Datasets:

Abirate
/

all_write_ups

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 3.85k rows

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.

Title of Competition stringclasses 343 values	Title of Writeup stringlengths 6 139	User stringclasses 1 value	Writeup stringlengths 419 80.4k
PetFinder.my - Pawpularity Contest	54th place – single model solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my - Pawpularity Contest <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thanks, yes with more time maybe I had done a multi model ensemble, but used the 15 TTA single model instead :)</p> <p>Congrats to the top solutions and a thanks for an important competition.<br> Also, a thanks to fellow Kagglers contributing with valuable insights, ideas, and notebooks.<br> I found inspiration and ideas from many notebooks, very useful as I entered the competition just 2 weeks ago.<br> Had this public notebook as the starting point <a href="https://www.kaggle.com/calvchen/swin-transformer-224-meta" target="_blank">https://www.kaggle.com/calvchen/swin-transformer-224-meta</a><br> Many public notebooks also had inspiration from this notebook <a href="https://www.kaggle.com/cdeotte/rapids-svr-boost-17-8" target="_blank">https://www.kaggle.com/cdeotte/rapids-svr-boost-17-8</a></p> <p>The solution summary:</p> <p>Data selection</p> <p>Used the clean dataset for training and all the metadata+dim.information and for inference I also used Yolov5+TTA to add the dog/cat features.</p> <p>Model selection</p> <p>I used FastAI with swin_large_patch4_window7_224 from timm library, tried many other both higher and lower in the model top lists, even tried the largest ones with stronger GPUs but the SwinL224 worked best.</p> <p>Training</p> <p>Used augmentation techniques Flip, Rotate, Zoom, Warp all with Reflection pad mode, also added Brightness and Contrast.<br> Started the 10 folds training with 1 epoch with freezing layers following with 5 epochs with unfreeze the layers, saved the best models using competition metric.<br> After finished the training of the main model I extracted the embedding and trained the extra regression head, for this I used a stacking approach – trained a stacked model with input from SVR, Ridge and LR.</p> <p>Inference</p> <p>15 TTA with some of the training augmentation techniques. For the final ensemble more weight was added to the stacked regression.<br> I also dropped the worst OOF folds, with the idea that they where badly trained instead of that the validation set were noisy.</p> <p>That’s it! 😊</p>
PetFinder.my - Pawpularity Contest	68th place solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my - Pawpularity Contest <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thank you<br> It's depending on model size in swin 224 i used 32 while in swin 384 i used 8</p> <p>thanks this makes me happy 😄</p> <p>you are welcome😊</p> <p>Thanks to organizers for this interesting challenge and congrats everyone who enjoyed it! </p> <p><strong>My approach</strong><br> <strong>1-</strong> I split the data into 5 folds using stratified kfold.<br> <strong>2-</strong> I used Brightness(), Contrast(), Hue() and Saturation() for augmentation.<br> <strong>3-</strong> I trained Swin-L 224x224 with 2e-5 learning rate for 10 epoch for each fold.<br> <strong>4-</strong> I applied tta on Swin-L 224x224 with n = 4.<br> <strong>5-</strong> I trained svr model on the activations of Swin-L 224x224 and metadata.<br> <strong>6-</strong> I trained Swin-L 384x384 with 2e-5 learning rate for 10 epoch for each fold.<br> <strong>7-</strong> I applied tta on Swin-L 384x384 with n = 4.<br> <strong>8-</strong> I trained aanother svr model on the activations of Swin-L 384x384 and metadata.<br> <strong>9-</strong> I combined the predictions of these four models with metadata using ridge regression model as blender.<br> <strong>10-</strong> I trained another Swin-L 224x224 model on the complete training data with testing data whose <br> predictions taken from the blender model for 17 epoch with 2e-5 learning rate and I've added <br> noise to the test predictions from ridge model during training within the range (-17.8, +17.8).<br> <strong>11-</strong> I did a weighted sum of blender preds and the last model preds on test data according to the LB <br> like this:-<br> final_preds = first_model_preds * 0.9 + last_model_preds * 0.1</p> <p>my kernel link:-<br> <a href="https://www.kaggle.com/riadalmadani/inf-final-blender-with-psudolabe-17-epoch/notebook" target="_blank">https://www.kaggle.com/riadalmadani/inf-final-blender-with-psudolabe-17-epoch/notebook</a></p>
PetFinder.my - Pawpularity Contest	6th Place - Multitask Learning	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my - Pawpularity Contest <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>The best answer is to try both (freeze and unfreeze) and see what achieves the best CV score. In general i will freeze less for classification tasks and more for regression.</p> <p>There are other factors too like how similar the competition images are to the pretraining images (which are imagenet collection). When they are similar we can freeze more. And when they are different we need to unfreeze more.</p> <p>After downloading a pretrained CNN from the internet, you can display the number of layers with <code>len( base_model.layers )</code> in TensorFlow. For example, InceptionV3 has 311 layers. Then you can freeze the first 90% of layers with </p> <pre><code>for layer in base_model.layers[:-28]: layer.trainable = False </code></pre> <p>There is a great public notebook which demonstrates this <a href="https://www.kaggle.com/vishalkarangale/inceptionv3-10folds-tensorflow" target="_blank">here</a>. (Note if you fork that notebook, FYI there is CV leak. You need to create a new <code>base_model</code> for each fold. That notebook uses the same <code>base_model</code> for each fold).</p> <p>Regarding the reason, this is frequently done when fine tuning models. From their pretraining, the early layers have already learned to extract features from images (or in NLP the early layers already understand words). When we fine tune, we can use a lower learning rate for early layers and/or freeze these layers to avoid losing the pretraining.</p> <p>Additionally when doing regression (which is different than classification), I have found that freezing is often required. I'm not sure why. I think it may be because of batch normalization layers and drop out layers. There is a discussion about this <a href="https://www.kaggle.com/c/commonlitreadabilityprize/discussion/260729#1442448" target="_blank">here</a> and blog <a href="https://towardsdatascience.com/pitfalls-with-dropout-and-batchnorm-in-regression-problems-39e02ce08e4d" target="_blank">here</a></p> <p>If image is <code>192x224</code>, then the largest square is <code>192x192</code>. First i randomly select a largest square from <code>192x224</code>, next I resize it to <code>224x224</code></p> <p>In general if dim is <code>A x B</code>, then largest square is <code>min(A,B) x min(A,B)</code>. Then resize to <code>224x224</code>. (Sometimes largest square is smaller than 224, sometimes larger).</p> <p>Thanks <a href="https://www.kaggle.com/mobassir" target="_blank">@mobassir</a>!</p> <p>Your second equation. But we only have 5 models if we kept every model from day 1 to 4. For example we may discard Day 3 like this:</p> <p>Day 1, CV = Day1. Keep Day 1<br> Day 2, CV = (Day1 + Day2)/2, CV increase. Keep Day 2<br> Day 3, CV = (Day1 + Day2 + Day3)/3. <strong>CV decrease. Discard Day 3</strong><br> Day 4, CV = (Day1 + Day2 + Day4)/3. CV increase. Keep Day 4<br> Day 5, CV = (Day1 + Day2 + Day4 + Day5)/4. etc etc</p> <p>And then every few days, I may remove each previous model one by one to make sure removing a previous day doesn't increase CV. Because sometimes an old model is no longer helpful.</p> <p>The reason I use equal weights is to guarantee that the newest model really helps. If we only add 1% like <code>CV = 0.33Day1 + 0.33Day2 + 0.33Day3 + 0.01Day4</code> then by pure luck adding Day4 may increase CV by a little bit. But if a model increase CV with equal weights that it really is helpful without luck.</p> <p>I don't think adding these features <strong>as input</strong> help the model because the model can already see this in the image. For example, an image model doesn't need to be told whether animal is dog or cat. It can learn this to nearly 99.9% accuracy on its own.</p> <p>What i think is helpful is adding the feature you suggest and others as multi task learning (i.e. <strong>as output</strong>). That encourages the model to work harder while it is learning Pawpularity and usually results in a smarter model.</p> <p>(This is what i think but i could be wrong)</p> <p>Here are some ideas that come to mind</p> <ul> <li>change batch size</li> <li>change input image size</li> <li>change backbone</li> <li>use CNN vs. Transformer</li> <li>use model pretrained on different original image datasets (1k image vs. 22k imagenet vs. etc)</li> <li>use imagenet pretrain vs. noisy-student vs. AP etc etc</li> <li>preprocess with square crops to keep aspect ratio vs. squish image into square</li> <li>data augmentation</li> <li>add meta as input</li> <li>add meta as auxiliary target</li> <li>change loss from regression, to BCE, to multiclass (put Pawpularity into buckets)</li> <li>initialize layers in different ways</li> <li>use different activation functions</li> <li>do regression with MAE instead of MSE</li> <li>have model predict breed as auxiliary</li> <li>have model predict dog cat as auxiliary</li> <li>use imagenet pretrain model <strong>without</strong> finetuning it. (Just train a head).</li> <li>modify architecture like removing downsampling from first EffNet layer</li> <li>make custom head like add spatial attention to CNN etc</li> <li>use different pooling, like attention pooling, max pooling, avg pooling, etc</li> <li>freeze model batch norms</li> <li>freeze different parts of backbone during train<br> etc etc</li> </ul> <p>Making models with the squish preprocess only improved my <code>best LB ensemble</code> (strangely). All models in my <code>best CV ensemble</code> use random square cropping (and no squish preprocess).</p> <p>This is why teaming up makes strong ensembles. Because although everyone is doing basically the same stuff, one teammate will do just one thing different. Like perhaps they will initialize the layers of their NN differently. Then when you ensemble their model with your team ensemble, it gets a surprising big boost.</p> <p>Thanks Mutopia</p> <p>Thanks <a href="https://www.kaggle.com/awsaf49" target="_blank">@awsaf49</a> . Your discussion and notebooks were very helpful. My earlier ensembles included your ViT model but as I continued to add more recent SOTA image transformers, using ViT no longer helped my CV LB. None-the-less, i learned from your code, discussion, and published experiments.</p> <p>UPDATE: I added emphasis that freezing batch norm layers was important when training EfficientNet to achieve good CV LB RSME. Team 42nd achieved CV 17.30 LB 17.92 with an EfficientNet explained <a href="https://www.kaggle.com/c/petfinder-pawpularity-score/discussion/301259" target="_blank">here</a>. Also Robin Smit's great notebook demonstrates it <a href="https://www.kaggle.com/rsmits/effnet-b2-feature-models-catboost" target="_blank">here</a>. These are awesome results!</p> <p>Yes. When you do KFolds CV, you get an OOF dataframe that is the same length as the length of train dataset. Your OOF has 1 prediction for every row of train data.</p> <pre><code>oof = np.array(len(train)) for fold in range(K): train_idx = train.loc[train.fold!=fold].index.values valid_idx = train.loc[train.fold==fold].index.values model.fit(train.loc[train_idx]) oof[valid_idx] = model.predict(train.loc[valid_idx]) </code></pre> <p>After Day 2, you have <code>oof_day1</code> Numpy array from Day 1 and <code>oof_day2</code> Numpy array from Day 2. The ensemble is </p> <pre><code>oof = (oof_day1 + oof_day2)/2 CV_RSME = np.sqrt(np.mean( (oof - train.y.values)**2.0 )) </code></pre> <p>So <strong>every time</strong>, you train a new model save the <code>oof</code> to disk as <code>numpy.save('oof_day1', oof_day1)</code>. Instead of doing a once per day, you can collect many <code>OOF</code> and then load all your Numpy OOF into my notebook <a href="https://www.kaggle.com/cdeotte/forward-selection-oof-ensemble-0-942-private" target="_blank">here</a>. (In code cell 5, change <code>TOL = 0.0003</code> to a meaningful value. In Pet Comp, i would use <code>TOL = 0.01</code> and only accept a new model if it increase RSME by <code>0.01</code>. To be more careful, you can set this higher like 0.015 or 0.02)</p> <p>Early in every competition, i decide whether I will pursue a single model or ensemble. If I believe single model is best then I spend most of my time improving the CV LB of one model (after exploring a bunch of different models). If I believe ensemble is best then I begin searching for diversity from the beginning. (And then yes, hill climbing is a great approach that i have used before. And when using ensemble, its important to read every pubic notebook to find diversity).</p> <p>The decision to chose single model or ensemble is as follows. If the data is very noisy and the CV changes when you train the same model with different seed. And the LB score is much different than CV score. And it feels difficult to improve a single model. And you're basically confused what to trust, then i go with ensemble.</p> <p>If CV score feels robust and behaves in consistent manner. If the LB score follows the CV score very closely. If I notice that changes to model improve CV and there is potential for more improvement. Then I will develop a single model. Of course in both cases (single model vs. ensemble), i will still spend time on each, but for each competition I will focus on one more than the other.</p> <p>If I have time, perhaps I will post. The idea is this. Choose a <strong>backbone</strong> and build a <strong>head</strong> with 12 outputs. Then train with previous images and predict the 12 meta targets from previous comp. This has the benefit of improving the backbone. Next remove the head, and keep the pretrained backbone. Next build a new head with 1 output for this comps Pawpularity. Next continue training with this comps images and this comp's target.</p> <p>Note that this is exactly how EffNet, Swin etc are pretrained. Someone else trained these backbones with a head with 1000 outputs. Then then trained with the images from imagenet collection which has 1000 targets. Then we download the backbone and remove head and add our new head.</p> <p>Congrats KhanhVD and team. 53th out of 3500 teams is excellent! Sorry to hear, you missed Gold. You'll get Gold next time! Good work.</p> <h1>Kaggle PetFinder Comp</h1> <p>Thanks Kaggle and Petfinder for hosting this wonderful competition. Building models to help pets was fun and rewarding. My solution is the combination of my shared discussion posts and notebooks.</p> <h1>Trust Your CV</h1> <p>For my final two submissions, I chose my best LB score (<strong>CV 17.15, LB 17.64</strong>) and my best CV score (<strong>CV 16.98, LB 17.78</strong>). My best CV score only had public leaderboard rank 120. It was <strong>scary</strong> to select it, but my final private LB 6th rank was this best CV submission!<br> <img src="https://raw.githubusercontent.com/cdeotte/Kaggle_Images/main/Jan-2022/catlb2.png" alt=""></p> <h1>Preprocess</h1> <p>There are at least two ways to resize images for our models. We can squish the original images into a square (distorting aspect ratio), or we can square crop (and preserve aspect ratio). I trained my models with random square crops. Not only does this create natural looking images that match how they are displayed at Petfinder website but it also provides data augmentation by using random crops during training (which helps our models generalize). <strong>This boosted CV LB</strong>.</p> <p><img src="https://raw.githubusercontent.com/cdeotte/Kaggle_Images/main/Jan-2022/preprocess.png" alt=""></p> <h1>Image CNN Training</h1> <p>Training CNNs to achieve a good CV LB RSME score was more difficult than training Transformers and it required freezing backbones to train the head (and/or freezing first 66% backbone to train last 33% backbone plus head). It's also <strong>important</strong> to freeze all batchnorm layers. (This was demonstrated <a href="https://www.kaggle.com/rsmits/effnet-b2-feature-models-catboost" target="_blank">here</a> and explained <a href="https://www.kaggle.com/c/petfinder-pawpularity-score/discussion/301259" target="_blank">here</a>). There are many great examples showing CNN training in public notebooks <a href="https://www.kaggle.com/rsmits/effnet-b2-feature-models-catboost" target="_blank">here</a>, <a href="https://www.kaggle.com/jackstapleton/d169-baseline-nc-inference" target="_blank">here</a>, and <a href="https://www.kaggle.com/vishalkarangale/inceptionv3-10folds-tensorflow" target="_blank">here</a>. My best single model CNN was <strong>EfficientNetB2 with RAPIDS SVR</strong> head. It achieved <strong>CV 17.55</strong></p> <p><img src="https://raw.githubusercontent.com/cdeotte/Kaggle_Images/main/Jan-2022/freeze2.png" alt=""></p> <h1>Image Transformer Training</h1> <p>There are dozens of image transformers. To view all available pretrained image transformers, just type <code>import timm; timm.list_models(pretrained=True)</code>. You will see <code>beit, cait, convit, crossvit, deit, levit, pit, swin, vit, xcit</code>! Next you can fork the public <code>fastai</code> notebook <a href="https://www.kaggle.com/tanlikesmath/petfinder-pawpularity-eda-fastai-starter" target="_blank">here</a> and run with each of these models! To evaluate which is the best, you must correct the RSME metric with <code>metrics=AccumMetric(func=petfinder_rmse)</code> as explained <a href="https://www.kaggle.com/c/petfinder-pawpularity-score/discussion/294054" target="_blank">here</a> (otherwise CV score is meaningless and wrong).</p> <p>My best single model was <code>BeIT</code> (not <code>Swin</code>). It's a transformer with BERT pretraining. Research paper <a href="https://arxiv.org/abs/2106.08254" target="_blank">here</a>. A <strong>single BeIT transformer</strong> achieved my best <strong>CV 17.42</strong> and best <strong>Private LB 17.00</strong> which is top 50 private LB!</p> <h1>Image Size and Data Augmentation</h1> <p>I trained models with image sizes 224, 384, 512. The advantage of a variety of image sizes is explained <a href="https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/160147" target="_blank">here</a>. For large image sizes more augmentation was required. When using image size 512, the following augmentations were applied; Random square crop, Horizontal Flip, Vertical Flip, Transpose, Brightness, Contrast, Hue, Saturation, Rotation, Random Erase. Some of the small 224 models, only used Random square crop and Flips.</p> <h1>Multi Task Learning</h1> <p>There are two ways to use meta data. We can input it into our model or use it as an additional output (i.e. auxiliary target) during training. The easiest way to train an image transformer with additional output is by modifying the PyTorch notebook <a href="https://www.kaggle.com/ytakayama/train-pytorch-swin-5fold-some-tips" target="_blank">here</a>. The meta data provided in this competition didn't help, but using Cat Dog labels as an additionally target from <a href="https://www.kaggle.com/chrisbradley/pawpularity-cat-or-dog" target="_blank">here</a> helped boost CV LB and add model diversity.</p> <p><img src="https://raw.githubusercontent.com/cdeotte/Kaggle_Images/main/Jan-2022/multitask.png" alt=""></p> <h1>Hill Climbing</h1> <p>My approach in this competition was to build an ensemble slow and steady. First I corrected my local RSME computation so that CV score was correct. Then each day during the competition, I would train a new model, save the OOF predictions, and <strong>ensemble it will equal weights</strong> to my previous best CV ensemble and previous best LB ensemble. If it increased CV then i added it to my best CV ensemble. If it increased LB then I added it to my best LB ensemble. </p> <p>I describe hill climbing more <a href="https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/175614" target="_blank">here</a>. The secret to hill climbing is to make new <strong>diverse</strong> models. (i.e. Try different image sizes, backbones, augmentations etc). Each new model doesn't necessarily need to have a great CV by itself. They <strong>need to be different</strong> than your previous models so that when you ensemble them to your previous models, the result is an improved CV LB. My two (hill climbing generated) ensembles contained different models. Here are the lists</p> <h3>Best CV Ensemble</h3> <p><strong>CV = 16.98, Public LB = 17.78, Private LB = 16.90</strong></p> <ul> <li>Swin 224 with Dog Cat auxiliary output</li> <li>Swin 384</li> <li>BeIT 224</li> <li>BeIT 384</li> <li>BeIT 512</li> <li>ConvNeXt 224</li> <li>ConvNeXt 384 with Dog Cat auxiliary output</li> <li>EfficientNetB2 with RAPIDS SVR head</li> </ul> <h3>Best LB Ensemble</h3> <p><strong>CV = 17.15, Public LB = 17.64, Private LB = 17.01</strong></p> <ul> <li>Swin 224 with Dog Cat auxiliary output</li> <li>Swin 384</li> <li>BeIT 224</li> <li>BeIT 384</li> <li>EfficientNetB2 with RAPIDS SVR head</li> <li>DenseNet 224 with MLP regression head</li> <li>Inception 229 with 66% frozen backbone</li> </ul> <h1>What Didn't Work</h1> <p>I tried to pseudo label external images but it didn't improve my CV nor LB. I also tried to use last year competition data by pretraining models with last year's images and last year's meta data. This did not help CV nor LB. I also tried to train models on last year data and infer new features on this year's data. But that didn't help either. <strong>In the end, I didn't use last year's comp data nor external data</strong>.</p> <p>After reading other winners' solutions; <a href="https://www.kaggle.com/c/petfinder-pawpularity-score/discussion/300929" target="_blank">2nd place here</a>, <a href="https://www.kaggle.com/c/petfinder-pawpularity-score/discussion/301072" target="_blank">4th place here</a>, <a href="https://www.kaggle.com/c/petfinder-pawpularity-score/discussion/300928" target="_blank">5th place here</a>, it appears that last year's data was actually very powerful. The way to use it was to find the duplicate images from last years data that were also used this year. Then for those images (which were about 33% of all images) concatenate last years meta features onto this year's data. Then train a model to use this year's image with the concatenated last years meta data. Apparently the meta data from the previous PetFinder competition was more powerful than the meta data provided this year.</p> <h1>The Power of SOTA Image Models!</h1> <p>It was shown from <a href="https://www.kaggle.com/c/petfinder-pawpularity-score/discussion/300938" target="_blank">1st place here</a>, <a href="https://www.kaggle.com/c/petfinder-pawpularity-score/discussion/301044" target="_blank">3rd place here</a> (and my 6th place solution) that using meta features from last year were not needed to achieve a powerful model. Today's SOTA image models are amazingly good at looking at images and extracting information!</p> <p>Thanks <a href="https://www.kaggle.com/crained" target="_blank">@crained</a> . Congratulations to you and your team for shaking upward on private LB and finishing in top 100. You must have had a good reliable local CV.</p> <p>Thanks <a href="https://www.kaggle.com/blankaf" target="_blank">@blankaf</a> . Congrats on finishing solo Silver 52nd out of 3550 teams. That's fantastic!</p>
