conversion. The following cell does that for you.If you are curious about how `yolo_head` is implemented, you can find the function definition in the file ['keras_yolo.py'](https://github.com/allanzelener/YAD2K/blob/master/yad2k/models/keras_yolo.py). The file is located in your workspace in this path 'yad2k/models/keras_yolo.py'. ###Code yolo_outputs = yolo_head(yolo_model.output, anchors, len(class_names)) ###Output _____no_output_____ ###Markdown You added `yolo_outputs` to your graph. This set of 4 tensors is ready to be used as input by your `yolo_eval` function. 3.4 - Filtering boxes`yolo_outputs` gave you all the predicted boxes of `yolo_model` in the correct format. You're now ready to perform filtering and select only the best boxes. Let's now call `yolo_eval`, which you had previously implemented, to do this. ###Code scores, boxes, classes = yolo_eval(yolo_outputs, image_shape) ###Output _____no_output_____ ###Markdown 3.5 - Run the graph on an imageLet the fun begin. You have created a graph that can be summarized as follows:1. yolo_model.input is given to `yolo_model`. The model is used to compute the output yolo_model.output 2. yolo_model.output is processed by `yolo_head`. It gives you yolo_outputs 3. yolo_outputs goes through a filtering function, `yolo_eval`. It outputs your predictions: scores, boxes, classes **Exercise**: Implement predict() which runs the graph to test YOLO on an image.You will need to run a TensorFlow session, to have it compute `scores, boxes, classes`.The code below also uses the following function:```pythonimage, image_data = preprocess_image("images/" + image_file, model_image_size = (608, 608))```which outputs:- image: a python (PIL) representation of your image used for drawing boxes. You won't need to use it.- image_data: a numpy-array representing the image. This will be the input to the CNN.**Important note**: when a model uses BatchNorm (as is the case in YOLO), you will need to pass an additional placeholder in the feed_dict {K.learning_phase(): 0}. Hint: Using the TensorFlow Session object* Recall that above, we called `K.get_Session()` and saved the Session object in `sess`.* To evaluate a list of tensors, we call `sess.run()` like this:```sess.run(fetches=[tensor1,tensor2,tensor3], feed_dict={yolo_model.input: the_input_variable, K.learning_phase():0 }```* Notice that the variables `scores, boxes, classes` are not passed into the `predict` function, but these are global variables that you will use within the `predict` function. ###Code def predict(sess, image_file): """ Runs the graph stored in "sess" to predict boxes for "image_file". Prints and plots the predictions. Arguments: sess -- your tensorflow/Keras session containing the YOLO graph image_file -- name of an image stored in the "images" folder. Returns: out_scores -- tensor of shape (None, ), scores of the predicted boxes out_boxes -- tensor of shape (None, 4), coordinates of the predicted boxes out_classes -- tensor of shape (None, ), class index of the predicted boxes Note: "None" actually represents the number of predicted boxes, it varies between 0 and max_boxes. """ # Preprocess your image image, image_data = preprocess_image("images/" + image_file, model_image_size = (608, 608)) # Run the session with the correct tensors and choose the correct placeholders in the feed_dict. # You'll need to use feed_dict={yolo_model.input: ... , K.learning_phase(): 0}) ### START CODE HERE ### (≈ 1 line) out_scores = tf.placeholder(tf.float32,shape=(None,)) out_boxes = tf.placeholder(tf.float32,shape=(None,4)) out_classes = tf.placeholder(tf.float32,shape=(None,)) ### END CODE HERE ### # Print predictions info print('Found {} boxes for {}'.format(len(out_boxes), image_file)) # Generate colors for drawing bounding boxes. colors = generate_colors(class_names) # Draw bounding boxes on the image file draw_boxes(image, out_scores, out_boxes, out_classes, class_names, colors) # Save the predicted bounding box on the image image.save(os.path.join("out", image_file), quality=90) # Display the results in the notebook output_image = scipy.misc.imread(os.path.join("out", image_file)) imshow(output_image) return out_scores, out_boxes, out_classes ###Output _____no_output_____ ###Markdown Run the following cell on the "test.jpg" image to verify that your function is correct. ###Code out_scores, out_boxes, out_classes = predict(sess, "test.jpg") ###Output _____no_output_____
Python Absolute Beginner/Module_1_2.3_Absolute_Beginner.ipynb
###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output It's time to save your code I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print("where's the homework?") # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein print('"Education is what remains after one has forgotten what one has learned in school" - Albert Einstein') ###Output "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alphabetical".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "are spaces and puncuation alphabetical?".isalpha() #false becasue of the spaces # [ ] initailize variable alpha_test with input alpha_test = input("What is your name?: ") # [ ] use .isalpha() on string variable alpha_test alpha_test.isalpha() ###Output What is your name?: cam ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output It's time to save your code I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print("Where's the homework?") # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein print('"Education is what remains after one has forgotten what one has learned in school" - Albert Einstein') ###Output "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stormy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alphabetical".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "Are spaces and punctuation Alphabetical?".isalpha() # [ ] initailize variable alpha_test with input alpha_test=input("Enter a what you wish to test for alphabetical: ") # [ ] use .isalpha() on string variable alpha_test alpha_test.isalpha() ###Output Enter a what you wish to test for alphabetical: Blah ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output It's time to save your code I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print("Where's the homework?") # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein person_name = "Albert Einstein" print('"Education is what remains after one has forgotten what one has learned in school" - ' + person_name) ###Output "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alphabetical".isalpha () # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "Are spaces and punctuation Aplhabetical?".isalpha() # [ ] initailize variable alpha_test with input alpha_test = 'Iamhungry' # [ ] use .isalpha() on string variable alpha_test print(alpha_test, 'is a', alpha_test.isalpha(), 'statement') ###Output Iamhungry is a True statement ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output It's time to save your code I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print ("Where's the homework?") # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein print ('"Education is what remains after one has forgotten what one has learned in school" - Albert Einstein') ###Output "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alphabetical".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "Are spaces and punctuation Alphaberical?".isalpha() # [ ] initailize variable alpha_test with input fir_name = input("enter first name: ") # [ ] use .isalpha() on string variable alpha_test fir_name.isalpha() ###Output enter first name: Tim ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output It's time to save your code I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print("Where's the homework?") # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein print('"Education is what remains after one has forgotten what one has learned in school" - Albert Eirnstein') ###Output "Education is what remains after one has forgotten what one has learned in school" - Albert Eirnstein ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alphabetical".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "Are spaces and punctuation Alphabetical?".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" # [ ] initailize variable alpha_test with input alpha_test = input() # [ ] use .isalpha() on string variable alpha_test alpha_test.isalpha() ###Output 3 ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output It's time to save your code I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print("Where's the homework?") # [ ] output with double quotes: "Education is what remains after one has #forgotten what one has learned in school" - Albert Einstein print('"Education is what remains after one has forgotten what one has learned in school" - Albert Einstein') ###Output "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alphabetical".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "Are spaces and punctuation Alphabetical?".isalpha() # [ ] initailize variable alpha_test with input alpha_test="alphabet test" # [ ] use .isalpha() on string variable alpha_test print("alphabet test:",alpha_test, "is all:" ,alpha_test.isalpha()) ###Output alphabet test: alphabet test is all: False ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output It's time to save your code I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print("Where's the homework?") # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein print('"Education is what remains after one has forgotten what one has learned in school" - Albert Einstein') ###Output "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alphabetical".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "Are spaces and punctuation Alphabetical?".isalpha() # [ ] initailize variable alpha_test with input alpha_test = input("Enter name: ") # [ ] use .isalpha() on string variable alpha_test alpha_test.isalpha() ###Output Enter name: Jacob ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output _____no_output_____ ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein ###Output _____no_output_____ ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" # [ ] initailize variable alpha_test with input # [ ] use .isalpha() on string variable alpha_test ###Output _____no_output_____ ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("'It's time to save your code'") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output 'It's time to save your code' I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print("'wheres the hw'") # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein print('"education is what remain after one has forgotten what one has learned"') ###Output "education is what remain after one has forgotten what one has learned" ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alpha".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "Are spaces and punctuation alpha?".isalpha # [ ] initailize variable alpha_test with input alpha_test=input() # [ ] use .isalpha() on string variable alpha_test alpha_test.isalpha() ###Output Aaron ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output It's time to save your code I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print("Where's the homework?") # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein print('"Education is what remains after one has forgotten what one has learned in school" - Albert Einstein') ###Output "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alphabetical".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "Are spaces and punctuation Alphabetical?".isalpha() # [ ] initailize variable alpha_test with input alpha_test = input("Enter an int, float, or string: ") # [ ] use .isalpha() on string variable alpha_test alpha_test.isalpha() ###Output Enter an int, float, or string: Yes ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output It's time to save your code I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print("Where's the homework?") # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein print('Albert states, "Education is what remains after one has forgotten what one has learned in school"') ###Output Albert states, "Education is what remains after one has forgotten what one has learned in school" ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".isupper()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alphabetical".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "Are spaces and punctuation alphabetical".isalpha() # [ ] initailize variable alpha_test with input alpha_test = input() # [ ] use .isalpha() on string variable alpha_test .isalpha() = alpha_test ###Output _____no_output_____ ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output It's time to save your code I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print("where's the homework?") # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein print('"Education is what remains after one has forgotten what ones has learned in school" - Albert Einstien') ###Output "Education is what remains after one has forgotten what ones has learned in school" - Albert Einstien ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alphabetical".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "Are spaces and punctuation Alphabetical?".isalpha() # [ ] initailize variable alpha_test with input alpha_test = input("Enter a word or phrase: ") # [ ] use .isalpha() on string variable alpha_test print("All alphabetical =", alpha_test.isalpha()) ###Output Enter a word or phrase: hello world All alphabetical = False ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output It's time to save your code I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print("Where's the homework?") # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein print('"Education is what remains after one has forgotten what one has learned in school" -Albert Einstein') ###Output "Education is what remains after one has forgotten what one has learned in school" -Albert Einstein ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alphabetical".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "Are spaces and punctuation Alphabetical?".isalpha() # [ ] initailize variable alpha_test with input alpha_test = input() # [ ] use .isalpha() on string variable alpha_test alpha_test.isalpha() ###Output alpha ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output It's time to save your code I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print("Where's the homework?") # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein print('"Education is what remains after one has forgotten what one has learned in school" - Albert Einstein') ###Output "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alphabetical".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "Are spaces and punctuation Alphabetical?".isalpha() # [ ] initailize variable alpha_test with input alpha_test = input() # [ ] use .isalpha() on string variable alpha_test alpha_test.isalpha() ###Output test ###Markdown 1-2.3 Intro Python Strings: input, testing, formatting- input() - gathering user input - print() formatting - **Quotes inside strings** - **Boolean string tests methods** - String formatting methods- Formatting string input()- Boolean œ`in` keyword -----> Student will be able to- gather, store and use string `input()` - format `print()` output - **test string characteristics** - format string output- search for a string in a string &nbsp; Concepts quotes inside strings [![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"http://jupyternootbookwams.streaming.mediaservices.windows.net/c21bfb29-21f6-4bef-b00e-25d2f7440153/Unit1_Section2-3-Quotes_in_Strings.vtt","srclang":"en","kind":"subtitles","label":"english"}]) single quotes in double quotesto display single quotes **`'`** in a string - double quotes can be used as the outer quotes: **`"it's time"`** double quotes in single quotesto display double quotes **`"`** in a string- single quotes can be used as the outer quotes: **`'Alton said "Hello"'`** &nbsp; Examples ###Code # review and run the code # Single quote surrounded by Double print("It's time to save your code") # Double quote surrounded by Single print('I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run"') ###Output It's time to save your code I said to the class "sometimes you need to shut down and restart a notebook when cells refuse to run" ###Markdown &nbsp; Task 1 - **[ ] `print()`** strings that display double and single quotation marks ###Code # [ ] using a print statement, display the text: Where's the homework? print("Where's the homework?") # [ ] output with double quotes: "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein print('"Education is what remains after one has forgotten what one has learned in school" - Albert Einstein') ###Output "Education is what remains after one has forgotten what one has learned in school" - Albert Einstein ###Markdown >**note:**: Quotes in quotes handles only simple cases of displaying quotation marks. More complex case are covered later under *escape sequences.* &nbsp; Concepts Boolean string tests[![view video](https://iajupyterprodblobs.blob.core.windows.net/imagecontainer/common/play_video.png)]( http://edxinteractivepage.blob.core.windows.net/edxpages/f7cff1a7-5601-48a1-95a6-fd1fdfabd20e.html?details=[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.ism/manifest","type":"application/vnd.ms-sstr+xml"}],[{"src":"https://jupyternootbookwams.streaming.mediaservices.windows.net/dfe6e85f-8022-471c-8d92-0b1d61ebffbd/Unit1_Section2-3-Boolean_String_Methods.vtt","srclang":"en","kind":"subtitles","label":"english"}])methods- .isalpha()- .isalnum()- .istitle()- .isdigit()- .islower()- .isupper()- .startswith()type **`str`** has methods that return a Boolean (True or False) for different tests on the properties of stings.>```python"Hello".isapha()```out:[ ] &nbsp; &nbsp; `True` `.isalpha()` returns True if all characters in the string ("Hello") are alphabetical, otherwise returns False &nbsp; Examples Boolean String Tests- **[ ] review and run code in each cell** ###Code "Python".isalpha() "3rd".isalnum() "A Cold Stromy Night".istitle() "1003".isdigit() cm_height = "176" print("cm height:",cm_height, "is all digits =",cm_height.isdigit()) print("SAVE".islower()) print("SAVE".isupper()) "Boolean".startswith("B") ###Output _____no_output_____ ###Markdown &nbsp; Task 2: multi-part test stings with **`.isalpha()`** ###Code # [ ] Use .isalpha() on the string "alphabetical" "alphabetical".isalpha() # [ ] Use .isalpha() on the string: "Are spaces and punctuation Alphabetical?" "Are spaces and punctuation Alphabetical?".isalpha() # [ ] initailize variable alpha_test with input alpha_test=input() # [ ] use .isalpha() on string variable alpha_test alpha_test.isalpha() ###Output 123
Iterables_and_iterators.ipynb
###Markdown IterableIn Python an Iterable is any object that implement the **Iterable protocol**. The requirement to comply with this protocol is to implement the `__iter__()` method and return an **Iterator**. IteratorAll **Iterators** must implement the **Iterable protocol** in addition to implement the `__next__()` method to retrieve elements from the **Iterator**. When there are no more elements available `next()` will raise the `StopIteration` exceptionAs an alternative, the **Iterator protocol** can be implemented with only the `__getitem__()` method that receives an index as parameter. It must return values for consecutive integers, starting from zero, as indexes. When the index is out of range of the data, it will raise the `IndexError` exception. ###Code class ExampleIterator: def __init__(self, data): self._index = 0 self._data = data def __iter__(self): return self def __next__(self): if self._index >= len(self._data): raise StopIteration() result = self._data[self._index] self._index += 1 return result class ExampleIterable: def __init__(self, data): self._data = data def __iter__(self): return ExampleIterator(self._data) sequence = ExampleIterable([1, 2, 3, 4, 5]) for i in sequence: print(i) [i * 2 for i in sequence] class AlternateIterable: def __init__(self, data): self._data = data def __getitem__(self, index): return self._data[index] sequence = AlternateIterable([1, 2, 3, 4, 5]) for i in sequence: print(i) [i * 3 for i in AlternateIterable([1, 2, 3, 4, 5])] ###Output _____no_output_____ ###Markdown `iter()` functionThis function is used implement the **Iterator protocol** for the **callable** that is passed as a parameter.`iter(callable, sentinel)`* callable: is an object that takes zero arguments* sentinel: it's the value used to stop the iterationThis is often used for creating **infite sequences** from existing functions ###Code from datetime import datetime as dt it = iter(dt.now, None) for i in range(10): print(next(it)) ###Output 2019-11-29 08:34:49.384085 2019-11-29 08:34:49.384311 2019-11-29 08:34:49.384351 2019-11-29 08:34:49.384387 2019-11-29 08:34:49.384422 2019-11-29 08:34:49.384456 2019-11-29 08:34:49.384490 2019-11-29 08:34:49.384525 2019-11-29 08:34:49.384559 2019-11-29 08:34:49.384594 ###Markdown Building-block functionsThe idea behind this functions was develop in the **functional programming** paradigm. All these functions implement the **Iterator protocol**. MapApply a function to every element in a sequence. It returns a new sequence with the result.In Python 3 `Map` has a **lazy** implementation, but in Python 2 has an **eager** implementation.It can accept **any number** of input sequences. The number of input sequences **must match** the number of function arguments ###Code def combine(size, colour, animal): return '{}, {}, {}'.format(size, colour, animal) sizes = ['small', 'medium', 'large'] colours = ['red','yellow','blue'] animals = ['dog','cat','duck'] list(map(combine, sizes, colours, animals)) import itertools def combine2(quantity, size, colour, animal): return '{}, {}, {}, {}'.format(quantity, size, colour, animal) list(map(combine2, itertools.count(), sizes, colours, animals)) ###Output _____no_output_____ ###Markdown FilterApply a function to each element in a sequence. It returns a new sequence with the elements for which the functions returns `True`In Python 3 `Filter` has a **lazy** implementation, but in Python 2 has an **eager** implementation.It can only accept a **single** input sequence. The function has to receive a single parameter too.Passing `None` as the first parameter to `Filter` in will return a new sequence without the elements for which the function evaluates to `False` ###Code list(filter(lambda x: x > 0, [1, 4, 7, -6, 0, 2, -7, 10, -55])) list(filter(None, [1, 4, 7, -6, 0, 2, -7, 10, -55])) list(filter(None, [0, 1, False, True, [], [1,2,3], '', 'hello'])) ###Output _____no_output_____ ###Markdown ReduceThe `Reduce` function is part of the `functools` module. It repeatedly apply a function to the elements of a sequence reducing them to a single value.The function provided to the `Reduce` function receives two parameters and must return another value, which it will be the first parameter in the following call to the function.If you pass a sequence with **only one element** to the `Reduce` function, the function provided **will never be called** and it will return the only element in the sequence as a result.The initial value of the accumulator can be passed as a third parameter to the `Reduce` function. Conceptually it is just added at the beginning of the sequence. ###Code from functools import reduce import operator reduce(operator.add, [1, 2, 3, 4, 5]) def mul(x, y): print('mul {} * {}'.format(x, y)) return x * y reduce(mul, [1, 2, 3, 4, 5]) reduce(mul, []) reduce(mul, [1]) reduce(mul, [1, 2, 3], 0) ###Output mul 0 * 1 mul 0 * 2 mul 0 * 3
notebook/MAP-to-NWB.ipynb
###Markdown Session and Subject ###Code session = (experiment.Session & session_key).fetch1() dj.ERD(lab.Subject) -1 + 1 subj = (lab.Subject * lab.CompleteGenotype & session_key).fetch1() session subj # -- NWB file - a NWB2.0 file for each session nwbfile = NWBFile( session_description='', identifier='_'.join( [str(session['subject_id']), str(session['session']), session['session_date'].strftime('%Y-%m-%d_%H-%M-%S')]), session_start_time=datetime.combine(session['session_date'], datetime.min.time()), file_create_date=datetime.now(tzlocal()), experimenter=session['username'], institution='Janelia Research Campus') # -- subject nwbfile.subject = pynwb.file.Subject( subject_id=str(subj['subject_id']), description=f'animal_source: {subj["animal_source"]}', genotype=subj['complete_genotype'], sex=subj['sex']) ###Output _____no_output_____ ###Markdown Units Electrode Group ###Code dj.ERD(ephys.ElectrodeGroup) - 1 + 1 probe = (ephys.ElectrodeGroup & session_key).fetch1() probe device = nwbfile.create_device(name = probe['probe_part_no']) electrode_group = nwbfile.create_electrode_group( name='', description = 'N/A', device = device, location = '') for chn in (ephys.ElectrodeGroup.Electrode & probe).fetch(as_dict=True): nwbfile.add_electrode(id=chn['electrode'], group=electrode_group, filtering='Bandpass filtered 300-6K Hz', imp=-1., location=electrode_group.location, x=np.nan, y=np.nan, z=np.nan) ###Output _____no_output_____ ###Markdown Units ###Code dj.ERD(ephys.Unit) + 1 - 1 units_view = ((ephys.Unit & session_key).aggr( ephys.UnitComment, *ephys.Unit.heading.names, comments='GROUP_CONCAT(unit_comment, "; ")', keep_all_rows=True).aggr( ephys.UnitCellType, "comments", *ephys.Unit.heading.names, cell_type='GROUP_CONCAT(cell_type, "; ")', keep_all_rows=True)) units_view additional_unit_columns = [{'name': tag, 'description': re.sub('\s+:|\s+', ' ', re.search( f'(?<={tag})(.*)', str(units_view.heading)).group())} for tag in units_view.heading.names if tag not in units_view.proj().heading.names + ['spike_times', 'waveform', 'electrode_group', 'electrode']] additional_unit_columns units_view.heading def select(d, *keys): return dict((k, v) for k, v in d.items() if k in keys) for unit in units_view.fetch(as_dict=True): # make an electrode table region (which electrode(s) is this unit coming from) nwbfile.add_unit(id=unit['unit'], electrode_group=electrode_group, **unit, waveform_mean=unit['waveform']) unit = units_view.fetch(as_dict=True)[0] unit['waveform'].shape ###Output _____no_output_____
