app.py

### ----------------------------- ###
###           libraries           ###
### ----------------------------- ###

import streamlit as st
import pickle as pkl
import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn import metrics


### ----------------------------- ###
###        interface setup        ###
### ----------------------------- ###

with open('styles.css') as f:
  st.markdown(f'<style>{f.read()}</style>', unsafe_allow_html=True)


### ------------------------------ ###
###       data transformation      ###
### ------------------------------ ###

# load dataset
uncleaned_data = pd.read_csv('data.csv')
  
# remove timestamp from dataset (always first column)
uncleaned_data = uncleaned_data.iloc[: , 1:]
data = pd.DataFrame()
  
# keep track of which columns are categorical and what 
# those columns' value mappings are
# structure: {colname1: {...}, colname2: {...} }
cat_value_dicts = {}
final_colname = uncleaned_data.columns[len(uncleaned_data.columns) - 1]
  
# for each column...
for (colname, colval) in uncleaned_data.iteritems():
  
  # check if col is already a number; if so, add col directly
  # to new dataframe and skip to next column
  if isinstance(colval.values[0], (np.integer, float)):
    data[colname] = uncleaned_data[colname].copy()
    continue
  
  # structure: {0: "lilac", 1: "blue", ...}
  new_dict = {}
  val = 0 # first index per column
  transformed_col_vals = [] # new numeric datapoints
  
  # if not, for each item in that column...
  for (row, item) in enumerate(colval.values):
      
    # if item is not in this col's dict...
    if item not in new_dict:
      new_dict[item] = val
      val += 1
      
    # then add numerical value to transformed dataframe
    transformed_col_vals.append(new_dict[item])
    
  # reverse dictionary only for final col (0, 1) => (vals)
  if colname == final_colname:
    new_dict = {value : key for (key, value) in new_dict.items()}
  
  cat_value_dicts[colname] = new_dict
  data[colname] = transformed_col_vals


### -------------------------------- ###
###           model training         ###
### -------------------------------- ###

def train_model():
  # select features and prediction; automatically selects last column as prediction
  cols = len(data.columns)
  num_features = cols - 1
  x = data.iloc[: , :num_features]
  y = data.iloc[: , num_features:]
  
  # split data into training and testing sets
  x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.25)
  
  # instantiate the model (using default parameters)
  model = LogisticRegression()
  model.fit(x_train, y_train.values.ravel())
  y_pred = model.predict(x_test)
  
  # save the model to file using the pickle package
  with open('model.pkl', 'wb') as f:
    pkl.dump(model, f)
  
  # save model accuracy to file using the pickle package
  with open('acc.txt', 'w+') as f:
    acc = metrics.accuracy_score(y_test, y_pred)
    f.write(str(round(acc * 100, 1)) + '%')
    
  return model

### -------------------------------- ###
###            rerun logic           ###
### -------------------------------- ###

# check to see if this is the first time running the script,
# if the model has already been trained and saved, load it
try:
  with open('model.pkl', 'rb') as f:
    model = pkl.load(f)
    
# if this is the first time running the script, train the model
# and save it to the file model.pkl
except FileNotFoundError as e:
  model = train_model()

# read the model accuracy from file
with open('acc.txt', 'r') as f:
  acc = f.read()


### ------------------------------- ###
###        interface creation       ###
### ------------------------------- ###

# uses the logistic regression to predict for a generic number
# of features
def general_predictor(input_list):
  features = []

  # transform categorical input
  for colname, input in zip(data.columns, input_list):
    if (colname in cat_value_dicts):
      features.append(cat_value_dicts[colname][input])
    else:
      features.append(input)

  # predict single datapoint
  new_input = [features]
  result = model.predict(new_input)
  return cat_value_dicts[final_colname][result[0]]

def get_feat():
  feats = [abs(x) for x in model.coef_[0]]
  max_val = max(feats)
  idx = feats.index(max_val)
  return data.columns[idx]

with open('info.md') as f:
  st.title(f.readline())
  st.subheader('Take the quiz to get a personalized recommendation using AI.')

form = st.form('ml-inputs')

# add data labels to replace those lost via star-args
inputls = []
for colname in data.columns:
  # skip last column
  if colname == final_colname:
    continue

  # access categories dict if data is categorical
  # otherwise, just use a number input
  if colname in cat_value_dicts:
    radio_options = list(cat_value_dicts[colname].keys())
    inputls.append(form.selectbox(colname, radio_options))
  else:
    # add numerical input
    inputls.append(form.number_imput(colname))

# generate gradio interface
if form.form_submit_button("Submit to get your recommendation!"):
  prediction = general_predictor(inputls)
  
  form.subheader(prediction)

col1, col2 = st.columns(2)
col1.metric("Number of Different Possible Results", len(cat_value_dicts[final_colname]))
col2.metric("Model Accuracy",  acc)
st.metric("Most Important Question", "")
st.subheader(get_feat())
st.text("")
st.markdown("***")

with open('info.md') as f:
  f.readline()
  st.markdown(f.read())