PetFinder.my - Pawpularity Contest	9th place solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my - Pawpularity Contest <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>It may be too late, but I have shared the code. Due to the environmental constraints of Kaggle Code, the model size, augmentation and number of epochs are limited, and the score is poor.</p> <p>I think mere classification of 100-class pawpularity without label distribution learning would not be able to learn well because the ordinality of the labels is not taken into account, so I didn't tried.<br> I thought this competition was similar to the age estimation task. Age is difficult to estimate in one-year increments, and it has the characteristic of being ordinal. The same is true for pet pawpularity. And I adopted DLDL-v2, one of the SOTAs from Papers with Code's <a href="https://paperswithcode.com/task/age-estimation" target="_blank">Age Estimation</a>.</p> <p>The <a href="https://github.com/PuNeal/DLDL-v2-PyTorch/blob/master/data.py" target="_blank">code</a> I referred to used std=2. I tried around 2 and 2 was the best.<br> (Fixed the reference DLDL-v2 link to the PyTorch version.)</p> <p>Thanks. ConvNeXt looks promising for future competitions too.</p> <p>I created normal distributions with the expected value of targets and the standard deviation of 2.0.<br> <img src="https://user-images.githubusercontent.com/75870915/149624023-4e6a27d3-9f76-4845-92d6-5f5a362fc4c7.jpg" alt=""></p> <p>My solution is a weighted average ensemble of the following three type of models. NN models are implemented by original PyTorch[1, 2] and fastai[3]. I did not use external data.</p> <ol> <li>Deep Label Distribution Learning (DLDL) models based on <a href="https://github.com/PuNeal/DLDL-v2-PyTorch" target="_blank">DLDL-v2-PyTorch</a> (<a href="https://www.ijcai.org/proceedings/2018/0099.pdf" target="_blank">Paper</a>) (3 models)</li> <li>Regression models trained as binary classification tasks (2 models)</li> <li>fastai models based on <a href="https://www.kaggle.com/ilovescience" target="_blank">@ilovescience</a>'s <a href="https://www.kaggle.com/tanlikesmath/petfinder-pawpularity-eda-fastai-starter" target="_blank">highly influential notebook</a> (2 models)</li> </ol> <p>The brief overview of each model is as follows.</p> <h3>[1] DLDL models</h3> <ul> <li>NN backbone : Swin Transformer L 384</li> <li>Augmentation : torchvision RandAugment(2, 10), mixup(alpha=0.2, p=0.5)</li> <li>Metadata : used tabular data (I think tabular metadata is useless, but I did not omit it.)</li> <li>Loss : (KLDivLoss and HuberLoss) or (FocalLoss and HuberLoss)</li> <li>Optimizer : <a href="https://github.com/lessw2020/Ranger21" target="_blank">Ranger21</a></li> <li>Scheduler : OneCycleLR</li> <li>Regularization : <a href="https://arxiv.org/abs/1909.11299" target="_blank">MixOut</a></li> <li>Data split : StratifiedKFold(n_splits=5)</li> <li>Additional head : SVR and Catboost</li> <li>TTA : 2 times [1: Resize(size)->CenterCrop(size) / 2: Resize(size, size)]<br> <img src="https://user-images.githubusercontent.com/75870915/149617749-fc23fab2-a420-4ae4-88f1-ee37a3addfcd.jpg" alt="DLDL model"></li> </ul> <h3>[2] Regression models</h3> <ul> <li>NN backbone : Swin Transformer L 384</li> <li>Augmentation : torchvision RandAugment(2, 10), mixup(alpha=0.2, p=0.5)</li> <li>Metadata : used tabular data</li> <li>Loss : BCEWithLogitsLoss()</li> <li>Optimizer : AdamW</li> <li>Scheduler : OneCycleLR</li> <li>Data split : StratifiedKFold(n_splits=5)</li> <li>Additional head : SVR and Catboost</li> <li>TTA : None [Resize(size)->CenterCrop(size)]</li> </ul> <h3>[3] fastai models</h3> <ul> <li>NN backbone : Swin Transformer L 224 / ConvNeXt XL 224</li> <li>Augmentation : Brightness(), Contrast(), Hue(), Saturation()</li> <li>Metadata : None</li> <li>Loss : BCEWithLogitsLossFlat()</li> <li>Additional head : None</li> <li>Data split : StratifiedKFold(n_splits=10) / StratifiedKFold(n_splits=5)</li> <li>TTA : 5 times / 1 time</li> </ul> <h3>Selected submissions for final score</h3> <ul> <li>Best public LB submission<ul> <li>Public LB : 17.74385 (33th)</li> <li>Private LB : 16.95033 (17th))</li></ul></li> <li>The only successful submission containing ConvNeXt (fastai model)<ul> <li>Public LB : 17.77487 (94th)</li> <li>Private LB : <strong>16.91650</strong> (9th)</li></ul></li> </ul> <p>As for the single ConvNeXt model , the CV was good, but the LB was not so good.<br> My best single NN model in private LB during competition was a DLDL model, its score is:</p> <ul> <li>CV : 17.36578</li> <li>Private LB : 17.01783 (2 TTA)</li> </ul> <p><strong>Update</strong> : I had completed training of a DLDL model of ConvNeXt XL about 9 hours before the competition end, but when I added it to the ensemble in a hurry and submitted it, I made a mistake in the weight calculation. I made a quick fix to the code and late submitted again, it scored <strong>16.89143</strong> in private. I just set the weights of weighted average by intuition, I think higher score will be obtained by hill climbing or stacking. I think the model to add into the ensemble need not be ConvNeXt. I should have tried wider variety of models.</p> <p><strong>Update2b</strong> : After competition end, using several single DLDL models, I tested the effectiveness of TTA and found that following 3 TTA using torchvision.transforms looks consistently work well in both private LB and public LB.</p> <ul> <li>[TTA1] <code>Resize(image_size), CenterCrop(image_size)</code> # preserve aspect ratio</li> <li>[TTA2] <code>Resize((image_size, image_size))</code> # squish the original images into a square</li> <li>[TTA3] <code>Resize(image_size), CenterCrop(image_size), RandomHorizontalFlip(p=1)</code></li> </ul> <p>All models are trained using random crop choice augmentation of TTA1 and TTA2, different seeds, and a little different hyper parameters.</p> <table> <thead> <tr> <th>model</th> <th>head</th> <th>TTA times</th> <th>Private LB</th> <th>Public LB</th> <th>CV</th> </tr> </thead> <tbody> <tr> <td>Swin L 224 DLDL #1</td> <td>NN only</td> <td>None (1)</td> <td>17.02085</td> <td>17.92344</td> <td>17.37</td> </tr> <tr> <td>^</td> <td>^</td> <td>3</td> <td><strong>17.00839</strong></td> <td>17.91271</td> <td></td> </tr> <tr> <td>^</td> <td>NN + SVR + CAT</td> <td>None (1)</td> <td>16.99053</td> <td>17.91827</td> <td></td> </tr> <tr> <td>^</td> <td>^</td> <td>3</td> <td><strong>16.97933</strong></td> <td>17.90832</td> <td></td> </tr> <tr> <td>Swin L 224 DLDL #2</td> <td>NN only</td> <td>None (1)</td> <td>17.13297</td> <td>17.92975</td> <td>17.37</td> </tr> <tr> <td>^</td> <td>^</td> <td>3</td> <td>17.09943</td> <td>17.90875</td> <td></td> </tr> <tr> <td>^</td> <td>NN + SVR + CAT</td> <td>None (1)</td> <td>17.04425</td> <td>17.82960</td> <td></td> </tr> <tr> <td>^</td> <td>^</td> <td>3</td> <td>17.00938</td> <td>17.81680</td> <td></td> </tr> <tr> <td>Swin L 224 DLDL #3</td> <td>NN only</td> <td>None (1)</td> <td>17.05831</td> <td>17.96708</td> <td>17.32</td> </tr> <tr> <td>^</td> <td>^</td> <td>3</td> <td>17.04335</td> <td>17.94935</td> <td></td> </tr> <tr> <td>^</td> <td>NN + SVR + CAT</td> <td>None (1)</td> <td>17.02163</td> <td>17.90633</td> <td></td> </tr> <tr> <td>^</td> <td>^</td> <td>3</td> <td>17.01208</td> <td>17.89366</td> <td></td> </tr> <tr> <td>ConvNeXt XL 224 DLDL #1</td> <td>NN only</td> <td>None (1)</td> <td>17.10452</td> <td>17.98510</td> <td>17.37</td> </tr> <tr> <td>^</td> <td>^</td> <td>3</td> <td>17.08961</td> <td>17.96937</td> <td></td> </tr> <tr> <td>^</td> <td>NN + SVR + CAT</td> <td>None (1)</td> <td>17.05040</td> <td>17.98521</td> <td></td> </tr> <tr> <td>^</td> <td>^</td> <td>3</td> <td>17.03909</td> <td>17.97758</td> <td></td> </tr> </tbody> </table> <p>The correlation between CV, Public LB, and Private LB is not high between other models, but within the same model, the improvement effect of TTA and additional head is consistently observed.<br> The idea and the hyper parameters of additional heads are based on these notebooks : <a href="https://www.kaggle.com/cdeotte/rapids-svr-boost-17-8" target="_blank">SVR</a>, <a href="https://www.kaggle.com/rsmits/effnet-b2-feature-models-catboost" target="_blank">Catboost</a>. Thanks to <a href="https://www.kaggle.com/cdeotte" target="_blank">@cdeotte</a> and <a href="https://www.kaggle.com/rsmits" target="_blank">@rsmits</a>.</p> <p><strong>Update3</strong> : I share the code example of DLDL model. <a href="https://www.kaggle.com/shigemitsutomizawa/9th-place-swinb224-dldl-nn-svr-cat?scriptVersionId=86009375" target="_blank">[9th place] SwinB224 DLDL NN+SVR+CAT</a></p>
PetFinder.my - Pawpularity Contest	Jumped from 45th to 42nd place? What happened?	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my - Pawpularity Contest <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>🤦‍♂️🤦‍♂️🤦‍♂️</p> <p>Why don't them remove them during the competition? This happens often?</p> <p>After the competition ended we were 45th on the private LB. Today I see we are 42nd.</p> <p>I see <a href="https://www.kaggle.com/cpmpml" target="_blank">@cpmpml</a> is still in his unlucky position which means none of the gold medal positions changed.</p> <p>Any idea what happened?</p> <p>Would that be your favorite gold then? 🤣</p>
PetFinder.my - Pawpularity Contest	Oopsie Doopsie, Apparently my public notebook got a silver medal (#145) 😆	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my - Pawpularity Contest <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Guys, </p> <p>I've made a bit of an oopsie 😆<br> Apparently, I accidentally got a silver medal with my "Nuclear Option" public notebook.</p> <blockquote> <p><a href="https://www.kaggle.com/yamqwe/the-nuclear-option-train" target="_blank">Here</a> If you are interested</p> </blockquote> <p><img src="https://i.ibb.co/gwpJXBm/lb.png" alt=""></p> <hr> <h4>The outline of the solution is simple:</h4> <p>I trained every single SOTA model from 2021 and then <strong>Nuked the leaderboard with 10 Billion parameters</strong> ensemble of ensembles. <br> The ensembles were also supplemented a bit with catboost 2nd stage models. <br> Just for the "why not". </p> <p>The approach: </p> <p><img src="https://i.ibb.co/McJ39mW/image-nuke.png" alt=""></p> <p>This stunt was done mainly for the purpose me catching up with the current most recent SOTA vision papers. </p> <p>I seriously didn't try to compete on the leaderboard and never had the intention of releasing a public notebook that actually gets a silver medal. Especially not so close to the competition deadline, I also didn't publish all the trained model weights. Just in case.</p> <p>This came as a complete surprise to me! <br> Hope the solution will be useful for many others in the future.</p> <p>If you got any questions or feedback, I'll be more than happy to discuss!</p>
PetFinder.my - Pawpularity Contest	Silver medal notebook (Private LB 17.04615, rank 126)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my - Pawpularity Contest <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>So because I used a two notebook structure for my submission, the submission attached to my final score doesn't do a good job of explaining exactly what I did.</p> <p>My main notebook was used to train the weights for my submission notebook (which just pulled in the trained weights and made predictions).</p> <p>I put together a complete notebook that walks you through the process I used to train and submit.</p> <p>Here is the link:<br> <a href="https://www.kaggle.com/malachymoran/the-general" target="_blank">The General</a></p> <p>Is the link in the bottom of the post not working?</p>
PetFinder.my - Pawpularity Contest	Tentative 2nd place solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my - Pawpularity Contest <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thanks! I answered the question in another comment. Please find it!</p> <p>Thanks! I answered the question in another comment. Please find it!</p> <p>We are also looking forward to discussing you and your team! btw, an email I got from Kaggle says</p> <blockquote> <p>Winner’s Information Gathering: Please use this form to learn about the materials and other information we need from you. Each member of your team should submit a form. <strong>There will be no winners call for this competition</strong>, so you may skip the steps calling for that presentation.</p> </blockquote> <p>I thought we won't have such a presentation opportunity. <a href="https://www.kaggle.com/kaggleteam" target="_blank">@kaggleteam</a>, what should we do next?</p> <p>Hi Jedi, <br> Thank you for hosting such a nice competition! We are preparing our solution to submit. Please let us know if there is something we can help to implement our solution.</p> <p>I think train data also has about 2,000 images overlapping. I am not sure if this can be the answer though.</p> <p>Thanks a lot!</p> <p>Thanks a lot!</p> <p>Yes, I just used this loss</p> <pre><code>poisson_loss = torch.nn.PoissonNLLLoss(log_input=False, full=False, size_average=None, eps=1e-08, reduce=None, reduction='mean') </code></pre> <p>For training image models, I did not do any special things. I start from this kernel <a href="https://www.kaggle.com/keagle/petfindder-updated-rules" target="_blank">https://www.kaggle.com/keagle/petfindder-updated-rules</a>. Then, I just changed the model name for timm…</p> <p>Hi all,<br> Congratulations to <a href="https://www.kaggle.com/titericz" target="_blank">@titericz</a>! Here is our brief solution write-up. We will publish all our codes and datasets. Please post any questions in this thread if you have any.</p> <p><strong>Update:</strong> our code is here! <a href="https://www.kaggle.com/ktakita/team-merged-ensemble-fixednewmodelsexp029nonmean" target="_blank">https://www.kaggle.com/ktakita/team-merged-ensemble-fixednewmodelsexp029nonmean</a></p> <h1>Models</h1> <p>We mainly used Swin Transformer as also used in top kernels. I just used the fastai pipeline that appears in many public kernels. My models and ensemble weights are following. Then, we ensemble each model from each member. </p> <table> <thead> <tr> <th>Modell</th> <th>Loss</th> <th>Weight</th> </tr> </thead> <tbody> <tr> <td>convnext_large</td> <td>BCE</td> <td>0.107571</td> </tr> <tr> <td>cait_m36_384</td> <td>BCE</td> <td>0.086765</td> </tr> <tr> <td>vit_base_patch16_224_miil_in21k</td> <td>Poisson</td> <td>0.084596</td> </tr> <tr> <td>cait_m36_384</td> <td>100-class CE</td> <td>0.076999</td> </tr> <tr> <td>xcit_small_24_p16_384_dist</td> <td>BCE</td> <td>0.076158</td> </tr> <tr> <td>xcit_small_24_p8_384_dist</td> <td>BCE</td> <td>0.073455</td> </tr> <tr> <td>crossvit_18_dagger_408</td> <td>BCE</td> <td>0.072553</td> </tr> <tr> <td>swin_large_patch4_window12_384_in22k</td> <td>BCE</td> <td>0.072102</td> </tr> <tr> <td>vit_base_patch16_224_miil_in21k</td> <td>BCE</td> <td>0.071953</td> </tr> <tr> <td>vit_base_r50_s16_384</td> <td>BCE</td> <td>0.058601</td> </tr> <tr> <td>swin_base_patch4_window12_384</td> <td>BCE</td> <td>0.057393</td> </tr> <tr> <td>vit_base_patch16_224_miil_in21k</td> <td>MSE</td> <td>0.056347</td> </tr> <tr> <td>jx_nest_base</td> <td>BCE</td> <td>0.055556</td> </tr> <tr> <td>vit_large_patch16_224</td> <td>Poisson</td> <td>0.054602</td> </tr> </tbody> </table> <p><a href="https://www.kaggle.com/ktakita" target="_blank">@ktakita</a> and <a href="https://www.kaggle.com/teyosan1229" target="_blank">@teyosan1229</a> parts are as follows:</p> <p>We used three types of the following data<br> ①Image data<br> T.Resize(image_size)<br> T.CenterCrop([image_size, image_size])<br> ②One hot data<br> 'Subject Focus', 'Eyes', 'Face', 'Near', 'Action', 'Accessory',<br> 'Group', 'Collage', 'Human', 'Occlusion', 'Info', 'Blur'<br> ③Image shape<br> 'height', 'width', 'aspect'</p> <p>For the models, we ensemble the following 6 models by Ridge.</p> <ul> <li><p>swin_large_patch4_window12_384<br> used data:①<br> image_size:384<br> loss:nn.BCEWithLogitsLoss</p></li> <li><p>poolformer_m36<br> used data:①, ②<br> image_size:224<br> loss:BCEWithLogitsLossFlat</p></li> <li><p>swin_large_patch4_window7_224_in22k<br> used data:①, ②, ③<br> image_size:224<br> loss:BCEWithLogitsLossFlat</p></li> <li><p>swin_large_patch4_window12_384<br> used data:①<br> image_size:384<br> loss:BCEWithLogitsLossFlat</p></li> <li><p>swin_large_patch4_window7_224<br> used data:①, ②<br> image_size:224<br> loss:BCEWithLogitsLossFlat</p></li> <li><p>tf_efficientnetv2_b1<br> used data:①, ②, ③<br> image_size:224<br> loss:BCEWithLogitsLossFlat<br> Kohei single model<br> swin_large_patch4_window12_384_in22k<br> used data:①, ②, ③<br> image_size:384<br> loss:BCEWithLogitsLossFlat<br> teyosan single model<br> dm_nfnet_f3<br> used data:①,②<br> image_size:512<br> loss:BCEWithLogitsLoss</p></li> </ul> <h1>Association with the data of the previous competition</h1> <p>As discussed, it is allowed to use the data of the previous competition (<a href="https://www.kaggle.com/c/petfinder-adoption-prediction)" target="_blank">https://www.kaggle.com/c/petfinder-adoption-prediction)</a>. We find that some images are overlapping between this and previous competitions. In the training data, there are about 1,600 images. We then guess that such kind of overlapping could happen also in test data. <br> Based on the finding, we join the metadata in the previous competition with this current competition data by using the image hash as a key. </p> <h1>2nd-stage GBDT model</h1> <p>To use the associated metadata, we trained a 2nd-stage GBDT model, which is also known as a “stacking model”. We use the prediction of the 1st-stage DNN models (e.g., SwinT) as a feature of the 2nd-stage GBDT model. The training code is available here (TBA).</p> <p><strong>Update:</strong> After discussion with other teams, one possible reason why we jumped to 2nd seems to be the training strategy that we train 2nd-stage GBDT model with the entire of this competition data, not only with the records which are associated with the previous competition. It may be helpful to preserve the distribution of predicting Pawpularity.</p> <p>These GBDT models easily got overfitted. We were not able to find a good combination of feature sets. In the previous competitions, there are a lot of effective features are reported, but the features we tested made LB score worse. There should be some effective features we have not tested due to the time limit. We suspect that this may be one of the reasons why we did not reach the top.</p> <p>It may only have a minor effect. I did not use it. How do you think, <a href="https://www.kaggle.com/ktakita" target="_blank">@ktakita</a> and <a href="https://www.kaggle.com/teyosan1229" target="_blank">@teyosan1229</a> ?</p> <p>Thanks! With CV prediction, I solved a least-square problem to minimize MSE as follows.</p> <pre><code>coef = np.linalg.lstsq(df[['pred1', 'pred2', 'pred3']], df['Pawpularity'] )[0] </code></pre> <p>Then, I iteratively remove models which have less than 5% contribution. </p> <p>The score of each model is as follows.</p> <table> <thead> <tr> <th></th> <th>RMSE</th> </tr> </thead> <tbody> <tr> <td>exp7_cait_m36_384</td> <td>17.661636</td> </tr> <tr> <td>exp7_swin_large_patch4_window12_384_in22k</td> <td>17.687250</td> </tr> <tr> <td>exp20_convnext_large</td> <td>17.741086</td> </tr> <tr> <td>exp7_vit_base_patch16_224_miil_in21k</td> <td>17.772114</td> </tr> <tr> <td>exp20_convnext_base</td> <td>17.823614</td> </tr> <tr> <td>exp4_xcit_small_24_p16_384_dist</td> <td>17.861272</td> </tr> <tr> <td>exp7_swin_base_patch4_window12_384</td> <td>17.864037</td> </tr> <tr> <td>exp7_jx_nest_base</td> <td>17.880826</td> </tr> <tr> <td>exp4_crossvit_18_dagger_408</td> <td>17.893266</td> </tr> <tr> <td>exp4_xcit_small_24_p8_384_dist</td> <td>17.953217</td> </tr> <tr> <td>exp8_vit_base_patch16_224_miil_in21k</td> <td>18.021636</td> </tr> <tr> <td>exp15_vit_base_patch16_224_miil_in21k</td> <td>18.085443</td> </tr> <tr> <td>exp7_vit_base_r50_s16_384</td> <td>18.115582</td> </tr> <tr> <td>exp15_vit_large_patch16_224</td> <td>18.223578</td> </tr> <tr> <td>exp9_cait_m36_384</td> <td>23.401867</td> </tr> </tbody> </table> <p>The ensembled score is eventually 17.21.</p> <p>I did not check the LB score, but you may try it with the code we shared. </p> <p>Thanks. I have made them public. I mainly used AMD Ryzen 9 3950X, 64GB mem, and RTX 3090 in my lab. </p> <p>Hi, I think my code is completely messed up and may be hard to understand. To submit our solution to the host, I have updated all code as Kaggle kernels as below. I checked that all code is executable. To reproduce our models, please increase the number of epochs to 20. Hope my code is easy to understand.</p> <ul> <li>1st stage models<ul> <li>Exp4: Trained with BCE Loss<ul> <li><a href="https://www.kaggle.com/tkm2261/petfinder-2nd-place-model-exp4" target="_blank">https://www.kaggle.com/tkm2261/petfinder-2nd-place-model-exp4</a> </li></ul></li> <li>Exp7: Trained with BCE Loss and larger batch size<ul> <li><a href="https://www.kaggle.com/tkm2261/petfinder-2nd-place-model-exp7" target="_blank">https://www.kaggle.com/tkm2261/petfinder-2nd-place-model-exp7</a> </li></ul></li> <li>Exp8: Trained with MSE Loss<ul> <li><a href="https://www.kaggle.com/tkm2261/petfinder-2nd-place-model-exp8" target="_blank">https://www.kaggle.com/tkm2261/petfinder-2nd-place-model-exp8</a> </li></ul></li> <li>Exp9: Trained with Cross-entropy loss as a 100-label problem<ul> <li><a href="https://www.kaggle.com/tkm2261/petfinder-2nd-place-model-exp9" target="_blank">https://www.kaggle.com/tkm2261/petfinder-2nd-place-model-exp9</a> </li></ul></li> <li>Exp15: Trained with Poisson NLL loss <ul> <li><a href="https://www.kaggle.com/tkm2261/petfinder-2nd-place-model-exp15" target="_blank">https://www.kaggle.com/tkm2261/petfinder-2nd-place-model-exp15</a> </li></ul></li> <li>Exp20: ConvNext training<ul> <li><a href="https://www.kaggle.com/tkm2261/petfinder-2nd-place-model-exp20" target="_blank">https://www.kaggle.com/tkm2261/petfinder-2nd-place-model-exp20</a> </li></ul></li></ul></li> <li>2nd stage models training with 1st stage prediction and metadata from the current and previous competitions<ul> <li><a href="https://www.kaggle.com/ktakita/0113-2nd-stage-gbdt-fixed-w-newmodels3" target="_blank">https://www.kaggle.com/ktakita/0113-2nd-stage-gbdt-fixed-w-newmodels3</a> </li></ul></li> <li>Inference<ul> <li><a href="https://www.kaggle.com/ktakita/team-merged-ensemble-fixednewmodelsexp029nonmean?scriptVersionId=85153314" target="_blank">https://www.kaggle.com/ktakita/team-merged-ensemble-fixednewmodelsexp029nonmean?scriptVersionId=85153314</a> </li></ul></li> </ul>