code/chap21.ipynb
###Markdown Modeling and Simulation in PythonChapter 21Copyright 2017 Allen DowneyLicense: [Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0) ###Code # Configure Jupyter so figures appear in the notebook %matplotlib inline # Configure Jupyter to display the assigned value after an assignment %config InteractiveShell.ast_node_interactivity='last_expr_or_assign' # import functions from the modsim.py module from modsim import * ###Output _____no_output_____ ###Markdown With air resistance Next we'll add air resistance using the [drag equation](https://en.wikipedia.org/wiki/Drag_equation) I'll start by getting the units we'll need from Pint. ###Code m = UNITS.meter s = UNITS.second kg = UNITS.kilogram ###Output _____no_output_____ ###Markdown Now I'll create a `Params` object to contain the quantities we need. Using a Params object is convenient for grouping the system parameters in a way that's easy to read (and double-check). ###Code params = Params(height = 381 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 2.5e-3 * kg, diameter = 19e-3 * m, rho = 1.2 * kg/m**3, v_term = 18 * m / s) ###Output _____no_output_____ ###Markdown Now we can pass the `Params` object `make_system` which computes some additional parameters and defines `init`.`make_system` uses the given radius to compute `area` and the given `v_term` to compute the drag coefficient `C_d`. ###Code def make_system(params): """Makes a System object for the given conditions. params: Params object returns: System object """ unpack(params) area = np.pi * (diameter/2)**2 C_d = 2 * mass * g / (rho * area * v_term**2) init = State(y=height, v=v_init) t_end = 30 * s return System(params, area=area, C_d=C_d, init=init, t_end=t_end) ###Output _____no_output_____ ###Markdown Let's make a `System` ###Code system = make_system(params) ###Output _____no_output_____ ###Markdown Here's the slope function, including acceleration due to gravity and drag. ###Code def slope_func(state, t, system): """Compute derivatives of the state. state: position, velocity t: time system: System object returns: derivatives of y and v """ y, v = state unpack(system) f_drag = rho * v**2 * C_d * area / 2 a_drag = f_drag / mass dydt = v dvdt = -g + a_drag return dydt, dvdt ###Output _____no_output_____ ###Markdown As always, let's test the slope function with the initial conditions. ###Code slope_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown We can use the same event function as in the previous chapter. ###Code def event_func(state, t, system): """Return the height of the penny above the sidewalk. """ y, v = state return y ###Output _____no_output_____ ###Markdown And then run the simulation. ###Code results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.5*s) details.message ###Output _____no_output_____ ###Markdown Here are the results. ###Code results ###Output _____no_output_____ ###Markdown The final height is close to 0, as expected.Interestingly, the final velocity is not exactly terminal velocity, which suggests that there are some numerical errors.We can get the flight time from `results`. ###Code t_sidewalk = get_last_label(results) ###Output _____no_output_____ ###Markdown Here's the plot of position as a function of time. ###Code def plot_position(results): plot(results.y) decorate(xlabel='Time (s)', ylabel='Position (m)') plot_position(results) savefig('figs/chap09-fig02.pdf') ###Output _____no_output_____ ###Markdown And velocity as a function of time: ###Code def plot_velocity(results): plot(results.v, color='C1', label='v') decorate(xlabel='Time (s)', ylabel='Velocity (m/s)') plot_velocity(results) ###Output _____no_output_____ ###Markdown From an initial velocity of 0, the penny accelerates downward until it reaches terminal velocity; after that, velocity is constant. **Exercise:** Run the simulation with an initial velocity, downward, that exceeds the penny's terminal velocity. Hint: You can create a new `Params` object based on an existing one, like this:`params = Params(params, v_init = -30 * m / s)`What do you expect to happen? Plot velocity and position as a function of time, and see if they are consistent with your prediction. ###Code # Solution goes here plot_position(results) # Solution goes here ###Output _____no_output_____ ###Markdown **Exercise:** Suppose we drop a quarter from the Empire State Building and find that its flight time is 19.1 seconds. Use this measurement to estimate the terminal velocity.1. You can get the relevant dimensions of a quarter from https://en.wikipedia.org/wiki/Quarter_(United_States_coin).2. Create a `Params` object with the system parameters. We don't know `v_term`, so we'll start with the inital guess `v_term = 18 * m / s`.3. Use `make_system` to create a `System` object.4. Call `run_ode_solver` to simulate the system. How does the flight time of the simulation compare to the measurement?5. Try a few different values of `t_term` and see if you can get the simulated flight time close to 19.1 seconds.6. Optionally, write an error function and use `fsolve` to improve your estimate.7. Use your best estimate of `v_term` to compute `C_d`.Note: I fabricated the observed flight time, so don't take the results of this exercise too seriously. ###Code # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown Bungee jumping Suppose you want to set the world record for the highest "bungee dunk", [as shown in this video](https://www.youtube.com/watch?v=UBf7WC19lpw). Since the record is 70 m, let's design a jump for 80 m.We'll make the following modeling assumptions:1. Initially the bungee cord hangs from a crane with the attachment point 80 m above a cup of tea.2. Until the cord is fully extended, it applies no force to the jumper. It turns out this might not be a good assumption; we will revisit it.3. After the cord is fully extended, it obeys [Hooke's Law](https://en.wikipedia.org/wiki/Hooke%27s_law); that is, it applies a force to the jumper proportional to the extension of the cord beyond its resting length.4. The jumper is subject to drag force proportional to the square of their velocity, in the opposite of their direction of motion.Our objective is to choose the length of the cord, `L`, and its spring constant, `k`, so that the jumper falls all the way to the tea cup, but no farther! First I'll create a `Param` object to contain the quantities we'll need:1. Let's assume that the jumper's mass is 75 kg.2. With a terminal velocity of 60 m/s.3. The length of the bungee cord is `L = 40 m`.4. The spring constant of the cord is `k = 20 N / m` when the cord is stretched, and 0 when it's compressed. ###Code m = UNITS.meter s = UNITS.second kg = UNITS.kilogram N = UNITS.newton params = Params(y_attach = 80 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 75 * kg, area = 1 * m**2, rho = 1.2 * kg/m**3, v_term = 60 * m / s, L = 25 * m, k = 40 * N / m) ###Output _____no_output_____ ###Markdown Now here's a version of `make_system` that takes a `Params` object as a parameter.`make_system` uses the given value of `v_term` to compute the drag coefficient `C_d`. ###Code def make_system(params): """Makes a System object for the given params. params: Params object returns: System object """ unpack(params) C_d = 2 * mass * g / (rho * area * v_term**2) init = State(y=y_attach, v=v_init) t_end = 20 * s return System(params, C_d=C_d, init=init, t_end=t_end) ###Output _____no_output_____ ###Markdown Let's make a `System` ###Code system = make_system(params) ###Output _____no_output_____ ###Markdown `spring_force` computes the force of the cord on the jumper: ###Code def spring_force(y, system): """Computes the force of the bungee cord on the jumper: y: height of the jumper Uses these variables from system| y_attach: height of the attachment point L: resting length of the cord k: spring constant of the cord returns: force in N """ unpack(system) distance_fallen = y_attach - y if distance_fallen <= L: return 0 * N extension = distance_fallen - L f_spring = k * extension return f_spring ###Output _____no_output_____ ###Markdown The spring force is 0 until the cord is fully extended. When it is extended 1 m, the spring force is 40 N. ###Code spring_force(80*m, system) spring_force(55*m, system) spring_force(54*m, system) ###Output _____no_output_____ ###Markdown `drag_force` computes drag as a function of velocity: ###Code def drag_force(v, system): """Computes drag force in the opposite direction of `v`. v: velocity system: System object returns: drag force """ unpack(system) f_drag = -np.sign(v) * rho * v**2 * C_d * area / 2 return f_drag ###Output _____no_output_____ ###Markdown Here's the drag force at 60 meters per second. ###Code v = -60 * m/s f_drag = drag_force(v, system) ###Output _____no_output_____ ###Markdown Acceleration due to drag at 60 m/s is approximately g, which confirms that 60 m/s is terminal velocity. ###Code a_drag = f_drag / system.mass ###Output _____no_output_____ ###Markdown Now here's the slope function: ###Code def slope_func(state, t, system): """Compute derivatives of the state. state: position, velocity t: time system: System object containing g, rho, C_d, area, and mass returns: derivatives of y and v """ y, v = state unpack(system) a_drag = drag_force(v, system) / mass a_spring = spring_force(y, system) / mass dvdt = -g + a_drag + a_spring return v, dvdt ###Output _____no_output_____ ###Markdown As always, let's test the slope function with the initial params. ###Code slope_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown And then run the simulation. ###Code results, details = run_ode_solver(system, slope_func, max_step=0.3*s) details ###Output _____no_output_____ ###Markdown Here's the plot of position as a function of time. ###Code plot_position(results) ###Output _____no_output_____ ###Markdown After reaching the lowest point, the jumper springs back almost to almost 70 m, and oscillates several times. That looks like more osciallation that we expect from an actual jump, which suggests that there some dissipation of energy in the real world that is not captured in our model. To improve the model, that might be a good thing to investigate.But since we are primarily interested in the initial descent, the model might be good enough for now.We can use `min` to find the lowest point: ###Code min(results.y) ###Output _____no_output_____ ###Markdown At the lowest point, the jumper is still too high, so we'll need to increase `L` or decrease `k`. Here's velocity as a function of time: ###Code plot_velocity(results) subplot(1, 2, 1) plot_position(results) subplot(1, 2, 2) plot_velocity(results) savefig('figs/chap09-fig03.pdf') ###Output _____no_output_____ ###Markdown Although we compute acceleration inside the slope function, we don't get acceleration as a result from `run_ode_solver`.We can approximate it by computing the numerical derivative of `ys`: ###Code a = gradient(results.v) plot(a) decorate(xlabel='Time (s)', ylabel='Acceleration (m/$s^2$)') ###Output _____no_output_____ ###Markdown And we can compute the maximum acceleration the jumper experiences: ###Code max_acceleration = max(a) * m/s**2 ###Output _____no_output_____ ###Markdown Relative to the acceleration of gravity, the jumper "pulls" about "1.7 g's". ###Code max_acceleration / g ###Output _____no_output_____ ###Markdown Under the hoodThe gradient function in `modsim.py` adapts the NumPy function of the same name so it works with `Series` objects. ###Code %psource gradient ###Output _____no_output_____ ###Markdown Solving for lengthAssuming that `k` is fixed, let's find the length `L` that makes the minimum altitude of the jumper exactly 0. The metric we are interested in is the lowest point of the first oscillation. For both efficiency and accuracy, it is better to stop the simulation when we reach this point, rather than run past it and the compute the minimum.Here's an event function that stops the simulation when velocity is 0. ###Code def event_func(state, t, system): """Return velocity. """ y, v = state return v ###Output _____no_output_____ ###Markdown As usual, we should test it with the initial conditions. ###Code event_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown And we see that we have a problem. Since the event function returns 0 under the initial conditions, the simulation would stop immediately. We can solve that problem by specifying the direction of the event function: ###Code event_func.direction = +1 ###Output _____no_output_____ ###Markdown When direction is positive, it only stops the simulation if the velocity is 0 and increasing, which is what we want. Now we can test it an confirm that it stops at the bottom of the jump. ###Code results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.3*s) plot_position(results) min(results.y) ###Output _____no_output_____ ###Markdown **Exercise:** Write an error function that takes `L` and `params` as arguments, simulates a bungee jump, and returns the lowest point.Test the error function with a guess of 25 m and confirm that the return value is about 5 meters.Use `fsolve` with your error function to find the value of `L` that yields a perfect bungee dunk.Run a simulation with the result from `fsolve` and confirm that it works. ###Code # Solution goes here # Solution goes here # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown **Optional exercise:** Search for the combination of length and spring constant that yields minimum height 0 while minimizing peak acceleration. ###Code # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown Modeling and Simulation in PythonChapter 21Copyright 2017 Allen DowneyLicense: [Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0) ###Code # Configure Jupyter so figures appear in the notebook %matplotlib inline # Configure Jupyter to display the assigned value after an assignment %config InteractiveShell.ast_node_interactivity='last_expr_or_assign' # import functions from the modsim.py module from modsim import * ###Output _____no_output_____ ###Markdown With air resistance Next we'll add air resistance using the [drag equation](https://en.wikipedia.org/wiki/Drag_equation) I'll start by getting the units we'll need from Pint. ###Code m = UNITS.meter s = UNITS.second kg = UNITS.kilogram ###Output _____no_output_____ ###Markdown Now I'll create a `Params` object to contain the quantities we need. Using a Params object is convenient for grouping the system parameters in a way that's easy to read (and double-check). ###Code params = Params(height = 381 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 2.5e-3 * kg, diameter = 19e-3 * m, rho = 1.2 * kg/m**3, v_term = 18 * m / s) ###Output _____no_output_____ ###Markdown Now we can pass the `Params` object `make_system` which computes some additional parameters and defines `init`.`make_system` uses the given radius to compute `area` and the given `v_term` to compute the drag coefficient `C_d`. ###Code def make_system(params): """Makes a System object for the given conditions. params: Params object returns: System object """ unpack(params) area = np.pi * (diameter/2)**2 C_d = 2 * mass * g / (rho * area * v_term**2) init = State(y=height, v=v_init) t_end = 30 * s return System(params, area=area, C_d=C_d, init=init, t_end=t_end) ###Output _____no_output_____ ###Markdown Let's make a `System` ###Code system = make_system(params) ###Output _____no_output_____ ###Markdown Here's the slope function, including acceleration due to gravity and drag. ###Code def slope_func(state, t, system): """Compute derivatives of the state. state: position, velocity t: time system: System object returns: derivatives of y and v """ y, v = state unpack(system) f_drag = rho * v**2 * C_d * area / 2 a_drag = f_drag / mass dydt = v dvdt = -g + a_drag return dydt, dvdt ###Output _____no_output_____ ###Markdown As always, let's test the slope function with the initial conditions. ###Code slope_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown We can use the same event function as in the previous chapter. ###Code def event_func(state, t, system): """Return the height of the penny above the sidewalk. """ y, v = state return y ###Output _____no_output_____ ###Markdown And then run the simulation. ###Code results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.5*s) details.message ###Output _____no_output_____ ###Markdown Here are the results. ###Code results ###Output _____no_output_____ ###Markdown The final height is close to 0, as expected.Interestingly, the final velocity is not exactly terminal velocity, which suggests that there are some numerical errors.We can get the flight time from `results`. ###Code t_sidewalk = get_last_label(results) ###Output _____no_output_____ ###Markdown Here's the plot of position as a function of time. ###Code def plot_position(results): plot(results.y) decorate(xlabel='Time (s)', ylabel='Position (m)') plot_position(results) savefig('figs/chap09-fig02.pdf') ###Output _____no_output_____ ###Markdown And velocity as a function of time: ###Code def plot_velocity(results): plot(results.v, color='C1', label='v') decorate(xlabel='Time (s)', ylabel='Velocity (m/s)') plot_velocity(results) ###Output _____no_output_____ ###Markdown From an initial velocity of 0, the penny accelerates downward until it reaches terminal velocity; after that, velocity is constant. **Exercise:** Run the simulation with an initial velocity, downward, that exceeds the penny's terminal velocity. Hint: You can create a new `Params` object based on an existing one, like this:`params = Params(params, v_init = -30 * m / s)`What do you expect to happen? Plot velocity and position as a function of time, and see if they are consistent with your prediction. ###Code # Solution goes here plot_position(results) # Solution goes here ###Output _____no_output_____ ###Markdown **Exercise:** Suppose we drop a quarter from the Empire State Building and find that its flight time is 19.1 seconds. Use this measurement to estimate the terminal velocity.1. You can get the relevant dimensions of a quarter from https://en.wikipedia.org/wiki/Quarter_(United_States_coin).2. Create a `Params` object with the system parameters. We don't know `v_term`, so we'll start with the inital guess `v_term = 18 * m / s`.3. Use `make_system` to create a `System` object.4. Call `run_ode_solver` to simulate the system. How does the flight time of the simulation compare to the measurement?5. Try a few different values of `t_term` and see if you can get the simulated flight time close to 19.1 seconds.6. Optionally, write an error function and use `fsolve` to improve your estimate.7. Use your best estimate of `v_term` to compute `C_d`.Note: I fabricated the observed flight time, so don't take the results of this exercise too seriously. ###Code # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown Bungee jumping Suppose you want to set the world record for the highest "bungee dunk", [as shown in this video](https://www.youtube.com/watch?v=UBf7WC19lpw). Since the record is 70 m, let's design a jump for 80 m.We'll make the following modeling assumptions:1. Initially the bungee cord hangs from a crane with the attachment point 80 m above a cup of tea.2. Until the cord is fully extended, it applies no force to the jumper. It turns out this might not be a good assumption; we will revisit it.3. After the cord is fully extended, it obeys [Hooke's Law](https://en.wikipedia.org/wiki/Hooke%27s_law); that is, it applies a force to the jumper proportional to the extension of the cord beyond its resting length.4. The jumper is subject to drag force proportional to the square of their velocity, in the opposite of their direction of motion.Our objective is to choose the length of the cord, `L`, and its spring constant, `k`, so that the jumper falls all the way to the tea cup, but no farther! First I'll create a `Param` object to contain the quantities we'll need:1. Let's assume that the jumper's mass is 75 kg.2. With a terminal velocity of 60 m/s.3. The length of the bungee cord is `L = 40 m`.4. The spring constant of the cord is `k = 20 N / m` when the cord is stretched, and 0 when it's compressed. ###Code m = UNITS.meter s = UNITS.second kg = UNITS.kilogram N = UNITS.newton params = Params(y_attach = 80 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 75 * kg, area = 1 * m**2, rho = 1.2 * kg/m**3, v_term = 60 * m / s, L = 25 * m, k = 40 * N / m) ###Output _____no_output_____ ###Markdown Now here's a version of `make_system` that takes a `Params` object as a parameter.`make_system` uses the given value of `v_term` to compute the drag coefficient `C_d`. ###Code def make_system(params): """Makes a System object for the given params. params: Params object returns: System object """ unpack(params) C_d = 2 * mass * g / (rho * area * v_term**2) init = State(y=y_attach, v=v_init) t_end = 20 * s return System(params, C_d=C_d, init=init, t_end=t_end) ###Output _____no_output_____ ###Markdown Let's make a `System` ###Code system = make_system(params) ###Output _____no_output_____ ###Markdown `spring_force` computes the force of the cord on the jumper: ###Code def spring_force(y, system): """Computes the force of the bungee cord on the jumper: y: height of the jumper Uses these variables from system| y_attach: height of the attachment point L: resting length of the cord k: spring constant of the cord returns: force in N """ unpack(system) distance_fallen = y_attach - y if distance_fallen <= L: return 0 * N extension = distance_fallen - L f_spring = k * extension return f_spring ###Output _____no_output_____ ###Markdown The spring force is 0 until the cord is fully extended. When it is extended 1 m, the spring force is 40 N. ###Code spring_force(80*m, system) spring_force(55*m, system) spring_force(54*m, system) ###Output _____no_output_____ ###Markdown `drag_force` computes drag as a function of velocity: ###Code def drag_force(v, system): """Computes drag force in the opposite direction of `v`. v: velocity system: System object returns: drag force """ unpack(system) f_drag = -np.sign(v) * rho * v**2 * C_d * area / 2 return f_drag ###Output _____no_output_____ ###Markdown Here's the drag force at 60 meters per second. ###Code v = -60 * m/s f_drag = drag_force(v, system) ###Output _____no_output_____ ###Markdown Acceleration due to drag at 60 m/s is approximately g, which confirms that 60 m/s is terminal velocity. ###Code a_drag = f_drag / system.mass ###Output _____no_output_____ ###Markdown Now here's the slope function: ###Code def slope_func(state, t, system): """Compute derivatives of the state. state: position, velocity t: time system: System object containing g, rho, C_d, area, and mass returns: derivatives of y and v """ y, v = state unpack(system) a_drag = drag_force(v, system) / mass a_spring = spring_force(y, system) / mass dvdt = -g + a_drag + a_spring return v, dvdt ###Output _____no_output_____ ###Markdown As always, let's test the slope function with the initial params. ###Code slope_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown And then run the simulation. ###Code results, details = run_ode_solver(system, slope_func, max_step=0.3*s) details ###Output _____no_output_____ ###Markdown Here's the plot of position as a function of time. ###Code plot_position(results) ###Output _____no_output_____ ###Markdown After reaching the lowest point, the jumper springs back almost to almost 70 m, and oscillates several times. That looks like more osciallation that we expect from an actual jump, which suggests that there some dissipation of energy in the real world that is not captured in our model. To improve the model, that might be a good thing to investigate.But since we are primarily interested in the initial descent, the model might be good enough for now.We can use `min` to find the lowest point: ###Code min(results.y) ###Output _____no_output_____ ###Markdown At the lowest point, the jumper is still too high, so we'll need to increase `L` or decrease `k`. Here's velocity as a function of time: ###Code plot_velocity(results) subplot(1, 2, 1) plot_position(results) subplot(1, 2, 2) plot_velocity(results) savefig('figs/chap09-fig03.pdf') ###Output _____no_output_____ ###Markdown Although we compute acceleration inside the slope function, we don't get acceleration as a result from `run_ode_solver`.We can approximate it by computing the numerical derivative of `ys`: ###Code a = gradient(results.v) plot(a) decorate(xlabel='Time (s)', ylabel='Acceleration (m/$s^2$)') ###Output _____no_output_____ ###Markdown And we can compute the maximum acceleration the jumper experiences: ###Code max_acceleration = max(a) * m/s**2 ###Output _____no_output_____ ###Markdown Relative to the acceleration of gravity, the jumper "pulls" about "1.7 g's". ###Code max_acceleration / g ###Output _____no_output_____ ###Markdown Under the hoodThe gradient function in `modsim.py` adapts the NumPy function of the same name so it works with `Series` objects. ###Code %psource gradient ###Output _____no_output_____ ###Markdown Solving for lengthAssuming that `k` is fixed, let's find the length `L` that makes the minimum altitude of the jumper exactly 0. The metric we are interested in is the lowest point of the first oscillation. For both efficiency and accuracy, it is better to stop the simulation when we reach this point, rather than run past it and the compute the minimum.Here's an event function that stops the simulation when velocity is 0. ###Code def event_func(state, t, system): """Return velocity. """ y, v = state return v ###Output _____no_output_____ ###Markdown As usual, we should test it with the initial conditions. ###Code event_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown And we see that we have a problem. Since the event function returns 0 under the initial conditions, the simulation would stop immediately. We can solve that problem by specifying the direction of the event function: ###Code event_func.direction = +1 ###Output _____no_output_____ ###Markdown When direction is positive, it only stops the simulation if the velocity is 0 and increasing, which is what we want. Now we can test it an confirm that it stops at the bottom of the jump. ###Code results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.3*s) plot_position(results) min(results.y) ###Output _____no_output_____ ###Markdown **Exercise:** Write an error function that takes `L` and `params` as arguments, simulates a bungee jump, and returns the lowest point.Test the error function with a guess of 25 m and confirm that the return value is about 5 meters.Use `fsolve` with your error function to find the value of `L` that yields a perfect bungee dunk.Run a simulation with the result from `fsolve` and confirm that it works. ###Code # Solution goes here # Solution goes here # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown **Optional exercise:** Search for the combination of length and spring constant that yields minimum height 0 while minimizing peak acceleration. ###Code # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown Modeling and Simulation in PythonChapter 21Copyright 2017 Allen DowneyLicense: [Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0) ###Code # Configure Jupyter so figures appear in the notebook %matplotlib inline # Configure Jupyter to display the assigned value after an assignment %config InteractiveShell.ast_node_interactivity='last_expr_or_assign' # import functions from the modsim.py module from modsim import * ###Output _____no_output_____ ###Markdown With air resistance Next we'll add air resistance using the [drag equation](https://en.wikipedia.org/wiki/Drag_equation) I'll start by getting the units we'll need from Pint. ###Code m = UNITS.meter s = UNITS.second kg = UNITS.kilogram ###Output _____no_output_____ ###Markdown Now I'll create a `Params` object to contain the quantities we need. Using a Params object is convenient for grouping the system parameters in a way that's easy to read (and double-check). ###Code params = Params(height = 381 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 2.5e-3 * kg, diameter = 19e-3 * m, rho = 1.2 * kg/m**3, v_term = 18 * m / s) ###Output _____no_output_____ ###Markdown Now we can pass the `Params` object `make_system` which computes some additional parameters and defines `init`.