PetFinder.my - Pawpularity Contest	Two main tricks helped us to avoid the shake	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my - Pawpularity Contest <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><h1>Prediction + Ranking Ensembling</h1> <p>We ranked each model's prediction ranking from 1 ~ N (including test data), effectively the features for the ensemble are doubled. </p> <pre><code>all_preds = np.concatenate([ensemble_oofs, ensemble_tests]) all_preds.shape all_ranks = np.zeros_like(all_preds) for i in range(all_preds.shape[1]): all_ranks[:, i] = rankdata(all_preds[:, i]) ensemble_oof_ranks = all_ranks[:num_trains] ensemble_test_ranks = all_ranks[num_trains:] ensemble_oofs = np.column_stack([ensemble_oofs, ensemble_oof_ranks]) ensemble_tests = np.column_stack([ensemble_tests, ensemble_test_ranks]) </code></pre> <h1>Cross Validate the Ensembling</h1> <p>By checking all of our historical submissions, it turns out switching to a CV-based (such as Leave one) ensembling is the main factor why our team can avoid the shake-up.</p> <p>Code for Leave one out ridge ensembling:</p> <pre><code>def leave_one_out_ensemble(oofs, labels, tests, search_range=np.arange(100, 10000, 0.5)): sc = StandardScaler() sc.fit(oofs) regcv = RidgeCV(alphas=search_range, cv=None, scoring="neg_root_mean_squared_error", normalize=False) regcv.fit(sc.transform(oofs), labels) print(f"regcv.alpha_: {regcv.alpha_}, regcv.best_score_: {regcv.best_score_}") </code></pre> <p>Also, H2O automl performs very well for ensembling as well. </p> <pre><code>aml1 = H2OAutoML(max_runtime_secs=3600, stopping_metric='rmse',seed=721) </code></pre>
PetFinder.my - Pawpularity Contest	[ 31th solution ] noisy student, knowledge distillation	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my - Pawpularity Contest <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>It's very interesting! I didn't test 3 folds.</p> <p>It's not dramatic. When I calculate CV score TTA7 is the best score. So when I ensemble and predict, I use TTA7.</p> <p>I used 5 fold. And I create 5 NN models by training OOF data.<br> In my case, the CV improved but the LB dropped too. So I didn't use this ensemble method. But in private LB, I found that this ensemble method achieve 16.90625. I'm sorry to didn't use that.</p> <p>It's simple. In fastai<br> <code>learn.tta(n=7)</code></p> <p>you means SVR ensemble? It's simple. Just extract each model's embedding and concentrate embeddings. Then train SVR <a href="https://www.kaggle.com/cdeotte/rapids-svr-boost-17-8" target="_blank">reference code</a></p> <p>Thank you all the people who participated in this competition, and the organizer that hosts such a sweet competition.</p> <p>I am glad to shake up to 31th place. And I can achieve competition master.<br> In this competition, i want to study multi-modal and noisy student and knowledge distillation.</p> <h1>Overview</h1> <p>For my solution, refer to this paper. <br> <a href="https://arxiv.org/abs/1911.04252" target="_blank">Self-training with Noisy Student improves ImageNet classification</a></p> <ol> <li>train single models</li> <li>ensemble models for predict and label external data</li> <li>train single models use external data + petfinder data</li> </ol> <h2>step1. train single models</h2> <h3>model architecture</h3> <p>I use fastai and timm library.<br> The model is as follows.</p> <p>class cust_fastai_model(nn.Module): </p> <pre><code>def __init__(self, model_name='swin_large_patch4_window7_224', emb=True): super().__init__() self.emb = emb self.backbone = create_model(model_name, pretrained=False, num_classes=0) image_num_features = self.backbone.num_features self.image_fc = nn.Sequential( nn.Dropout(0.3), nn.Linear(image_num_features, 128) ) self.custom_head = nn.Sequential( nn.Dropout(0.4), nn.Linear(128, 1) ) def forward(self, image): emb = self.backbone(image) emb = self.image_fc(emb) out = self.custom_head(emb) if self.emb: return torch.cat([out, emb], dim=1) else: return out </code></pre> <h3>single model performance</h3> <table> <thead> <tr> <th>model name</th> <th>image size</th> <th>CV</th> </tr> </thead> <tbody> <tr> <td>swin_large_patch4_window7_224</td> <td>224</td> <td>17.4291008</td> </tr> <tr> <td>swin_base_patch4_window7_224</td> <td>224</td> <td>17.62332115</td> </tr> <tr> <td>swin_large_patch4_window12_384</td> <td>384</td> <td>17.51918907</td> </tr> <tr> <td>tf_efficientnet_b5_ns</td> <td>456</td> <td>17.70378685</td> </tr> <tr> <td>vit_large_patch16_224</td> <td>224</td> <td>17.63189049</td> </tr> <tr> <td>eca_nfnet_l2</td> <td>384</td> <td>17.58191376</td> </tr> </tbody> </table> <h2>step2. ensemble models for predict and label external data.</h2> <h3>calculate CV about tta</h3> <table> <thead> <tr> <th>TTA</th> <th>CV</th> </tr> </thead> <tbody> <tr> <td>1</td> <td>17.22635913</td> </tr> <tr> <td>2</td> <td>17.20699514</td> </tr> <tr> <td>3</td> <td>17.2165909</td> </tr> <tr> <td>4</td> <td>17.20069566</td> </tr> <tr> <td>5</td> <td>17.21008584</td> </tr> <tr> <td>6</td> <td>17.20981791</td> </tr> <tr> <td>7</td> <td>17.1992845</td> </tr> <tr> <td>8</td> <td>17.19996605</td> </tr> <tr> <td>9</td> <td>17.21487459</td> </tr> <tr> <td>10</td> <td>17.2102797</td> </tr> </tbody> </table> <p>Then, I use tta7 to predict and label external data.</p> <h3>External data</h3> <ol> <li><a href="https://cocodataset.org/#home" target="_blank">COCO cat and dog</a></li> <li><a href="https://www.kaggle.com/c/petfinder-adoption-prediction" target="_blank">Petfinder adopt dataset</a></li> <li><a href="https://www.kaggle.com/c/dogs-vs-cats" target="_blank">kaggle cat and dog dataset</a></li> </ol> <h2>step3. train single models by use external data</h2> <h3>single models performance</h3> <table> <thead> <tr> <th>model name</th> <th>image size</th> <th>CV</th> </tr> </thead> <tbody> <tr> <td>swin_large_patch4_window7_224</td> <td>224</td> <td>17.32193184</td> </tr> <tr> <td>swin_base_patch4_window7_224</td> <td>224</td> <td>17.34374313</td> </tr> <tr> <td>swin_large_patch4_window12_384</td> <td>384</td> <td>17.28282471</td> </tr> <tr> <td>tf_efficientnet_b5_ns</td> <td>456</td> <td>17.25554123</td> </tr> <tr> <td>vit_large_patch16_224</td> <td>224</td> <td>17.37071266 </td> </tr> <tr> <td>eca_nfnet_l2</td> <td>384</td> <td>-</td> </tr> </tbody> </table> <p>In this experiment, eca_nfnnet_l2 need a lot of times, so i didn't train this model.</p> <p>And I can found out efficientnet can learn more than other models.</p> <h3>ensemble performance</h3> <table> <thead> <tr> <th>TTA</th> <th>CV</th> </tr> </thead> <tbody> <tr> <td>1</td> <td>17.19522166</td> </tr> <tr> <td>2</td> <td>17.18533359</td> </tr> <tr> <td>3</td> <td>17.19308663</td> </tr> <tr> <td>4</td> <td>17.18103402 </td> </tr> <tr> <td>5</td> <td>17.18683142</td> </tr> <tr> <td>6</td> <td>17.18794017</td> </tr> <tr> <td>7</td> <td>17.17808405</td> </tr> <tr> <td>8</td> <td>17.18468712</td> </tr> <tr> <td>9</td> <td>17.18554875</td> </tr> <tr> <td>10</td> <td>17.18290917</td> </tr> </tbody> </table> <p>I can improve CV a little. And i have a problem about notebook execution time. When i use TTA7.</p> <h3>predict and train again.</h3> <p>In noisy student paper, they can perform more when they predict and train again. so predict and train again.<br> And ensemble performance is as follow.</p> <table> <thead> <tr> <th>TTA</th> <th>CV</th> </tr> </thead> <tbody> <tr> <td>1</td> <td>17.17037105</td> </tr> <tr> <td>2</td> <td>17.17464033</td> </tr> <tr> <td>3</td> <td>17.17522471</td> </tr> <tr> <td>4</td> <td>17.1676797</td> </tr> <tr> <td>5</td> <td>17.17269141</td> </tr> <tr> <td>6</td> <td>17.1767502</td> </tr> <tr> <td>7</td> <td>17.16777866</td> </tr> <tr> <td>8</td> <td>17.17085731</td> </tr> <tr> <td>9</td> <td>17.17535145</td> </tr> <tr> <td>10</td> <td>17.17163645 </td> </tr> </tbody> </table> <p>In this experminet I can improve CV a little and I don't have to TTA much. So I use TTA2. And I can use all of this models.</p> <h3>more models</h3> <p>Looking at the experiment above, I was able to save time. So I train more models by 2nd pseudo label.<br> I train 5models.</p> <ol> <li>regnet</li> <li>resnext</li> <li>convnext base 384</li> <li>convnext large 224</li> <li>convnext xlarge 224<br> And regnet and resnext can't achieve high CV. So I use only convnext models.</li> </ol> <h2>Final models</h2> <p>I use models as follow</p> <table> <thead> <tr> <th>model name</th> <th>dataset</th> <th>image size</th> <th>CV</th> </tr> </thead> <tbody> <tr> <td>swin_large_patch4_window7_224</td> <td>2nd pseudo label</td> <td>224</td> <td>17.31553802</td> </tr> <tr> <td>swin_base_patch4_window7_224</td> <td>2nd pseudo label</td> <td>224</td> <td>17.36221161</td> </tr> <tr> <td>swin_large_patch4_window12_384</td> <td>2nd pseudo label</td> <td>384</td> <td>17.31249619</td> </tr> <tr> <td>tf_efficientnet_b5_ns</td> <td>2nd pseudo label</td> <td>456</td> <td>17.31544266</td> </tr> <tr> <td>vit_large_patch16_224</td> <td>2nd pseudo label</td> <td>224</td> <td>17.36038437</td> </tr> <tr> <td>eca_nfnet_l2</td> <td>origin dataset</td> <td>384</td> <td>17.58191376</td> </tr> <tr> <td>swin_large_patch4_window7_224_pse</td> <td>1st pseudo label</td> <td>224</td> <td>17.32193184</td> </tr> <tr> <td>swin_large_patch4_window12_384</td> <td>1st pseudo label</td> <td>384</td> <td>17.28282471</td> </tr> <tr> <td>tf_efficientnet_b5_ns</td> <td>1st pseudo label</td> <td>456</td> <td>17.25554123</td> </tr> <tr> <td>convnext-base</td> <td>2st pseudo label</td> <td>384</td> <td>soon update</td> </tr> <tr> <td>convnext-large</td> <td>2st pseudo label</td> <td>224</td> <td>soon update</td> </tr> <tr> <td>convnext-xlarge</td> <td>2st pseudo label</td> <td>224</td> <td>soon update</td> </tr> </tbody> </table> <h2>Other things</h2> <p>I use SVR to ensemble.<br> <img src="https://user-images.githubusercontent.com/24911666/149612546-8f4f8988-06eb-4e24-aab8-89a635c70d98.png" alt="ensemble"><br> It can improve CV. But decrease LB a lot. So i didn't use. But in private LB,</p> <table> <thead> <tr> <th>model</th> <th>dataset</th> <th>private LB</th> </tr> </thead> <tbody> <tr> <td>ensemble SVR</td> <td>origin dataset</td> <td>16.90625</td> </tr> <tr> <td>ensemble 6 models</td> <td>origin dataset</td> <td>17.02122</td> </tr> <tr> <td>my final model 1</td> <td>pseudo labeling</td> <td>16.99605</td> </tr> <tr> <td>my final model 2</td> <td>pseudo labeling</td> <td>16.99253</td> </tr> <tr> <td>I'm sorry to didn't use SVR.</td> <td></td> <td></td> </tr> </tbody> </table> <h1>What I can learn in this competition</h1> <ul> <li>use noisy student, I can improve sinlge model's performance.</li> <li>use noisy student, I didn't need to many TTA.</li> <li>We can use SVR to ensemble models.</li> <li>Trust your CV.</li> </ul> <p>Everyone did great! See you at the next competition.</p>
PetFinder.my - Pawpularity Contest	[114th] Ensemble 7 models (pseudo label and not pseudo label)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my - Pawpularity Contest <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>First of all, congrats to all winners and thanks to organizers. I have learned a lot from this competition, so I would like to share my solution here!</p> <p>Inference code and the detail of my solution:<br> <a href="https://www.kaggle.com/ttkagglett/119th-ensemble-7models-pseudo-and-not-pseudo" target="_blank">https://www.kaggle.com/ttkagglett/119th-ensemble-7models-pseudo-and-not-pseudo</a></p> <p>Make pseudo label dataset code:<br> <a href="https://www.kaggle.com/ttkagglett/petfinder-make-pseudolabel-dataset" target="_blank">https://www.kaggle.com/ttkagglett/petfinder-make-pseudolabel-dataset</a></p> <p>Semi-Supervised Learning (Training with pseudo labeled data) code:<br> <a href="https://www.kaggle.com/ttkagglett/petfinder-cnn-semi-supervised-training-fastai" target="_blank">https://www.kaggle.com/ttkagglett/petfinder-cnn-semi-supervised-training-fastai</a></p> <p>Training code:<br> <a href="https://www.kaggle.com/ttkagglett/petfinder-cnn-training-for-pseudolabel-2-fastai" target="_blank">https://www.kaggle.com/ttkagglett/petfinder-cnn-training-for-pseudolabel-2-fastai</a><br> I made many cloned notebooks from this to make various types of models.</p> <p>I used Kaggle environment only (1 GPU including GPU time limitation). Many of you can not use GPUs as you want, but we can train models! <br> I hope some Kagglers find new ideas from my solution. Please let me know if you have any questions.</p>
PetFinder.my - Pawpularity Contest	[Public 1st/ Private 5th] Solution Overview	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my - Pawpularity Contest <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>I'm grad to receive comment from legend kaggler as you.<br> I always learn from your solution and discussion comment !!</p> <p>Thank you all teams who competed with us, all the people who participated in this competition, and the organizer that hosts such a great competition.</p> <p>I'm very sorry to shake down and missed 1st … !</p> <p>Our final notebooks are available.</p> <ul> <li>sub1 : <a href="https://www.kaggle.com/kaerunantoka/offense-final?scriptVersionId=85192352" target="_blank">https://www.kaggle.com/kaerunantoka/offense-final?scriptVersionId=85192352</a></li> <li>sub2 : <a href="https://www.kaggle.com/yururoi/stacking-pp-avg-pp-34-defensive-breed-group?scriptVersionId=84896876" target="_blank">https://www.kaggle.com/yururoi/stacking-pp-avg-pp-34-defensive-breed-group?scriptVersionId=84896876</a></li> <li>2nd stage training code : <a href="https://www.kaggle.com/yururoi/offense-stacking-pp-avg-pp-36-b285-b103" target="_blank">https://www.kaggle.com/yururoi/offense-stacking-pp-avg-pp-36-b285-b103</a></li> </ul> <h1>Overview</h1> <p>our solution has 3 steps.</p> <ul> <li><p>1st : Make a lot of single models. (our final submittion has 14 single models)</p></li> <li><p>2nd : Calculate similarity score to detect the same images, previous competiton's train+test images (about 70000 images) and this competition's images.</p></li> <li><p>Final : On overthreshold data, We can use previous competitions meta data. So, we made models for each adoptionspeed as post proccessing.</p></li> </ul> <p><img alt="Screen Shot 2022-01-15 at 7 23 54" src="https://user-images.githubusercontent.com/38504303/149600083-089fab1c-94fa-4065-997f-e90896c080b3.png"></p> <h1>Single Models</h1> <p>We use below models.</p> <ul> <li>All models are trained with Only Images. (Without meta data!)</li> <li>Cut long side to make the aspect ratio fixed in Dataset class.</li> </ul> <table> <thead> <tr> <th>Model</th> <th>Loss func</th> <th>CV</th> <th>LB (if any)</th> </tr> </thead> <tbody> <tr> <td>swin_large_patch4_window12_384</td> <td>BCE</td> <td>17.484603</td> <td>18.02737</td> </tr> <tr> <td>swin_large_patch4_window12_384</td> <td>BCE</td> <td>17.3944997416399</td> <td>-</td> </tr> <tr> <td>swin_large_patch4_window12_384</td> <td>BCE</td> <td>17.7691509886505</td> <td>17.98811</td> </tr> <tr> <td>swin_large_patch4_window7_224</td> <td>BCE</td> <td>17.7937685311947</td> <td>17.97379</td> </tr> <tr> <td>swin_large_patch4_window7_224</td> <td>BCE</td> <td>17.7916477565051</td> <td>17.94652</td> </tr> <tr> <td>swin_large_patch4_window7_224</td> <td>BCE</td> <td>17.458288</td> <td>-</td> </tr> <tr> <td>swin_large_patch4_window7_224</td> <td>BCE</td> <td>17.5616612122726</td> <td>17.96412</td> </tr> <tr> <td>swin_large_patch4_window7_224</td> <td>BCE</td> <td>17.6927744617933</td> <td>17.98374</td> </tr> <tr> <td>swin_large_patch4_window7_224</td> <td>BCE</td> <td>17.52656</td> <td>17.90651</td> </tr> <tr> <td>vit_base_r50_s16_384</td> <td>BCE</td> <td>17.9787132114343</td> <td>-</td> </tr> <tr> <td>vit_large_r50_s32_384</td> <td>BCE</td> <td>17.705468</td> <td>-</td> </tr> <tr> <td>densenet121</td> <td>MSE</td> <td>18.811589</td> <td>-</td> </tr> <tr> <td>tf_efficientnet_b0_ns</td> <td>MSE</td> <td>18.1307282713156</td> <td>-</td> </tr> <tr> <td>resnext101_32x8d</td> <td>BCE</td> <td>18.4265347215127</td> <td>-</td> </tr> </tbody> </table> <h1>Validation Strategy</h1> <p>We used below 2 method</p> <ul> <li><p>GroupKfold … Add id to the same images, this id is used as group.</p></li> <li><p>StratifiedKfold … Drop the same images, and split fold.</p></li> </ul> <h1>DataAugmentation</h1> <pre><code>CFG.get_transforms = { 'train' : A.Compose([ A.OneOf([ A.RandomResizedCrop(CFG.IMG_SIZE, CFG.IMG_SIZE, p=0.3, scale=(0.85, 0.95)), A.Resize(CFG.IMG_SIZE, CFG.IMG_SIZE, p=0.4), A.Compose([ A.Resize(int(CFG.IMG_SIZE * 1.5), int(CFG.IMG_SIZE * 1.5), p=1.0), A.CenterCrop(p=1.0, height=CFG.IMG_SIZE, width=CFG.IMG_SIZE), ], p=0.3), ], p=1.0), A.HorizontalFlip(p=0.5), A.VerticalFlip(p=0.5), A.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225), max_pixel_value=255.0, p=1.0,), ], p=1.0), 'valid' : A.Compose([ A.Resize(CFG.IMG_SIZE, CFG.IMG_SIZE, p=1), A.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225), max_pixel_value=255.0, p=1.0,), ], p=1.0), } </code></pre> <h1>Detect similar (maybe the same) image</h1> <ul> <li>Feature Extractor : EfficientNetB0</li> <li>Calculate cosine similarity</li> </ul> <h1>Stacking</h1> <ul> <li>Model … BayesianRidge</li> <li>fold … 4fold</li> </ul> <h1>Post Processing</h1> <ul> <li>Model … 5 BayesianRidge Models for each AdoptionSpeed (0~4)</li> <li>fold … 4fold</li> <li>Feature … oof, Age, Breed, TextLength, and so on.</li> </ul> <h1>What didn't work</h1> <ul> <li>Use Metadata</li> <li>Cropped images with Yolo v5</li> <li>Data augmentations on CNN models</li> <li>Image stats features</li> </ul> <p>Thank you!</p> <p>Thank you for comment! お疲れ様でした！</p>
PetFinder.my Adoption Prediction	10th Solution Summary	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my Adoption Prediction <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thank you for reading</p> <p>Thank you for pointing it out. Sorry,I make a mistake here,LGB 6(multi-classify/regression) is more suitable .I do not combine reg and classify. </p> <p>muitl-classify details: I use multi-classify to get 5 dims prob-matrix, then sum([0,1,2,3,4]prob-matrix) to get one col value and finally use optR to predict it as five classes. </p> <p>Thank you for reading</p> <p>First of all, thanks to Petfinder.my and kaggle for hosting this great competition. And congratulations to the winners!Thanks to my teammates for their efforts. Here is our solution.</p> <h2>General</h2> <p>As our team name,Stacking is all you need. We have 4 group features:</p> <h3>features one:</h3> <ul> <li>1）clean breed ```python def deal_breed(df): if df['Breed1']==df['Breed2']: df['Breed2']=0 if df['Breed1']!=307 and df['Breed2']==307: temp=df["Breed1"] df['Breed1']=df['Breed2'] df['Breed2']=temp return df</li> </ul> <p><code> * 2）res features: </code>python def get_res_feat(df): temp=pd.DataFrame(index=range(1)) temp['RescuerID']=df['RescuerID'].values[0] temp['res_type_cnt']=len(df['Type'].unique()) temp['res_breed_cnt']=len(df['Breed1'].unique()) temp['res_breed_mode']=df['Breed1'].mode() temp['res_fee_mean']=df['Fee'].mean() temp['res_Quantity_sum']=df['Quantity'].sum() temp['res_MaturitySize_mean']=df['MaturitySize'].mean() temp['res_Description_unique']=len(df['Description'].unique()) return temp ``` * 3)meta features from public kernel * 4)Description features: * tfidf+svg * desc+type+breed+color->tfidf+svg/nmf/lda * desc+type+breed+color->countvec+svg/lda * desc->wordbatch+svg * 5)category_col onehot+svg * 6)densenet121 extract img features+svg * 7)state features (external data) * state population density * state_rank(according to state population density) * 8)mean target encode with breed(breed=breed1+breed2) boost 0.007(from 0.463 to 0.470) * 9)linear model oof * use 1-7 features to build some leaner model ,get linear model oof.</p> <p>with thease 101 dim features and multi-classify LGB (BEST SINGLE MODEL LB 0.471)</p> <h3>note:</h3> <p>get multi-classify result and optimize it: sum([0,1,2,3,4]prob_matrix) <code>python class_list=[0,1,2,3,4] pred_test_y=np.array([sum(pred_test_y[ix]class_list) for ix in range(len(pred_test_y[:,0]))]) </code></p> <h3>features two</h3> <p>In the early competition, I forked from the public kernel（lb 0.444）</p> <h3>features three <a href="/zhouqingsongct">@zhouqingsongct</a></h3> <p>use featuretools auto extract features </p> <h3>features four [for build lgb6]</h3> <p>from my teammate <a href="/amgis3">@amgis3</a>(lb 0.470)</p> <h3>models</h3> <p>and we stacking them with these models: * LGB 6LGB 6(multi-classify/regression) * CAT 2(multi-classify/regression) NN 3(multi-classify/regression) linear model(regression)</p> <h3>NN <a href="/gmhost">@gmhost</a>(NN structures reference to <a href="https://www.kaggle.com/c/avito-demand-prediction/discussion/59880#latest-450523">https://www.kaggle.com/c/avito-demand-prediction/discussion/59880#latest-450523</a>)</h3> <p>we do not use public Embeddings.we just use train+test desc to pretain a new w2v model(think of that are many chinese and Malay.) maybe we are wrong, our best nn is only near lb 0.44 </p> <h2>structures</h2> <p><img src="https://storage.googleapis.com/kaggle-forum-message-attachments/513498/12922/img.jpg" alt="img"></p> <h3>optR</h3> <p>The open source optR predicts very few zeros, so I manually divide the smallest part of the value to 0 (0.95* len_0) ( len_0 is the size of 0 in trainset) ```python</p> <p>def predict(self, X, coef,len_0): X_p = np.copy(X) temp = sorted(list(X_p)) threshold=temp[int(0.95*len_0)-1] for i, pred in enumerate(X_p): if pred < threshold: X_p[i] = 0 elif pred >= threshold and pred < coef[1]: X_p[i] = 1 elif pred >= coef[1] and pred < coef[2]: X_p[i] = 2 elif pred >= coef[2] and pred < coef[3]: X_p[i] = 3 else: X_p[i] = 4 return X_p ``` That's all,thank you <a href="https://www.kaggle.com/chizhu2018/final-submit-two-10th-solution-private-0-442">code</a> is here <a href="https://github.com/chizhu/kaggle-petfinder-adoption-prediction-10th-solution">github</a></p> <p>Thank you for figuring out this error.</p> <p>Thank you</p> <p>Thank you!congratulations for your solo gold medal.</p>
PetFinder.my Adoption Prediction	13-th place solution summary 0.44091 (65-th on public LB: 0.459~0.467)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my Adoption Prediction <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thank you! It seems I had a verrrry good luck in this competition:)</p> <p>I published the code :) <a href="https://www.kaggle.com/corochann/13-th-place-solution-ensemble-of-5-models">13-th place solution: ensemble of 5 models</a></p> <p>It's similar to one-hot encoding, but it is not one-hot vector but 2 places are set 1 when both breed1 & breed2 are set. I applied cutoff, all minor breeds are dealt as "unknown" breed" so breed have only 18 dim vector.</p> <p>Thank you for your interest!</p> <p>I permute only tabular input feature, target value is kept same value. Every time when I sample minibatch, I select 15 % of columns randomly to permute then I permute rows within minibatch for these 15% of columns. Thus 85% of columns are remain same value, and target value is also remain same.</p> <p>Actually I also tried mixup & adding gaussian noise on input feature, but both approach do not seem to contribute to performance so much. And I did not adopt them in the final submission.</p> <p>Thank you for the interest. Currently the code is quite messy but I will consider :) </p> <p>First of all, thank you for organizers. It was my first time to seriously enter kernel competition, I enjoyed it. I will write up my approach to summarize.</p> <p><strong>[UPDATED] I published the kernel code</strong> - <a href="https://www.kaggle.com/corochann/13-th-place-solution-ensemble-of-5-models">13-th place solution: ensemble of 5 models</a></p> <h2>Feature Engineering</h2> <p>Many feature are adopted from <a href="https://www.kaggle.com/ranjoranjan/single-xgboost-model">Single XGBoost model</a> as a baseline. I will only write the additional feature engineering from this kernel.</p> <h3>Tabular data</h3> <p>Applied cutoff for "Age", "Quantity", "VideoAmt" and "PhotoAmt".</p> <p>Additional information I used is: - state gdp info (ref <a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/78040">The GDP and population of Malaysia states</a>) - breed ratings (ref <a href="https://www.kaggle.com/hocop1/cat-and-dog-breeds-parameters">Cat and dog breeds parameters</a>): performance does not change so much. I adopted only 4 keys which are included both dogs & cats. - language: English, Malay and Chinese are mainly used in Malaysia. Most of the description are English but some are Malay or Chinese. I detected language using <code>langdetect</code> library, and added these language as categorical feature. Detect language is meaningful, because Malay people are mainly Muslim and do not have dogs (detail info can be found in <a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/86581">Stray animals in Malaysia: the Reality I Saw Travelling There For the Past Months</a>).</p> <h3>Text data</h3> <p>TFIDF --> SVD feature extraction is executed for 3 text type 'Description', 'metadata_annots_top_desc', 'sentiment_entities' as same with Simple XGBoost model kernel.</p> <p>Additionally, glove word embedding feature is extracted <strong>only for "metadata_annots_top_desc"</strong>. When I tried to extract word embedding for description or sentiment_entities performance was worse. I tried both <a href="https://www.kaggle.com/rtatman/glove-global-vectors-for-word-representation">glove</a> and <a href="https://www.kaggle.com/facebook/fatsttext-common-crawl">fasttext</a> word embedding and <code>glove.6B.200d.txt</code> was the best performance which I adopted as final submission.</p> <h3>Image data</h3> <p>Final submission used <code>densenet</code> feature extraction (See <a href="https://www.kaggle.com/christofhenkel/extract-image-features-from-pretrained-nn">Extract Image features from pretrained NN</a>). Difference is that - I did not perform AveragePooling1D, and obtrain 1024 feature. I applied SVD to this 1024-dim vector. - More than 1 image may be contained in each pet, I calculated "mean" of image embedding (I calculated image embeddings up to 10 images per pet) feature when more than 1 image is contained.</p> <p>Although I tested a lot of different models supported in ChainerCV (see <a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/75943#486597">here</a> for supported models), I could not get better score than densenet and the final model used only the densenet. I was thinking Densenet performed better because it contains shallow & deep information at once.</p> <h2>Validation strategy</h2> <p>I applied GroupKFold with rescuerid, instead of StratifiedKFold done in many public kernels. (Refer <a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/81809">this discussion</a>).</p> <p>When I check train and test data, it seems train and test rescuers are not overlapped. GroupKFold shows much "worse" score in local experiment, but I think this is more proper setting.</p> <p>I used 4-fold during fast experiment on local, and applied 10-fold for final submission to create 10 models for each architecture.</p> <h2>Model</h2> <h3>XGBoost</h3> <p>It was quite fast to train models among these 5 models when using GPU, and I mainly tested my feature engineering only with XGB through this competition. Since XGBoost cannot handle categorical data as opposed to LightGBM or CatBoost, I converted categorical values to numeric values by following - one hot encoding: Type, Gender, Vaccinated, Dewormed, Sterilized, State, FurLength, Health - breed encoding: Minor breeds are truncated as "unknown" and added breed1 & breed2 as same field. - color encoding: Same with breeds, color1 color2 and color3 are added as same field.</p> <h3>LightGBM</h3> <p>Since it can handle categorical feature, no additional feature engineering is performed.</p> <h3>CatBoost</h3> <p>Same with LightGBM.</p> <h3>xlearn</h3> <p>As written in these kernels (<a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/81597">The Hitchhiker's Guide to the PetFinder Competition</a>, <a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/80937">https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/80937</a>), using FFM for sparse category dataset seems to good to try. It was a problem that official kaggle kernel does not support <code>xlearn</code> library to use FFM, but <a href="/bminixhofer">@bminixhofer</a> showed how to use it by adding library as a dataset, thank you! (Ref: <a href="https://www.kaggle.com/bminixhofer/xlearn">xlearn</a>)</p> <p>It basically handles category data, so I made discretize to bin for the numeric data such as age & fee. Other features (PhotoAmt etc) are simply used as categorical data.</p> <h3>Neural Network (XDeepFM)</h3> <p>At first, I used simple MLP model but the performance is very bad. During investigating <code>xlearn</code> approach, I found there are research which utilizes "FM" idea into neural network. <a href="https://arxiv.org/abs/1803.05170">xDeepFM</a> is one of the model which aims to apply many sparse categorical data. The model is also applied for <a href="https://www.kaggle.com/c/criteo-display-ad-challenge">criteo dataset</a> in the paper. I implemented & used this model.</p> <p>Main problem for this competition is that model easily overfits. Other than architecture, below Technics are used to reduce overfitting.</p> <ul> <li>permutation augmentation: to reduce overfitting, I permute some column's value between rows for data augmentation purpose.</li> <li><a href="https://openreview.net/forum?id=B1QRgziT-">spectral normalization</a></li> <li>dropout, weight decay</li> <li>weight tying: same weights are used among the depth for CIN network inside xDeepFM.</li> </ul> <p>I used most of the time for developing neural network, but sadly the performance was worst compared to the other models. However it is still effective to include for ensemble model, so it was not vain.</p> <p>I used <a href="https://github.com/pfnet/optuna">optuna</a> for hyper parameter tuning in the beginning-middle of the stage for neural network part. But I could not manage time to apply it for GBM & xlearn models, so score might improve little bit more.</p> <h3>Ensemble</h3> <p>Ensemble is performed by simply taking mean of each models. Each model is trained with 10 models (using 10-folded GroupKFold), so final prediction is made by 10 * 5 = 50 models.</p> <h2>Model performance summary</h2> <p>Below RMSE & QWK are the value for validation data of <code>train</code> dataset, calculated by 10 GroupKFold on RescuerID.</p> <p>\| Model \| RMSE \| QWK (after optimized by nelder-mead) \| \| --- \| --- \| --- \| \| XGBoost \| 1.041 \| 0.448 \| \| LightGBM \| 1.047 \| 0.439\| \| CatBoost \| 1.043 \| 0.447\| \| xlearn \| 1.055 \| 0.432 \| \| NN(XDeepFM) \| 1.065 \| 0.414 \| \| Ensembled \| 1.038 \| 0.457 \|</p> <h2>Final submission</h2> <p>How to determine threshold for regression model is quite important, and optimal way is not trivial in this competition (some <a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/83870">discussion</a> can also be found).</p> <p>I could consider following 4 options.</p> <p>[Determine threshold] 1.Determine threshold to align with train histogram. 2.Determine threshold to align with histogram which gets good score on public LB data.</p> <p>[Fix final prediction's histogram, instead of setting threshold.] 3.To align with train histogram. 4.To align with histogram which gets good score on public LB data.</p> <p>I submitted A: 3 (0.459 on public LB) and B: 2 (0.467 on public LB) as final submission.</p> <h3>[Updated] private leader board result</h3> <p>As I expected, threshold tuning for training histogram, method A, got high score = 0.44091 on private leader board which achieved me to reach gold medal in this competition. method B got score 0.43866.</p> <h2>Discussion</h2> <p>I have no idea how top kagglers got more than 0.470 scores and I am interested for those posts after the competition ends!</p> <p>One interesting approach which I could not try during competition is to predict "ranking" rather than actual value. Since threshold is adjusted in post-processing, proper ranking is quite important to get high score in this competition.</p> <p>Thanks for your post, it is interesting to know the background of competition :) </p> <p>Thanks for comment, the score on the table is calculated using validation data of <code>train</code> data. I could not submit each model separately.</p> <p>I can say that since "Single XGBoost model" public kernel already achieves 0.455 LB score with local validation 0.437, I could not get much improvement on LB even if RMSE & QWK score improves on local <code>train</code> validation.</p> <p>Calculating image embedding does not take so much time, about 6 min for all images (including train & test up to 10). The code is almost same with "Extract Image features from pretrained NN" kernel. But I skipped calculating embedding when image does not exist. When 0 image exist in pet zeros array is used as embedding. </p>
PetFinder.my Adoption Prediction	15th solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my Adoption Prediction <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>thank you very much</p> <p>thank you dofu.</p> <p>Very pitty to stand on the side of solo gold. Here is the solution. Thanks for opensource in kernels. It help me start in the last 10 days.</p> <p><a href="https://www.kaggle.com/baomengjiao/fork-of-v5-fork-v4-change-gpu-add-name-feature">https://www.kaggle.com/baomengjiao/fork-of-v5-fork-v4-change-gpu-add-name-feature</a></p>
PetFinder.my Adoption Prediction	1st Place Solution Summary	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my Adoption Prediction <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thanks! I am very happy to hear that.</p> <p>The answers are as follows:</p> <ol> <li><p>Dropping by adversarial validation was a pretty offensive strategy. According to the experiment in the 1st stage, CV & LB score had a certain fluctuation by the dropping rate. Therefore, we decided to adopt this method to only one model.</p></li> <li><p>Managing kernel source codes is difficult for us. We gave editing authority each other, and forked the other's kernel. In order to prevent confusion, we regularly create figure of kernel relationships and share them within the team.</p></li> </ol> <p>Thanks!</p> <p>Simply because I love cats more than dogs! (Of course I love both.)</p> <p>[edit: 2020-01-11] Updated title following <a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/125436">this discussion</a> [/edit]</p> <p>We got 2nd place in the private leaderboard! Very happy to be rewarded with lots of trials and errors. I'd like to thank everyone, and give special thanks for my great teammates <a href="https://www.kaggle.com/gegege">gege</a>, <a href="https://www.kaggle.com/takuok">takuoko</a>, <a href="https://www.kaggle.com/naka2ka">ynktk</a>, <a href="https://www.kaggle.com/kaerunantoka">kaerururu</a> !</p> <p>This post is a brief summary, and more detailed information will be updated or posted as a new topic by my teammates.</p> <ul> <li><a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/88812">2nd Place Solution about k_features and LGBM2</a></li> <li><a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/88963">g_features + XGB (part of 2nd place solution)</a></li> <li><a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/89042">2nd solution tyu part</a></li> </ul> <h1>Our Model Pipeline</h1> <p>Our model pipeline is as follows:</p> <p><img src="https://storage.googleapis.com/kaggle-forum-message-attachments/511979/12897/wodori-2nd-place.png" alt=""></p> <h2>Feature Engineering</h2> <p>We created 3 types of features.</p> <ul> <li>tyu_features: mainly by takuoko, ynktk and u++</li> <li>k_features: mainly by kaerururu</li> <li>g_features: mainly by gege</li> </ul> <h3>tyu_features</h3> <p>The size of tyu_features is over 5000, and we selected 1000 features based on lgbm feature importance. As a result of trying various thresholds, this threshold was determined by execution times and cv & public lb scores</p> <h3>k_features</h3> <p>k_features and LGBM2 is based on <a href="https://www.kaggle.com/wrosinski/baselinemodeling">wrosinski’s public kernel</a>. The size of k_features is not so big.</p> <p>You can see <a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/88812">2nd Place Solution about k_features and LGBM2</a> for more detail.</p> <h3>g_features</h3> <p>g_features and XGB is based on <a href="https://www.kaggle.com/reppy4620/xgboost">reppy4620’s public kernel</a>. What’s surprising for me is that 15% of features are dropped by adversarial validation. Adversarial validation tell us similarities of train datasets to test datasets. Dropping dissimilar train datasets gave us robustness and diversity, I suppose.</p> <p>You can see <a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/88963">g_features + XGB (part of 2nd place solution)</a> for more detail.</p> <h2>Validation Strategy</h2> <p>GroupKFold by Rescure_ID is used in the best score kernel. </p> <p>We first used GroupKFold, and from around mid-March we implemented and started using “StratifiedGroupKFold” like <a href="https://www.kaggle.com/jakubwasikowski/stratified-group-k-fold-cross-validation">jakubwasikowski’s public kernel</a>. </p> <p>As final submissions, one kernel uses GroupKFold and another uses StratifiedGroupKFold.</p> <h2>Models and Stacking</h2> <p>We used 4 models and stacking. Since this is a 2 stage competition, we didn't put much effort into hyperparameter tuning for boosting models. </p> <p>NN didn't work very well at first, but ynktk, takuoko and kaerururu tried a lot. Finally we can get NN model worth putting into the pipeline. You can see the information of NN <a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/89042">here</a>.</p> <ul> <li>NN</li> <li>LGBM1</li> <li>LGBM2</li> <li>XGB</li> </ul> <p>Of course, we tried the other models like catboost, xlearn, linier regression and so on. But we decided not to use them due to executions time and cv & public lb scores.</p> <h2>Threshold Optimizer</h2> <p>In this competition, we have to submit the discrete values (from 0 to 4). We used <a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/76107">OptimizedRounder given by hocop1</a>. </p> <p>For additional improvement, we use rank averaging:</p> <ol> <li>Convert predicted value of each model to rank</li> <li>Converting predicted values to rank even with stacking</li> <li>Determine the threshold by OptimizedRounder()</li> </ol> <p>The idea of rank averaging comes from <a href="https://www.kaggle.com/c/quora-insincere-questions-classification/discussion/80568">Quora 1st place solution</a>.</p> <h2>Post Processing</h2> <p>We found image duplications and overwrote the predictions as post processing. The idea is based on <a href="https://www.kaggle.com/appian/let-s-find-out-duplicate-images-with-imagehash">appian’s wonderful public kernel</a>. The original Kernel searchd only the first image in each pet, but ynktk extended the search range to the second. We tried to extend the range to the fifth, but execution time was not enough.</p> <p>ynktk contributed to the model pipeline comprehensively for reducing execution time and handling errors. There may be no success in the 2nd stage without his effort.</p> <h1>Conclusion</h1> <p>In this post, the brief summary of our model pipeline was described. The more detailed information will be updated or posted as a new topic by my great teammates. </p> <h1>Appendix</h1> <h2>Data Cleaning (by u++)</h2> <p>We created “typo dictionary” for the text data. I don’t know how this contribute to the scores, but there is a few typos in the text data as follows. I checked all the words that were out of vocabulary manually, and updated the dictionary.</p> <p><code> u"whatapp": u"whatapps", u"whatssapp": u"whatapps", u"whatssap": u"whatapps", u"whatspp": u"whatapps", u"whastapp": u"whatapps", u"whatsap": u"whatapps", u"whassap": u"whatapps", u"watapps": u"whatapps", </code></p> <p><code> u"sherpherd": u"shepherd", u"sherphed": u"shepherd", u"sherperd": u"shepherd", u"sherpard": u"shepherd", </code></p> <p>I was also in charge of preparing the external data, some operations to unify BreedName were necessary there.</p> <p>These are a bit of a hard works, so I went to the cat cafe to get healing (and domain knowledge). 😺</p> <p><img src="https://pbs.twimg.com/media/D12QmZaUgAA83Ga?format=jpg&name=4096x4096" alt=""></p> <p>Thank you, and congrats also to you!</p> <p>Your understanding is a bit wrong. We removed not features but the part of train dataset. </p> <ul> <li>Shape of g_feature is around (14993, 200)</li> <li>Shape of g'_feature is around (14993 * 0.85, 200)</li> </ul> <p>Which train dataset to remove is decided by the classifier as you said.</p> <p>Thank you and congrats to you! We thought that the Quora competition's solution would be helpful in that we need to determine the threshold.</p> <p>Maybe <a href="https://www.kaggle.com/takuok">takuoko</a> will post a NN solution!</p> <p>Your understanding seems correct. We don't know how it works in stage 2, but we can get about +0.002 improvement in public lb.</p> <p>Each of us selected features by watching the score of CV & public lb.</p> <p>When we stack, we only use 85% of train data.</p> <p><code> if T_flag and K_flag and G_flag: with timer('stacking'): X = np.concatenate([y_pred_t[extract_idx].reshape(-1, 1), y_pred_t_nn[extract_idx].reshape(-1, 1), y_pred_k[extract_idx].reshape(-1, 1), y_pred_g.reshape(-1, 1), ], axis=1) X_test = np.concatenate([y_test_t.reshape(-1, 1), y_test_t_nn.reshape(-1, 1), y_test_k.reshape(-1, 1), y_test_g.reshape(-1, 1), ], axis=1) </code></p> <p>Notebook has been published, and you can see the full code <a href="https://www.kaggle.com/naka2ka/stack-480-speedup-groupkfold-with-no-dict?scriptVersionId=12166931">here</a>.</p> <p>We utilized <a href="/hocop1">@hocop1</a> 's implementation, and don't know enough for the intention. <a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/76107">https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/76107</a></p>