`make_system` uses the given radius to compute `area` and the given `v_term` to compute the drag coefficient `C_d`. ###Code def make_system(params): """Makes a System object for the given conditions. params: Params object returns: System object """ unpack(params) area = np.pi * (diameter/2)**2 C_d = 2 * mass * g / (rho * area * v_term**2) init = State(y=height, v=v_init) t_end = 30 * s return System(params, area=area, C_d=C_d, init=init, t_end=t_end) ###Output _____no_output_____ ###Markdown Let's make a `System` ###Code system = make_system(params) ###Output _____no_output_____ ###Markdown Here's the slope function, including acceleration due to gravity and drag. ###Code def slope_func(state, t, system): """Compute derivatives of the state. state: position, velocity t: time system: System object returns: derivatives of y and v """ y, v = state unpack(system) f_drag = rho * v**2 * C_d * area / 2 a_drag = f_drag / mass dydt = v dvdt = -g + a_drag return dydt, dvdt ###Output _____no_output_____ ###Markdown As always, let's test the slope function with the initial conditions. ###Code slope_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown We can use the same event function as in the previous chapter. ###Code def event_func(state, t, system): """Return the height of the penny above the sidewalk. """ y, v = state return y ###Output _____no_output_____ ###Markdown And then run the simulation. ###Code results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.5*s) details.message ###Output _____no_output_____ ###Markdown Here are the results. ###Code results ###Output _____no_output_____ ###Markdown The final height is close to 0, as expected.Interestingly, the final velocity is not exactly terminal velocity, which suggests that there are some numerical errors.We can get the flight time from `results`. ###Code t_sidewalk = get_last_label(results) ###Output _____no_output_____ ###Markdown Here's the plot of position as a function of time. ###Code def plot_position(results): plot(results.y) decorate(xlabel='Time (s)', ylabel='Position (m)') plot_position(results) savefig('figs/chap09-fig02.pdf') ###Output _____no_output_____ ###Markdown And velocity as a function of time: ###Code def plot_velocity(results): plot(results.v, color='C1', label='v') decorate(xlabel='Time (s)', ylabel='Velocity (m/s)') plot_velocity(results) ###Output _____no_output_____ ###Markdown From an initial velocity of 0, the penny accelerates downward until it reaches terminal velocity; after that, velocity is constant. **Exercise:** Run the simulation with an initial velocity, downward, that exceeds the penny's terminal velocity. Hint: You can create a new `Params` object based on an existing one, like this:`params = Params(params, v_init = -30 * m / s)`What do you expect to happen? Plot velocity and position as a function of time, and see if they are consistent with your prediction. ###Code # Solution goes here plot_position(results) # Solution goes here ###Output _____no_output_____ ###Markdown **Exercise:** Suppose we drop a quarter from the Empire State Building and find that its flight time is 19.1 seconds. Use this measurement to estimate the terminal velocity.1. You can get the relevant dimensions of a quarter from https://en.wikipedia.org/wiki/Quarter_(United_States_coin).2. Create a `Params` object with the system parameters. We don't know `v_term`, so we'll start with the inital guess `v_term = 18 * m / s`.3. Use `make_system` to create a `System` object.4. Call `run_ode_solver` to simulate the system. How does the flight time of the simulation compare to the measurement?5. Try a few different values of `t_term` and see if you can get the simulated flight time close to 19.1 seconds.6. Optionally, write an error function and use `fsolve` to improve your estimate.7. Use your best estimate of `v_term` to compute `C_d`.Note: I fabricated the observed flight time, so don't take the results of this exercise too seriously. ###Code # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown Bungee jumping Suppose you want to set the world record for the highest "bungee dunk", [as shown in this video](https://www.youtube.com/watch?v=UBf7WC19lpw). Since the record is 70 m, let's design a jump for 80 m.We'll make the following modeling assumptions:1. Initially the bungee cord hangs from a crane with the attachment point 80 m above a cup of tea.2. Until the cord is fully extended, it applies no force to the jumper. It turns out this might not be a good assumption; we will revisit it.3. After the cord is fully extended, it obeys [Hooke's Law](https://en.wikipedia.org/wiki/Hooke%27s_law); that is, it applies a force to the jumper proportional to the extension of the cord beyond its resting length.4. The jumper is subject to drag force proportional to the square of their velocity, in the opposite of their direction of motion.Our objective is to choose the length of the cord, `L`, and its spring constant, `k`, so that the jumper falls all the way to the tea cup, but no farther! First I'll create a `Param` object to contain the quantities we'll need:1. Let's assume that the jumper's mass is 75 kg.2. With a terminal velocity of 60 m/s.3. The length of the bungee cord is `L = 40 m`.4. The spring constant of the cord is `k = 20 N / m` when the cord is stretched, and 0 when it's compressed. ###Code m = UNITS.meter s = UNITS.second kg = UNITS.kilogram N = UNITS.newton params = Params(y_attach = 80 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 75 * kg, area = 1 * m**2, rho = 1.2 * kg/m**3, v_term = 60 * m / s, L = 25 * m, k = 40 * N / m) ###Output _____no_output_____ ###Markdown Now here's a version of `make_system` that takes a `Params` object as a parameter.`make_system` uses the given value of `v_term` to compute the drag coefficient `C_d`. ###Code def make_system(params): """Makes a System object for the given params. params: Params object returns: System object """ unpack(params) C_d = 2 * mass * g / (rho * area * v_term**2) init = State(y=y_attach, v=v_init) t_end = 20 * s return System(params, C_d=C_d, init=init, t_end=t_end) ###Output _____no_output_____ ###Markdown Let's make a `System` ###Code system = make_system(params) ###Output _____no_output_____ ###Markdown `spring_force` computes the force of the cord on the jumper: ###Code def spring_force(y, system): """Computes the force of the bungee cord on the jumper: y: height of the jumper Uses these variables from system| y_attach: height of the attachment point L: resting length of the cord k: spring constant of the cord returns: force in N """ unpack(system) distance_fallen = y_attach - y if distance_fallen <= L: return 0 * N extension = distance_fallen - L f_spring = k * extension return f_spring ###Output _____no_output_____ ###Markdown The spring force is 0 until the cord is fully extended. When it is extended 1 m, the spring force is 40 N. ###Code spring_force(80*m, system) spring_force(55*m, system) spring_force(54*m, system) ###Output _____no_output_____ ###Markdown `drag_force` computes drag as a function of velocity: ###Code def drag_force(v, system): """Computes drag force in the opposite direction of `v`. v: velocity system: System object returns: drag force """ unpack(system) f_drag = -np.sign(v) * rho * v**2 * C_d * area / 2 return f_drag ###Output _____no_output_____ ###Markdown Here's the drag force at 60 meters per second. ###Code v = -60 * m/s f_drag = drag_force(v, system) ###Output _____no_output_____ ###Markdown Acceleration due to drag at 60 m/s is approximately g, which confirms that 60 m/s is terminal velocity. ###Code a_drag = f_drag / system.mass ###Output _____no_output_____ ###Markdown Now here's the slope function: ###Code def slope_func(state, t, system): """Compute derivatives of the state. state: position, velocity t: time system: System object containing g, rho, C_d, area, and mass returns: derivatives of y and v """ y, v = state unpack(system) a_drag = drag_force(v, system) / mass a_spring = spring_force(y, system) / mass dvdt = -g + a_drag + a_spring return v, dvdt ###Output _____no_output_____ ###Markdown As always, let's test the slope function with the initial params. ###Code slope_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown And then run the simulation. ###Code results, details = run_ode_solver(system, slope_func, max_step=0.3*s) details ###Output _____no_output_____ ###Markdown Here's the plot of position as a function of time. ###Code plot_position(results) ###Output _____no_output_____ ###Markdown After reaching the lowest point, the jumper springs back almost to almost 70 m, and oscillates several times. That looks like more osciallation that we expect from an actual jump, which suggests that there some dissipation of energy in the real world that is not captured in our model. To improve the model, that might be a good thing to investigate.But since we are primarily interested in the initial descent, the model might be good enough for now.We can use `min` to find the lowest point: ###Code min(results.y) ###Output _____no_output_____ ###Markdown At the lowest point, the jumper is still too high, so we'll need to increase `L` or decrease `k`. Here's velocity as a function of time: ###Code plot_velocity(results) subplot(1, 2, 1) plot_position(results) subplot(1, 2, 2) plot_velocity(results) savefig('figs/chap09-fig03.pdf') ###Output _____no_output_____ ###Markdown Although we compute acceleration inside the slope function, we don't get acceleration as a result from `run_ode_solver`.We can approximate it by computing the numerical derivative of `ys`: ###Code a = gradient(results.v) plot(a) decorate(xlabel='Time (s)', ylabel='Acceleration (m/$s^2$)') ###Output _____no_output_____ ###Markdown And we can compute the maximum acceleration the jumper experiences: ###Code max_acceleration = max(a) * m/s**2 ###Output _____no_output_____ ###Markdown Relative to the acceleration of gravity, the jumper "pulls" about "1.7 g's". ###Code max_acceleration / g ###Output _____no_output_____ ###Markdown Under the hoodThe gradient function in `modsim.py` adapts the NumPy function of the same name so it works with `Series` objects. ###Code %psource gradient ###Output _____no_output_____ ###Markdown Solving for lengthAssuming that `k` is fixed, let's find the length `L` that makes the minimum altitude of the jumper exactly 0. The metric we are interested in is the lowest point of the first oscillation. For both efficiency and accuracy, it is better to stop the simulation when we reach this point, rather than run past it and the compute the minimum.Here's an event function that stops the simulation when velocity is 0. ###Code def event_func(state, t, system): """Return velocity. """ y, v = state return v ###Output _____no_output_____ ###Markdown As usual, we should test it with the initial conditions. ###Code event_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown And we see that we have a problem. Since the event function returns 0 under the initial conditions, the simulation would stop immediately. We can solve that problem by specifying the direction of the event function: ###Code event_func.direction = +1 ###Output _____no_output_____ ###Markdown When direction is positive, it only stops the simulation if the velocity is 0 and increasing, which is what we want. Now we can test it an confirm that it stops at the bottom of the jump. ###Code results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.3*s) plot_position(results) min(results.y) ###Output _____no_output_____ ###Markdown **Exercise:** Write an error function that takes `L` and `params` as arguments, simulates a bungee jump, and returns the lowest point.Test the error function with a guess of 25 m and confirm that the return value is about 5 meters.Use `fsolve` with your error function to find the value of `L` that yields a perfect bungee dunk.Run a simulation with the result from `fsolve` and confirm that it works. ###Code # Solution goes here # Solution goes here # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown **Optional exercise:** Search for the combination of length and spring constant that yields minimum height 0 while minimizing peak acceleration. ###Code # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown Modeling and Simulation in PythonChapter 21Copyright 2017 Allen DowneyLicense: [Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0) ###Code # Configure Jupyter so figures appear in the notebook %matplotlib inline # Configure Jupyter to display the assigned value after an assignment %config InteractiveShell.ast_node_interactivity='last_expr_or_assign' # import functions from the modsim.py module from modsim import * ###Output _____no_output_____ ###Markdown With air resistance Next we'll add air resistance using the [drag equation](https://en.wikipedia.org/wiki/Drag_equation) I'll start by getting the units we'll need from Pint. ###Code m = UNITS.meter s = UNITS.second kg = UNITS.kilogram ###Output _____no_output_____ ###Markdown Now I'll create a `Params` object to contain the quantities we need. Using a Params object is convenient for grouping the system parameters in a way that's easy to read (and double-check). ###Code params = Params(height = 381 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 2.5e-3 * kg, diameter = 19e-3 * m, rho = 1.2 * kg/m**3, v_term = 18 * m / s) ###Output _____no_output_____ ###Markdown Now we can pass the `Params` object `make_system` which computes some additional parameters and defines `init`.`make_system` uses the given radius to compute `area` and the given `v_term` to compute the drag coefficient `C_d`. ###Code def make_system(params): """Makes a System object for the given conditions. params: Params object returns: System object """ unpack(params) area = np.pi * (diameter/2)**2 C_d = 2 * mass * g / (rho * area * v_term**2) init = State(y=height, v=v_init) t_end = 30 * s return System(params, area=area, C_d=C_d, init=init, t_end=t_end) ###Output _____no_output_____ ###Markdown Let's make a `System` ###Code system = make_system(params) ###Output _____no_output_____ ###Markdown Here's the slope function, including acceleration due to gravity and drag. ###Code def slope_func(state, t, system): """Compute derivatives of the state. state: position, velocity t: time system: System object returns: derivatives of y and v """ y, v = state unpack(system) f_drag = rho * v**2 * C_d * area / 2 a_drag = f_drag / mass dydt = v dvdt = -g + a_drag return dydt, dvdt ###Output _____no_output_____ ###Markdown As always, let's test the slope function with the initial conditions. ###Code slope_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown We can use the same event function as in the previous chapter. ###Code def event_func(state, t, system): """Return the height of the penny above the sidewalk. """ y, v = state return y ###Output _____no_output_____ ###Markdown And then run the simulation. ###Code results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.5*s) details.message ###Output _____no_output_____ ###Markdown Here are the results. ###Code results ###Output _____no_output_____ ###Markdown The final height is close to 0, as expected.Interestingly, the final velocity is not exactly terminal velocity, which suggests that there are some numerical errors.We can get the flight time from `results`. ###Code t_sidewalk = get_last_label(results) ###Output _____no_output_____ ###Markdown Here's the plot of position as a function of time. ###Code def plot_position(results): plot(results.y) decorate(xlabel='Time (s)', ylabel='Position (m)') plot_position(results) savefig('figs/chap09-fig02.pdf') ###Output Saving figure to file figs/chap09-fig02.pdf ###Markdown And velocity as a function of time: ###Code def plot_velocity(results): plot(results.v, color='C1', label='v') decorate(xlabel='Time (s)', ylabel='Velocity (m/s)') plot_velocity(results) ###Output _____no_output_____ ###Markdown From an initial velocity of 0, the penny accelerates downward until it reaches terminal velocity; after that, velocity is constant. **Exercise:** Run the simulation with an initial velocity, downward, that exceeds the penny's terminal velocity. Hint: You can create a new `Params` object based on an existing one, like this:`params = Params(params, v_init = -30 * m / s)`What do you expect to happen? Plot velocity and position as a function of time, and see if they are consistent with your prediction. ###Code # Solution goes here plot_position(results) # Solution goes here ###Output _____no_output_____ ###Markdown **Exercise:** Suppose we drop a quarter from the Empire State Building and find that its flight time is 19.1 seconds. Use this measurement to estimate the terminal velocity.1. You can get the relevant dimensions of a quarter from https://en.wikipedia.org/wiki/Quarter_(United_States_coin).2. Create a `Params` object with the system parameters. We don't know `v_term`, so we'll start with the inital guess `v_term = 18 * m / s`.3. Use `make_system` to create a `System` object.4. Call `run_ode_solver` to simulate the system. How does the flight time of the simulation compare to the measurement?5. Try a few different values of `t_term` and see if you can get the simulated flight time close to 19.1 seconds.6. Optionally, write an error function and use `fsolve` to improve your estimate.7. Use your best estimate of `v_term` to compute `C_d`.Note: I fabricated the observed flight time, so don't take the results of this exercise too seriously. ###Code # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown Bungee jumping Suppose you want to set the world record for the highest "bungee dunk", [as shown in this video](https://www.youtube.com/watch?v=UBf7WC19lpw). Since the record is 70 m, let's design a jump for 80 m.We'll make the following modeling assumptions:1. Initially the bungee cord hangs from a crane with the attachment point 80 m above a cup of tea.2. Until the cord is fully extended, it applies no force to the jumper. It turns out this might not be a good assumption; we will revisit it.3. After the cord is fully extended, it obeys [Hooke's Law](https://en.wikipedia.org/wiki/Hooke%27s_law); that is, it applies a force to the jumper proportional to the extension of the cord beyond its resting length.4. The jumper is subject to drag force proportional to the square of their velocity, in the opposite of their direction of motion.Our objective is to choose the length of the cord, `L`, and its spring constant, `k`, so that the jumper falls all the way to the tea cup, but no farther! First I'll create a `Param` object to contain the quantities we'll need:1. Let's assume that the jumper's mass is 75 kg.2. With a terminal velocity of 60 m/s.3. The length of the bungee cord is `L = 40 m`.4. The spring constant of the cord is `k = 20 N / m` when the cord is stretched, and 0 when it's compressed. ###Code m = UNITS.meter s = UNITS.second kg = UNITS.kilogram N = UNITS.newton params = Params(y_attach = 80 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 75 * kg, area = 1 * m**2, rho = 1.2 * kg/m**3, v_term = 60 * m / s, L = 25 * m, k = 40 * N / m) ###Output _____no_output_____ ###Markdown Now here's a version of `make_system` that takes a `Params` object as a parameter.`make_system` uses the given value of `v_term` to compute the drag coefficient `C_d`. ###Code def make_system(params): """Makes a System object for the given params. params: Params object returns: System object """ unpack(params) C_d = 2 * mass * g / (rho * area * v_term**2) init = State(y=y_attach, v=v_init) t_end = 20 * s return System(params, C_d=C_d, init=init, t_end=t_end) ###Output _____no_output_____ ###Markdown Let's make a `System` ###Code system = make_system(params) ###Output _____no_output_____ ###Markdown `spring_force` computes the force of the cord on the jumper: ###Code def spring_force(y, system): """Computes the force of the bungee cord on the jumper: y: height of the jumper Uses these variables from system| y_attach: height of the attachment point L: resting length of the cord k: spring constant of the cord returns: force in N """ unpack(system) distance_fallen = y_attach - y if distance_fallen <= L: return 0 * N extension = distance_fallen - L f_spring = k * extension return f_spring ###Output _____no_output_____ ###Markdown The spring force is 0 until the cord is fully extended. When it is extended 1 m, the spring force is 40 N. ###Code spring_force(80*m, system) spring_force(55*m, system) spring_force(54*m, system) ###Output _____no_output_____ ###Markdown `drag_force` computes drag as a function of velocity: ###Code def drag_force(v, system): """Computes drag force in the opposite direction of `v`. v: velocity system: System object returns: drag force """ unpack(system) f_drag = -np.sign(v) * rho * v**2 * C_d * area / 2 return f_drag ###Output _____no_output_____ ###Markdown Here's the drag force at 60 meters per second. ###Code v = -60 * m/s f_drag = drag_force(v, system) ###Output _____no_output_____ ###Markdown Acceleration due to drag at 60 m/s is approximately g, which confirms that 60 m/s is terminal velocity. ###Code a_drag = f_drag / system.mass ###Output _____no_output_____ ###Markdown Now here's the slope function: ###Code def slope_func(state, t, system): """Compute derivatives of the state. state: position, velocity t: time system: System object containing g, rho, C_d, area, and mass returns: derivatives of y and v """ y, v = state unpack(system) a_drag = drag_force(v, system) / mass a_spring = spring_force(y, system) / mass dvdt = -g + a_drag + a_spring return v, dvdt ###Output _____no_output_____ ###Markdown As always, let's test the slope function with the initial params. ###Code slope_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown And then run the simulation. ###Code results, details = run_ode_solver(system, slope_func, max_step=0.3*s) details ###Output _____no_output_____ ###Markdown Here's the plot of position as a function of time. ###Code plot_position(results) ###Output _____no_output_____ ###Markdown After reaching the lowest point, the jumper springs back almost to almost 70 m, and oscillates several times. That looks like more osciallation that we expect from an actual jump, which suggests that there some dissipation of energy in the real world that is not captured in our model. To improve the model, that might be a good thing to investigate.But since we are primarily interested in the initial descent, the model might be good enough for now.We can use `min` to find the lowest point: ###Code min(results.y) ###Output _____no_output_____ ###Markdown At the lowest point, the jumper is still too high, so we'll need to increase `L` or decrease `k`. Here's velocity as a function of time: ###Code plot_velocity(results) subplot(1, 2, 1) plot_position(results) subplot(1, 2, 2) plot_velocity(results) savefig('figs/chap09-fig03.pdf') ###Output _____no_output_____ ###Markdown Although we compute acceleration inside the slope function, we don't get acceleration as a result from `run_ode_solver`.We can approximate it by computing the numerical derivative of `ys`: ###Code a = gradient(results.v) plot(a) decorate(xlabel='Time (s)', ylabel='Acceleration (m/$s^2$)') ###Output _____no_output_____ ###Markdown And we can compute the maximum acceleration the jumper experiences: ###Code max_acceleration = max(a) * m/s**2 ###Output _____no_output_____ ###Markdown Relative to the acceleration of gravity, the jumper "pulls" about "1.7 g's". ###Code max_acceleration / g ###Output _____no_output_____ ###Markdown Under the hoodThe gradient function in `modsim.py` adapts the NumPy function of the same name so it works with `Series` objects. ###Code %psource gradient ###Output _____no_output_____ ###Markdown Solving for lengthAssuming that `k` is fixed, let's find the length `L` that makes the minimum altitude of the jumper exactly 0. The metric we are interested in is the lowest point of the first oscillation. For both efficiency and accuracy, it is better to stop the simulation when we reach this point, rather than run past it and the compute the minimum.Here's an event function that stops the simulation when velocity is 0. ###Code def event_func(state, t, system): """Return velocity. """ y, v = state return v ###Output _____no_output_____ ###Markdown As usual, we should test it with the initial conditions. ###Code event_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown And we see that we have a problem. Since the event function returns 0 under the initial conditions, the simulation would stop immediately. We can solve that problem by specifying the direction of the event function: ###Code event_func.direction = +1 ###Output _____no_output_____ ###Markdown When direction is positive, it only stops the simulation if the velocity is 0 and increasing, which is what we want. Now we can test it an confirm that it stops at the bottom of the jump. ###Code results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.3*s) plot_position(results) min(results.y) ###Output _____no_output_____ ###Markdown **Exercise:** Write an error function that takes `L` and `params` as arguments, simulates a bungee jump, and returns the lowest point.Test the error function with a guess of 25 m and confirm that the return value is about 5 meters.Use `fsolve` with your error function to find the value of `L` that yields a perfect bungee dunk.Run a simulation with the result from `fsolve` and confirm that it works. ###Code # Solution goes here # Solution goes here # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown **Optional exercise:** Search for the combination of length and spring constant that yields minimum height 0 while minimizing peak acceleration. ###Code # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown Modeling and Simulation in PythonChapter 21Copyright 2017 Allen DowneyLicense: [Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0) ###Code # Configure Jupyter so figures appear in the notebook %matplotlib inline # Configure Jupyter to display the assigned value after an assignment %config InteractiveShell.ast_node_interactivity='last_expr_or_assign' # import functions from the modsim.py module from modsim import * ###Output _____no_output_____ ###Markdown With air resistance Next we'll add air resistance using the [drag equation](https://en.wikipedia.org/wiki/Drag_equation) I'll start by getting the units we'll need from Pint. ###Code m = UNITS.meter s = UNITS.second kg = UNITS.kilogram ###Output _____no_output_____ ###Markdown Now I'll create a `Params` object to contain the quantities we need. Using a Params object is convenient for grouping the system parameters in a way that's easy to read (and double-check). ###Code params = Params(height = 381 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 2.5e-3 * kg, diameter = 19e-3 * m, rho = 1.2 * kg/m**3, v_term = 18 * m / s) ###Output _____no_output_____ ###Markdown Now we can pass the `Params` object `make_system` which computes some additional parameters and defines `init`.`make_system` uses the given radius to compute `area` and the given `v_term` to compute the drag coefficient `C_d`. ###Code def make_system(params): """Makes a System object for the given conditions. params: Params object returns: System object """ unpack(params) area = np.pi * (diameter/2)**2 C_d = 2 * mass * g / (rho * area * v_term**2) init = State(y=height, v=v_init) t_end = 30 * s return System(params, area=area, C_d=C_d, init=init, t_end=t_end) ###Output _____no_output_____ ###Markdown Let's make a `System` ###Code system = make_system(params) ###Output _____no_output_____ ###Markdown Here's the slope function, including acceleration due to gravity and drag. ###Code def slope_func(state, t, system): """Compute derivatives of the state. state: position, velocity t: time system: System object returns: derivatives of y and v """ y, v = state unpack(system) f_drag = rho * v**2 * C_d * area / 2 a_drag = f_drag / mass dydt = v dvdt = -g + a_drag return dydt, dvdt ###Output _____no_output_____ ###Markdown As always, let's test the slope function with the initial conditions. ###Code slope_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown We can use the same event function as in the previous chapter. ###Code def event_func(state, t, system): """Return the height of the penny above the sidewalk. """ y, v = state return y ###Output _____no_output_____ ###Markdown And then run the simulation. ###Code results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.5*s) details.message ###Output _____no_output_____ ###Markdown Here are the results. ###Code results ###Output _____no_output_____ ###Markdown The final height is close to 0, as expected.Interestingly, the final velocity is not exactly terminal velocity, which suggests that there are some numerical errors.We can get the flight time from `results`. ###Code t_sidewalk = get_last_label(results) ###Output _____no_output_____ ###Markdown Here's the plot of position as a function of time. ###Code def plot_position(results): plot(results.y) decorate(xlabel='Time (s)', ylabel='Position (m)') plot_position(results) savefig('figs/chap09-fig02.pdf') ###Output Saving figure to file figs/chap09-fig02.pdf ###Markdown And velocity as a function of time: ###Code def plot_velocity(results): plot(results.v, color='C1', label='v') decorate(xlabel='Time (s)', ylabel='Velocity (m/s)') plot_velocity(results) ###Output _____no_output_____ ###Markdown From an initial velocity of 0, the penny accelerates downward until it reaches terminal velocity; after that, velocity is constant. **Exercise:** Run the simulation with an initial velocity, downward, that exceeds the penny's terminal velocity. Hint: You can create a new `Params` object based on an existing one, like this:`params = Params(params, v_init = -30 * m / s)`What do you expect to happen? Plot velocity and position as a function of time, and see if they are consistent with your prediction. ###Code # Solution goes here plot_position(results) # Solution goes here ###Output _____no_output_____ ###Markdown **Exercise:** Suppose we drop a quarter from the Empire State Building and find that its flight time is 19.1 seconds. Use this measurement to estimate the terminal velocity.1. You can get the relevant dimensions of a quarter from https://en.wikipedia.org/wiki/Quarter_(United_States_coin).2. Create a `Params` object with the system parameters. We don't know `v_term`, so we'll start with the inital guess `v_term = 18 * m / s`.3. Use `make_system` to create a `System` object.4. Call `run_ode_solver` to simulate the system. How does the flight time of the simulation compare to the measurement?5. Try a few different values of `t_term` and see if you can get the simulated flight time close to 19.1 seconds.6. Optionally, write an error function and use `fsolve` to improve your estimate.7. Use your best estimate of `v_term` to compute `C_d`.Note: I fabricated the observed flight time, so don't take the results of this exercise too seriously. ###Code # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown Bungee jumping Suppose you want to set the world record for the highest "bungee dunk", [as shown in this video](https://www.youtube.com/watch?v=UBf7WC19lpw). Since the record is 70 m, let's design a jump for 80 m.We'll make the following modeling assumptions:1. Initially the bungee cord hangs from a crane with the attachment point 80 m above a cup of tea.2. Until the cord is fully extended, it applies no force to the jumper. It turns out this might not be a good assumption; we will revisit it.3. After the cord is fully extended, it obeys [Hooke's Law](https://en.wikipedia.org/wiki/Hooke%27s_law); that is, it applies a force to the jumper proportional to the extension of the cord beyond its resting length.4. The jumper is subject to drag force proportional to the square of their velocity, in the opposite of their direction of motion.Our objective is to choose the length of the cord, `L`, and its spring constant, `k`, so that the jumper falls all the way to the tea cup, but no farther! First I'll create a `Param` object to contain the quantities we'll need:1. Let's assume that the jumper's mass is 75 kg.2. With a terminal velocity of 60 m/s.3. The length of the bungee cord is `L = 40 m`.4. The spring constant of the cord is `k = 20 N / m` when the cord is stretched, and 0 when it's compressed. ###Code m = UNITS.meter s = UNITS.second kg = UNITS.kilogram N = UNITS.newton params = Params(y_attach = 80 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 75 * kg, area = 1 * m**2, rho = 1.2 * kg/m**3, v_term = 60 * m / s, L = 25 * m, k = 40 * N / m) ###Output _____no_output_____ ###Markdown Now here's a version of `make_system` that takes a `Params` object as a parameter.`make_system` uses the given value of `v_term` to compute the drag coefficient `C_d`. ###Code def make_system(params): """Makes a System object for the given params. params: Params object returns: System object """ unpack(params) C_d = 2 * mass * g / (rho * area * v_term**2) init = State(y=y_attach, v=v_init) t_end = 20 * s return System(params, C_d=C_d, init=init, t_end=t_end) ###Output _____no_output_____ ###Markdown Let's make a `System` ###Code system = make_system(params) ###Output _____no_output_____ ###Markdown `spring_force` computes the force of the cord on the jumper: ###Code def spring_force(y, system): """Computes the force of the bungee cord on the jumper: y: height of the jumper Uses these variables from system| y_attach: height of the attachment point L: resting length of the cord k: spring constant of the cord returns: force in N """ unpack(system) distance_fallen = y_attach - y if distance_fallen <= L: return 0 * N extension = distance_fallen - L f_spring = k * extension return f_spring ###Output _____no_output_____ ###Markdown The spring force is 0 until the cord is fully extended. When it is extended 1 m, the spring force is 40 N. ###Code spring_force(80*m, system) spring_force(55*m, system) spring_force(54*m, system) ###Output _____no_output_____ ###Markdown `drag_force` computes drag as a function of velocity: ###Code def drag_force(v, system): """Computes drag force in the opposite direction of `v`. v: velocity system: System object returns: drag force """ unpack(system) f_drag = -np.sign(v) * rho * v**2 * C_d * area / 2 return f_drag ###Output _____no_output_____ ###Markdown Here's the drag force at 60 meters per second. ###Code v = -60 * m/s f_drag = drag_force(v, system) ###Output _____no_output_____ ###Markdown Acceleration due to drag at 60 m/s is approximately g, which confirms that 60 m/s is terminal velocity. ###Code a_drag = f_drag / system.mass ###Output _____no_output_____ ###Markdown Now here's the slope function: ###Code def slope_func(state, t, system): """Compute derivatives of the state. state: position, velocity t: time system: System object containing g, rho, C_d, area, and mass returns: derivatives of y and v """ y, v = state unpack(system) a_drag = drag_force(v, system) / mass a_spring = spring_force(y, system) / mass dvdt = -g + a_drag + a_spring return v, dvdt ###Output _____no_output_____ ###Markdown As always, let's test the slope function with the initial params. ###Code slope_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown And then run the simulation. ###Code results, details = run_ode_solver(system, slope_func, max_step=0.3*s) details ###Output _____no_output_____ ###Markdown Here's the plot of position as a function of time. ###Code plot_position(results) ###Output _____no_output_____ ###Markdown After reaching the lowest point, the jumper springs back almost to almost 70 m, and oscillates several times. That looks like more osciallation that we expect from an actual jump, which suggests that there some dissipation of energy in the real world that is not captured in our model. To improve the model, that might be a good thing to investigate.But since we are primarily interested in the initial descent, the model might be good enough for now.We can use `min` to find the lowest point: ###Code min(results.y) ###Output _____no_output_____ ###Markdown At the lowest point, the jumper is still too high, so we'll need to increase `L` or decrease `k`. Here's velocity as a function of time: ###Code plot_velocity(results) subplot(1, 2, 1) plot_position(results) subplot(1, 2, 2) plot_velocity(results) savefig('figs/chap09-fig03.pdf') ###Output _____no_output_____ ###Markdown Although we compute acceleration inside the slope function, we don't get acceleration as a result from `run_ode_solver`.We can approximate it by computing the numerical derivative of `ys`: ###Code a = gradient(results.v) plot(a) decorate(xlabel='Time (s)', ylabel='Acceleration (m/$s^2$)') ###Output _____no_output_____ ###Markdown And we can compute the maximum acceleration the jumper experiences: ###Code max_acceleration = max(a) * m/s**2 ###Output _____no_output_____ ###Markdown Relative to the acceleration of gravity, the jumper "pulls" about "1.7 g's". ###Code max_acceleration / g ###Output _____no_output_____ ###Markdown Under the hoodThe gradient function in `modsim.py` adapts the NumPy function of the same name so it works with `Series` objects. ###Code %psource gradient ###Output _____no_output_____ ###Markdown Solving for lengthAssuming that `k` is fixed, let's find the length `L` that makes the minimum altitude of the jumper exactly 0. The metric we are interested in is the lowest point of the first oscillation. For both efficiency and accuracy, it is better to stop the simulation when we reach this point, rather than run past it and the compute the minimum.Here's an event function that stops the simulation when velocity is 0. ###Code def event_func(state, t, system): """Return velocity. """ y, v = state return v ###Output _____no_output_____ ###Markdown As usual, we should test it with the initial conditions. ###Code event_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown And we see that we have a problem. Since the event function returns 0 under the initial conditions, the simulation would stop immediately. We can solve that problem by specifying the direction of the event function: ###Code event_func.direction = +1 ###Output _____no_output_____ ###Markdown When direction is positive, it only stops the simulation if the velocity is 0 and increasing, which is what we want. Now we can test it an confirm that it stops at the bottom of the jump. ###Code results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.3*s) plot_position(results) min(results.y) ###Output _____no_output_____ ###Markdown **Exercise:** Write an error function that takes `L` and `params` as arguments, simulates a bungee jump, and returns the lowest point.Test the error function with a guess of 25 m and confirm that the return value is about 5 meters.Use `fsolve` with your error function to find the value of `L` that yields a perfect bungee dunk.Run a simulation with the result from `fsolve` and confirm that it works. ###Code # Solution goes here # Solution goes here # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown **Optional exercise:** Search for the combination of length and spring constant that yields minimum height 0 while minimizing peak acceleration. ###Code # Solution goes here # Solution goes here ###Output _____no_output_____ ###Markdown Modeling and Simulation in PythonChapter 21Copyright 2017 Allen DowneyLicense: [Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0) ###Code # Configure Jupyter so figures appear in the notebook %matplotlib inline # Configure Jupyter to display the assigned value after an assignment %config InteractiveShell.ast_node_interactivity='last_expr_or_assign' # import functions from the modsim.py module from modsim import * ###Output _____no_output_____ ###Markdown With air resistance Next we'll add air resistance using the [drag equation](https://en.wikipedia.org/wiki/Drag_equation) I'll start by getting the units we'll need from Pint. ###Code m = UNITS.meter s = UNITS.second kg = UNITS.kilogram ###Output _____no_output_____ ###Markdown Now I'll create a `Params` object to contain the quantities we need. Using a Params object is convenient for grouping the system parameters in a way that's easy to read (and double-check). ###Code params = Params(height = 381 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 2.5e-3 * kg, diameter = 19e-3 * m, rho = 1.2 * kg/m**3, v_term = 18 * m / s) ###Output _____no_output_____ ###Markdown Now we can pass the `Params` object `make_system` which computes some additional parameters and defines `init`.`make_system` uses the given radius to compute `area` and the given `v_term` to compute the drag coefficient `C_d`. ###Code def make_system(params): """Makes a System object for the given conditions. params: Params object returns: System object """ unpack(params) area = np.pi * (diameter/2)**2 C_d = 2 * mass * g / (rho * area * v_term**2) init = State(y=height, v=v_init) t_end = 30 * s return System(params, area=area, C_d=C_d, init=init, t_end=t_end) ###Output _____no_output_____ ###Markdown Let's make a `System` ###Code system = make_system(params) ###Output _____no_output_____ ###Markdown Here's the slope function, including acceleration due to gravity and drag. ###Code def slope_func(state, t, system): """Compute derivatives of the state. state: position, velocity t: time system: System object returns: derivatives of y and v """ y, v = state unpack(system) f_drag = rho * v**2 * C_d * area / 2 a_drag = f_drag / mass dydt = v dvdt = -g + a_drag return dydt, dvdt ###Output _____no_output_____ ###Markdown As always, let's test the slope function with the initial conditions. ###Code slope_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown We can use the same event function as in the previous chapter. ###Code def event_func(state, t, system): """Return the height of the penny above the sidewalk. """ y, v = state return y ###Output _____no_output_____ ###Markdown And then run the simulation. ###Code results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.5*s) details.message ###Output _____no_output_____ ###Markdown Here are the results. ###Code results ###Output _____no_output_____ ###Markdown The final height is close to 0, as expected.Interestingly, the final velocity is not exactly terminal velocity, which suggests that there are some numerical errors.We can get the flight time from `results`. ###Code t_sidewalk = get_last_label(results) ###Output _____no_output_____ ###Markdown Here's the plot of position as a function of time. ###Code def plot_position(results): plot(results.y) decorate(xlabel='Time (s)', ylabel='Position (m)') plot_position(results) savefig('figs/chap09-fig02.pdf') ###Output Saving figure to file figs/chap09-fig02.pdf ###Markdown And velocity as a function of time: ###Code def plot_velocity(results): plot(results.v, color='C1', label='v') decorate(xlabel='Time (s)', ylabel='Velocity (m/s)') plot_velocity(results) ###Output _____no_output_____ ###Markdown From an initial velocity of 0, the penny accelerates downward until it reaches terminal velocity; after that, velocity is constant. **Exercise:** Run the simulation with an initial velocity, downward, that exceeds the penny's terminal velocity. Hint: You can create a new `Params` object based on an existing one, like this:`params = Params(params, v_init = -30 * m / s)`What do you expect to happen? Plot velocity and position as a function of time, and see if they are consistent with your prediction. ###Code params = Params(params, v_init = -30 * m / s) system = make_system(params) results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.5*s) details.message ###Output _____no_output_____ ###Markdown I expect the velocity to approach terminal velocity and the penny to follow a similar linear descent ###Code plot_position(results) plot_velocity(results) ###Output _____no_output_____ ###Markdown **Exercise:** Suppose we drop a quarter from the Empire State Building and find that its flight time is 19.1 seconds. Use this measurement to estimate the terminal velocity.1. You can get the relevant dimensions of a quarter from https://en.wikipedia.org/wiki/Quarter_(United_States_coin).2. Create a `Params` object with the system parameters. We don't know `v_term`, so we'll start with the inital guess `v_term = 18 * m / s`.3. Use `make_system` to create a `System` object.4. Call `run_ode_solver` to simulate the system. How does the flight time of the simulation compare to the measurement?5. Try a few different values of `t_term` and see if you can get the simulated flight time close to 19.1 seconds.6. Optionally, write an error function and use `fsolve` to improve your estimate.7. Use your best estimate of `v_term` to compute `C_d`.Note: I fabricated the observed flight time, so don't take the results of this exercise too seriously. ###Code params = Params(height = 381 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 6.25e-3 * kg, diameter = 24.26e-3 * m, rho = 1.2 * kg/m**3, v_term = 21.65 * m / s); system = make_system(params); results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.5*s) details.message print(str(results.last_valid_index()), "is when the quarter hits the ground") subplot(2,1,1) plot_position(results) subplot(2,1,2) plot_velocity(results) def error_func(v_term): params = Params(height = 381 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 6.25e-3 * kg, diameter = 24.26e-3 * m, rho = 1.2 * kg/m**3, v_term =v_term* m / s); system = make_system(params); results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.5*s) return results.last_valid_index() - 19.1 v_term1 = fsolve(error_func, 21)[0] unpack(params) C_d = 2* mass * g / (rho * v_term1**2 *pi*(diameter/2)**2) # Solution goes here ###Output _____no_output_____ ###Markdown Bungee jumping Suppose you want to set the world record for the highest "bungee dunk", [as shown in this video](https://www.youtube.com/watch?v=UBf7WC19lpw). Since the record is 70 m, let's design a jump for 80 m.We'll make the following modeling assumptions:1. Initially the bungee cord hangs from a crane with the attachment point 80 m above a cup of tea.2. Until the cord is fully extended, it applies no force to the jumper. It turns out this might not be a good assumption; we will revisit it.3. After the cord is fully extended, it obeys [Hooke's Law](https://en.wikipedia.org/wiki/Hooke%27s_law); that is, it applies a force to the jumper proportional to the extension of the cord beyond its resting length.4. The jumper is subject to drag force proportional to the square of their velocity, in the opposite of their direction of motion.Our objective is to choose the length of the cord, `L`, and its spring constant, `k`, so that the jumper falls all the way to the tea cup, but no farther! First I'll create a `Param` object to contain the quantities we'll need:1. Let's assume that the jumper's mass is 75 kg.2. With a terminal velocity of 60 m/s.3. The length of the bungee cord is `L = 40 m`.4. The spring constant of the cord is `k = 20 N / m` when the cord is stretched, and 0 when it's compressed. ###Code m = UNITS.meter s = UNITS.second kg = UNITS.kilogram N = UNITS.newton params = Params(y_attach = 80 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 75 * kg, area = 1 * m**2, rho = 1.2 * kg/m**3, v_term = 60 * m / s, L = 25 * m, k = 40 * N / m) ###Output _____no_output_____ ###Markdown Now here's a version of `make_system` that takes a `Params` object as a parameter.