PetFinder.my Adoption Prediction	1st place Solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my Adoption Prediction <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>gone but with $$$; the next action by the host will be interesting 🤔 🤔 > 1st Place - $ 10,000</p> <p>Letz hope they reveal their 'magic' soon !!!</p> <p>Am I the only one desperate for the 1st place(and other top holders) solution write up? </p>
PetFinder.my Adoption Prediction	5th Place Solution Summary	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my Adoption Prediction <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>You're welcome! At the moment I'm kind of busy, but I'm planning to take part in a competition in early 2020 (thinking February or March) if any interest me at that time.</p> <p>Thanks for letting me know! The image hosting service I used apparently doesn't exist anymore. I fixed it. Great to know people still find my solution useful :)</p> <p>I only used the first image except in my LGB model where I used mean, sum and variance of image metadata features over all images. This is one area of my solution which could probably be improved upon.</p> <p>Thanks! And yes, I had a good XGB model (around the same score as LGB on the public LB) which also had reasonably low correlation with LGB. On the public LB, it significantly decreased my score though.</p> <p>I submitted it in my CV-tuned submission together with some changes in the NN architecture but it turned out worse than the LB-tuned models there too.</p> <p>Thanks! We both experienced the same situation, leading for such a long time and then falling back a little in the end. Good luck on the private LB in the santander competition though!</p> <p>You can see the exact weights <a href="https://www.kaggle.com/bminixhofer/6th-place-solution-code">here</a>. I'd say NN and LGB were very important (my score with only LGB on the public LB was ~ 0.467). FFM helped but didn't contribute that much.</p> <p>Thanks! Also congratulations on Gold :) </p> <p><strong>Edit: <a href="https://www.kaggle.com/bminixhofer/6th-place-solution-code">The code is now public.</a></strong></p> <p>First of all, thanks to Petfinder.my for hosting this great, meaningful competition. And congratulations to the winners! I did not expect to do this well in it. I started because I wanted to investigate the mismatch between CV and LB, but quickly managed to come up with a good solution and become first on the public LB. So naturally, I decided to stick with it. "Defending" this position was a challenging task and it motivated me to work a lot on this competition. And although I did not manage to do that in the end it was a great experience and I learnt amazingly much.</p> <p>As a student, one challenge was the time I could spend on this competition. I tried to work on it for 1h - 2h everyday during the week, more on weekends, but that was not nearly enough to try everything I wanted to try.</p> <p>I'll get right to summarizing my solution. You can see the code <a href="https://www.kaggle.com/bminixhofer/6th-place-solution-code">here</a>. I'd recommend reading this summary first.</p> <h1>General</h1> <p>I have been stacking 5 models:</p> <ul> <li>3 NNs using PyTorch</li> <li>one LightGBM model</li> <li>one FFM with xlearn</li> </ul> <p>I used repeated stratified cross-validation with 5 splits and 5 repeats in my submission to reduce the effect of randomness, so each model was run 25 times overall.</p> <p>I used another strategy for evaluating changes in the models.</p> <h1>Validation Strategy</h1> <p>Like in <a href="https://www.kaggle.com/bminixhofer/a-validation-framework-impact-of-the-random-seed">my kernel about validation</a> in the recent quora competition, I used a kind of nested K-Fold CV. In an outer split, I split the dataset into a local test set and a train+valid set. The train+valid set is then also split into K-Folds and a model is trained on each fold. The predictions of each model on the data in the local test set are then averaged to get a QWK score.</p> <p>This is repeated for every outer split. I made a figure to illustrate this.</p> <p><img src="https://i.imgur.com/8NdTYuN.png" alt="Validation Strategy"></p> <p>The advantage of this strategy over others is that it takes into account the effect of averaging the test predictions of each fold. The same is done in the actual submission and because the correlation between folds has been low (especially of the NNs), the score would not be accurately estimated when using regular K-Fold CV.</p> <h1>Features</h1> <p>I used all categorical and numerical features available in the .csv files, the document sentiment magnitude and score, and the image metadata features commonly used in public kernels. Additionally, I:</p> <ul> <li>created count features of Breed1, Breed2 and State. When using them in the NNs, they were logarithmically scaled.</li> <li>used GDP per capita of the state.</li> <li>binned RescuerID count into 10 quantile bins. This greatly increased my score (almost 0.01 on the public LB) and intuitively makes sense because rather than the exact amount of pets rescued, the <em>type</em> of rescuer should matter (e. g. whether it is corporation or a single person).</li> </ul> <h1>Neural Networks</h1> <p>All my neural networks use the same basic structure. To make the neural networks as diverse as possible, I modified:</p> <ul> <li>The loss function</li> <li>The image activations</li> <li>How the network treats text</li> </ul> <p>You can see the shared structure between networks below. The black arrows represent connections in every network. The red, green and blue arrows represent connections only made in NN1, NN2 and NN3, respectively.</p> <p><img src="https://i.imgur.com/A8HHxXa.png" alt="Neural Network Overview"></p> <p>All network were trained using Cyclic LR and the Adam optimizer. I achieved the best results training for 10 epochs.</p> <h2>NN 1 - TfIdf matrix</h2> <p>My first and best performing network uses a kind of ridiculous structure. Text is encoded using 3-grams of all words with document frequency >= 2. That amounts to ~120k features. It is fed to the NN through a linear layer with 8 neurons, so there are 8 weights per 3-gram which amounts to ~1 million weights just in the text layer. To prevent this from miserably overfitting I used dropout of 80% and completely dropped the text features for 40% of all samples.</p> <p>I used MSE loss to train it. </p> <h2>NN 2 - Non-trainable 300d crawl embeddings</h2> <p>The second network uses a regular RNN structure: Text is fed to the NN via an embedding layer pretrained on crawl embeddings and non-trainable. It is then processed using a bidirectional LSTM. Max-pooled output and output of an attention layer is then concatenated with the other features.</p> <p>This network is also trained using MSE loss.</p> <h2>NN 3 - Trainable 200d GloVe embeddings + flipped image activations</h2> <p>The third network uses the exact same architecture as the second, but with 200d GloVe embeddings that I froze for the first 6 epochs and started training afterwards.</p> <p>Also, instead of using the regular image features I used image features extracted from the same densenet121 model on flipped images to increase diversity. I also tried multiple other augmentations (rotating, zooming, ...) but none of them worked. Flipping the images significantly increased my score though.</p> <p>This network was trained using SmoothL1Loss.</p> <h1>Other models</h1> <p>Additionally, I used LightGBM and xlearn. Text is encoded using SVD of a TfIdf matrix with 3-grams. The SVD has 10 components for xlearn and 120 components for LightGBM.</p> <p>I found that raw image activations were too high dimensional to work well, so I trained an image activation extractor NN model. This model has the top layer activations from Densenet121 as input and tries to predict the Type, Age and Breed1 of the pet. It is trained on the train + test set.</p> <p>This model has two hidden layers with 64 and 16 neurons. The activations of the first hidden layer are used as input for LightGBM, the activations of the second layer for xlearn. This greatly increased my score.</p> <p>For xlearn, it worked best to treat all features as categorical, so I binned the image extractor activations and the SVD representation into 10 quantile bins. You can see the structure of the input features for the two models below.</p> <p><img src="https://i.imgur.com/TsBN6D1.png" alt="Other models overview"></p> <h1>Stacking</h1> <p>At first, I used a linear regression to stack the models, but it did not work well. There were mainly 2 problems:</p> <ol> <li>Taking into account the increase when averaging predictions of each fold from the NN.</li> <li>Mismatch between train and test distribution.</li> </ol> <p>I solved the first problem by dividing the coefficients of the linear regression by the squared correlation between folds of the model. This somewhat accounts for the increase in score from averaging the predictions. I chose to use the squared correlation instead of just dividing by correlation because it slightly increased local CV score.</p> <p>The second problem was that neural networks were weighted significantly lower than expected. Of course, I didn't want to overfit to the public LB so adjusting the coefficients manually was definitely a bad idea. What i came up with is setting the sample weights of the linear regression to the out-of-fold predictions of an adversarial validation model. As such, the samples which were seen as closer to the test distribution were weighted higher and the coefficients of the neural networks increased strongly.</p> <p>When these two problems were solved, I arrived at a solid stacking strategy in which CV and LB score were closely related.</p> <p>To convert the output of the stacking model to classes, I just followed the train distribution of classes. I tried the public OptimizedRounder and various genetic algorithms but they performed on par with following the train distribution, so I just sticked with it.</p> <h1>Trouble with the Leaderboard</h1> <p>Like everyone else, I had problems with the mismatch between CV and LB. My validation strategy largely solved mismatches when just tuning hyperparameters of the models, but some significant discrepancies remained.</p> <h2>Binning Age</h2> <p>Binning the pet age into 20 quantile bins greatly increased my CV score and it intuitively makes sense to me because rather than looking at the exact age of a pet, adopters will typically divide pets into categories like young, middle aged or old. I guess there is little difference between chance for a pet that is e. g. 6 months old and another one that is 8 months old. However, this decreased my score on the public LB significantly.</p> <h2>Extreme sensivity to small changes</h2> <p>Another problem I faced was that the LB score was extremely sensitive to small changes in input data. Just changing the number of SVD components from 120 to 256 decreased my LB score by 0.007, although it slightly increased CV score.</p> <h2>Using ordinal regression</h2> <p>This competition is an ordinal regression problem. There are different classes but the classes have an ordinal relation. So why not treat is as such? Neural networks can directly handle ordinal regression by using 4 output neurons and encoding classes as:</p> <ul> <li>0: [0 0 0 0]</li> <li>1: [1 0 0 0]</li> <li>2: [1 1 0 0]</li> <li>3: [1 1 1 0]</li> <li>4: [1 1 1 1]</li> </ul> <p>More information on this kind of encoding <a href="https://stats.stackexchange.com/questions/140061/how-to-set-up-neural-network-to-output-ordinal-data">here</a>.</p> <p>The model can then be trained using binary cross entropy loss. The 4 outputs are summed to convert them back to a single scalar for each pet. This also greatly increased my CV score (by about 0.01) but decreased the public LB score.</p> <h2>A mistake increasing the LB score</h2> <p>For a long time, I used the following two lines of code in my kernel to create the gdp per capita feature: <code> train_df["gdp_vs_population"] = train_df["state_gdp"] / train_df["state_population"] test_df["gdp_vs_population"] = test_df["state_gdp"] / train_df["state_population"] </code></p> <p>When taking a closer look at this, you can see that it is wrong because the state GDP of pets in the test dataframe is divided by the state population of pets in the train dataframe. I didn't notice this until the last day in the competition. And I discovered that this wrong code gave me boost of an insane 0.007 on the leaderboard. So the LB score of my best correct submission was 0.485. This must have been just noise and it surely wouln't have held up in the private LB.</p> <p>But there is another strange thing about this. Why did the code even run? At first glance, I would think that it has to fail because the test dataframe and train dataframe have a different length so division can not be done correctly. It turns out that pandas automatically pads the series with NaN values to allow for a division to work. And that it then only uses the elements of this series up to the length of the test dataframe to make the assignment operation work. In my opinion, that is confusing behaviour and should probably be removed. </p> <p>Anyway, this is a testament for how noisy to public LB was, and how noisy the private LB probably also is because of the very little data that is available.</p> <p>Thanks. And I agree, I couldn't trust my CV that much after all. It is worth noting that my CV-tuned submission scored significantly worse than my LB-tuned submission on the private LB.</p> <p>Thanks! I'm working on making the kernel public. I had to run it again in a separate kernel because I chose my final solution from my main kernel and it is not the most recent version there, so it would be confusing to publish it. But it should be finished soon.</p> <p>I'm looking forward to your solution too if you'll publish it :)</p> <p>Thanks! And great question. One problem with my validation strategy is that in the case of 5 test splits and 5 "regular" splits I only use 80% * 80% = 64% of the data to train each model compared to the 80% of the data which the model in my submission is trained on in each fold. This might account for some mismatches.</p> <p>However, especially in this competition, there is a lot of unavoidable noise because we have very few samples to work with. </p> <p>In a future competition, when you find that CV and LB are not consistent, it is worth putting a lot of effort in making your CV as consistent as possible. Although my CV-tuned submission scored worse in the end, I only arrived at my "LB-tuned" submission because I was able to reduce the mismatches between CV and LB to a minimum and thus get a - in most cases - relatively accurate estimate of my LB score through local validation.</p> <p>It's important to note that when writing of "LB-tuned" I mean a submission which scored well in CV and LB, as opposed to CV-tuned where the submission scored <em>better</em> in CV but worse on LB. I did not just submit to the leaderboard and see what works.</p> <p>Thanks! The densenet model was not trained, I extracted the activations once for regular images and once for horizontally flipped images. It took about 20 minutes total. That might be a little unclear from the images. I'll update them.</p> <p>The rest of my models all take less than 30 seconds to train, except of LightGBM which takes about 1min 20s so time was not an issue.</p> <p>Edit: The images are now updated.</p> <p>Thanks! </p> <p>We average the <em>test</em> predictions, but not the <em>validation</em> predictions in normal K-Fold CV.</p> <p>When submitting to the Leaderboard, the averaged predictions on the test set of each fold are submitted. But when validating our model using regular K-Fold CV, the CV score is given by the mean of the validation scores of each fold. There is no averaging of predictions involved there.</p> <p>Yes. I explicitly specify categorical features when constructing the LightGBM dataset:</p> <pre><code>dtrain = lgb.Dataset(x_train_proc, y_train, feature_name=feature_names, categorical_feature=cat_features) </code></pre> <p>However, it even works well without that if the feature has low cardinality because LightGBM constructs bins of features internally anyway.</p>
PetFinder.my Adoption Prediction	7th Place Solution Summary	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my Adoption Prediction <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thanks all! After reading all the sols shared till now, I feel we have the most simplest one :(</p> <p>Good for me! Steep Learning lies ahead:)</p> <h3><a href="https://www.kaggle.com/adityaecdrid/best-sub-selected-re-run-478">The code Is public now</a></h3> <p>First of all, thanks to Petfinder.my and Kaggle for hosting this great and meaningful competition. And congratulations to the winners! And it's the first time for many team members to grab the Gold!! Thanks All :)</p> <p>>PS Pardon me if it isn't upto the mark, as it's my first time writing the solution desc..</p> <p><a href="https://www.kaggle.com/adityaecdrid/best-sub-selected-re-run-478">Kernel link</a></p> <p>The main issue with the PetFinder JSON Parsing script shared on public was that it wasn't accounting for many other json keys that were there and when you will add this, the script will break and this is what Shaz meant... (though the file having such contents were quite less, but we still added them because in case the test set had them in depth). Also we used all the files, not just the $-1$ tagged for everything (images, sentiments and metadata iirc)</p> <h2>FE's (Mainly, handcrafted)</h2> <p>Now coming to FE's, well we did a lot of work on that as well. - We tried cleaning the names col - Added Relative age feature <a href="https://www.kaggle.com/adityaecdrid/best-sub-selected-re-run-478#L362">line</a> - Also many interaction b/w Health, Sterelized etc cols as well :) - We also tried to mimic the way PetFinder.my adds the rating to their images(i created an account on their website to try to see it myself as well), we called it <a href="https://www.kaggle.com/adityaecdrid/best-sub-selected-re-run-478#L380">SEO Features</a> - After that we also ranked them accordingly as if someone visits the website, the pets are ranked somehow, so we just tried to mimic the same <a href="https://www.kaggle.com/adityaecdrid/best-sub-selected-re-run-478#L399">rankbyG</a> - Also we added a lot of aggregates as well on basis of rescuer id/state as well ( i guess this worked for the ID because since the rescuer id was auto-generated, it had the TS effect to it hidden as well because if you will select first 4-5 chars and sort the df, you will see the same, like earlier joinied users might save more etc kinda analogy) - Next we added <a href="https://www.kaggle.com/adityaecdrid/best-sub-selected-re-run-478#L556">nlp feats</a> as well which were borrowed from Quora's Shaz's work.. We tried POS tagging as well but didn't add it in the final solution as it was time consuming (but it helped improved the model) - You can see <a href="https://www.kaggle.com/adityaecdrid/best-sub-selected-re-run-478#L804">this</a> for many basic features we added as well, basically adding many groupby feats, states stats, cat/dogs breeds stats, ranking, insta features, year, top breeds, pets alllowed according to malaysian laws etc.. - Next we also had <a href="https://www.kaggle.com/adityaecdrid/best-sub-selected-re-run-478#L993">image_dims_feats</a> - We also added <a href="https://arxiv.org/abs/1709.05424">NIA</a> features as well (image quality, this ranked way higher for us) - Added features to indicate the dullness/brightness (refer Avito Comp, Peter's solution and great kernels their) - Also Lucas did a great job by verifying the CV with double CV strategy as well..</p> <ul> <li><p>Modelling includes lgb+xgb with different CV strategy and different STD as well for better ensembles.</p></li> <li><p>And there are ton of things which didn't work as well....</p></li> </ul> <p>(Will be updated with the CV used and other things as well by team mates as well)</p> <p>Thanks And Happy Kaggling:)</p> <p>> Would request team-mates( <a href="/shaz13">@shaz13</a>, <a href="/lucamassaron">@lucamassaron</a>, <a href="/init27">@init27</a>, <a href="/init927">@init927</a>, <a href="/backaggle">@backaggle</a>) (in no order) to drop whatever i missed</p> <p>Check this out mlcourse.ai </p>
PetFinder.my Adoption Prediction	99th Solution: Ensemble of 8 models	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my Adoption Prediction <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Congratulations all winners of this wonderful competition. This is the 2nd time silver medal for me. <a href="https://www.kaggle.com/bigswimatom/ensemble-8-model">https://www.kaggle.com/bigswimatom/ensemble-8-model</a></p> <p>I'm appreciate all kagglers who published following kernels. - <a href="https://www.kaggle.com/rookzeno/ensemble-other-people-kernel">https://www.kaggle.com/rookzeno/ensemble-other-people-kernel</a> @rookzeno - <a href="https://www.kaggle.com/ranjoranjan/single-xgboost-model">https://www.kaggle.com/ranjoranjan/single-xgboost-model</a> @ranjoranjan - <a href="https://www.kaggle.com/reppy4620/xgboost">https://www.kaggle.com/reppy4620/xgboost</a> @reppy4620 - <a href="https://www.kaggle.com/wakamezake/lightgbm-image-features-etc">https://www.kaggle.com/wakamezake/lightgbm-image-features-etc</a> @wakamezake</p>