`make_system` uses the given value of `v_term` to compute the drag coefficient `C_d`. ###Code def make_system(params): """Makes a System object for the given params. params: Params object returns: System object """ unpack(params) C_d = 2 * mass * g / (rho * area * v_term**2) init = State(y=y_attach, v=v_init) t_end = 20 * s return System(params, C_d=C_d, init=init, t_end=t_end) ###Output _____no_output_____ ###Markdown Let's make a `System` ###Code system = make_system(params) ###Output _____no_output_____ ###Markdown `spring_force` computes the force of the cord on the jumper: ###Code def spring_force(y, system): """Computes the force of the bungee cord on the jumper: y: height of the jumper Uses these variables from system| y_attach: height of the attachment point L: resting length of the cord k: spring constant of the cord returns: force in N """ unpack(system) distance_fallen = y_attach - y if distance_fallen <= L: return 0 * N extension = distance_fallen - L f_spring = k * extension return f_spring ###Output _____no_output_____ ###Markdown The spring force is 0 until the cord is fully extended. When it is extended 1 m, the spring force is 40 N. ###Code spring_force(80*m, system) spring_force(55*m, system) spring_force(54*m, system) ###Output _____no_output_____ ###Markdown `drag_force` computes drag as a function of velocity: ###Code def drag_force(v, system): """Computes drag force in the opposite direction of `v`. v: velocity system: System object returns: drag force """ unpack(system) f_drag = -np.sign(v) * rho * v**2 * C_d * area / 2 return f_drag ###Output _____no_output_____ ###Markdown Here's the drag force at 60 meters per second. ###Code v = -60 * m/s f_drag = drag_force(v, system) ###Output _____no_output_____ ###Markdown Acceleration due to drag at 60 m/s is approximately g, which confirms that 60 m/s is terminal velocity. ###Code a_drag = f_drag / system.mass ###Output _____no_output_____ ###Markdown Now here's the slope function: ###Code def slope_func(state, t, system): """Compute derivatives of the state. state: position, velocity t: time system: System object containing g, rho, C_d, area, and mass returns: derivatives of y and v """ y, v = state unpack(system) a_drag = drag_force(v, system) / mass a_spring = spring_force(y, system) / mass dvdt = -g + a_drag + a_spring return v, dvdt ###Output _____no_output_____ ###Markdown As always, let's test the slope function with the initial params. ###Code slope_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown And then run the simulation. ###Code results, details = run_ode_solver(system, slope_func, max_step=0.3*s) details ###Output _____no_output_____ ###Markdown Here's the plot of position as a function of time. ###Code plot_position(results) ###Output _____no_output_____ ###Markdown After reaching the lowest point, the jumper springs back almost to almost 70 m, and oscillates several times. That looks like more osciallation that we expect from an actual jump, which suggests that there some dissipation of energy in the real world that is not captured in our model. To improve the model, that might be a good thing to investigate.But since we are primarily interested in the initial descent, the model might be good enough for now.We can use `min` to find the lowest point: ###Code min(results.y) ###Output _____no_output_____ ###Markdown At the lowest point, the jumper is still too high, so we'll need to increase `L` or decrease `k`. Here's velocity as a function of time: ###Code plot_velocity(results) subplot(1, 2, 1) plot_position(results) subplot(1, 2, 2) plot_velocity(results) savefig('figs/chap09-fig03.pdf') ###Output Saving figure to file figs/chap09-fig03.pdf ###Markdown Although we compute acceleration inside the slope function, we don't get acceleration as a result from `run_ode_solver`.We can approximate it by computing the numerical derivative of `ys`: ###Code a = gradient(results.v) plot(a) decorate(xlabel='Time (s)', ylabel='Acceleration (m/$s^2$)') ###Output _____no_output_____ ###Markdown And we can compute the maximum acceleration the jumper experiences: ###Code max_acceleration = max(a) * m/s**2 ###Output _____no_output_____ ###Markdown Relative to the acceleration of gravity, the jumper "pulls" about "1.7 g's". ###Code max_acceleration / g ###Output _____no_output_____ ###Markdown Under the hoodThe gradient function in `modsim.py` adapts the NumPy function of the same name so it works with `Series` objects. ###Code %psource gradient ###Output _____no_output_____ ###Markdown Solving for lengthAssuming that `k` is fixed, let's find the length `L` that makes the minimum altitude of the jumper exactly 0. The metric we are interested in is the lowest point of the first oscillation. For both efficiency and accuracy, it is better to stop the simulation when we reach this point, rather than run past it and the compute the minimum.Here's an event function that stops the simulation when velocity is 0. ###Code def event_func(state, t, system): """Return velocity. """ y, v = state return v ###Output _____no_output_____ ###Markdown As usual, we should test it with the initial conditions. ###Code event_func(system.init, 0, system) ###Output _____no_output_____ ###Markdown And we see that we have a problem. Since the event function returns 0 under the initial conditions, the simulation would stop immediately. We can solve that problem by specifying the direction of the event function: ###Code event_func.direction = +1 ###Output _____no_output_____ ###Markdown When direction is positive, it only stops the simulation if the velocity is 0 and increasing, which is what we want. Now we can test it an confirm that it stops at the bottom of the jump. ###Code results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.3*s) plot_position(results) min(results.y) ###Output _____no_output_____ ###Markdown **Exercise:** Write an error function that takes `L` and `params` as arguments, simulates a bungee jump, and returns the lowest point.Test the error function with a guess of 25 m and confirm that the return value is about 5 meters.Use `fsolve` with your error function to find the value of `L` that yields a perfect bungee dunk.Run a simulation with the result from `fsolve` and confirm that it works. ###Code params = Params(y_attach = 80 * m, v_init = 0 * m / s, g = 9.8 * m/s**2, mass = 75 * kg, area = 1 * m**2, rho = 1.2 * kg/m**3, v_term = 60 * m / s, L = 25 * m, k = 40 * N / m) def error_func2(L, params): params1 = Params(params, L = L) system = make_system(params1) results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.3*s) return min(results.y) error_func2(25,params) params1 = Params(params, L = L_best) L_best = fsolve(error_func2, 25, params1)[0] params1 = Params(params, L = L_best, k = 50) system = make_system(params1) results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.3*s) subplot(2, 1, 1) plot_position(results) subplot(2, 1, 2) plot_velocity(results) ###Output _____no_output_____ ###Markdown **Optional exercise:** Search for the combination of length and spring constant that yields minimum height 0 while minimizing peak acceleration. ###Code def error_func3(k,params): params1 = Params(params, k = k*N/m) print(k) L_best = fsolve(error_func2, 30, params1)[0] print("yo", L_best) params2 = Params(params1, L =L_best) system = make_system(params2) results, details = run_ode_solver(system, slope_func, events=event_func, max_step=0.3*s) a = gradient(results.v) print(2, max(a)) return max(a) fsolve(error_func3, 40, params) # Solution goes here ###Output _____no_output_____
modules/module-05/mod5_nb1_pandas_foundations.ipynb
###Markdown Runtime Dependencies: Must Run First! ###Code import numpy as np import pandas as pd # ### Bonus: Multiple Outputs Per Cell from IPython.core.interactiveshell import InteractiveShell InteractiveShell.ast_node_interactivity = "all" ###Output _____no_output_____ ###Markdown Module 5: Pandas **Importing Pandas:**```pythonimport pandas as pd```Pandas is an essential data science package for Python that brings DataFrame functionality into Python. DataFrames are built upon NumPy arrays, and allow us to set custom indicies, add column and row labels, and bring SQL-like and Excel-like operations possible. Module 5.1: Series & DataFrames SeriesLet's start with a Series. A series is essentially a NumPy array underneath, with some added features. Let's go through some of them! ###Code arr = np.array(range(50,71,2)) arr ser = pd.Series(arr) ser ###Output _____no_output_____ ###Markdown As you can see, this looks a bit different than the array output. We now have data being formatted vertically in a table with the index on the right hand side.By default, Pandas will assign a zero-indexed index if none is supplied. However, we can start changing some of these properties quickly.Essential keyword arguments upon creation:1) data2) index4) nameWe've already used the data keyword to tell Python what data needs to be imported, and now we're going to add the name property. This is essentially a label for the data. These will come back in DataFrames as column names. ###Code lst = ["Apples","Bananas","Grapes","Mangos","Avocados"] ser = pd.Series(lst, name="Produce") ser ###Output _____no_output_____ ###Markdown And we can even set a custom index if we don't want to use the one automatically generated!Let's say we have a small list of employees with their ID number and name. It would make sense to have the ID be the index! ###Code loc = "https://raw.githubusercontent.com/mhall-simon/python/main/data/misc/arr-ID-and-names.csv" arr = np.genfromtxt(loc, dtype=str, delimiter=',', skip_header=1) arr ###Output _____no_output_____ ###Markdown Now, let's create the series with the custom index and employee names: ###Code ser = pd.Series(arr[:,1], index=arr[:,0], name="Employees by ID") ser ###Output _____no_output_____ ###Markdown DataFrameRight now, we can create an indexed series (with one columnn of data and an index), but sometimes we have complex datasets that can't fit in a single column.This is where the DataFrame comes in handy. We can now store and process more complex data!Let's start with a really simple inventory example: ###Code dct = {'partNum': [104,105,106], 'Quantity': [415,346,98], 'Price': [1.24, 2.15, 5.98], 'Cost': [0.78,1.56,3.12]} df = pd.DataFrame(dct) df ###Output _____no_output_____ ###Markdown Now, when we visualize the data structure you can see the column labels! Since no index was created, Pandas will automatically assign one!In section 3, I'm going to cover how to generate DataFrames from various sources. Some DataFrame Attributes *Pulled From Documentation or Summarized Desc*| Attribute | Description || :---: | --- || `.index` | The Index (row labels) of the DataFrame || `.columns` | The column labels of the DataFrame || `.shape` | Dimensions of DataFrame | ###Code df.index df.columns df.shape ###Output _____no_output_____ ###Markdown Some DataFrame Methods*Pulled From Documentation or Summarized Desc*| Method | Description || :---: | --- || `.info()` | Prints out info for DataFrame || `.describe()` | Prints out summary statistics || `.head()` | Print out top n rows of DataFrame, default to 5 || `.tail()` | Print out last n rows of DataFrame, default to 5 |Just keep in mind that sometimes a summary statistic for a column doesn't make sense! ###Code df.info() df.describe() ###Output _____no_output_____ ###Markdown Module 5.2: Indexing DataFrames In Pandas, there's 3 main methods for indexing a DataFrame.The first one allows us to easily subset a column (or more) from our DataFrame.To start, I'm going to import a data set about COVID 19 vaccines. This CSV has vaccination info from the CDC, with the report being pulled on February 8th, 2021.**Keep In Mind That You Can Always Reference This. You Don't Need to Memorize, But You Should Be Familiar Enough to Know Which One to Choose** ###Code loc = "https://raw.githubusercontent.com/mhall-simon/python/main/data/covid/covid19_vaccinations_in_the_united_states.csv" df = pd.read_csv(loc, header=2, index_col="State/Territory/Federal Entity") df ###Output _____no_output_____ ###Markdown Easily Accessing Columns This is our first indexing method, to pull columns of data, and there are a few syntax styles: Pulling Single Column ###Code df['Total Delivered'] ###Output _____no_output_____ ###Markdown Pulling Multiple Columns ###Code df[['Total Delivered','Total Administered']] ###Output _____no_output_____ ###Markdown Note About Single ColumnIf our column names have no spaces, we can quickly access a column like this too:```pythondf.col_name``` Indexing with `.loc[]` We can also use the method `.loc[]` to index our DataFrames. This method uses the labels of rows/columns to index.Just like 2D NumPy arrays, loc takes the arguments as `[row:row,col:col]` Pulling a Row with `.loc()` ###Code df.loc['Wisconsin'] ###Output _____no_output_____ ###Markdown Pulling Range of Rows - All Columns ###Code df.loc['Oklahoma':'Wisconsin',:] ###Output _____no_output_____ ###Markdown Pulling Range of Rows, Select Columns ###Code df.loc['Oregon':'Texas','Total Delivered':'Total Administered'] ###Output _____no_output_____ ###Markdown Indexing with `.iloc[]` Iloc uses purely integer-based indexing to find values for us! This means that we need to make sure our dataset is in an expected order. Rows: Scalar ###Code df.iloc[5] ###Output _____no_output_____ ###Markdown Rows: List ###Code df.iloc[[2,4,8]] ###Output _____no_output_____ ###Markdown Rows: Slicing ###Code df.iloc[2:5] ###Output _____no_output_____ ###Markdown Rows & Cols: Scalars ###Code df.iloc[0,0] ###Output _____no_output_____ ###Markdown Rows & Cols: List ###Code df.iloc[[12,18], [0,2]] ###Output _____no_output_____ ###Markdown Rows & Cols: Slicing ###Code df.iloc[12:22, 0:4] ###Output _____no_output_____ ###Markdown Module 5.3: Importing Data Sets Overview You've already seen some import statements above, but now I'm going to cover them on a high level!Dig through the documentation to learn about all the little tweaks and tricks you can use upon import.2 Essential Methods:- Read CSV- Read ExcelThe most important thing to note about importing data is to look through the file in Excel!You cannot properly import something if you don't know what it is!I'm only going to cover the CSV method, as the two are very similar and have only slightly different tweaks. Using Read CSV In GitHub, I have a CSV file downloaded from Yahoo Finance about Tesla's monthly stock performance over the last 5 years.We can examine it here: https://github.com/mhall-simon/python/blob/main/data/misc/TSLA.csvAnd the GitHub raw link is: https://raw.githubusercontent.com/mhall-simon/python/main/data/misc/TSLA.csvLet's import it by just targeting a link: ###Code loc = "https://raw.githubusercontent.com/mhall-simon/python/main/data/misc/TSLA.csv" df = pd.read_csv(loc) df.head() ###Output _____no_output_____ ###Markdown Pretty neat! However, a useful feature would be to index by Date, as that is a unique key for the data set.Since the Date column is first, we're going to use `index_col=0` as a keyword argument.Also, since it's Date/Time data, we should parse it into the proper format.We do this by providing the argument `parse_dates=True` ###Code loc = "https://raw.githubusercontent.com/mhall-simon/python/main/data/misc/TSLA.csv" df = pd.read_csv(loc, index_col=0, parse_dates=True) df.head() ###Output _____no_output_____ ###Markdown Let's inspect the DataFrame now using `.info()` ###Code df.info() ###Output <class 'pandas.core.frame.DataFrame'> DatetimeIndex: 61 entries, 2016-03-01 to 2021-02-12 Data columns (total 6 columns): # Column Non-Null Count Dtype --- ------ -------------- ----- 0 Open 61 non-null float64 1 High 61 non-null float64 2 Low 61 non-null float64 3 Close 61 non-null float64 4 Adj Close 61 non-null float64 5 Volume 61 non-null int64 dtypes: float64(5), int64(1) memory usage: 3.3 KB ###Markdown There's a ton of other arguments, some being more useful than the others. As we keep progressing through training, try to see what keyword arguments I use to import data! Module 5.4: Date & Time Overview Pandas also brings with it a new data type: DateTime.This is very useful for using time series data, as we don't need to map information out into multiple columns.If the DateTime data is the index, it makes processing time series data in Python very efficient!Let's start with the same data set as above for Tesla stock: ###Code loc = "https://raw.githubusercontent.com/mhall-simon/python/main/data/misc/TSLA.csv" df = pd.read_csv(loc, index_col=0, parse_dates=True) df.head() ###Output _____no_output_____ ###Markdown Extracting D-M-Y Information ###Code df.index.day df.index.month df.index.year ###Output _____no_output_____ ###Markdown These properties will become very useful when we want to index, slice, or even run SQL functionality on our time series data! Module 5.5: Multi-Indexed Data Sets Above, you've only seen data with a single index. However, you can also have a multi-index!These are excellent for when you have multiple indicies, and don't want to store information in columns. Importing Regularly: ###Code loc = "https://github.com/mhall-simon/python/blob/main/data/misc/inventory-multi-index.xlsx?raw=true" df = pd.read_excel(loc, parse_dates=True) df ###Output _____no_output_____ ###Markdown Importing Multi-Index: ###Code loc = "https://github.com/mhall-simon/python/blob/main/data/misc/inventory-multi-index.xlsx?raw=true" df = pd.read_excel(loc, parse_dates=True, index_col=[0,1]) df ###Output _____no_output_____ ###Markdown Now, our data is formatted better! It's hierarchical between the date and part number!I'm not going to go too in depth with Multi-Indexed Data Sets, but you should know that they exist! Accessing Row in Multi-Indexed Dataset (Outer Most Group) ###Code df.loc['2020-03-01'] ###Output _____no_output_____ ###Markdown Accessing Row in Multi-Indexed Dataset (Inner Most Group) ###Code df.loc[('2020-03-01','A-01')] ###Output _____no_output_____ ###Markdown Tip: Multi-Indexed Datasets behave very-similarly to groupby objects! Module 5.6: Broadcasting Scalars Just like with NumPy, we can broadcast scalars across our dataframe!Since we're already familiar with how it works, we just need the general formula:1. Index Subset (Optional)2. Broadcast ScalarIt's pretty easy! And allows us to update information stored in DataFrames without creating new rows / cols!Above, we imported a dataset of inventory prices and amount sold, let's do some basic operations on them! ###Code loc = "https://github.com/mhall-simon/python/blob/main/data/misc/inventory-multi-index.xlsx?raw=true" df = pd.read_excel(loc, parse_dates=True, index_col=[0,1]) df ###Output _____no_output_____ ###Markdown Broadcast Across Slice - Column Oh no! Our costs were too low because we forgot tax. Let's add 8% to the cost column of the dataset: ###Code df.Cost = df.Cost*1.08 df ###Output _____no_output_____ ###Markdown Broadcast Across Slice - Row Oh no! We forgot we have a 5% discount on the first day of the month. Let's easily correct this. ###Code df.loc['2020-03-01','Price'] = df.loc['2020-03-01','Price']*0.95 df ###Output _____no_output_____ ###Markdown These operations are easy—it's really just a question of knowing how to slice your data! Module 5.7: Row & Column-wise Operations Just like with NumPy, we can also do element-wise operations with our DataFrames. Again, it's really just a practice of knowing how to properly slice your data! Creating New Column Based Upon Other Columns Right now, we have a data frame with prices, quantity, and cost. Let's start working towards profitability for each part per day.Here's how we would build a column named Gross, which is Price * Number Sold. ###Code loc = "https://github.com/mhall-simon/python/blob/main/data/misc/inventory-multi-index.xlsx?raw=true" df = pd.read_excel(loc, parse_dates=True, index_col=[0,1]) df df['Gross'] = df.Price * df.Sold df ###Output _____no_output_____ ###Markdown Now, let's calculate Net, which is going to be Gross minus our total cost for the day. ###Code df['Net'] = df.Gross - (df.Cost * df.Sold) df ###Output _____no_output_____ ###Markdown Now, we can also figure out our margin. We'll calculate our margin as Net / Gross, and leave it in decimal terms (don't multiply by 100) ###Code df['Margin'] = df.Net / df.Gross df ###Output _____no_output_____ ###Markdown If you're thinking that right now we're really only calculating new columns, you'd be right! We usually have our data formatted as new entries going vertically, within the columns.Most of our row-wise operations come in SQL-functionality, when we calculate statistics for our columns or run groups, filters, and more! Applying Custom FunctionsUsually, we want to keep our operations to run as broadcasting or element-wise operations with subsets and slices. However, sometimes we need to apply a function!These are going to be slower, but when they're needed they're necessary!We can write them with lambda functions. Let's mark our products with the Note "High" to denote a high-margin product. Management defines any high margin product as one with at least a 65% profit margin. ###Code df['Note'] = df.apply(lambda row: 'High' if row.Margin >= 0.65 else 'Low', axis=1) df ###Output _____no_output_____ ###Markdown Bonus Box: Nested Lambdas for Multiple Logic Checks We can "nest" lambda functions to do multiple logic checks! Essentially, the else block is going to run another lambda!Let's redo the above example and cast 3 margins. One for high (>65%), one for medium (>40%), and one for low. ###Code df['Note'] = df.apply(lambda row: 'High' if row.Margin >= 0.65 else ('Medium' if row.Margin >= 0.4 else 'Low'), axis=1) df ###Output _____no_output_____
data-science/scikit-learn/08/01-Implement-Confusion-Mafrix.ipynb
###Markdown 实现混淆矩阵, 精准率和召回率 ###Code import numpy as np from sklearn import datasets digits = datasets.load_digits() X = digits.data y = digits.target.copy() y[digits.target == 9] = 1 y[digits.target != 9] = 0 from sklearn.model_selection import train_test_split X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=666) from sklearn.linear_model import LogisticRegression log_reg = LogisticRegression() log_reg.fit(X_train, y_train) log_reg.score(X_test, y_test) y_log_predict = log_reg.predict(X_test) # True Negative def TN(y_true, y_predict): assert len(y_true) == len(y_predict) return np.sum((y_true == 0) * (y_predict == 0)) TN(y_test, y_log_predict) # False Positive def FP(y_true, y_predict): assert len(y_true) == len(y_predict) return np.sum((y_true == 0) * (y_predict == 1)) FP(y_test, y_log_predict) # False Negative def FN(y_true, y_predict): assert len(y_true) == len(y_predict) return np.sum((y_true == 1) * (y_predict == 0)) FN(y_test, y_log_predict) # True Positive def TP(y_true, y_predict): assert len(y_true) == len(y_predict) return np.sum((y_true == 1) * (y_predict == 1)) TP(y_test, y_log_predict) # 混淆矩阵 def confusion_matrix(y_true, y_predict): return np.array([ [TN(y_true, y_predict), FP(y_true, y_predict)], [FN(y_true, y_predict), TP(y_true, y_predict)] ]) confusion_matrix(y_test, y_log_predict) # 准确率 def precision_score(y_true, y_predict): tp = TP(y_true, y_predict) fp = FP(y_true, y_predict) try: return tp / (tp + fp) except: return 0.0 precision_score(y_test, y_log_predict) # 召回率 def recall_score(y_true, y_predict): tp = TP(y_true, y_predict) fn = FN(y_true, y_predict) try: return tp / (tp + fn) except: return 0.0 recall_score(y_test, y_log_predict) ###Output _____no_output_____ ###Markdown scikit-learn中的混淆矩阵, 精准率和召回率 ###Code from sklearn.metrics import confusion_matrix confusion_matrix(y_test, y_log_predict) from sklearn.metrics import precision_score precision_score(y_test, y_log_predict) from sklearn.metrics import recall_score recall_score(y_test, y_log_predict) ###Output _____no_output_____
6 PySpark KMeans.ipynb
###Markdown Welcome to exercise two of week three of “Apache Spark for Scalable Machine Learning on BigData”. In this exercise we’ll work on clustering.Let’s create our DataFrame again: ###Code %sh # delete files from previous runs #rm -f hmp.parquet* # download the file containing the data in PARQUET format #wget https://github.com/IBM/coursera/raw/master/hmp.parquet #ls -ltr /databricks/driver #mkdir /dbfs/tmp/HMPPARQ/ #cp /databricks/driver/hmp.parquet /dbfs/tmp/HMPPARQ/ ls -ltr /dbfs/tmp/HMPPARQ/ # create a dataframe out of it df = spark.read.parquet('/tmp/HMPPARQ/hmp.parquet') # register a corresponding query table df.createOrReplaceTempView('df') df.select("class").distinct().show() ###Output _____no_output_____ ###Markdown Let’s reuse our feature engineering pipeline. ###Code from pyspark.ml.feature import OneHotEncoder, StringIndexer, VectorAssembler, Normalizer from pyspark.ml.linalg import Vectors from pyspark.ml import Pipeline indexer = StringIndexer(inputCol="class", outputCol="classIndex") encoder = OneHotEncoder(inputCol="classIndex", outputCol="categoryVec") vectorAssembler = VectorAssembler(inputCols=["x","y","z"], outputCol="features") normalizer = Normalizer(inputCol="features", outputCol="features_norm", p=1.0) pipeline = Pipeline(stages=[indexer, encoder, vectorAssembler, normalizer]) model = pipeline.fit(df) prediction = model.transform(df) prediction.show() ###Output _____no_output_____ ###Markdown Now let’s create a new pipeline for kmeans. ###Code from pyspark.ml.clustering import KMeans from pyspark.ml.feature import OneHotEncoder, StringIndexer, VectorAssembler, Normalizer from pyspark.ml.evaluation import ClusteringEvaluator from pyspark.ml import Pipeline kmeans = KMeans(featuresCol="features_norm").setK(14).setSeed(1) model = kmeans.fit(prediction) predictions = model.transform(prediction) evaluator = ClusteringEvaluator() silhouette = evaluator.evaluate(predictions) print("Silhouette with squared euclidean distance = " + str(silhouette)) centers = model.clusterCenters() print("Cluster Centers: ") for center in centers: print(center) ###Output _____no_output_____ ###Markdown We have 14 different movement patterns in the dataset, so setting K of KMeans to 14 is a good idea. But please experiment with different values for K, do you find a sweet spot? The closer Silhouette gets to 1, the better.https://en.wikipedia.org/wiki/Silhouette_(clustering) ###Code # please change the pipeline the check performance for different K, feel free to use a loop ###Output _____no_output_____
flowws/spheres/Spheres.ipynb
###Markdown In this notebook we run a simple system of purely repulsive spheres using a [Weeks-Chandler-Andersen potential](https://hoomd-blue.readthedocs.io/en/stable/module-md-pair.htmlhoomd.md.pair.lj). We visualize the spheres using the povray backend and color them by distance to their nearest neighbor. ###Code import flowws import gtar from hoomd_flowws.Init import Init from hoomd_flowws.Interaction import Interaction from hoomd_flowws.Run import Run import plato, plato.draw.povray as draw import freud import numpy as np import IPython import ipywidgets storage = flowws.DirectoryStorage() stages = [ Init(number=128), Interaction( type='lj', global_params=[('r_cut', 2**(1./6))], pair_params=[('A', 'A', 'epsilon', 1), ('A', 'A', 'sigma', 1)]), Run(steps=1e3, integrator='langevin'), Run(steps=1e4, integrator='langevin', compress_to=.57, dump_period=1e3), ] flowws.Workflow(stages, storage).run(); num_frames = 0 def get_frame(frame=-1): global num_frames with gtar.GTAR('dump.sqlite', 'r') as traj: (posRec, boxRec), frames = traj.framesWithRecordsNamed(['position', 'box']) num_frames = len(frames) positions = traj.getRecord(posRec, frames[frame]) box = traj.getRecord(boxRec, frames[frame]) return positions, box def update(scene, frame=-1): (positions, box) = get_frame(frame) # get nearest-neighbor distance, rescaled to go from 0-1, as cval fbox = freud.box.Box.from_box(box) nn = freud.locality.AABBQuery(fbox, positions) nlist = nn.query(positions, dict(exclude_ii=True, num_neighbors=1)).toNeighborList(True) cval = nlist.distances.copy() cval -= np.min(cval) cval /= np.max(cval) colors = plato.cmap.cubehelix(.25 + .5*cval) for prim in scene: prim.colors = colors prim.positions = positions prim.diameters = np.ones(len(positions)) prim = draw.Spheres() features = dict(ambient_light=.4) scene = draw.Scene(prim, features=features, zoom=4.8) update(scene) target = '../../gallery/flowws_spheres_povray.png' scene.save(target) IPython.display.Image(filename=target) import plato.draw.vispy as interactive live_scene = scene.convert(interactive) live_scene.show() @ipywidgets.interact(frame=(0, num_frames - 1)) def plot(frame=0): update(live_scene, frame) live_scene.render() ###Output _____no_output_____
Jupyter-notebook/URL-categorization-Jupyter-Notebook.ipynb
###Markdown Import librariesThese libraries will be used for our URL_classification project. ###Code import datetime import csv import nltk import numpy as np import pandas as pd import ast from urllib.request import urlopen from bs4 import BeautifulSoup import os.path print(datetime.datetime.now().time()) ###Output 14:23:30.843081 ###Markdown Use this command if you have any errors on importing nltk library. It will open a nltk meniu with download and update options. If it's still missing some libraries, it needs to install manually by writing nltk.download('library name') where library name is missing library name which asserts error message. ###Code nltk.download('stopwords') nltk.download('words') nltk.download('punkt') def conctruct_dataset(): file = 'URL-categorization-DFE.csv' df = pd.read_csv(file)[['main_category', 'main_category:confidence', 'url']] df = df[(df['main_category'] != 'Not_working') & (df['main_category:confidence'] > 0.5)] char_blacklist = list(chr(i) for i in range(32, 127) if i <= 64 or i >= 91 and i <= 96 or i >= 123) stopwords = nltk.corpus.stopwords.words('english') stopwords.extend(char_blacklist) language_whitelist = ['en'] english_vocab = set(w.lower() for w in nltk.corpus.words.words()) blacklist_domain = ['.it', '.ru', '.cn', '.jp', '.tw', '.de', '.pl', '.fr', '.hu', '.bg', '.nl'] df = df[~df['url'].str.endswith(tuple(blacklist_domain))] df['tokenized_words'] = '' counter = 0 for i, row in df.iterrows(): counter += 1 if counter >= 50: break print("{}, {}/{}".format(row['url'], counter, len(df))) try: html = urlopen('http://' + row['url'], timeout=1).read() except: continue soup = BeautifulSoup(html, "html.parser") for script in soup(["script", "style"]): script.extract() text = soup.get_text() lines = (line.strip() for line in text.splitlines()) chunks = (phrase.strip() for line in lines for phrase in line.split(" ")) text = '\n'.join(chunk.lower() for chunk in chunks if chunk) filter_text = " ".join(w for w in nltk.word_tokenize(text) \ if w.lower() in english_vocab) tokens = nltk.word_tokenize(filter_text) allWordExceptStopDist = nltk.FreqDist( w.lower() for w in tokens if w not in stopwords and len(w) >= 3 and w[0] not in char_blacklist) all_words = [i for i in allWordExceptStopDist] if len(all_words) > 0: continue df.at[i, 'tokenized_words'] = all_words df = df[df['tokenized_words'] != ''] def train_machine(): char_blacklist = list(chr(i) for i in range(32, 127) if i <= 64 or i >= 91 and i <= 96 or i >= 123) stopwords = nltk.corpus.stopwords.words('english') stopwords.extend(char_blacklist) df = pd.read_csv('cleaned_data.csv') top = 50 words_frequency = {} for category in set(df['main_category'].values): all_words = [] for row in df[df['main_category'] == category]['tokenized_words'].tolist(): for word in ast.literal_eval(row): all_words.append(word) allWordExceptStopDist = nltk.FreqDist( w.lower() for w in all_words if w not in stopwords and len(w) >= 3 and w[0] not in char_blacklist) most_common = allWordExceptStopDist.most_common(top) words_frequency[category] = most_common for category in set(df['main_category'].values): words_frequency[category] = [word for word, number in words_frequency[category]] from collections import Counter features = np.zeros(df.shape[0] * top).reshape(df.shape[0], top) labels = np.zeros(df.shape[0]) counter = 0 for i, row in df.iterrows(): c = [word for word, word_count in Counter(ast.literal_eval(row['tokenized_words'])).most_common(top)] labels[counter] = list(set(df['main_category'].values)).index(row['main_category']) for word in c: if word in words_frequency[row['main_category']]: features[counter][words_frequency[row['main_category']].index(word)] = 1 counter += 1 return labels, features def no_filter_data(): file = 'URL-categorization-DFE.csv' df = pd.read_csv(file)[['main_category', 'main_category:confidence', 'url']] df = df[(df['main_category'] != 'Not_working') & (df['main_category:confidence'] > 0.5)] df['tokenized_words'] = '' counter = 0 for i, row in df.iterrows(): counter += 1 print("{}, {}/{}".format(row['url'], counter, len(df))) try: html = urlopen('http://' + row['url'], timeout=1).read() except: continue soup = BeautifulSoup(html, "html.parser") for script in soup(["script", "style"]): script.extract() text = soup.get_text() lines = (line.strip() for line in text.splitlines()) chunks = (phrase.strip() for line in lines for phrase in line.split(" ")) text = '\n'.join(chunk.lower() for chunk in chunks if chunk) tokens = nltk.word_tokenize(text) df.at[i, 'tokenized_words'] = tokens if len(tokens) > 0 else '' df = df[df['tokenized_words'] != ''] return df # if os.path.isfile("cleaned_data.csv"): # labels, features = train_machine() # else: # conctruct_dataset() # labels, features = train_machine() if not os.path.isfile("Datasets/full_data.csv"): df = no_filter_data() df.shape from sklearn.metrics import accuracy_score from scipy.sparse import coo_matrix X_sparse = coo_matrix(features) from sklearn.utils import shuffle X, X_sparse, y = shuffle(features, X_sparse, labels, random_state=0) from sklearn.model_selection import train_test_split X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33, random_state=42) from sklearn.linear_model import LogisticRegression lr = LogisticRegression() lr.fit(X_train, y_train) predictions = lr.predict(X_test) score = lr.score(X_test, y_test) print(predictions) print(score) from sklearn.tree import DecisionTreeClassifier dtc = DecisionTreeClassifier() dtc.fit(X_train, y_train) predictions = dtc.predict(X_test) score = dtc.score(X_test, y_test) print(predictions) print(score) from sklearn import svm svm = svm.SVC() svm.fit(X_train, y_train) predictions = svm.predict(X_test) score = svm.score(X_test, y_test) print(predictions) print(score) ###Output [12. 12. 15. ... 1. 1. 18.] 0.3989290495314592
Hilbert Transform and Analytic Representation.ipynb
###Markdown The Hilbert Transform and Analytic Representation of Signals Hilbert TransformThe Hilbert transform of a function $u(t)$ is the covolution of the function with $\frac{1}{{\pi}t}$ (the Cauchy Kernel), or $u(t)*\frac{1}{{\pi}t}$:Alternatively, an easier concept to grasp is that we are simply applying a $90^{\circ}$ phase shift to **all sinusoids in the Fourier series** that comprises the signal.1. Take the FFT of a signal2. Rotate phase of Fourier coefficients by $90^\circ$, or $\frac{\pi}{2} rad$ + Positive frequencies are shifted by $\frac{\pi}{2}$ (rotate in counterclockwise direction) + Negative frequencies are shifted by $-\frac{\pi}{2}$ (rotate in clockwise direction)>*NOTE: a $90^\circ$ phase shift of a sinusoid is easily accomplished by multiplying the Fourier coefficient by $j$. To see this, apply a $\frac{\pi}{2}$ phase shift to a complex exponential sinusoid:*>>$$e^{j(\omega{t} + \frac{\pi}{2})} = e^{j\omega{t}}{\cdot}e^{j\frac{\pi}{2}}$$>>...and since $e^{j\frac{\pi}{2}} = j$, the phase-shifted sinusoid will be: $je^{j\omega{t}}$ (or $-je^{-j\omega{t}}$ for negative $\omega$). Voila!3. Take iFFT of the rotated Fourier coefficients ###Code import numpy as np from scipy import signal import matplotlib.pyplot as plt from mpl_toolkits.mplot3d import Axes3D import seaborn as sns sns.set() def hilbert_t(signal): # take FFT to compute complex fourier coefficients: fft = np.fft.fft(signal) # rotate coefficients by +/-90 degrees: if len(fft)%2 == 0: neg = -1j*fft[:int(len(fft)/2)] # coefficients of negative harmonics pos = 1j*fft[int(len(fft)/2):] # coefficients of positive harmonics shifted_fft = np.concatenate((neg, pos)) else: neg = -1j*fft[:int(np.floor(len(fft)/2))] pos = 1j*fft[int(np.floor(len(fft)/2))+1:] shifted_fft = np.concatenate((neg, np.array([fft[int(np.floor(len(fft)/2))]]), pos)) assert len(shifted_fft) == len(fft), str(len(shifted_fft)) + " does not equal " + str(len(fft)) # apply iFFT: ifft = np.fft.ifft(shifted_fft) # add to original (real) signal values: return np.real(signal) + 1j*np.real(ifft) num_samples = 500 t = np.linspace(0,5*2*np.pi,num_samples) test_signal = np.sin(t) + np.sin(2*t) #test_signal = np.random.random(num_samples) test_signal_ht = hilbert_t(test_signal) fig, axs = plt.subplots(nrows=2, ncols=1, figsize=(15,5)) axs[0].plot(t, np.real(test_signal)) axs[0].set_title("Original Signal x(t) (Real-valued)") axs[1].plot(t, np.imag(test_signal_ht)) axs[1].set_title("Hilbert Transform, (Imaginary Values Plotted)") fig.tight_layout(pad=1) ###Output _____no_output_____ ###Markdown Analytic Representation of a SignalThe **Analytic Representation** of a real-valued signal is the original signal added to its Hilbert transform. This results in a complex-valued signal where:1. The real values are the same as the real values of the original signal2. The imaginary values are the values provided by the Hilbert transform (i.e., its fourier series's sinusoids shifted 90 deg)The AR is used because it discards the "negative" frequency components, representing them instead as the complex part of the fourier coefficients. This is possible because the Hilbert transform is **conjugate symmetric**, (i.e. has the property of **Hermitian symmetry**).+ A signal $x(t)$ is **conjugate symmetric** if $x(t) = x^\ast(-t)$, that is, if its real part is **even** and it's imaginary part is **odd** about the origin + see https://www.youtube.com/watch?v=O1HRtZmkI4E+ Practically, this is accomplished by adding the "negative" frequency components to the "positive", resulting in a complex fourier coefficient:$$e^{j(\omega{t}+({-\omega})t)} = e^{j\omega{t}-j\omega{t}} = e^{j\omega{t}}e^{-j\omega{t}} \rightarrow$$$$(real - j\cdot{imaginary})e^{j\omega{t}}$$and since $-j = e^{-j\frac{\pi}{2}}$:$$real{\cdot}e^{j\omega{t}} + e^{-j\frac{\pi}{2}}imaginary{\cdot}e^{j\omega{t}}$$$$real{\cdot}e^{j\omega{t}} + imaginary{\cdot}e^{-j\frac{\pi}{2} + j\omega{t}}$$$$real{\cdot}e^{j\omega{t}} + imaginary{\cdot}e^{j(\omega{t} - \frac{\pi}{2})}$$ ###Code gridsize = (3, 2) # overall grid layout for the figure fig = plt.figure(figsize=(12, 8)) ax1 = plt.subplot2grid(gridsize, (0,0)) # top left in the 3x2 grid ax2 = plt.subplot2grid(gridsize, (0,1)) # top right in the 3x2 grid ax3 = plt.subplot2grid(gridsize, (1, 0), colspan=2, rowspan=2, projection='3d') # this one takes up 2 rows and 2 columns # manipulate the little graphs: ax1.plot(t, np.real(test_signal)) ax1.set_title("Original Signal x(t) (Real-valued)") ax2.plot(t, np.imag(test_signal_ht)) ax2.set_title("Hilbert Transform, (Imaginary Values Plotted)") # manipulate the big graph: #ax3.plot(projection='3d') ax3.set_title('3-Dimensional View of Analytic Representation') #ax3.plot(xs=t, ys=np.imag(test_signal_ht), zs=test_signal, zdir='z', label='Analytic Representation') ax3.scatter(xs=t, ys=np.imag(test_signal_ht), zs=test_signal, zdir='z', label='Analytic Representation') fig.tight_layout(pad=1) # do this to prevent xlabels on upper plots overlapping title on lower plot # The scipy hilbert transform function produces the analytic representation of the input: # 1. real values are the original, real-valued input signal # 2. imaginary values are the hilbert-transformed signal (phase-shifted 90 degrees) test_signal_scipy_ht = signal.hilbert(np.real(test_signal)) fig, ax = plt.subplots() ax.plot(np.imag(test_signal_scipy_ht), label="scipy hilbert transform") ax.plot(np.imag(test_signal_ht), label="my hilbert transform") ax.legend(loc="lower right") %timeit temp = hilbert_t(test_signal) %timeit temp = signal.hilbert(test_signal) ###Output 30 µs ± 101 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) ###Markdown Normalized Analytic Respresentation Scale the AR to "normalize", so that the sum of all the squared magnitudes is 1 Probability Amplitude Function and Probability Density FunctionNotes from:+ Funkhouser Scott, Suski William and Winn Andrew, "Waveform information from quantum mechanical entropy", *Proceedings of the Royal Society,* Vol 472, Issue 2190, June 2016. ###Code num_samples = 200 t = np.linspace(-np.pi,np.pi,num_samples) # from Funkhouser, Winn, Suski paper: d = 1 theta = 10 c = (np.power((8/np.pi),0.25)*(1/np.sqrt(d))) test_signal = c*np.exp(-np.power(t,2)/np.power(d,2))*np.cos(theta*t) test_signal_ht = signal.hilbert(test_signal) gridsize = (3, 2) # overall grid layout for the figure fig = plt.figure(figsize=(12, 8)) ax1 = plt.subplot2grid(gridsize, (0,0)) # top left in the 3x2 grid ax2 = plt.subplot2grid(gridsize, (0,1)) # top right in the 3x2 grid ax3 = plt.subplot2grid(gridsize, (1, 0), colspan=2, rowspan=2, projection='3d') # this one takes up 2 rows and 2 columns # manipulate the little graphs: ax1.plot(t, np.real(test_signal)) ax1.set_title("Original Signal x(t) (Real-valued)") ax2.plot(t, np.imag(test_signal_ht)) ax2.set_title("Hilbert Transform, (Imaginary Values Plotted)") # manipulate the big graph: #ax3.plot(projection='3d') ax3.set_title('3-Dimensional View of Analytic Representation') #ax3.plot(xs=t, ys=np.imag(test_signal_ht), zs=test_signal, zdir='z', label='Analytic Representation') ax3.scatter(xs=t, ys=np.imag(test_signal_ht), zs=test_signal, zdir='z', label='Analytic Representation') fig.tight_layout(pad=1) # do this to prevent xlabels on upper plots overlapping title on lower plot fig, ax = plt.subplots() ax.plot(t, np.real(test_signal_ht), label='real') ax.plot(t, np.imag(test_signal_ht), label='imaginary', color='green') ax.plot(t, np.abs(test_signal_ht), label='envelope', color='red') ax.legend(loc='upper right') norm = np.sum(np.abs(test_signal_ht)) normalized_test_signal_ht = test_signal_ht/norm np.sum(np.abs(normalized_test_signal_ht)) fig, ax = plt.subplots() ax.plot(t, np.real(normalized_test_signal_ht), label='real') ax.plot(t, np.imag(normalized_test_signal_ht), label='imaginary', color='green') ax.plot(t, np.abs(normalized_test_signal_ht), label='envelope', color='red') ax.legend(loc='upper right') time_domain_squared_amp = np.square(np.abs(test_signal_ht)) freq_domain_squared_amp = np.square(np.abs(np.fft.fft(test_signal_ht))) fig, ax = plt.subplots(ncols=2, figsize=(14, 4)) ax[0].plot(t, time_domain_squared_amp) ax[0].set_title('Time Domain Squared Amplitude') ax[1].plot(freq_domain_squared_amp) ax[1].set_title('Frequency Domain Squared Amplitude') ###Output _____no_output_____
Practice6/3_char_rnn_train.ipynb
###Markdown chars[0] converts index to char vocab['a'] converts char to index ###Code # Now convert all text to index using vocab! corpus = np.array(list(map(vocab.get, data))) print ("Type of 'corpus' is %s, shape is %s, and length is %d" % (type(corpus), corpus.shape, len(corpus))) check_len = 10 print ("\n'corpus' looks like %s" % (corpus[0:check_len])) for i in range(check_len): _wordidx = corpus[i] print ("[%d/%d] chars[%02d] corresponds to '%s'" % (i, check_len, _wordidx, chars[_wordidx])) # Generate batch data batch_size = 50 seq_length = 200 num_batches = int(corpus.size / (batch_size * seq_length)) # First, reduce the length of corpus to fit batch_size corpus_reduced = corpus[:(num_batches*batch_size*seq_length)] xdata = corpus_reduced ydata = np.copy(xdata) ydata[:-1] = xdata[1:] ydata[-1] = xdata[0] print ('xdata is ... %s and length is %d' % (xdata, xdata.size)) print ('ydata is ... %s and length is %d' % (ydata, xdata.size)) print ("") # Second, make batch xbatches = np.split(xdata.reshape(batch_size, -1), num_batches, 1) ybatches = np.split(ydata.reshape(batch_size, -1), num_batches, 1) print ("Type of 'xbatches' is %s and length is %d" % (type(xbatches), len(xbatches))) print ("Type of 'ybatches' is %s and length is %d" % (type(ybatches), len(ybatches))) print ("") # How can we access to xbatches?? nbatch = 5 temp = xbatches[0:nbatch] print ("Type of 'temp' is %s and length is %d" % (type(temp), len(temp))) for i in range(nbatch): temp2 = temp[i] print ("Type of 'temp[%d]' is %s and shape is %s" % (i, type(temp2), temp2.shape,)) ###Output xdata is ... [36 22 7 ..., 11 25 3] and length is 1700000 ydata is ... [22 7 0 ..., 25 3 36] and length is 1700000 Type of 'xbatches' is <type 'list'> and length is 170 Type of 'ybatches' is <type 'list'> and length is 170 Type of 'temp' is <type 'list'> and length is 5 Type of 'temp[0]' is <type 'numpy.ndarray'> and shape is (50, 200) Type of 'temp[1]' is <type 'numpy.ndarray'> and shape is (50, 200) Type of 'temp[2]' is <type 'numpy.ndarray'> and shape is (50, 200) Type of 'temp[3]' is <type 'numpy.ndarray'> and shape is (50, 200) Type of 'temp[4]' is <type 'numpy.ndarray'> and shape is (50, 200) ###Markdown Now, we are ready to make our RNN model with seq2seq ###Code # Important RNN parameters vocab_size = len(vocab) rnn_size = 128 num_layers = 2 grad_clip = 5. def unit_cell(): return tf.contrib.rnn.BasicLSTMCell(rnn_size,state_is_tuple=True,reuse=tf.get_variable_scope().reuse) cell = tf.contrib.rnn.MultiRNNCell([unit_cell() for _ in range(num_layers)]) input_data = tf.placeholder(tf.int32, [batch_size, seq_length]) targets = tf.placeholder(tf.int32, [batch_size, seq_length]) istate = cell.zero_state(batch_size, tf.float32) # Weigths with tf.variable_scope('rnnlm'): softmax_w = tf.get_variable("softmax_w", [rnn_size, vocab_size]) softmax_b = tf.get_variable("softmax_b", [vocab_size]) with tf.device("/cpu:0"): embedding = tf.get_variable("embedding", [vocab_size, rnn_size]) inputs = tf.split(tf.nn.embedding_lookup(embedding, input_data), seq_length, 1) inputs = [tf.squeeze(_input, [1]) for _input in inputs] # Output def loop(prev, _): prev = tf.nn.xw_plus_b(prev, softmax_w, softmax_b) prev_symbol = tf.stop_gradient(tf.argmax(prev, 1)) return tf.nn.embedding_lookup(embedding, prev_symbol) """ loop_function: If not None, this function will be applied to the i-th output in order to generate the i+1-st input, and decoder_inputs will be ignored, except for the first element ("GO" symbol). """ outputs, last_state = tf.contrib.rnn.static_rnn(cell, inputs, istate , scope='rnnlm') output = tf.reshape(tf.concat(outputs, 1), [-1, rnn_size]) logits = tf.nn.xw_plus_b(output, softmax_w, softmax_b) probs = tf.nn.softmax(logits) # Loss loss = tf.contrib.legacy_seq2seq.sequence_loss_by_example([logits], # Input [tf.reshape(targets, [-1])], # Target [tf.ones([batch_size * seq_length])], # Weight vocab_size) # Optimizer cost = tf.reduce_sum(loss) / batch_size / seq_length final_state = last_state lr = tf.Variable(0.0, trainable=False) tvars = tf.trainable_variables() grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars), grad_clip) _optm = tf.train.AdamOptimizer(lr) optm = _optm.apply_gradients(zip(grads, tvars)) print ("Network Ready") # Train the model! num_epochs = 50 save_every = 500 learning_rate = 0.002 decay_rate = 0.97 sess = tf.Session() sess.run(tf.initialize_all_variables()) summary_writer = tf.summary.FileWriter(save_dir, graph=sess.graph) saver = tf.train.Saver(tf.all_variables()) init_time = time.time() for epoch in range(num_epochs): # Learning rate scheduling sess.run(tf.assign(lr, learning_rate * (decay_rate ** epoch))) state = sess.run(istate) batchidx = 0 for iteration in range(num_batches): start_time = time.time() randbatchidx = np.random.randint(num_batches) xbatch = xbatches[batchidx] ybatch = ybatches[batchidx] batchidx = batchidx + 1 # Note that, num_batches = len(xbatches) # Train! train_loss, state, _ = sess.run([cost, final_state, optm] , feed_dict={input_data: xbatch, targets: ybatch, istate: state}) total_iter = epoch*num_batches + iteration end_time = time.time(); duration = end_time - start_time if total_iter % 100 == 0: print ("[%d/%d] cost: %.4f / Each batch learning took %.4f sec" % (total_iter, num_epochs*num_batches, train_loss, duration)) if total_iter % save_every == 0: ckpt_path = os.path.join(save_dir, 'model.ckpt') saver.save(sess, ckpt_path, global_step = total_iter) # Save network! print("model saved to '%s'" % (ckpt_path)) ###Output WARNING:tensorflow:From <ipython-input-9-c08af8068626>:8: initialize_all_variables (from tensorflow.python.ops.variables) is deprecated and will be removed after 2017-03-02. Instructions for updating: Use `tf.global_variables_initializer` instead. WARNING:tensorflow:From <ipython-input-9-c08af8068626>:10: all_variables (from tensorflow.python.ops.variables) is deprecated and will be removed after 2017-03-02. Instructions for updating: Please use tf.global_variables instead. [0/8500] cost: 4.6006 / Each batch learning took 2.2222 sec model saved to 'data/linux_kernel/model.ckpt' [100/8500] cost: 3.1259 / Each batch learning took 0.3366 sec [200/8500] cost: 2.5992 / Each batch learning took 0.3258 sec [300/8500] cost: 2.4603 / Each batch learning took 0.3260 sec [400/8500] cost: 2.2591 / Each batch learning took 0.3136 sec [500/8500] cost: 2.0035 / Each batch learning took 0.3140 sec model saved to 'data/linux_kernel/model.ckpt' [600/8500] cost: 1.9589 / Each batch learning took 0.3695 sec [700/8500] cost: 1.8066 / Each batch learning took 0.3130 sec [800/8500] cost: 1.7801 / Each batch learning took 0.3119 sec [900/8500] cost: 1.7433 / Each batch learning took 0.4185 sec [1000/8500] cost: 1.6289 / Each batch learning took 0.3153 sec model saved to 'data/linux_kernel/model.ckpt' [1100/8500] cost: 1.6194 / Each batch learning took 0.3388 sec [1200/8500] cost: 1.4603 / Each batch learning took 0.3129 sec [1300/8500] cost: 1.5877 / Each batch learning took 0.3141 sec [1400/8500] cost: 1.5235 / Each batch learning took 0.3087 sec [1500/8500] cost: 1.5317 / Each batch learning took 0.3440 sec model saved to 'data/linux_kernel/model.ckpt' [1600/8500] cost: 1.5362 / Each batch learning took 0.4632 sec [1700/8500] cost: 1.4946 / Each batch learning took 0.3351 sec [1800/8500] cost: 1.4392 / Each batch learning took 0.3374 sec [1900/8500] cost: 1.4224 / Each batch learning took 0.3323 sec [2000/8500] cost: 1.4797 / Each batch learning took 0.3115 sec model saved to 'data/linux_kernel/model.ckpt' [2100/8500] cost: 1.4381 / Each batch learning took 0.3863 sec [2200/8500] cost: 1.3570 / Each batch learning took 0.3080 sec [2300/8500] cost: 1.3689 / Each batch learning took 0.3120 sec [2400/8500] cost: 1.3241 / Each batch learning took 0.3174 sec [2500/8500] cost: 1.3431 / Each batch learning took 0.3326 sec model saved to 'data/linux_kernel/model.ckpt' [2600/8500] cost: 1.3311 / Each batch learning took 0.4586 sec [2700/8500] cost: 1.2888 / Each batch learning took 0.3147 sec [2800/8500] cost: 1.3359 / Each batch learning took 0.3262 sec [2900/8500] cost: 1.1899 / Each batch learning took 0.3310 sec [3000/8500] cost: 1.3265 / Each batch learning took 0.3324 sec model saved to 'data/linux_kernel/model.ckpt' [3100/8500] cost: 1.2806 / Each batch learning took 0.5395 sec [3200/8500] cost: 1.3113 / Each batch learning took 0.3448 sec [3300/8500] cost: 1.3262 / Each batch learning took 0.3422 sec [3400/8500] cost: 1.3011 / Each batch learning took 0.3195 sec [3500/8500] cost: 1.2781 / Each batch learning took 0.3138 sec model saved to 'data/linux_kernel/model.ckpt' [3600/8500] cost: 1.2607 / Each batch learning took 0.3156 sec [3700/8500] cost: 1.2897 / Each batch learning took 0.4064 sec [3800/8500] cost: 1.2809 / Each batch learning took 0.3063 sec [3900/8500] cost: 1.2301 / Each batch learning took 0.3330 sec [4000/8500] cost: 1.2372 / Each batch learning took 0.3157 sec model saved to 'data/linux_kernel/model.ckpt' [4100/8500] cost: 1.2088 / Each batch learning took 0.3536 sec [4200/8500] cost: 1.2277 / Each batch learning took 0.3146 sec [4300/8500] cost: 1.2095 / Each batch learning took 0.3148 sec [4400/8500] cost: 1.1840 / Each batch learning took 0.3425 sec [4500/8500] cost: 1.2459 / Each batch learning took 0.3368 sec model saved to 'data/linux_kernel/model.ckpt' [4600/8500] cost: 1.0941 / Each batch learning took 0.4124 sec [4700/8500] cost: 1.2265 / Each batch learning took 0.3164 sec [4800/8500] cost: 1.1862 / Each batch learning took 0.3307 sec [4900/8500] cost: 1.2198 / Each batch learning took 0.3371 sec [5000/8500] cost: 1.2345 / Each batch learning took 0.3298 sec model saved to 'data/linux_kernel/model.ckpt' [5100/8500] cost: 1.2081 / Each batch learning took 0.3418 sec [5200/8500] cost: 1.2043 / Each batch learning took 0.3105 sec [5300/8500] cost: 1.1929 / Each batch learning took 0.3377 sec [5400/8500] cost: 1.2155 / Each batch learning took 0.3373 sec [5500/8500] cost: 1.2052 / Each batch learning took 0.3908 sec model saved to 'data/linux_kernel/model.ckpt' [5600/8500] cost: 1.1683 / Each batch learning took 0.3207 sec [5700/8500] cost: 1.1695 / Each batch learning took 0.3358 sec [5800/8500] cost: 1.1485 / Each batch learning took 0.3392 sec [5900/8500] cost: 1.1671 / Each batch learning took 0.3451 sec [6000/8500] cost: 1.1481 / Each batch learning took 0.3391 sec model saved to 'data/linux_kernel/model.ckpt' [6100/8500] cost: 1.1262 / Each batch learning took 0.3186 sec [6200/8500] cost: 1.1943 / Each batch learning took 0.4622 sec [6300/8500] cost: 1.0425 / Each batch learning took 0.3805 sec [6400/8500] cost: 1.1697 / Each batch learning took 0.3373 sec [6500/8500] cost: 1.1365 / Each batch learning took 0.3838 sec model saved to 'data/linux_kernel/model.ckpt' [6600/8500] cost: 1.1704 / Each batch learning took 0.3196 sec [6700/8500] cost: 1.1841 / Each batch learning took 0.3364 sec [6800/8500] cost: 1.1521 / Each batch learning took 0.3404 sec [6900/8500] cost: 1.1598 / Each batch learning took 0.3631 sec [7000/8500] cost: 1.1523 / Each batch learning took 0.3372 sec model saved to 'data/linux_kernel/model.ckpt' [7100/8500] cost: 1.1689 / Each batch learning took 0.3289 sec [7200/8500] cost: 1.1579 / Each batch learning took 0.3935 sec [7300/8500] cost: 1.1316 / Each batch learning took 0.3154 sec [7400/8500] cost: 1.1284 / Each batch learning took 0.3672 sec [7500/8500] cost: 1.1087 / Each batch learning took 0.3855 sec model saved to 'data/linux_kernel/model.ckpt' [7600/8500] cost: 1.1276 / Each batch learning took 0.5031 sec [7700/8500] cost: 1.1090 / Each batch learning took 0.3774 sec [7800/8500] cost: 1.0901 / Each batch learning took 0.4558 sec [7900/8500] cost: 1.1609 / Each batch learning took 0.4658 sec [8000/8500] cost: 1.0116 / Each batch learning took 0.3612 sec model saved to 'data/linux_kernel/model.ckpt' [8100/8500] cost: 1.1309 / Each batch learning took 0.3854 sec [8200/8500] cost: 1.1066 / Each batch learning took 0.3160 sec [8300/8500] cost: 1.1417 / Each batch learning took 0.3196 sec [8400/8500] cost: 1.1568 / Each batch learning took 0.3270 sec ###Markdown Run the command line tensorboard --logdir=/tmp/tf_logs/char_rnn_tutorial Open http://localhost:6006/ into your web browser ###Code print ("Done!! It took %.4f second. " %(time.time() - init_time)) ###Output Done!! It took 5238.4040 second.