PetFinder.my Adoption Prediction	[22nd place] From 325th in Public LB to 22nd in Private LB	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my Adoption Prediction <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Hi Kagglers, this is my first competition and I am excited to share our solution. Please provide feedback if you spot anything!</p> <p>We’d like to thank for Kaggle community and Petfinder.my for this learning experience. We learned a lot from those who share their knowledge and want to share our method with the hope of contributing to the community.</p> <p><a href="https://www.kaggle.com/enisimsar/22nd-place-solution">See our kernel here</a></p> <h1>Data Loading and Feature Extraction</h1> <h2>Metadata, Sentiment, Image and Their Aggregations</h2> <ul> <li>We took this part from <a href="https://www.kaggle.com/ranjoranjan/single-xgboost-model">Single XGBoost model</a> </li> <li>Addition to kernel above, we added min, max and std to its aggregations. Image number for each pet is also added. <h2>Breed Features</h2></li> </ul> <h3>Handling Mix Breed</h3> <p>We want to detect mixed breed instances and put specific(values other than 0-unknown- or 307-mixed-) breed information to <em>Breed1</em> column. Here is our strategy. 1. If any of <em>Breed1</em> or <em>Breed2</em> is mix breed(denoted by 307), we set mixed_breed column to 1. Then we change 307 value with 0. 2. For each mixed breed instance we set Breed1 as Breed1 + Breed2. Aim is to hold specific breed information in the <em>Breed1</em> column as much as possible. 3. If <em>Breed1</em> and <em>Breed2</em> columns are different than 0 for an instance, we set mixed_breed column to 1.</p> <h3>Breed Prediction</h3> <p>For some instances we do not have specific breed type(<em>Breed1 == 0</em> and mixed_breed=1). We’ve applied breed prediction only for dogs since we were only be able to find proper dataset for dogs. </p> <p><a href="https://github.com/sez-ai/Dog-Breed-Classifier">See our dog breed type prediction code here</a> <a href="https://www.kaggle.com/enisimsar/dog-identification-pretrained">Get pretrained model here</a></p> <p>We insert a column pred_breed, and fill that column with predicted values. If we do not predict breed since that instance already has, we simply copy <em>Breed1</em> value to pred_breed.</p> <h3>Creating pseudo-feature from breed names</h3> <p>Search breed names, if they contain any of the adjective below, add it as a categorical feature. [‘hair’, ’domestic’, ‘short’, ‘medium’, ‘retriever’, ‘terrier’, ‘tabby’, ‘long’]</p> <h2>Age</h2> <p>9 Bins are created for age column. See code for details.</p> <h2>Name</h2> <p>Following features are extracted from names. [‘Name_len’,’Name_isna’,”Name_with_numbers”]</p> <h2>Fee</h2> <p>5 Bins are created for fee column. See code for details.</p> <h2>State</h2> <p>We used state_gdp and state_population just like everyone else.</p> <h2>Word - Sentence Embeddings</h2> <p>We used Spacy’s <code>en_vectors_web_lg</code> and applied SVD with n_components 32.</p> <h2>Image Features</h2> <p>We were too lazy to use all images thus only used first photos of pets… Bottleneck image features are extracted from DenseNet121, Inception V3 and Xception with applying SVD for each one.</p> <h2>External Data for Breed</h2> <ul> <li>Scrapped from petfinder.com, note that it’s not petfinder.my. <a href="https://www.kaggle.com/enisimsar/petfindercomexternal">petfinder.com External Data</a> </li> <li><a href="https://www.kaggle.com/hocop1/cat-and-dog-breeds-parameters">Cat and dog breeds parameters</a></li> <li><a href="https://www.kaggle.com/rturley/pet-breed-characteristics">Pet Breed Characteristics</a> <h2>Rescuer ID</h2></li> </ul> <p>There were many discussion about Rescuer ID, we also observed it causes overfitting. Thus one of our kernel was playing with rescuer ID while other was dropping this feature. Well, sometimes overfitting wins…</p> <p><a href="https://www.kaggle.com/enisimsar/final-submission-with-rescuer-features#Rescuer-ID">See code for details</a></p> <h2>Aggregations</h2> <p>To enrich features even more, we have extracted many aggregations. <a href="https://www.kaggle.com/enisimsar/final-submission-with-rescuer-features#Aggregations">See code for details</a></p> <h2>Freq Encoder Categorical Features</h2> <p>Thanks to <a href="https://www.kaggle.com/c/avito-demand-prediction/discussion/55521">Sample categorical feature encoding methods</a></p> <h1>Model</h1> <p>We used 2 models, XGBoost and LightGBM. Golden Section Search method is used for OptimizedRounder. Thanks to @hocop1 at <a href="https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/76107">How to use regression here? \| Kaggle</a>. We’ve called OptimizedRounder in every fold and stored coefficients for each model. Then took mean of that coefficients for final use. </p> <p>As for final submission we’ve simply averaged coefficients and regression output of LGBM and XGBoost. </p>
PetFinder.my Adoption Prediction	[72nd place solution] Stacking Madness (only CPU)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my Adoption Prediction <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>I used DenseNet121 only for XXX-1.jpg. Then it takes only 4500 sec for DenseNet121.</p> <p>As it takes time to translate the text about my model into English, I will only share the diagrams for now.</p> <p>I wrote 2 kernels. Please check and upvote them. 1. <a href="https://www.kaggle.com/nadare/72nd-place-key-points">essence of my solution</a> 2. <a href="https://www.kaggle.com/nadare/xlearn-model-cv-42-lb">xlearn example for PetFinder</a></p> <p>CPU: 17701.8 sec <img src="https://pbs.twimg.com/media/D30E-vzV4AAxdyP.png" alt="72nd place solution"></p>
PetFinder.my Adoption Prediction	part of 9th place (denoising auto-encoder NN)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: PetFinder.my Adoption Prediction <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>First I want to thank petfinder.my for hosting such a challenging competitions. Also thanks to my wonderful teammates @peterhurford @rdizzl3 . Our solution (which we will share in more detail in the coming days) consists mainly of 4 NNs and several XGBs. In our main summary we will also address feature engineering, cv-scheme, post processing, and more details on pretrained image NNs. Here I want to give a first glimpse how the general architecture of our NNs look like. They are all different and quite low correlated but follow the same idea: using denoising auto-encoders for automatic feature engineering as used by @mjahrer in the Porto Seguro’s Safe Driver Prediction and explained <a href="https://www.kaggle.com/c/porto-seguro-safe-driver-prediction/discussion/44629#250927">here</a> Let me show an illustration of one of our NNs for illustration:</p> <p><img src="https://i.imgur.com/gWDC532.jpg" alt=""></p> <p>The training consists of two stages. </p> <p>First train an auto-encoder on train+test to predict the numerical and categorical variables also using text + image. Then in a second stage use the encoder part of that auto encoder to encode your variable and add a classification head. To be more specific:</p> <ol> <li><p>Stage: Training the auto encoder Use tfidf for feature extraction of text data(sentiment.json + label.json + description text), one hot encode of categorical and a pretrained image model (we used different ones) to get image features. Concatenate everything in one vector and use several fully connected layers to train the reconstruction of the numerical and categorical features. In that way the auto encode will learn the interactions between features which I think was the most important task in this competition. A nice side effect is, since the auto encoder works unsupervised, you can also leverage from test data. </p></li> <li><p>Stage Training the classifier Throw away the reconstruction head of the auto encoder you end up with a NN that is able to interpret the given features and their interaction. This now can be used to perform the classification task. One thing (besides many others) that I learned from @mjahrer post was to concatenate the main part of the auto-encoder to leverage even more from its encoding. I also realized since I only trained the auto encoder to reconstruct categorical and numerical features, it might help to again feed the image features to the classification head.</p></li> </ol> <p>I summary, the main advantage (at least for myself) comparing to other NN architectures, was that an auto encoder was easier to tune to to handle overfitting. If you for example dropout some of the input features for the auto encoder it can practically never overfit, and as a side effect it even learns to reconstruct variables from others. E.g. reconstruct type from image. As mentioned above I will tell more details n our overall summary, but I already wanted to give you a glimpse of what we did.</p> <p>Yes, I used swap noise. That worked best here.</p> <p>we ignored multiple images</p> <p>everything is shared apart from the classification head. The filled yellow layers are re-ordered (concatenated) but in the classification task hasve still the same weights they learned in autoencoder training</p> <p>you are funny</p>
Planet: Understanding the Amazon from Space	0.93017 solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Planet: Understanding the Amazon from Space <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>I've just pushed pytorch code <a href="https://github.com/EdwardTyantov/pytorch-kaggle-amazon-space">https://github.com/EdwardTyantov/pytorch-kaggle-amazon-space</a> Nothing special inside, but code is very clean & structured, so it can help someone.</p> <p><strong>Results</strong>:</p> <ul> <li>Best single model: mixnet_v6: public: 0.92905, private: 0.93071</li> <li>Best blended ensemble: public: 0.93015, private: 0.93217</li> <li>Best submit during competition: 0.93023: 0.93168 (also ensemble)</li> </ul> <p><strong>Quick overview of the solution</strong>:</p> <p>Best single model (mixnetv6) solution consist of following tricks:</p> <ul> <li>6-channel input (3-jpg channel, NIR-chanell, NDWI-index, SAVI-index)</li> <li>model: resnet18 on jpg, resnet 18 on nir+indexes, concat -> 256 embedding FC + final FC <ul><li>jpg branch lr modifier 0.05 to base LR, for nir branch - layer{3,4} - 1.0, layer{2,3} - 0.1, FC - 1.0 </li></ul></li> <li>plateau scheduller on val loss (cross-entropy), patience=3</li> <li>early stopping: 6 epochs</li> <li>train time augmentation: shift, flip, scale, rotate, transpose</li> <li>test time augmentation, 6x: as-is, rotate 90{1,2,3}, flip x, flip y</li> <li>standart for this challenge searching thresholds for F2 (firstly I implement per class search - it is more consistent, but default on scale is better)</li> </ul> <p>Best Ensemble:</p> <ul> <li>trained various models on jpg, mix channels <ul><li>models: <ul><li>densenet{121,169} on jpg, 5 folds</li> <li>mixnetv6, mixnetv3 (different LRs) 5,6,7 folds</li> <li>wideresnet on 6-channel (for mix branch unpretrained WideResNet), 7 folds</li> <li>resnet18 + embedding FC on 8 folds</li></ul></li></ul></li> <li>best submit was based on weighting predictions using holdout F2 score, weight=((score - min_score)/max_score)*0.5</li> </ul> <p>Slides about challenge and my solution <a href="https://www.slideshare.net/Eduardyantov/kaggle-reviewplanet-understanding-the-amazon-from-space">https://www.slideshare.net/Eduardyantov/kaggle-reviewplanet-understanding-the-amazon-from-space</a></p>
Planet: Understanding the Amazon from Space	3rd place Solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Planet: Understanding the Amazon from Space <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>There was some modification of bisection method which was performed on validation data. Obtained thresholds then used for test data. </p> <p>On first step we search for single threshold for all classes. On the second step search optimal threshold independently for each class, while fix all other thresholds. Step 2 can be repeat several times since it keep increasing accuracy.</p> <p>There are many syntax and parameters changes in Keras 2.0. Not all the networks worked fine in 2.0 version while I solve Amazon problem. But I plan to move on Keras 2.0 in next projects, since all CNNs now fixed to work with 2.0.</p> <h1>Solution summary</h1> <p><strong>Software</strong></p> <ul> <li>Windows 10, Python 3.4 + 3.5</li> <li>Keras 1.2.1 - main framework for CNN training</li> <li>Theano 0.9.2 - main backend</li> <li>Tensorflow - additional backend for neural nets which is not supported by Theano (ResNet152 and XCeption)</li> <li>XGboost, LightGBM, Keras - main classificators for final ensemble</li> <li>OpenCV - image preprocessing</li> </ul> <p><strong>Hardware</strong></p> <ul> <li>We used 6 GPUs: 21080Ti + 41080. </li> <li>Parallelization was made using either different fold on different GPUs, or different CNNs on different GPUs.</li> </ul> <p><strong>Training Time</strong></p> <ul> <li>On single GPU around a month for full pipeline. On 6 GPUs it’s possible to finish in around 1 week.</li> </ul> <h1>Dataflow</h1> <p><img src="https://storage.googleapis.com/kaggle-forum-message-attachments/217773/7225/Dataflow.png" alt="Dataflow" title=""></p> <h1>Set of CNN based models</h1> <p>Due to problem with TIFFs all CNN models used only JPEG images</p> <ul> <li>11 models for different CNNs with 17 output neurons with sigmoid activation. Set of CNNs: INCEPTION_V3, INCEPTION_V4, DENSENET_121, DENSENET_169, DENSENET_161, RESNET50, RESNET101, VGG16, VGG19, RESNET152, XCEPTION</li> <li>1 DENSENET_121 model with weather classes only: ['clear', 'partly_cloudy', 'haze', 'cloudy']. “Softmax” activation because classes is mutually exclusive - 4 output neurons.</li> <li>1 DENSENET_121 model with land classes only - 13 output neurons.</li> <li>17 DENSENET_121 models for single classe. Each model has 1 output neuron to predict single class. 50/50 batches, where half of images have class presented.</li> </ul> <p>We used 5 KFold Cross Validation. So each model has 5 different weights set. 30 CNN models in total. 150 weight files</p> <h1>Training Process and Data Augmentation</h1> <ul> <li>Keras module for Python with Theano or Tensorflow backend.</li> <li>Typical batch size around 20 images for small nets, 16 images for large nets due to GPU memory limitation</li> <li>Batches created on the fly, so no need to store all the data in memory</li> <li>As optimizer we used Adam, with learning rate ~0.00003</li> <li>We used “logloss” loss function, since direct usage of F2-score in training process gave worse results.</li> <li>Single class models (for rare classes) used 50/50 batches, where half of images have class presented</li> </ul> <p><strong>Data augmentation includes</strong></p> <ul> <li>Random crops: chose some part of picture and resize it to CNN input shape (224x224 or 299x299)</li> <li>Random mirrors or 90 degrees rotations (8 possibilities in total)</li> <li>Random intensity changes</li> </ul> <h1>Validation and Test Time Augmentation (TTA)</h1> <ul> <li>We used 5 KFold, so to process Train images for validation we process it one time with corresponding fold model. In total 40K images</li> <li>To process test images each image must be processed with each fold model. 5 times in total, so for test data we process 300K images.</li> <li>To increase the accuracy we used TTA. This mean we process each image several times with some changes. We used 32 variation of single image, predict each and use mean as final prediction: -- For Train images 1.2M predictions -- For Test images 9.6M predictions</li> </ul> <p>After this we got 2 text CSV files with probabilities for each image to have predicted class. One CSV file for train images, second for test images. These files used later in second level classifiers.</p> <h1>Neighbors</h1> <p>More details in this thread: <a href="https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/36738">https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/36738</a></p> <p>Based on neighbours IDs, panorama IDs and CNN predictions we created the following set of features:</p> <ul> <li>Average CNN predictions for 4 neighbours</li> <li>Average CNN predictions for 8 neighbours (including diagonal elements)</li> <li>Average CNN predictions for whole Panorama ID</li> </ul> <p>Each image got “panorama ID” and “panorama Size“ as features for second level models or “N/A” depending on following rules:</p> <ul> <li>panorama must has more than one element</li> <li>panorama must contain at least one element from train and at least one element from test parts of data</li> </ul> <h1>2nd level models: XGBoost blender</h1> <p>We have large amount of CNN predictions and features. Now we need to some methods to join them altogether for more accurate predictions. One of methods is to use many random XGboost runs with random parameters.</p> <p>XGboost can predict only one class at the time. So each run create 17 different XGBoost models, each predict its own class. On the input we had ~300 features. What we variate from run to run:</p> <ul> <li>Folds number [4 - 10]</li> <li>Learning rate [0.06 - 0.45]</li> <li>Max Depth [2 - 5]</li> <li>Subsample [0.6 - 0.99]</li> <li>Colsample by tree [0.6 - 0.99]</li> </ul> <h1>2nd level models: Keras blender</h1> <p>Keras neural net classification models are good for following reasons:</p> <ul> <li>They’re the very different from XGboost models</li> <li>They can predict all 17 classes at once, catching the mutual influence of classes The bad thing is that it in most cases has slightly worse accuracy than XGboost or LightGBM.</li> </ul> <p><img src="https://storage.googleapis.com/kaggle-forum-message-attachments/217773/7230/Keras-blender.png" alt="Keras blender" title=""></p> <p>Variation parameters:</p> <ul> <li>Number of neurons on each level: 1st: 400 - 700, 2nd: 350 - 600, 3rd: 200 - 500</li> <li>Dropout value</li> <li>Random choice of activation: ’RELU’, ‘ELU’, ‘PRELU’</li> <li>Batch size: 200-1000</li> <li>Learning rate: 1e-5 - 1e-3</li> <li>Early stopping: 50 - 150</li> <li>Number of folds: 4 - 10</li> </ul> <h1>Final Ensemble</h1> <p><img src="https://storage.googleapis.com/kaggle-forum-message-attachments/217773/7229/Ensemble.png" alt="Ensemble" title=""></p> <p><img src="https://storage.googleapis.com/kaggle-forum-message-attachments/217773/7228/Score1.png" alt="Score" title=""></p> <p>It's possible to get 2nd place with this solution:</p> <p><img src="https://storage.googleapis.com/kaggle-forum-message-attachments/217773/7232/2nd-place.png" alt="2nd place" title=""></p> <h1>Code</h1> <p><strong><a href="https://github.com/ZFTurbo/Kaggle-Planet-Understanding-the-Amazon-from-Space">Github repo</a></strong></p> <p>Trial and error. Also based on previous experience. I think LR can be tuned further for faster performance with same accuracy.</p>