stats-newtextbook-python/samples/6-3-ロジスティック回帰.ipynb
###Markdown 第6部 一般化線形モデル|Pythonで学ぶ統計学入門 3章 ロジスティック回帰 分析の準備 ###Code # 数値計算に使うライブラリ import numpy as np import pandas as pd import scipy as sp from scipy import stats # グラフを描画するライブラリ from matplotlib import pyplot as plt import seaborn as sns sns.set() # 統計モデルを推定するライブラリ(ワーニングが出ることもあります) import statsmodels.formula.api as smf import statsmodels.api as sm # 表示桁数の指定 %precision 3 # グラフをjupyter Notebook内に表示させるための指定 %matplotlib inline ###Output _____no_output_____ ###Markdown 実装:データの読み込みと図示 ###Code # データの読み込み test_result = pd.read_csv("6-3-1-logistic-regression.csv") print(test_result.head(3)) # データの図示 sns.barplot(x = "hours",y = "result", data = test_result, palette='gray_r') # 勉強時間ごとの合格率 print(test_result.groupby("hours").mean()) ###Output result hours 0 0.0 1 0.0 2 0.1 3 0.1 4 0.4 5 0.4 6 0.9 7 0.8 8 0.9 9 1.0 ###Markdown 実装:ロジスティック回帰 ###Code # モデル化 mod_glm = smf.glm(formula = "result ~ hours", data = test_result, family=sm.families.Binomial()).fit() # 参考:リンク関数を指定する logistic_reg = smf.glm(formula = "result ~ hours", data = test_result, family=sm.families.Binomial(link=sm.families.links.logit)).fit() ###Output _____no_output_____ ###Markdown 実装:ロジスティック回帰の結果の出力 ###Code # 結果の出力 mod_glm.summary() ###Output _____no_output_____ ###Markdown 実装:AICによるモデル選択 ###Code # Nullモデル mod_glm_null = smf.glm( "result ~ 1", data = test_result, family=sm.families.Binomial()).fit() # AICの比較 print("Nullモデル :", mod_glm_null.aic.round(3)) print("変数入りモデル:", mod_glm.aic.round(3)) ###Output Nullモデル : 139.989 変数入りモデル: 72.028 ###Markdown 実装:ロジスティック回帰曲線の図示 ###Code # lmplotでロジスティック回帰曲線を図示する sns.lmplot(x = "hours", y = "result", data = test_result, logistic = True, scatter_kws = {"color": "black"}, line_kws = {"color": "black"}, x_jitter = 0.1, y_jitter = 0.02) ###Output _____no_output_____ ###Markdown 実装:成功確率の予測 ###Code # 0~9まで1ずつ増える等差数列 exp_val = pd.DataFrame({ "hours": np.arange(0, 10, 1) }) # 成功確率の予測値 pred = mod_glm.predict(exp_val) pred ###Output _____no_output_____ ###Markdown ロジスティック回帰の係数とオッズ比の関係 ###Code # 勉強時間が1時間の時の合格率 exp_val_1 = pd.DataFrame({"hours": [1]}) pred_1 = mod_glm.predict(exp_val_1) # 勉強時間が2時間の時の合格率 exp_val_2 = pd.DataFrame({"hours": [2]}) pred_2 = mod_glm.predict(exp_val_2) # オッズ odds_1 = pred_1 / (1 - pred_1) odds_2 = pred_2 / (1 - pred_2) # 対数オッズ比 sp.log(odds_2 / odds_1) # 係数 mod_glm.params["hours"] # 補足:オッズ比に戻す sp.exp(mod_glm.params["hours"]) ###Output _____no_output_____
notebooks/08_extended_refactored.ipynb
###Markdown Deep Reinforcement Learning in Action - Chaper 8 - Intrinsic Curiosity Module - Refactor and Extended - save model, load model- save plots of losses and episode lengths- bug fixes including fixing large memory footprint from saving tensors with gradients to a list- exposes additional hparams through the param dictionary ###Code import gym from nes_py.wrappers import JoypadSpace # A import gym_super_mario_bros from gym_super_mario_bros.actions import SIMPLE_MOVEMENT, COMPLEX_MOVEMENT # B import torch from torch import nn from torch import optim import torch.nn.functional as F from collections import deque from tqdm.notebook import trange from random import shuffle import matplotlib.pyplot as plt from skimage.transform import resize # A import numpy as np ###Output _____no_output_____ ###Markdown Random Agent Mario ###Code env = gym_super_mario_bros.make('SuperMarioBros-v0') env = JoypadSpace(env, COMPLEX_MOVEMENT) # C done = True for step in range(2500): # D if done: state = env.reset() state, reward, done, info = env.step(env.action_space.sample()) env.render() env.close() ###Output _____no_output_____ ###Markdown Downscaling ###Code # code intentionally duplicated in training code block def downscale_obs(obs, new_size=(42, 42), to_gray=True): if to_gray: return resize(obs, new_size, anti_aliasing=True).max(axis=2) # B else: return resize(obs, new_size, anti_aliasing=True) env = gym_super_mario_bros.make('SuperMarioBros-v0') plt.imshow(env.render("rgb_array")) plt.imshow(downscale_obs(env.render("rgb_array"))) env.close() ###Output _____no_output_____ ###Markdown Train Agent ###Code # Curiosity-driven Exploration by Self-supervised Prediction # https://pathak22.github.io/noreward-rl/resources/icml17.pdf # https://github.com/pathak22/noreward-rl/blob/master/src/constants.py # TODO: distributional, n-step import logging import sys from einops import rearrange LOGGER = logging.getLogger(__name__) LOGGER.setLevel(logging.INFO) handler = logging.StreamHandler(sys.stdout) handler.setLevel(logging.DEBUG) formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s') handler.setFormatter(formatter) LOGGER.addHandler(handler) device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') LOGGER.info('Pytorch using device: %s', device) # Listing 8.9 default_params_book = { 'batch_size': 150, 'beta': 0.2, 'lambda': 0.1, 'eta': 1.0, 'gamma': 0.2, 'max_episode_len': 100, # this is really the max episode length without exceeding min_progress 'min_progress': 15, 'action_repeats': 6, # something is hard-coded where it will throw an error if < 3 'frames_per_state': 3, # book uses 3 and hardcoded it into the models 'epochs': 5000, 'eps': 0.15, 'test_eps': 0.15, 'switch_to_eps_greedy': 1000, 'use_extrinsic': False, 'experience_replay_length': 1000, 'learning_rate': 0.001, 'include_LSTM': False, } default_params_paper = { 'batch_size': 1000, 'beta': 0.2, 'lambda': 1.0, 'eta': 0.01, 'gamma': 0.99, 'max_episode_len': 100, # this is really the max episode length without exceeding min_progress 'min_progress': 15, # something about the creation of state2 on first pass causes it to throw an error if < frames_per_state # now fixed 'action_repeats': 6, 'frames_per_state': 6, # book uses 3 and hardcoded it into the models 'epochs': 6000, 'eps': 0.15, 'test_eps': 0.15, 'switch_to_eps_greedy': 1000, 'use_extrinsic': True, 'experience_replay_length': 1000, 'learning_rate': 0.001, 'include_LSTM': True, } params = { 'batch_size': 1000, 'beta': 0.2, 'lambda': 1.0, 'eta': 0.01, 'gamma': 0.99, 'max_episode_len': 100, # this is really the max episode length without exceeding min_progress 'min_progress': 15, # something about the creation of state2 on first pass causes it to throw an error if < frames_per_state # now fixed 'action_repeats': 6, 'frames_per_state': 6, # book uses 3 and hardcoded it into the models 'epochs': 6000, 'eps': 0.15, 'test_eps': 0.15, 'switch_to_eps_greedy': 1000, 'use_extrinsic': True, 'experience_replay_length': 1000, 'learning_rate': 0.001, 'include_LSTM': True, } # Listing 8.2 def downscale_obs(obs, new_size=(42, 42), to_gray=True): if to_gray: return resize(obs, new_size, anti_aliasing=True).max(axis=2) # B else: return resize(obs, new_size, anti_aliasing=True) # Listing 8.4 def prepare_state(state): # A return torch.from_numpy(downscale_obs(state, to_gray=True)).float().unsqueeze(dim=0) def prepare_multi_state(state1, state2): # B state1 = state1.clone() tmp = torch.from_numpy(downscale_obs(state2, to_gray=True)).float() state1[0][0] = state1[0][1] state1[0][1] = state1[0][2] state1[0][2] = tmp return state1 def prepare_initial_state(state, N=params['frames_per_state']): # C state_ = torch.from_numpy(downscale_obs(state, to_gray=True)).float() tmp = state_.repeat((N, 1, 1)) return tmp.unsqueeze(dim=0) # Listing 8.5 def policy(qvalues, eps=None): # A if eps is not None: if torch.rand(1) < eps: return torch.randint(low=0, high=7, size=(1,)) else: return torch.argmax(qvalues) else: # if epsilon (eps) is not provided, use a softmax policy by sampling from the softmax # using the torch.multinomial function LOGGER.debug("q values shape: %s", qvalues.shape) return torch.multinomial(F.softmax(F.normalize(qvalues), dim=1), num_samples=1) # Listing 8.6 class ExperienceReplay: def __init__(self, N=500, batch_size=100): self.N = N # A self.batch_size = batch_size # B self.memory = [] self.counter = 0 def add_memory(self, state1, action, reward, state2): self.counter += 1 if self.counter % 500 == 0: # C self.shuffle_memory() if len(self.memory) < self.N: # D self.memory.append((state1, action, reward, state2)) else: rand_index = np.random.randint(0, self.N-1) self.memory[rand_index] = (state1, action, reward, state2) def shuffle_memory(self): # E shuffle(self.memory) def get_batch(self): # F if len(self.memory) < self.batch_size: batch_size = len(self.memory) else: batch_size = self.batch_size if len(self.memory) < 1: LOGGER.error("Error: No data in memory.") return None # G ind = np.random.choice(np.arange(len(self.memory)), batch_size, replace=False) batch = [self.memory[i] for i in ind] # batch is a list of tuples state1_batch = torch.stack([x[0].squeeze(dim=0) for x in batch], dim=0) action_batch = torch.Tensor([x[1] for x in batch]).long() reward_batch = torch.Tensor([x[2] for x in batch]) state2_batch = torch.stack([x[3].squeeze(dim=0) for x in batch], dim=0) LOGGER.debug(state2_batch.shape) return state1_batch, action_batch, reward_batch, state2_batch # Listing 8.7 class Phi(nn.Module): # A def __init__(self): super(Phi, self).__init__() self.conv1 = nn.Conv2d( params['frames_per_state'], 32, kernel_size=(3, 3), stride=2, padding=1) self.conv2 = nn.Conv2d(32, 32, kernel_size=(3, 3), stride=2, padding=1) self.conv3 = nn.Conv2d(32, 32, kernel_size=(3, 3), stride=2, padding=1) self.conv4 = nn.Conv2d(32, 32, kernel_size=(3, 3), stride=2, padding=1) def forward(self, x): x = F.normalize(x) y = F.elu(self.conv1(x)) y = F.elu(self.conv2(y)) y = F.elu(self.conv3(y)) y = F.elu(self.conv4(y)) # size [1, 32, 3, 3] batch, channels, 3 x 3 y = y.flatten(start_dim=1) # size N, 288 return y class Gnet(nn.Module): # B def __init__(self): super(Gnet, self).__init__() self.linear1 = nn.Linear(576, 256) self.linear2 = nn.Linear(256, 12) def forward(self, state1, state2): x = torch.cat((state1, state2), dim=1) y = F.relu(self.linear1(x)) y = self.linear2(y) y = F.softmax(y, dim=1) return y class Fnet(nn.Module): # C def __init__(self): super(Fnet, self).__init__() self.linear1 = nn.Linear(300, 256) self.linear2 = nn.Linear(256, 288) def forward(self, state, action): action_ = torch.zeros(action.shape[0], 12) # D indices = torch.stack( (torch.arange(action.shape[0]), action.squeeze()), dim=0) indices = indices.tolist() action_[indices] = 1. x = torch.cat((state, action_), dim=1) y = F.relu(self.linear1(x)) y = self.linear2(y) return y # Listing 8.8 class Qnetwork(nn.Module): def __init__(self): super(Qnetwork, self).__init__() # in_channels, out_channels, kernel_size, stride=1, padding=0 self.conv1 = nn.Conv2d(in_channels=params['frames_per_state'], out_channels=32, kernel_size=( 3, 3), stride=2, padding=1) self.conv2 = nn.Conv2d(32, 32, kernel_size=(3, 3), stride=2, padding=1) self.conv3 = nn.Conv2d(32, 32, kernel_size=(3, 3), stride=2, padding=1) self.conv4 = nn.Conv2d(32, 32, kernel_size=(3, 3), stride=2, padding=1) self.linear1 = nn.Linear(288, 100) self.linear2 = nn.Linear(100, 12) self.batchnorm1 = nn.BatchNorm2d(32) self.batchnorm2 = nn.BatchNorm2d(32) self.batchnorm3 = nn.BatchNorm2d(32) self.batchnorm4 = nn.BatchNorm2d(32) self.lstm = nn.LSTM(input_size=9, hidden_size=288, batch_first=True) def forward(self, x): x = F.normalize(x) LOGGER.debug("DQN input shape: %s", x.shape) y = F.elu(self.conv1(x)) LOGGER.debug("DQN conv1 output shape: %s", y.shape) y = self.batchnorm1(y) y = F.elu(self.conv2(y)) y = self.batchnorm2(y) y = F.elu(self.conv3(y)) y = self.batchnorm3(y) y = F.elu(self.conv4(y)) y = self.batchnorm4(y) LOGGER.debug("shape before lstm: %s", y.shape) if params['include_LSTM'] == True: y = rearrange(y, 'batch channels x y -> batch channels (x y)') _, (_, y) = self.lstm(y) LOGGER.debug("shape after lstm: %s", y.shape) y = rearrange( y, 'd batch hidden -> (d batch) hidden', d=1, hidden=288) else: y = y.flatten(start_dim=2) y = y.view(y.shape[0], -1, 32) y = y.flatten(start_dim=1) y = F.elu(self.linear1(y)) y = self.linear2(y) # size N, 12 return y # Listing 8.9 replay = ExperienceReplay( N=params['experience_replay_length'], batch_size=params['batch_size']) Qmodel = Qnetwork() encoder = Phi() forward_model = Fnet() inverse_model = Gnet() forward_loss = nn.MSELoss(reduction='none') inverse_loss = nn.CrossEntropyLoss(reduction='none') qloss = nn.MSELoss() all_model_params = list(Qmodel.parameters()) + list(encoder.parameters()) # A all_model_params += list(forward_model.parameters()) + \ list(inverse_model.parameters()) opt = optim.Adam(lr=params['learning_rate'], params=all_model_params) # Listing 8.10 def loss_fn(q_loss, inverse_loss, forward_loss): """ book: minimize[λ × Qloss + (1 – β)Floss + β × Gloss] paper: minimize[λ × Qloss + (1 – β)I_loss + (β)F_loss] forward model is F inverse model is G """ loss_ = (1 - params['beta']) * inverse_loss loss_ += params['beta'] * forward_loss loss_ = loss_.sum() / loss_.flatten().shape[0] loss = loss_ + params['lambda'] * q_loss return loss def reset_env(): """ Reset the environment and return a new initial state """ env.reset() state1 = prepare_initial_state(env.render('rgb_array')) return state1 # Listing 8.11 def ICM(state1, action, state2, forward_scale=1., inverse_scale=1e4): state1_hat = encoder(state1) # A state2_hat = encoder(state2) state2_hat_pred = forward_model(state1_hat.detach(), action.detach()) # B forward_pred_err = forward_scale * forward_loss(state2_hat_pred, state2_hat.detach()).sum(dim=1).unsqueeze(dim=1) pred_action = inverse_model(state1_hat, state2_hat) # C inverse_pred_err = inverse_scale * inverse_loss(pred_action, action.detach().flatten()).unsqueeze(dim=1) return forward_pred_err, inverse_pred_err # Listing 8.12 def minibatch_train(use_extrinsic=True): state1_batch, action_batch, reward_batch, state2_batch = replay.get_batch() action_batch = action_batch.view(action_batch.shape[0], 1) # A reward_batch = reward_batch.view(reward_batch.shape[0], 1) forward_pred_err, inverse_pred_err = ICM( state1_batch, action_batch, state2_batch) # B i_reward = (1. / params['eta']) * forward_pred_err # C reward = i_reward.detach() # D if use_extrinsic: # E reward += reward_batch qvals = Qmodel(state2_batch) # F reward += params['gamma'] * torch.max(qvals) reward_pred = Qmodel(state1_batch) reward_target = reward_pred.clone() indices = torch.stack((torch.arange(action_batch.shape[0]), action_batch.squeeze()), dim=0) indices = indices.tolist() reward_target[indices] = reward.squeeze() q_loss = 1e5 * qloss(F.normalize(reward_pred), F.normalize(reward_target.detach())) return forward_pred_err, inverse_pred_err, q_loss # Listing 8.13 env = gym_super_mario_bros.make('SuperMarioBros-v0') env = JoypadSpace(env, COMPLEX_MOVEMENT) # C env.reset() state1 = prepare_initial_state(env.render('rgb_array')) losses = [] episode_length = 0 # this prevents storing more action repeats than frames_per_state state_deque = deque(maxlen=params['frames_per_state']) e_reward = 0. # keep track of the last x position in order to reset it if there's no forward progress last_x_pos = env.env.env._x_position ep_lengths = [] for i in trange(params['epochs']): opt.zero_grad() episode_length += 1 # run DQN forward to get action-value predictions q_val_pred = Qmodel(state1) # after x epochs, switch to the epsilon-greedy policy if i > params['switch_to_eps_greedy']: action = int(policy(q_val_pred, params['eps'])) else: action = int(policy(q_val_pred)) # repeat action selected by policy x times to speed up learning for j in range(params['action_repeats']): state2, e_reward_, done, info = env.step(action) last_x_pos = info['x_pos'] if done: state1 = reset_env() break e_reward += e_reward_ state_deque.append(prepare_state(state2)) while len(state_deque) < params['frames_per_state']: print("adding to queue") # this should prevent the error caused when action_repeats < frames_per_state state_deque.append(prepare_state(state2)) # convert deque object into a tensor state2 = torch.stack(list(state_deque), dim=1) LOGGER.debug("state2.shape: %s", state2.shape) # add single experience to replay buffer replay.add_memory(state1, action, e_reward, state2) e_reward = 0 # if Mario is not making enough forward progress, restart game if episode_length > params['max_episode_len']: if (info['x_pos'] - last_x_pos) < params['min_progress']: done = True else: last_x_pos = info['x_pos'] if done: ep_lengths.append(info['x_pos']) state1 = reset_env() last_x_pos = env.env.env._x_position episode_length = 0 else: # this causes an error when action_repeats < frames_per_state LOGGER.debug("state2.shape: %s", state2.shape) state1 = state2 if params['experience_replay_length'] < params['batch_size']: raise ValueError( "params['experience_replay_length'] < params['batch_size'] will cause model not to train") if len(replay.memory) < params['batch_size']: continue forward_pred_err, inverse_pred_err, q_loss = minibatch_train( use_extrinsic=params['use_extrinsic']) # get errors for one mini-batch of data from replay buffer # compute overall loss loss = loss_fn(q_loss=q_loss, inverse_loss=inverse_pred_err, forward_loss=forward_pred_err) # it's important to not save the losses to a list directly because then you're saving the gradients as well loss_list = (q_loss.mean().detach().numpy(), forward_pred_err.flatten().mean().detach().numpy(), inverse_pred_err.flatten().mean().detach().numpy()) LOGGER.debug("loss list: %s", loss_list) losses.append(loss_list) loss.backward() opt.step() env.close() ###Output 2021-11-01 09:51:21,027 - __main__ - INFO - Pytorch using device: cuda ###Markdown Name model for saving files ###Code model_version_name = 'mario_curiosity_model_with_extrinsic_reward_long_frames_per_state_v5' ###Output _____no_output_____ ###Markdown Plot losses and episode length ###Code ## now handling this inside the training function # losses_ = np.array([tuple([loss.detach().numpy() # for loss in loss_tensor_tuple]) for loss_tensor_tuple in losses]) losses_ = np.array(losses) plt.figure(figsize=(8, 6)) plt.plot(np.log(losses_[:, 0]), label='Q loss') plt.plot(np.log(losses_[:, 1]), label='Forward loss') plt.plot(np.log(losses_[:, 2]), label='Inverse loss') plt.legend() plt.savefig(f'{model_version_name}_losses.png') plt.show() # https://stackoverflow.com/questions/55466298/pytorch-cant-call-numpy-on-variable-that-requires-grad-use-var-detach-num # https://stackoverflow.com/questions/16940293/why-is-there-no-tuple-comprehension-in-python plt.figure() plt.plot(np.array(ep_lengths), label='Episode length') # plt.xlabel('Training time') # plt.ylabel('Episode length') plt.legend() plt.savefig(f'{model_version_name}_episode_lengths.png') plt.show() ###Output _____no_output_____ ###Markdown save model ###Code import json torch.save(Qmodel, f'{model_version_name}.pt') with open(f'{model_version_name}_config.json', 'w') as outfile: json.dump(params, outfile) ###Output _____no_output_____ ###Markdown load model ###Code Qmodel = torch.load(f'{model_version_name}.pt') Qmodel.eval() # needed for dropout and/or batchnorm layers to function correctly for inference ###Output _____no_output_____ ###Markdown Test Trained Agent ###Code save_video = True render_also = True if save_video: from gym.wrappers.monitoring.video_recorder import VideoRecorder env = gym_super_mario_bros.make('SuperMarioBros-v0') env = JoypadSpace(env, COMPLEX_MOVEMENT) # C if save_video: video_recorder = None video_recorder = VideoRecorder( env, f'{model_version_name}_gameplay.mp4', enabled=True) done = True state_deque = deque(maxlen=params['frames_per_state']) for step in range(10000): if done: env.reset() state1 = prepare_initial_state(env.render('rgb_array')) q_val_pred = Qmodel(state1) action = int(policy(q_val_pred, params['test_eps'])) state2, reward, done, info = env.step(action) state2 = prepare_multi_state(state1, state2) state1 = state2 if save_video: video_recorder.capture_frame() if render_also: env.render() else: env.render() if save_video: video_recorder.close() video_recorder.enabled = False env.close() # # if you're done, run env.close() # env.close() ###Output _____no_output_____
Real_world_examples/Mining_rehabilitation.ipynb
###Markdown Tracking rehabilitation of mines * **Compatability:** Notebook currently compatible with both the `NCI` and `DEA Sandbox` environments* **Products used:** [ls8_fc_albers](https://explorer.sandbox.dea.ga.gov.au/ls8_fc_albers), [wofs_albers](https://explorer.sandbox.dea.ga.gov.au/wofs_albers) BackgroundLand rehabilitation is an important aspect of responsible mining.For example, The Department of Mines, Industry Regulation and Safety (DMIRS) maintain a Mining Rehabilitation Fund (MRF) that Western Australian mining operators contribute to.The fund is used to rehabilitate abandoned and legacy mines, and is underpinned by the Mining Rehabilitation Fund Act 2012.As part of the fund, tenement holders report ground disturbance, which can help DMIRS monitor how a mine's rehabilitation is going, as well as major disurbance events related to mining activity. At the moment, most mining organisations only review disturbance annually, often contracting out the service to third party surveying and ecological consulting agencies.While these providers generally provide excellent information, there are two main issues:* Annual visits give a very coarse view of how the mine is changing in time.* There is no way to validate or sanity check consultants reports without a site visit. Digital Earth Australia use caseRehabilitation and land disturbance can be monitored through satellite data by tracking the amount of vegetation and bare ground on the site compared with surrounding areas.A decrease in bare ground and increase in vegetation can be linked to positive rehabilitation.A slow increase or sharp spike in the amount of bare ground over a mining site may indicate increased disturbance, which is against the trend expected during rehabilitation efforts.This tracking can be achieved using the Fractional Cover data product from the Joint Remote Sensing Research Program, which is provided through DEA.Fractional Cover is derived from Landsat data, which has a revisit time of around two weeks for Australia, providing regular insight to a given mine's rehabilitation.This would allow companies to identify any disturbance events early in the year and take corrective action before the yearly reporting.It would also allow DMIRS to keep detailed records of how the mines they monitor are changing in time.Fractional Cover can also be used to validate the field reporting from surveying and ecological consultants before submitting reports.While reports from field surveys will provide more detail than most Earth Observation data products, such products provide the ability to provide context and validation of reports.For example, if the survey detects a disturbance, it may be hard to detect a reason.Fractional Cover can be used to identify the point in time, and possibly the cause of each disturbance event. DescriptionIn this example, the Landsat 8 Fractional Cover product is used to assess how land cover (specifically bare soil, green vegetation and non-green vegetation) is changing over time.The worked example below takes users through the code required to* Create a time series data cube over a mine site.* Create graphs to identify rehabilitation trends and disturbance events.* Interpret the results.*** Getting started**To run this analysis**, run all the cells in the notebook, starting with the "Load packages and apps" cell. Load packages and appsThis notebook works via two functions, which are referred to as apps: `load_miningrehab_data` and `run_miningrehab_app`.The apps allow the majority of the analysis code to be stored in another file, making the notebook easy to use and run.To view the code behind the apps, open the [notebookapp_miningrehab.py](../Scripts/notebookapp_miningrehab.py) file. ###Code %matplotlib inline import sys import datacube sys.path.append("../Scripts") from notebookapp_miningrehab import load_miningrehab_data from notebookapp_miningrehab import run_miningrehab_app ###Output _____no_output_____ ###Markdown Load the dataThe `load_miningrehab_data()` command performs several key steps:* Load Fractional Cover and Water Observations from Space (WOfS) data for the study area.* Match the datasets to only retain data with the same time stamps.* Mask areas that are classified as water using WOfS.* Resample the masked Fractional Cover to get monthly average values.* Return the masked data for analysis.The masked data is stored in the `dataset_fc` object.As the command runs, feedback will be provided below the cell.**Please be patient**.The load is complete when the cell status goes from `[*]` to `[number]`. ###Code dataset_fc = load_miningrehab_data() ###Output Loading Fractional Cover for Landsat 8 Loading WoFS for Landsat 8 ###Markdown Run the mining appThe `run_mining_app()` command launches an interactive map.Drawing polygons within the boundary (which represents the area covered by the loaded data) will result in plots of the average bare, green and non-green cover in that area.Draw polygons by clicking the &11039; symbol in the app.The app works by taking the loaded data `dataset_fc` as an argument. > **Note:** When drawing polygons, draw one over the mine and one over the forest nearby, then the fractional cover values can be compared on the produced plot. ###Code run_miningrehab_app(dataset_fc) ###Output _____no_output_____ ###Markdown Drawing conclusionsHere are some questions to think about:* Rehabilitation can be indicated by either a decrease in bare cover, or an increase in either green or non-green cover. Can you find any evidence that rehabilitation is occurring?* What differences are there between polygons drawn over the mine site and those drawn over the forest? What similarities are there? *** Additional information**License:** The code in this notebook is licensed under the [Apache License, Version 2.0](https://www.apache.org/licenses/LICENSE-2.0). Digital Earth Australia data is licensed under the [Creative Commons by Attribution 4.0](https://creativecommons.org/licenses/by/4.0/) license.**Contact:** If you need assistance, please post a question on the [Open Data Cube Slack channel](http://slack.opendatacube.org/) or on the [GIS Stack Exchange](https://gis.stackexchange.com/questions/ask?tags=open-data-cube) using the `open-data-cube` tag (you can view previously asked questions [here](https://gis.stackexchange.com/questions/tagged/open-data-cube)).If you would like to report an issue with this notebook, you can file one on [Github](https://github.com/GeoscienceAustralia/dea-notebooks).**Last modified:** January 2020**Compatible datacube version:** ###Code print(datacube.__version__) ###Output 1.7 ###Markdown TagsBrowse all available tags on the DEA User Guide's [Tags Index](https://docs.dea.ga.gov.au/genindex.html) ###Code **Tags**: :index:`sandbox compatible`, :index:`NCI compatible`, :index:`landsat 8`, :index:`fractional cover`, :index:`WOfS`, :index:`real world`, :index:`mining`, :index:`time series`, :index:`interactive`, :index:`widgets` ###Output _____no_output_____ ###Markdown Tracking rehabilitation of mines * [**Sign up to the DEA Sandbox**](https://docs.dea.ga.gov.au/setup/sandbox.html) to run this notebook interactively from a browser* **Compatibility:** Notebook currently compatible with both the `NCI` and `DEA Sandbox` environments* **Products used:** [ls8_fc_albers](https://explorer.sandbox.dea.ga.gov.au/ls8_fc_albers), [wofs_albers](https://explorer.sandbox.dea.ga.gov.au/wofs_albers) BackgroundLand rehabilitation is an important aspect of responsible mining.For example, The Department of Mines, Industry Regulation and Safety (DMIRS) maintain a Mining Rehabilitation Fund (MRF) that Western Australian mining operators contribute to.The fund is used to rehabilitate abandoned and legacy mines, and is underpinned by the Mining Rehabilitation Fund Act 2012.As part of the fund, tenement holders report ground disturbance, which can help DMIRS monitor how a mine's rehabilitation is going, as well as major disurbance events related to mining activity. At the moment, most mining organisations only review disturbance annually, often contracting out the service to third party surveying and ecological consulting agencies.While these providers generally provide excellent information, there are two main issues:* Annual visits give a very coarse view of how the mine is changing in time.