Planet: Understanding the Amazon from Space	9th place solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Planet: Understanding the Amazon from Space <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Oh you could definitly have achieved similiar results with keras+tf or any other neural networks library :)</p> <p>I optimised the threshold on the validation. :/</p> <p>I used the same binary-crossentropy for both.</p> <p>Good explanation! I will definitely try more sophisticated research of semisupervised learning the next time, but in this competition to be honest I was just happy I could increase my score and I didn't think much further than that :)</p> <p>I have no experience with "classic" classifiers like SVMs or random forests, but still wanted the best possible classifier. XGBoost seems to win most of the competitions on Kaggle, so I just chose it for that reason.</p> <p>For me it was hard to find a good α(t) so I just fixed it to 1. I used 25% of the test data for semi-supervised learning. Instead of using α(t) used the running average and only started to adapt the labels after a few epochs. Keep in mind this system is a not a result of structured research. I just tried things and if they worked I build upon these. I think there are better ways to implement Pseudo-Labeling!</p> <p>Yes. 20%. No stratification.</p> <p>You are right. I will upload the code soon and then you can just look it up there.</p> <p>I like the low-level approach. In the end Pytorch is a tensor library like tensorflow. I never used such libraries, only keras. So I appreciated how easy it is to do custom implementations :)</p> <p>Code here: <a href="https://github.com/Mctigger/KagglePlanetPytorch">https://github.com/Mctigger/KagglePlanetPytorch</a> This code is not meant to reproduce the 9th place submissions, but to give you a basic idea of our approach.</p> <h1>CCT: 9th place solution</h1> <p>This was my first serious Kaggle competition and it has been an amazing experience. I went from merely understanding basic convolutional networks to understanding them a little better (in a practical way at least).</p> <h2>1. First attempts</h2> <p>In the beginning I tried custom convolutional neural networks. I got inspiration from VGG-like networks and just tried different architectures to get a hang of architectures and their performance. Since I started with "only" a GTX1060 my archiectures where limited to small to medium sized models. When I saw Heng CherKengs repository I tried custom implementations of DenseNets and ResNets in Keras, which has been a great learning experience, but did not yield results better than 0.927 on the public leaderboard.</p> <h2>2. Getting serious</h2> <p>Heng CherKeng convinced me to use Pytorch instead of Keras and I instantly learned to love it. Also a big advantage is the availability of many pretrained models, which have been key to improving my score. Although I was lacking any structure approach to this competition at this time I ranked quite good on the leaderboard. For my score of ~0.9315 I used an ensemble of the following models: - ResNets: 34, 50, 101, 152 - DenseNets: 121, 163, 169, 201</p> <p>I used SGD with momentum and weight-decay. The learning-rate scheduling was manually determined on based on the local validation loss and the leaderboard f2-score. I used data-augmentation for training which included arbirtrary rotations, flipping, translation and zooming. I also tried blurring and distortion which did not change much. Also I did not see a difference between 90° rotations and arbitrary rotations. To continue training models, which had been pretrained on Imagenet, I just removed the last layer and replaced it by a simple fully connected layer with sigmoids. I did not use combinations like sigmoids + softmax (for weather) since I expected misslabeling of the data in the test set and also for f2-score I would rather risk some false positives than false negatives. I used block-wise learning rates: - For DenseNets I used a constant learning rate for all layers, but trained the classifier alone for a few epochs in the beginning. - For ResNets I locked all froze all weights and unfroze them every 2 epochs until the classifier, 4 convolutional blocks and the low-level filter weights were unfrozen. From the on I continued to train on all layers.</p> <p>I trained exclusivly on JPG-files. The persons who labeled would have also only seen the JPG-files. To get hard labels for our submissions we used the f2-score optimiser from the kernels. Also my 8 models were trained on the full dataset and I used many submissions to find out the best epoch of each model. This obviously mean I probably overfit the public leaderboard pretty hard at that point in time.</p> <h2>3. Teaming up</h2> <p>Finally approximatly one month before the competition end I teamed up with Cpruce (<a href="https://www.kaggle.com/cpruce">https://www.kaggle.com/cpruce</a>) and Chris Chross (<a href="https://www.kaggle.com/crailtap">https://www.kaggle.com/crailtap</a>). While Cpruce had the same leaderboard score as me and the same approach, ChrisChross used 5-fold cross validtion for his models and replaced the final fully connected layers of his ResNets by 3 fully connected layers with dropout! This led to a leaderboard score of ~0.93177 for an averaged-ensemble of 5 folds. His overall leaderboard score was ~0.9325 and by combining our models with weighted voting we got to ~0.9330.</p> <h2>4. A structured approach</h2> <h3>Cross-validation</h3> <p>Now our final approach. So far I had never cross validated my models, since I thought this would take to much time, but thankfully I got an GTX1080TI and 5-fold-cv did not seem to time consuming anymore. First I just tried 5-fold-cv on all my existing models, averaged the results and added them to the ensemble which did not increase our score substantially. The best 5-fold-cv ensembles scored ~0.93050-0.93100.</p> <h3>Semisupervised learning</h3> <p>Based on Sander Dielemans competition reports (<a href="http://benanne.github.io/2015/03/17/plankton.html">http://benanne.github.io/2015/03/17/plankton.html</a>) I tried pseudo-labeling to fight overfitting of my models. I added a part of the test-data to the training-procedure and used the (soft-) outputs of the network as labels. I initialized the labels for the test-data with our best leaderboard submission and used a running average of the network outputs and the initialization as label. This led to a minor increase on model performance for DenseNets, but ResNet-50 and ResNet-101 improved a lot.</p> <h3>XGBoosted TTA</h3> <p>This has probably been the most important part of our final approach. We used 8-14 different transformations for test-time-augmentation (TTA). 14 had the best performance, but in the end there was not enough time to rerun these for all existing models. First we combined our TTAs by averaging, which was already quite good. When I read about model-stacking I got the idea to train a classifier to combine TTAs the most efficient way. Also since we already used 5-fold cross validation it was trivial to implement stacking. As classifier we used XGBoost even though we had never worked with it before. Now we got 0.93200 for ResNet-50 and ResNet-101 and big jumps for the worse models (for example ResNet-34 from ~0.9307 to ~0.9312. We combined the new models with weighted voting (based on leaderboard-f2) and removed most of the old models from out ensemble. This gave us our final public leaderboard score of 0.93340 and private leaderboard score of 0.93220</p> XGBoosted Ensembling <p>We tried the same procedure we did for TTA for ensembling all TTAs of all models, but we got very bad results (0.900). Either an implementation error or a systematic error, we do not know yet.</p> <p>We will upload our code soon!</p> <p>Hi, what do you mean by unsupervised features? SOTA is just the currently best submission we had on the public leaderboard at that point. We used it as labels for the test data. :) Btw: I did not try normal pseudo-labeling so I do not even know whether this strategy increases performance!</p> <p><a href="https://github.com/Mctigger/KagglePlanetPytorch">https://github.com/Mctigger/KagglePlanetPytorch</a></p>
Planet: Understanding the Amazon from Space	Code from a record breaking 903 place drop	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Planet: Understanding the Amazon from Space <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>FYI, the best single model from that codebase was a fbresnet-200 that hit 0.93143 on the public LB and 0.92915 on the private with 8x flip/transpose test-time augmentation enabled. My teammate had a few models that bested that score on the private LB almost cracking .93 with ResNet-50. </p> <p>So, instead of rounding out the bottom of the top 15 with a private LB score of 0.93202 and public 0.93386, we ended up in the bottom 50 after a 903 place drop thanks to an undetected issue jumbling filenames of just private LB in one of the scripts we were using to process and ensemble our submissions. Damn!</p> <p><a href="https://github.com/rwightman/pytorch-planet-amazon">https://github.com/rwightman/pytorch-planet-amazon</a></p> <p>Anyways, I wrote the code I used for this competition to be re-usable in the future. Nothing super fancy here, just a good collection of models (ResNets, DenseNets, fbresnet200, Inception-V4, Inception-Resnet-V2, Wide-Resnet 50, ResNext) pulled together from a variety of sources with a lot of different training and tweaking options, a convenient cross validation setup for training, and some test-time augmentation.</p> <p>My teammate also used Pytorch for his codebase so we had some model overlap but ended up favouring different hyper parameters and models so our ensembles worked well together. Most of our models were trained at 256x256 resolution with quite a bit of augmentation but we also had several at 224 and 299. </p> <p>F2 threshold optimization was used (thanks anokas!) and additional manual thresholding was done at ensembling time. We definitely overfit a bit to the leaderboard for the 'best' submissions as many did. Several earlier submissions that I unjumbled and re-verified since competition close performed better on the private than our best public scores. </p> <p>Note: the fateful CSV handling code is not in the above repo, sorry ;) </p> <p>Thanks! Hehe, you're right, I could still get a medal. Anyone who likes the code, give this a bump :) </p> <p>It was a complete fail... I saw your comments in their github issue tracker. It was very unstable, and within a few epochs the loss started skyrocketing. I tried clipping gradients and fiddling with the parameters but got nowhere so I abandoned that idea pretty quickly.</p> <p>This makes for a more amusing story :) </p> <p>Yes, that was the goal, and a very inefficient/brute force way of trying to achieve it. The .2 was a fudge factor to give a wider margin of acceptance so it had a hope of succeeding in a reasonable time. For the most part I used a 5-fold and 10-fold set, I was going to experiment with a more granular 12 fold and use two of the 12 for different tiers of CV but never got around to running that is it is. </p>
Planet: Understanding the Amazon from Space	My brief overview of my solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Planet: Understanding the Amazon from Space <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Congratulations to Team plant and Team Russian Bears.<br> Special thanks to @Heng CherKeng for your share on forum,your codes and reports are very helpful,and the competition would not be so interesting without your generous contributions. <br>Here is the brief overview of my solution.<br> </p><h2>Challenges</h2><br> 1.Submissions are evaluated based on F2 score,so we must search thresholds after predictions and we can not find perfect thresholds.<br> 2.Label noise.We can only control the side effect of the noise.<br> 3.The public score are very close,and the division of the test data can cause shakeup..<br> <br> <h2>Solution</h2><br> My solution are similar to Heng’s,so I will only introduce the differences<br> 1.I write a F2 loss function in pytorch,(thanks to @Bruno G. do Amaral,<a href="https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/34484">https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/34484</a>) and add it to logloss,and I found it was helpful,but I did not check how much improvement we can get because I only compared them by using resnet 34 with 15 epochs and did not predict and submit.<br><br> 2.I used ＂Single Image Haze Removal using Dark Channel Prior＂ from <a href="http://kaiminghe.com/">http://kaiminghe.com/</a> ,I found that it was very useful,my networks can '<b>see</b>' the images more clearly.I hesitated to use it on weather labels,but I found that it’s quite good on all kinds of labels,especially road/water/habitation... <br><br> 3.I trained a hard example mining network on simple-net (based on heng’s code ) ,I selected 1/3 examples with largest loss and do back-propagation,I also trained other hard example methods but they did not help a lot ,so I did not use it.<br><br> 4.I trained different networks with 6464 224224 256256 inputs and I used dilation of the in a resnet 34 network.Different networks have different capabilities on different labels.For example,SimpleNet 64 have good performance on Label:clear.<br><br> 5.Because I have networks with difference capabilities,so I do Ridge regression on them to predict each label separately,I mean,I have 17 regression models.<br><br> 6.My final submission are based on some strong models(based on CV),My best submission are voted based 9 models.<br><br> 7.I selected tiff images with correct labels (20000+ in train set),although they are more clear but I found they have little value ,I guess my dark channel pre-processing were strong enough. And I was afraid of the differences of distribution of train and test set will be harmful to F2 thresholds,so I did not use them.<br><br> 8.I also found I can get big images like ZFturbo did(<a href="https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/36738#latest-205616">https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/36738#latest-205616</a>) but I did not use it because I thought it will be of little value because convolutional network models are strong enough.<br><br> 9.I tried bayesian inference but I found it was not helpful.<br><br> 10.I compared the perfect F2 thresholds with thresholds searched,this can lead to 0.001 difference .What is perfect F2 thresholds?If the distribution of test data is exactly same as train set ,then the thresholds are perfect.I tried a lot of ways(such as normalization the prediction,controlling ratio of the labels of the test set)But I found that I could not get good result,and it seemed to be very difficult to reach perfect thresholds for all of us, so I just let it be.<br><br><p></p> <p></p><h2>Shake up</h2><br> I tried my best to evaluate the possible shake up in final stage of this competition.<br> I simulated the test set use half of train set,and I split the simulated test set to 66:34 using different random seeds,I found as the seed change ,the value of public-private score will change 0.001 to 0.0025.<br> <img src="https://raw.githubusercontent.com/bestfitting/kaggle/master/amazon_demos/gap.jpg" alt="gap" title=""><br> the x is seed index,y is simulated public-private (I sort the seed by the value:public-private) <img src="https://raw.githubusercontent.com/bestfitting/kaggle/master/amazon_demos/dist2.png" alt="dist" title=""> the distributions of simulated public-private value <br><br> As we can see,the shake up happened on real Leardboard. <br> I thought carefully,I checked public LB,I must persuade myself that the public scores were not perfect metric of the capabilities of models,it was difficult,because at first glance,we have 40000+ images in public test set,it should be stable.<br><br> But<br> 1.The gap between us were very small(0.0005-0.001)<br> 2.I thought the public-private score changed along with seed is caused by Hard Examples,and these Hard Examples can only be random guess,as they are caused by label noise.<br><br> Finally,I persuaded myself by a guess:If we label all the images ourselves 3 times,some of then images will have different labels,because we can not say for sure some images are haze or cloudy,road or water,slash burn or not,blooming or not.....so the noise are random noises.<br> And If the noises are random,the score get from public LB will loss in private LB. <br> So, I adjusted my goal to keep myself in TOP 10,and decided not to care about public LB in last week and tried to seek most stable way to ensemble my models,I threw away any models may cause over-fitting,and I used just vote and ridge regression.And the final result is a big surprise to me.<br> <br><br> By the way,I entered this competition partly because I used Google satellite map for years when I preparing my Mountain bike trip,it is very useful when we cross mountains.I once thought I can benefit from my capabilities of distinguish water/road,habitation or not,but I found this kind of skills can only be used in my trip.:)<br> Thanks to kaggle and organizer !<p></p> <p>We have only probablities of each class when train and we can not use thresholds on them,so we can not use the F2 function directly,and most important reason is we must make it differentiable and back-propagate the loss .</p> <p>Thanks,waiting for your solution,I think it will be very good one.</p> <p>HI,Alex,<br> I have no time to clean my codes in next weeks,and the good news is heng will collect all solution and re-implement,at the moment, you can try the schedule like this:</p> <pre> def schedule(self,net,epoch,all_data=False): self._sgd = optim.SGD(net.parameters(), lr=0, momentum=0.9, weight_decay=0.0005) self._cur_optimizer = self._sgd lr = 0.005 if epoch > 5: lr = 0.001 if epoch > 10: lr = 0.0001 if epoch > 15: lr = -1 self._cur_lr = lr return self._cur_optimizer,self._cur_lr </pre> <p>If you want your model to be stable,you can use n-folds training.n can be 4,5,or even 10.<br></p> <p>Hi,Jeremy<br> Happy to relpy to your question!<br> I used and only used simplenet version hard example in final ensemble.I tried to use hard example on other net,my local CV decreased,so I removed them.My ensemble strategy is very simple,just let Ridge model to select correct models to predict a label.<br> For dilation,I changed the restnet34 model: self.layer4 = self._make_layer(block, 512, layers[3], stride=1,dilation=2)<br> As I use pytorch, the dilation is a parameter of conv2d,so it's very easy.<br></p> <pre> if dilation>1: self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1,dilation=dilation, padding=2, bias=False) else: self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) </pre> <p><br> And as to end to end solution,to be frank,I am not very statisfied with my solution in this competition.I prepared to find a end to end one at the early stage but I entered NOAA sealion competition at the same time (I managed to find a end to end way in that competition)and when I was back ,there were less 3 weeks left only,so I can only train a lot of models and ensemble them :) <br> I think we can find a good network structure with high performace to sovle this problem.<br> If we use resnet,we can use 256256 input,we split the network from conv4 and split labels to some group,the weather with large stride ,the blooming ,road,water with dilation and add some hard-example loss on them. I hope I have time to have a try but I am exhausted now after these months.</p> <p>Hi,Mark<br> I did not submit any single model based submission after very early stage of this competition.<br> But I check my local validation of my models,I think the CV F2 scores are good.<br> densenet161_224 0.933014<br> resnet_50_256 0.932966<br> resnet_34_256 0.932912<br> resnet_dilation 0.932307<br> My ensemble model's validation F2 score are 0.9344x,and the gap to public LB score are 0.0005 to 0.0008,So I think the dataset of this competition is balanced among train set,public LB test set and private LB test set.We can trust our local CV.</p> <p>HI,Gene,<br>If we have 3 models such us resnet50,densnet121,simplenet,they have 17 predictions (probabilities) each,and we have 17 true value of 17 tags.<br>we can train 17 Ridge models,each of them predict 1 tag.<br>for example,for tag:Water,<br>X is [water_prob_of_resnet50, water_prob_of_densnet121, water_prob_of_simplenet]<br> y is [True_tag_of_water],<br>So we find optimized combine of three models for predicting water tag.</p> <p>Hi,Alex,<br> I use part of heng's argument methods,and I don't think too much argument will help.I don't want to use crop,translation argumentations when train,because it will change the image too much ,it will move important objects(cloud,road....) out of images.And use too much test time tta will slow down prediction,andI used only 4 basic argumentations provided by heng in some of my models.And I have emailed some key codes(haze removal,F2 loss function..)of my own to heng.</p> <p>Zsolt,I split the samples randomly:<br></p> <pre>df_train = pd.read_csv('/data4/data/amazon/input/train_v2.csv') kf = KFold(n_splits=n_splits, shuffle=True, random_state=random_state) for i, (train, test) in enumerate(kf.split(df_train)): print(train.shape, test.shape) </pre> <p>please check the topic:<a href="https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/36743">https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/36743</a></p> <p>Fan Fei Chong,<br> 1.I fixed my random seed during the competition,I fixed my random seed to 100 in all competitions indeed. :) I simulated the train and test data set split by using different seeds to evaluate the influence of the randomness. 2.Yes,I used the same train-val split during the competition.<br> 3.We can split train set into 4 parts,train on 3 three of them and validate on the left one ,it is very common practice in Machine Learning competitions,but in Deep Learning competitions,we must take the performance into account,the dataset of this competition is not so large so we can use 4 folds or even 5 folds.But if the datasets are two large we can not use this method ,we must find a good validate method ,the key is keep the train-validate-test dataset in nearly same distribution.</p> <p>Jeremy,I replied Zolst's questions about local-cv strategy above <a href="https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/36809#206724">https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/36809#206724</a> ,the reason,the codes,the rule of how to use local-CV,if you have any more question please let me know.</p> <p>Zsolt,<br> Simple random is a good way ,because the images are splited to train-test set in the same way!As you konw,I recovered the images from 256256 to 25602560 images,and I can say for sure that the train and test set images are random selected.What's more,I checked train-val-test predictions carefully to confirm the distribution.As we must search the thresholds for each label,we must make sure we keep the balance between train and test<br> I chose models by local-cv of ensemble , when I found the local-cv high and public-LB was not so good or too good,then I decided not use newly added models(Too good LB was also harmful,and I confirmed after the post competition submissions ).I tried to keep public-LB and local-CV in a reasonable range.<br> And as we discussed in Fisheries competition,that competition was a little special,it was a two-stage one,we could not build a good CV strategy,we must guess the distribution of stage 2 dataset.If I knew all the images were from new boat,I could use by-boat split....I did so indeed and found the local-cv scores were OK after competition but I was afraid that some new images were from train-set boats...<br> The reason use ridge regression was quite simple,I wanted to find importance of each model,and I did not want to use complex one.And I also tried lasso but I found it had higher local-score and a little lower public score,so I did not use it.I guess lasso weights are more sparse and is not good for F2 thresholds,we need more continuous value space.<br></p> <p>Yes,4 folds means 4X time to train and predict.<br> As to the worries about the performace,there were a lot of discussios(criticisms) after kaggle competitions,here are my opinions:<br> 1.Performance is important only when it's a really problem.In this competion,the models are compared to human,If the models are accurate,I don't care if they are predicted in relatively long time.<br> 2.Kaggle competitions are attractive because of their clear and fair rules:the private LB,I some case,performance is important,kaggle will improve their rules,for example,In Two Sigma Competition,all models must be finished in one hour(train and predict).In NOAA sealion competition there are 18000 very large images(>50005000) which prevent anyone from using too complex models.<br> 3.As Andrew Ng said in CS229,when the train-data obtained from months hard work,we must make full use of them.If I work for Planet,I will crop more images and use all top models to predict labels and ensemble them,and correct the mislabeled images by human,and add them to train set,and I will select less models iteratively by CV to get balance among human labor,computer resources,performace,accuracy.<br> 4.In most of kaggle competitions,winner often have top single model too.<br> 5.When we have a look at the improvement from RCNN,Fast-RCNN to Faster-RCNN and SSD,RFCN, we can see ,we can find a way to better models/algorithms when we get more understanding of the problem and use different models in different situations.As to this competition,I think we also can find an end-to-end simple model<br></p> <p>I hope it will be helpful to you.<br></p> <pre><code>def f2_loss(logits, labels): __small_value=1e-6 beta = 2 batch_size = logits.size()[0] p = F.sigmoid(logits) l = labels num_pos = torch.sum(p, 1) + __small_value num_pos_hat = torch.sum(l, 1) + __small_value tp = torch.sum(l p, 1) precise = tp / num_pos recall = tp / num_pos_hat fs = (1 + beta * beta) * precise * recall / (beta * beta * precise + recall + __small_value) loss = fs.sum() / batch_size return (1 - loss) </code></pre> <p>Yes,your pseudo codes are clear,we have 17 models.In this competition,we can use the ground truth of labels when we train NN models,so their correlations are considered,I tried to used Baysian theory during the competition,however, it did not help me,so I used Ridge Regression,I found it helpful in a lot of competitions.<br> Although we did not use all the true labels when we train Ridge models,but the Ridge models are trained to decide whether we should choose some label when some labels coexist under some probabilities in our predictions.</p> <p><a href="https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/32402">https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/32402</a><br> <a href="https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/33559">https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/33559</a><br></p> <p>You can sort the topic by most voted then you can find most valuable information of a competition.:)</p>