* There is no way to validate or sanity check consultants reports without a site visit. Digital Earth Australia use caseRehabilitation and land disturbance can be monitored through satellite data by tracking the amount of vegetation and bare ground on the site compared with surrounding areas.A decrease in bare ground and increase in vegetation can be linked to positive rehabilitation.A slow increase or sharp spike in the amount of bare ground over a mining site may indicate increased disturbance, which is against the trend expected during rehabilitation efforts.This tracking can be achieved using the Fractional Cover data product from the Joint Remote Sensing Research Program, which is provided through DEA.Fractional Cover is derived from Landsat data, which has a revisit time of around two weeks for Australia, providing regular insight to a given mine's rehabilitation.This would allow companies to identify any disturbance events early in the year and take corrective action before the yearly reporting.It would also allow DMIRS to keep detailed records of how the mines they monitor are changing in time.Fractional Cover can also be used to validate the field reporting from surveying and ecological consultants before submitting reports.While reports from field surveys will provide more detail than most Earth Observation data products, such products provide the ability to provide context and validation of reports.For example, if the survey detects a disturbance, it may be hard to detect a reason.Fractional Cover can be used to identify the point in time, and possibly the cause of each disturbance event. DescriptionIn this example, the Landsat 8 Fractional Cover product is used to assess how land cover (specifically bare soil, green vegetation and non-green vegetation) is changing over time.The worked example below takes users through the code required to* Create a time series data cube over a mine site.* Create graphs to identify rehabilitation trends and disturbance events.* Interpret the results.*** Getting started**To run this analysis**, run all the cells in the notebook, starting with the "Load packages and apps" cell. Load packages and appsThis notebook works via two functions, which are referred to as apps: `load_miningrehab_data` and `run_miningrehab_app`.The apps allow the majority of the analysis code to be stored in another file, making the notebook easy to use and run.To view the code behind the apps, open the [notebookapp_miningrehab.py](../Scripts/notebookapp_miningrehab.py) file. ###Code %matplotlib inline import sys import datacube sys.path.append("../Scripts") from notebookapp_miningrehab import load_miningrehab_data from notebookapp_miningrehab import run_miningrehab_app ###Output /env/lib/python3.6/site-packages/datacube/storage/masking.py:4: DeprecationWarning: datacube.storage.masking has moved to datacube.utils.masking category=DeprecationWarning) ###Markdown Load the dataThe `load_miningrehab_data()` command performs several key steps:* Load Fractional Cover and Water Observations from Space (WOfS) data for the study area.* Match the datasets to only retain data with the same time stamps.* Mask areas that are classified as water using WOfS.* Resample the masked Fractional Cover to get monthly average values.* Return the masked data for analysis.The masked data is stored in the `dataset_fc` object.As the command runs, feedback will be provided below the cell.**Please be patient**.The load is complete when the cell status goes from `[*]` to `[number]`. ###Code dataset_fc = load_miningrehab_data() ###Output Loading Fractional Cover for Landsat 8 Loading WoFS for Landsat 8 ###Markdown Run the mining appThe `run_mining_app()` command launches an interactive map.Drawing polygons within the boundary (which represents the area covered by the loaded data) will result in plots of the average bare, green and non-green cover in that area.Draw polygons by clicking the &11039; symbol in the app.The app works by taking the loaded data `dataset_fc` as an argument. > **Note:** When drawing polygons, draw one over the mine and one over the forest nearby, then the fractional cover values can be compared on the produced plot. ###Code run_miningrehab_app(dataset_fc) ###Output _____no_output_____ ###Markdown Drawing conclusionsHere are some questions to think about:* Rehabilitation can be indicated by either a decrease in bare cover, or an increase in either green or non-green cover. Can you find any evidence that rehabilitation is occurring?* What differences are there between polygons drawn over the mine site and those drawn over the forest? What similarities are there? *** Additional information**License:** The code in this notebook is licensed under the [Apache License, Version 2.0](https://www.apache.org/licenses/LICENSE-2.0). Digital Earth Australia data is licensed under the [Creative Commons by Attribution 4.0](https://creativecommons.org/licenses/by/4.0/) license.**Contact:** If you need assistance, please post a question on the [Open Data Cube Slack channel](http://slack.opendatacube.org/) or on the [GIS Stack Exchange](https://gis.stackexchange.com/questions/ask?tags=open-data-cube) using the `open-data-cube` tag (you can view previously asked questions [here](https://gis.stackexchange.com/questions/tagged/open-data-cube)).If you would like to report an issue with this notebook, you can file one on [Github](https://github.com/GeoscienceAustralia/dea-notebooks).**Last modified:** January 2020**Compatible datacube version:** ###Code print(datacube.__version__) ###Output 1.8.0b7.dev35+g5023dada ###Markdown TagsBrowse all available tags on the DEA User Guide's [Tags Index](https://docs.dea.ga.gov.au/genindex.html) ###Code **Tags**: :index:`sandbox compatible`, :index:`NCI compatible`, :index:`landsat 8`, :index:`fractional cover`, :index:`WOfS`, :index:`real world`, :index:`mining`, :index:`time series`, :index:`interactive`, :index:`widgets`, :index:`no_testing` ###Output _____no_output_____ ###Markdown Tracking rehabilitation of mines * [**Sign up to the DEA Sandbox**](https://docs.dea.ga.gov.au/setup/sandbox.html) to run this notebook interactively from a browser* **Compatibility:** Notebook currently compatible with both the `NCI` and `DEA Sandbox` environments* **Products used:** [ga_ls_fc_3](https://explorer.sandbox.dea.ga.gov.au/ga_ls_fc_3), [ga_ls_wo_3](https://explorer.sandbox.dea.ga.gov.au/ga_ls_wo_3) BackgroundLand rehabilitation is an important aspect of responsible mining.For example, The Department of Mines, Industry Regulation and Safety (DMIRS) maintain a Mining Rehabilitation Fund (MRF) that Western Australian mining operators contribute to.The fund is used to rehabilitate abandoned and legacy mines, and is underpinned by the Mining Rehabilitation Fund Act 2012.As part of the fund, tenement holders report ground disturbance, which can help DMIRS monitor how a mine's rehabilitation is going, as well as major disurbance events related to mining activity. At the moment, most mining organisations only review disturbance annually, often contracting out the service to third party surveying and ecological consulting agencies.While these providers generally provide excellent information, there are two main issues:* Annual visits give a very coarse view of how the mine is changing in time.* There is no way to validate or sanity check consultants reports without a site visit. Digital Earth Australia use caseRehabilitation and land disturbance can be monitored through satellite data by tracking the amount of vegetation and bare ground on the site compared with surrounding areas.A decrease in bare ground and increase in vegetation can be linked to positive rehabilitation.A slow increase or sharp spike in the amount of bare ground over a mining site may indicate increased disturbance, which is against the trend expected during rehabilitation efforts.This tracking can be achieved using the Fractional Cover data product from the Joint Remote Sensing Research Program, which is provided through DEA.Fractional Cover is derived from Landsat data, which has a revisit time of around two weeks for Australia, providing regular insight to a given mine's rehabilitation.This would allow companies to identify any disturbance events early in the year and take corrective action before the yearly reporting.It would also allow DMIRS to keep detailed records of how the mines they monitor are changing in time.Fractional Cover can also be used to validate the field reporting from surveying and ecological consultants before submitting reports.While reports from field surveys will provide more detail than most Earth Observation data products, such products provide the ability to provide context and validation of reports.For example, if the survey detects a disturbance, it may be hard to detect a reason.Fractional Cover can be used to identify the point in time, and possibly the cause of each disturbance event. DescriptionIn this example, the [DEA Fractional Cover](../DEA_datasets/DEA_Fractional_Cover.ipynb) product is used to assess how land cover (specifically bare soil, green vegetation and non-green vegetation) is changing over time.The worked example below takes users through the code required to* Create a time series data cube over a mine site.* Create graphs to identify rehabilitation trends and disturbance events.* Interpret the results.*** Getting started**To run this analysis**, run all the cells in the notebook, starting with the "Load packages and apps" cell. Load packages and appsThis notebook works via two functions, which are referred to as apps: `load_miningrehab_data` and `run_miningrehab_app`.The apps allow the majority of the analysis code to be stored in another file, making the notebook easy to use and run.To view the code behind the apps, open the [notebookapp_miningrehab.py](../Scripts/notebookapp_miningrehab.py) file. ###Code %matplotlib inline import sys import datacube sys.path.append("../Scripts") from notebookapp_miningrehab import load_miningrehab_data from notebookapp_miningrehab import run_miningrehab_app ###Output /env/lib/python3.6/site-packages/geopandas/_compat.py:110: UserWarning: The Shapely GEOS version (3.7.2-CAPI-1.11.0 ) is incompatible with the GEOS version PyGEOS was compiled with (3.9.1-CAPI-1.14.2). Conversions between both will be slow. shapely_geos_version, geos_capi_version_string ###Markdown Load the dataThe `load_miningrehab_data()` command performs several key steps:* Load [DEA Fractional Cover (FC)](../DEA_datasets/DEA_Fractional_Cover.ipynb) and [DEA Water Observations (WO)](../DEA_datasets/DEA_Water_Observations.ipynb) data for the study area.* Match the datasets to only retain data with the same time stamps.* Mask areas that are classified as water using WOs.* Resample the masked FC to get monthly average values.* Return the masked data for analysis.The masked data is stored in the `dataset_fc` object.As the command runs, feedback will be provided below the cell.**Please be patient**.The load is complete when the cell status goes from `[*]` to `[number]`. ###Code dataset_fc = load_miningrehab_data() ###Output Loading DEA Fractional Cover Loading DEA Water Observations ###Markdown Run the mining appThe `run_mining_app()` command launches an interactive map.Drawing polygons within the boundary (which represents the area covered by the loaded data) will result in plots of the average bare, green and non-green cover in that area. Draw polygons by clicking the &11039; symbol in the app.The app works by taking the loaded data `dataset_fc` as an argument. > **Note:** When drawing polygons, draw one over the mine and one over the forest nearby, then the fractional cover values can be compared on the produced plot. ###Code run_miningrehab_app(dataset_fc) ###Output _____no_output_____ ###Markdown Drawing conclusionsHere are some questions to think about:* Rehabilitation can be indicated by either a decrease in bare cover, or an increase in either green or non-green cover. Can you find any evidence that rehabilitation is occurring?* What differences are there between polygons drawn over the mine site and those drawn over the forest? What similarities are there? *** Additional information**License:** The code in this notebook is licensed under the [Apache License, Version 2.0](https://www.apache.org/licenses/LICENSE-2.0). Digital Earth Australia data is licensed under the [Creative Commons by Attribution 4.0](https://creativecommons.org/licenses/by/4.0/) license.**Contact:** If you need assistance, please post a question on the [Open Data Cube Slack channel](http://slack.opendatacube.org/) or on the [GIS Stack Exchange](https://gis.stackexchange.com/questions/ask?tags=open-data-cube) using the `open-data-cube` tag (you can view previously asked questions [here](https://gis.stackexchange.com/questions/tagged/open-data-cube)).If you would like to report an issue with this notebook, you can file one on [Github](https://github.com/GeoscienceAustralia/dea-notebooks).**Last modified:** July 2021**Compatible datacube version:** ###Code print(datacube.__version__) ###Output 1.8.4.dev81+g80d466a2 ###Markdown TagsBrowse all available tags on the DEA User Guide's [Tags Index](https://docs.dea.ga.gov.au/genindex.html) ###Code **Tags**: :index:`sandbox compatible`, :index:`NCI compatible`, :index:``, :index:`fractional cover`, :index:`water observations`, :index:`real world`, :index:`mining`, :index:`time series`, :index:`interactive`, :index:`widgets`, :index:`no_testing` ###Output _____no_output_____ ###Markdown Tracking rehabilitation of mines * [**Sign up to the DEA Sandbox**](https://docs.dea.ga.gov.au/setup/sandbox.html) to run this notebook interactively from a browser* **Compatibility:** Notebook currently compatible with both the `NCI` and `DEA Sandbox` environments* **Products used:** [ga_ls_fc_3](https://explorer.sandbox.dea.ga.gov.au/ga_ls_fc_3), [ga_ls_wo_3](https://explorer.sandbox.dea.ga.gov.au/ga_ls_wo_3) BackgroundLand rehabilitation is an important aspect of responsible mining.For example, The Department of Mines, Industry Regulation and Safety (DMIRS) maintain a Mining Rehabilitation Fund (MRF) that Western Australian mining operators contribute to.The fund is used to rehabilitate abandoned and legacy mines, and is underpinned by the Mining Rehabilitation Fund Act 2012.As part of the fund, tenement holders report ground disturbance, which can help DMIRS monitor how a mine's rehabilitation is going, as well as major disurbance events related to mining activity. At the moment, most mining organisations only review disturbance annually, often contracting out the service to third party surveying and ecological consulting agencies.While these providers generally provide excellent information, there are two main issues:* Annual visits give a very coarse view of how the mine is changing in time.* There is no way to validate or sanity check consultants reports without a site visit. Digital Earth Australia use caseRehabilitation and land disturbance can be monitored through satellite data by tracking the amount of vegetation and bare ground on the site compared with surrounding areas.A decrease in bare ground and increase in vegetation can be linked to positive rehabilitation.A slow increase or sharp spike in the amount of bare ground over a mining site may indicate increased disturbance, which is against the trend expected during rehabilitation efforts.This tracking can be achieved using the Fractional Cover data product from the Joint Remote Sensing Research Program, which is provided through DEA.Fractional Cover is derived from Landsat data, which has a revisit time of around two weeks for Australia, providing regular insight to a given mine's rehabilitation.This would allow companies to identify any disturbance events early in the year and take corrective action before the yearly reporting.It would also allow DMIRS to keep detailed records of how the mines they monitor are changing in time.Fractional Cover can also be used to validate the field reporting from surveying and ecological consultants before submitting reports.While reports from field surveys will provide more detail than most Earth Observation data products, such products provide the ability to provide context and validation of reports.For example, if the survey detects a disturbance, it may be hard to detect a reason.Fractional Cover can be used to identify the point in time, and possibly the cause of each disturbance event. DescriptionIn this example, the [DEA Fractional Cover](../DEA_datasets/DEA_Fractional_Cover.ipynb) product is used to assess how land cover (specifically bare soil, green vegetation and non-green vegetation) is changing over time.The worked example below takes users through the code required to* Create a time series data cube over a mine site.* Create graphs to identify rehabilitation trends and disturbance events.* Interpret the results.*** Getting started**To run this analysis**, run all the cells in the notebook, starting with the "Load packages and apps" cell. Load packages and appsThis notebook works via two functions, which are referred to as apps: `load_miningrehab_data` and `run_miningrehab_app`.The apps allow the majority of the analysis code to be stored in another file, making the notebook easy to use and run.To view the code behind the apps, open the [notebookapp_miningrehab.py](../Scripts/notebookapp_miningrehab.py) file. ###Code %matplotlib inline import datacube import sys sys.path.insert(1, '../Tools/') from dea_tools.app import miningrehab ###Output _____no_output_____ ###Markdown Load the dataThe `load_miningrehab_data()` command performs several key steps:* Load [DEA Fractional Cover (FC)](../DEA_datasets/DEA_Fractional_Cover.ipynb) and [DEA Water Observations (WO)](../DEA_datasets/DEA_Water_Observations.ipynb) data for the study area.* Match the datasets to only retain data with the same time stamps.* Mask areas that are classified as water using WOs.* Resample the masked FC to get monthly average values.* Return the masked data for analysis.The masked data is stored in the `dataset_fc` object.As the command runs, feedback will be provided below the cell.**Please be patient**.The load is complete when the cell status goes from `[*]` to `[number]`. ###Code dataset_fc = miningrehab.load_miningrehab_data() ###Output Loading DEA Fractional Cover Loading DEA Water Observations ###Markdown Run the mining appThe `run_mining_app()` command launches an interactive map.Drawing polygons within the boundary (which represents the area covered by the loaded data) will result in plots of the average bare, green and non-green cover in that area. Draw polygons by clicking the &11039; symbol in the app.The app works by taking the loaded data `dataset_fc` as an argument. > **Note:** When drawing polygons, draw one over the mine and one over the forest nearby, then the fractional cover values can be compared on the produced plot. ###Code miningrehab.run_miningrehab_app(dataset_fc) ###Output _____no_output_____ ###Markdown Drawing conclusionsHere are some questions to think about:* Rehabilitation can be indicated by either a decrease in bare cover, or an increase in either green or non-green cover. Can you find any evidence that rehabilitation is occurring?* What differences are there between polygons drawn over the mine site and those drawn over the forest? What similarities are there? *** Additional information**License:** The code in this notebook is licensed under the [Apache License, Version 2.0](https://www.apache.org/licenses/LICENSE-2.0). Digital Earth Australia data is licensed under the [Creative Commons by Attribution 4.0](https://creativecommons.org/licenses/by/4.0/) license.**Contact:** If you need assistance, please post a question on the [Open Data Cube Slack channel](http://slack.opendatacube.org/) or on the [GIS Stack Exchange](https://gis.stackexchange.com/questions/ask?tags=open-data-cube) using the `open-data-cube` tag (you can view previously asked questions [here](https://gis.stackexchange.com/questions/tagged/open-data-cube)).If you would like to report an issue with this notebook, you can file one on [Github](https://github.com/GeoscienceAustralia/dea-notebooks).**Last modified:** January 2022**Compatible datacube version:** ###Code print(datacube.__version__) ###Output 1.8.6 ###Markdown TagsBrowse all available tags on the DEA User Guide's [Tags Index](https://docs.dea.ga.gov.au/genindex.html) ###Code **Tags**: :index:`sandbox compatible`, :index:`NCI compatible`, :index:``, :index:`fractional cover`, :index:`water observations`, :index:`real world`, :index:`mining`, :index:`time series`, :index:`interactive`, :index:`widgets`, :index:`no_testing` ###Output _____no_output_____
Paper/Mel Explain.ipynb
###Markdown Analyzing Mel ###Code for i in mel_fb: print(np.nonzero(i)) mel_frequencies(n_mels+2, fmax=fmax, htk=htk) ###Output _____no_output_____
MLP_final_test/MLP_from_data.ipynb
###Markdown This code shows an example for using the imported data from a modified .mat file into a artificial neural network and its training ###Code import numpy as np from sklearn.neural_network import MLPRegressor from sklearn import preprocessing from sklearn.cross_validation import train_test_split import matplotlib.pyplot as plt import matplotlib.patches as mpatches from sklearn.metrics import r2_score # in order to test the results from sklearn.grid_search import GridSearchCV # looking for parameters import pickle #saving to file ###Output _____no_output_____ ###Markdown Importing preprocessing data ###Code #this function reads the file def read_data(archive, rows, columns): data = open(archive, 'r') mylist = data.read().split() data.close() myarray = np.array(mylist).reshape(( rows, columns)).astype(float) return myarray data = read_data('../get_data_example/set.txt',72, 12) X = data[:, [0, 2, 4, 6, 7, 8, 9, 10, 11]] #print pre_X.shape, data.shape y = data[:,1] #print y.shape #getting the time vector for plotting purposes time_stamp = np.zeros(data.shape[0]) for i in xrange(data.shape[0]): time_stamp[i] = i*(1.0/60.0) #print X.shape, time_stamp.shape X = np.hstack((X, time_stamp.reshape((X.shape[0], 1)))) print X.shape X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=0) t_test = X_test[:,-1] t_train = X_train[:, -1] X_train_std = preprocessing.scale(X_train[:,0:-1]) X_test_std = preprocessing.scale(X_test[:, 0:-1]) ###Output (72, 10) ###Markdown Sorting out data (for plotting purposes) ###Code #Here comes the way to sort out the data according to one the elements of it test_sorted = np.hstack( (t_test.reshape(X_test_std.shape[0], 1), X_test_std, y_test.reshape(X_test_std.shape[0], 1))) test_sorted = test_sorted[np.argsort(test_sorted[:,0])] #modified train_sorted = np.hstack((t_train.reshape(t_train.shape[0], 1), y_train.reshape(y_train.shape[0], 1) )) train_sorted = train_sorted[np.argsort(train_sorted[:,0])] ###Output _____no_output_____ ###Markdown Artificial Neural Network (Gridsearch, DO NOT RUN) ###Code #Grid search, random state =0: same beginning for all alpha1 = np.linspace(0.001,0.9, 9).tolist() momentum1 = np.linspace(0.3,0.9, 9).tolist() params_dist = {"hidden_layer_sizes":[(20, 40), (15, 40), (10,15), (15, 15, 10), (15, 10), (15, 5)], "activation":['tanh','logistic'],"algorithm":['sgd', 'l-bfgs'], "alpha":alpha1, "learning_rate":['constant'],"max_iter":[500], "random_state":[0], "verbose": [False], "warm_start":[False], "momentum":momentum1} grid = GridSearchCV(MLPRegressor(), param_grid=params_dist) grid.fit(X_train_std, y_train) print "Best score:", grid.best_score_ print "Best parameter's set found:\n" print grid.best_params_ reg = MLPRegressor(warm_start = grid.best_params_['warm_start'], verbose= grid.best_params_['verbose'], algorithm= grid.best_params_['algorithm'],hidden_layer_sizes=grid.best_params_['hidden_layer_sizes'], activation= grid.best_params_['activation'], max_iter= grid.best_params_['max_iter'], random_state= None,alpha= grid.best_params_['alpha'], learning_rate= grid.best_params_['learning_rate'], momentum= grid.best_params_['momentum']) reg.fit(X_train_std, y_train) ###Output _____no_output_____ ###Markdown Plotting ###Code %matplotlib inline results = reg.predict(test_sorted[:, 1:-1]) plt.plot(test_sorted[:, 0], results, c='r') # ( sorted time, results) plt.plot(train_sorted[:, 0], train_sorted[:,1], c='b' ) #expected plt.scatter(time_stamp, y, c='k') plt.xlabel("Time(s)") plt.ylabel("Angular velocities(rad/s)") red_patch = mpatches.Patch(color='red', label='Predicted') blue_patch = mpatches.Patch(color='blue', label ='Expected') black_patch = mpatches.Patch(color='black', label ='Original') plt.legend(handles=[red_patch, blue_patch, black_patch]) plt.title("MLP results vs Expected values") plt.show() print "Accuracy:", reg.score(X_test_std, y_test) #print "Accuracy test 2", r2_score(test_sorted[:,-1], results) ###Output _____no_output_____ ###Markdown Saving ANN to file through pickle (and using it later) ###Code #This prevents the user from losing a previous important result def save_it(ans): if ans == "yes": f = open('data.ann', 'w') mem = pickle.dumps(grid) f.write(mem) f.close() else: print "Nothing to save" save_it("no") #Loading a successful ANN f = open('data.ann', 'r') nw = f.read() saved_ann = pickle.loads(nw) print "Just the accuracy:", saved_ann.score(X_test_std, y_test), "\n" print "Parameters:" print saved_ann.get_params(), "\n" print "Loss:", saved_ann.loss_ print "Total of layers:", saved_ann.n_layers_ print "Total of iterations:", saved_ann.n_iter_ #print from previously saved data %matplotlib inline results = saved_ann.predict(test_sorted[:, 1:-1]) plt.plot(test_sorted[:, 0], results, c='r') # ( sorted time, results) plt.plot(train_sorted[:, 0], train_sorted[:,1], c='b' ) #expected plt.scatter(time_stamp, y, c='k') plt.xlabel("Time(s)") plt.ylabel("Angular velocities(rad/s)") red_patch = mpatches.Patch(color='red', label='Predicted') blue_patch = mpatches.Patch(color='blue', label ='Expected') black_patch = mpatches.Patch(color='black', label ='Original') plt.legend(handles=[red_patch, blue_patch, black_patch]) plt.title("MLP results vs Expected values (Loaded from file)") plt.show() print " Accuracy:", saved_ann.score(X_test_std, y_test) plt.plot(time_stamp, y,'--.', c='r') plt.xlabel("Time(s)") plt.ylabel("Angular velocities(rad/s)") plt.title("Resuts from patient:\n" " Angular velocities for the right knee") plt.show() #print "Accuracy test 2", r2_score(test_sorted[:,-1], results) print max(y), saved_ann.predict(X_train_std[y_train.tolist().index(max(y_train)),:].reshape((1,9))) ###Output 3.67175193015 [ 3.68474801]