Planet: Understanding the Amazon from Space	New to 14th in 1 week	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Planet: Understanding the Amazon from Space <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Hi all, I wanted to make a thread to share some of the things that I have learned throughout this competition, and hopefully to help newer Kagglers have a course of action and game-plan for competitions. I'm quite new to machine learning, python, and data science, and I had never touched image recognition before entering this competition with a little over a week left to go. So, before I go into details on my solution (which is quite similar to Team Urucu) I am going to give a list of things that I found helpful during this competition as a newbie. </p> <p>1) The forums are your best friend. Before you tackle a competition, read any EDA's, discussion topics with more than 5 votes, and look through the code on starter kernels to familiarize yourself with the important packages and techniques and how they work. Our amazon hero, Heng, is the only reason that I knew enough about image processing to even participate in this competition. There are a lot of smart Kagglers out there to learn from!</p> <p>2) Don't reinvent the wheel. There are a lot of great resources out there for Kagglers. While it is important to understand how things work, you don't need to implement everything from scratch. A few that I have found helpful which I don't see very often: - <a href="https://github.com/Far0n/kaggletils">https://github.com/Far0n/kaggletils</a> (shoutout to @Faron) - past competition solutions. Reading those is a great way to generate ideas to apply to the current competition.</p> <p>3) Don't be afraid to try things. You will probably make some very time-consuming mistakes at the start, and that's OK . It does help to understand enough of the theory behind and inner workings of algorithms/libraries to understand when you are making mistakes though. On this competition, my worst moment was the first time I ran a pretrained model. I got a MemoryExhaustedError 8 hours into training, and had not been saving models. Ouch.</p> <p>4) Setting up good cross-validation is key. Don't pay too much attention to the leaderboard- focus on making your solution stable. On this competition, I actually had submissions that dropped by as little as .0005 (eg. went from .9268 public -> .9215 private, but alas, hindsight is 20/20), and none of my submissions in the final two days dropped by significantly more than .001. </p> <hr> <p>On to the solution. Public LB: 0.93288 (Not my best public LB) Private LB: 0.93201 Drop: .00087 Since I had limited computation time, I only used 5 base models in my final submission. They were trained on 5-fold CV, and the test set predictions were bagged predictions of the 5 models trained during cross validation. The models were all keras pre-trained models: VGG-19 (128x128 VGG-16 (96x96) InceptionV3 (160x160) Xception (192x192) Resnet50 (256x256)</p> <p>For data augmentation, i used all possible random rotations/flips with equal probability, and with a small probability (1/4) I zoomed to a random value in (.85, 1.15).</p> <p>For the classification layer(s), I looked at the model base code in Keras, and used the same layers (scaled appropriately for the no. of classes), and this gave me fairly good results. (.928-.931 CV per model). Peter G. tipped me off to this strategy when he did so for VGG and scored ~.931 on the leaderboard. </p> <p>Something I noticed during my first day on the competition is that when stacking models, using models with different input shapes made the std. dev on stacking CV lower and also made the gains from stacking larger, and I think this was the key point in making my solutions stable. In addition, I didn't have time to run all of my models on a 256x256 input (ResNet50 took me almost 35 hours to run 5 fold) so that contributed to the decision as well. </p> <p>Then, I used the OOF predictions (with OOF-time augmentation on all rotations, just like TTA) to train xgboost, lightgbm, and catboost classifiers (one classifier for each label), and again with 5-fold CV and bagging test predictions. I optimized f1 on each model/class combo separately, which lead to quite big LB gains (.005). My final model was a voting ensemble of the second level classifiers.</p> <p>I think there was still a lot of room to improve this solution, but there was not time for me to run models on larger input sizes in addition to the smaller ones. I'd estimate at least .001 could be gained by adding all of the architectures with size 228x228 (center cropped) or 256x256, or by adding more sensitive models (maybe trained only on one class with a revised dataset, for example).</p> <p>You can actually just automatically reload at each point by iterating through the values that you want to do for epochs and learning rates, and then at the end of the loop reloading. This allows you to start with a lower learn rate from your best previous iteration (although I got mixed results with it- a few models it did much better and others it was somewhat of a wash). Here is how it would work: </p> <blockquote> </blockquote> <pre><code> model=your_model() epochs_arr = [10, 4,4,4,4] learn_rates = [0.0001, .00001, .000005, .000001, .0000005] kfold_weights_path=your_filepath for learn_rate, epochs in zip(learn_rates, epochs_arr): opt = optimizers.Adam(lr=learn_rate) model.compile(loss=your_loss_function optimizer=opt, metrics=[your_metrics]) callbacks = [EarlyStopping(monitor='val_loss', patience=2, verbose=0), ModelCheckpoint(kfold_weights_path, monitor='val_loss', save_best_only=True, verbose=0)] model.fit_generator(generator=train_generator(), steps_per_epoch=(len(X_train) // batch_size) + 1, epochs=epochs, verbose=2, callbacks=callbacks, validation_data=valid_generator(), validation_steps=(len(X_valid) // batch_size) + 1) if os.path.isfile(kfold_weights_path): model.load_weights(kfold_weights_path) </code></pre>
Planet: Understanding the Amazon from Space	Our approach to PUB-LB 0.93444 #3 / PRIV-LB 0.93206 #13	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Planet: Understanding the Amazon from Space <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Here is an image generated from noise using this AC-GAN. Every four images, from left to right and top to bottom, should match the labels 'clear', 'cloudy', 'haze', and'partly_cloudy'.</p> <p><img src="https://storage.googleapis.com/kaggle-forum-message-attachments/207064/6910/fake_samples_epoch_137.png" alt="AC-GAN" title=""></p> <p>Hi Miha!</p> <p>Unfortunately, I did this implementation on the last days of competition and have no time to analyze how it contributes to the final score. My hope with this implementation is that it would add diversity to the ensemble.</p> <p>I based my code on <a href="https://github.com/kimhc6028/acgan-pytorch">this implementation</a>. The only difference is that I used images from test-jpg to train true/fake output and images from train-jpg to train the classifier (so I have two forwards on Discriminator). My thesis was that this would lead to better generalization on unseen data (though I have no evidence to support this claim).</p> <p>By the end, the generated samples were far from perfect and have low relation to original classes.</p> <p>Hi Tim! Indeed, it is a brute-force-like approach.</p> <p>XGBoost was used to join predictions from all base-models (same approach until step <a href="http://blog.kaggle.com/2016/12/27/a-kagglers-guide-to-model-stacking-in-practice/">3.2.1 here</a>).</p> <p>The difference is that <strong>we didn't fit any model to the whole dataset</strong> (steps 4 and 5 from previous link). Instead, we use the same 4-fold scheme from base-models to train the XGB. This generates 4 out-of-fold predictions, which we individually tuned for best thresholds (targeting the best f2-score).</p> <p>Regarding our rank drop, we did a little investigation to find the reason for such loss. We concludes that we may used wrong TIF files for some of the base models. Furthermore, we find that some models had bad class-correlation between train and test, which indicates some bug in any part of our implementation (I find one input-normalization issue, for example). Anyway, it is not perfect clear to us the LB drop.</p> <p><strong>Summary of Urucu team solution</strong></p> <p>Public LB: 0.93444 #3 Private LB: 0.93206 #13</p> <p>Team members: Bruno, Gilberto, Paulo and Yang</p> <p>Our solutions is based on an ensemble of convolutional neural networks (CNN), trained on 4-fold cross validation. The final ensemble included 25 models in 4-fold, in a total of 100 trained neural networks!</p> <p>One important step to reach this number was to start early in the competition, as each train/predict would take up to 3 days to run (on a 980ti). By the end, Yang joined forces with a 4-TitanX which accelerates this process.</p> <p>To develop the models, we use both Keras and Pytorch with (mostly) pre-trained weights on imagenet. Architectures includes VGG[16\|19], Resnet50, Densenet121, InceptionV3 and custom convolutional neural networks, trained from scratch. By the end, we still included one AC-GAN, trained only on weather classes (in this case, using a softmax output).</p> <p>Our best single model, a VGG19, has an average logloss below 0.085 (f2 ~ 0.931 CV):</p> <p><img src="https://storage.googleapis.com/kaggle-forum-message-attachments/206968/6908/Screen%20Shot%202017-07-25%20at%2010.44.47.png" alt="Train loss" title=""></p> <p>Some models were trained in only one class. We resampled the dataset to ensure 50% of the trained samples were from the trained class. This produces a biased and sensitive model, but with great power to ensemble.</p> <p>After training, we start the predict step. In this phase, we augment each image 8 times (4 rotations and flip). Then we compute the mean, min, max and std of activations for each output class for that model.</p> <p>The ensemble is a combination of all features from all models generated on previous steps. For this we used a XGBoost, trained for each label on each fold (!). The number of trees is too tuned for each label. Also some feature selection techniques was applied for each individual label.</p> <p>The final step is to find the best thresholds that maximizes the f2-score. An initial approach was to find a global threshold for each individual label, but we notice that tuning it individually leads to better results. Also threshold are optimized per fold. Also a smoothing filter was applied to select best threshold and avoid local minimas.</p> <p>To build the submission, we included each label that was predicted in more than 2 of our folds.</p> <p>The final submission, made 20 minutes before deadline, was a Level 3 average of previous Level 2 models soft predictions (before threshold). This scored 0.9345 locally(CV), 0.93444 on public LB (but only 0.93206 on private). Anyway, this was our best private submission.</p>
Planet: Understanding the Amazon from Space	San Francisco. August 8. Meetup. Team ods.ai 7th place solution.	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Planet: Understanding the Amazon from Space <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>I really enjoyed discussions and overall atmosphere at the meetup in which I presented our 3rd place solution for the Dstl Satellite Imagery Feature Detection competition.</p> <p>I would like to do something similar and I am going to present our 7th place solution on August 8 at the meetup in San Francisco. It would be great if you will join. </p> <p>P.S. Beer and pizza provided.</p> <p><a href="https://www.meetup.com/San-Francisco-Kagglers/events/241919255/">https://www.meetup.com/San-Francisco-Kagglers/events/241919255/</a></p> <p>Are you also coming this time?</p> <p>Let's get this topic close to the top :)</p> <p>Reminder, tomorrow at Adroll in San Francisco, I will talk about our solution to this problem, guys from Planet will talk on how this competition looked from their side, and I hope Wendy will stop by and say a couple words from a Kaggle perspective.</p>
Planet: Understanding the Amazon from Space	a very simple brute force solution(public LB 0.9334,private LB 0.9318)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Planet: Understanding the Amazon from Space <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>my main ideal is come from @Peter Giannakopoulos in this topic <a href="https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/discussion/36091">Struggling to achieve LB 0.93 with fine tune pretrained model</a>,it really give me a geat help,Thanks a lot.as for a new bee, I have no ideal how to train a best network,I spent a lot of time to train a single model,but it just can't over 0.929.So i deside to using model combination.</p> <h1>1. some nets I used</h1> <pre><code>1) vgg19 2) densenet121,densenet169,densenet201 3) inceptionv3 4) renset50,resnet101,resnet152 </code></pre> <h1>2. training pipline</h1> <pre><code>1) resize picture to 128 * 128 2) DA by rotate90,horizontal and vertical flips,transposition. 3) using adam Opt, LR start by 1e-4,when vaild loss stop decresse 2 epoch devide LR by 10 until LR small than 1e-6. 3) train 10 times every net, and random split data set to train(0.8) and valid(0.2) set everytimes. 4) everytime I choose 2 snapshot one has best training set F2 score,anthor has best valid set F2 score,so every net I can get 15 ~ 20 models. 5) every model using a optimal thresh to predict,and through vote to generate result. </code></pre> <h1>3. some result</h1> <p><img src="https://storage.googleapis.com/kaggle-forum-message-attachments/206078/6896/vgg.png" alt="" title=""> <img src="https://storage.googleapis.com/kaggle-forum-message-attachments/206078/6897/vgg_den.png" alt="" title=""> <img src="https://storage.googleapis.com/kaggle-forum-message-attachments/206078/6898/full.png" alt="" title=""></p> <h1>4. maybe improve</h1> <pre><code>1) There are no single model can get 0.93 LB stable,so maybe it can get a higher LB using stable 0.93 models. 2) Because time limit I can't do muti-scale train,maybe different models come from different scale can get better result. 3) I wasn't using CV to choose best combine of these nets,because I didn't split a test set from the data set,when I want combine there has no times. </code></pre> <h1>4. some puzzled</h1> <pre><code>1) I try to using some models whose F2 &gt; 0.93 to generate result,but is get a worse LB,I don't know if it is because overfit. 2) DA such as shear,bright,color sames no benefit for the train,even I think flip and rotate is useless,doesn't somebody know why or I am wrong? </code></pre> <h1>Hope my method can help somebody~~~</h1>
Plant Pathology 2020 - FGVC7	0.98454(>1st) solution (RegNetY)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Plant Pathology 2020 - FGVC7 <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>This is NOT the 11th solution. I did not choose it.</p> <p>Model RegNetY-32GF 　・newest model from FAIR, this is powerful. 　・<a href="https://github.com/facebookresearch/pycls">https://github.com/facebookresearch/pycls</a></p> <p>Data augmentation (This idea almost is from <a href="https://www.kaggle.com/akasharidas/plant-pathology-2020-in-pytorch">https://www.kaggle.com/akasharidas/plant-pathology-2020-in-pytorch</a>) 　Horizontal and Vertical Flip, 　ShiftScaleRotate, 　RandomBrightness, 　IAAEmboss, IAASharpen,Blur, 　ElasticTransform, 　IAAPiecewiseAffine</p> <p>Imagesize 　512x512</p> <p>scheduler 　cosine annealing</p> <p>optimizer 　SGD+momentum</p> <p>loss 　Focal loss</p> <p>Inference 　one fold, no TTA</p> <p>other 　Pseudo labeling (0.98>confidence) 　This idea from my teammate xiaopeng! </p> <p>Thank you so much for letting me join the team.😃 </p> <p>I used note pc with no gpu. I trained all model on google colaboratory.</p> <p>Sorry for a late😨 </p>
Plant Pathology 2020 - FGVC7	10th place - missed 30+ better submissions :)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Plant Pathology 2020 - FGVC7 <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>:)</p> <p>you are welcome, any time</p> <p>it's simple but it worked on public</p> <p>it's pretty much based on this <a href="https://www.kaggle.com/dimakyn/classification-densenet201-efficientnetb7">kernel</a> with some minor changes like keeping aspect ratio while resizing (600x850) and using whole training without cross validation; added TTA with flips (0, H, V, HV). </p> <p>I did this for B5, B6, B7 and IRv2. The best was IRv2 with 0.981 on public and ensemble gave 0.984. </p> <p>Then added pseudo labels and got single IRv2 0.984 and ensemble 0.988.</p> <p>No fancy augmentation or loss functions, i did not treat imbalance. Still i ended up getting many high scores on both public and private. But it was hard to pick the best one as I had no validation set and both test and training sets are small. </p> <p>So I took the best public submission and randomly chose one out of 100 best submission :). The random one ended up above my best public. </p> <p>Also, I missed 30+ better submissions than my current score :D</p> <p><img src="https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1556833%2Ff36bff29350340c3669cdad0a7dd203d%2Fbad_luck.png?generation=1590563501929665&alt=media" alt=""></p> <p>thx</p>
Plant Pathology 2020 - FGVC7	1st place solution(single model)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Plant Pathology 2020 - FGVC7 <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>I used cross entropy loss.</p> <p>I resized the image to 320x512 and I don't use crop.</p> <p>Congrats to all. Many thanks to Kaggle and the organizers for holding this competition. And special thanks to my teammates.</p> <h2>Problems with the data</h2> <p>We noticed that there are different labels on the same images, for example, Train_379 and Train_1173 are exactly the same image while they have different labels. <img src="https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2186594%2Fcb794f440db244c11e0e78966e2202ae%2Fimage.png?generation=1590583887495151&alt=media" alt=""> Some images in the training dataset are generated by one image, but they have different labels. We can tell from the details of image Train_171 and Train_1 that they are generated by the same one image, but are labelled differently.<img src="https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2186594%2F7e3d1d198e843a907b52e61b0a9a4e14%2Fimage1.png?generation=1590583924522781&alt=media" alt=""> <img src="https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2186594%2Fa953717fab66a2fe1ce37bc3bb646691%2Fimage2.png?generation=1590583935937220&alt=media" alt=""> Therefore, we mainly want to solve the problem of noisy labels in the training dataset. We used the knowledge distillation method, first train a 5-fold models and get out-of-fold results about valid dataset, and then mix the out-of-fold results and ground truth by 3: 7 as the labels of a new training model.</p> <h2>Model</h2> <h3>Data augmentation</h3> <ul> <li>Brightness, Contrast</li> <li>Blur</li> <li>Flip</li> <li><p>Shift, Scale, Rotate</p> <h3>Backbone</h3></li> <li><p>seresnextnet50</p> <h3>Imagesize</h3></li> <li><p>320x512</p> <h3>loss</h3></li> <li><p>Cross entropy loss</p> <h3>Inference</h3></li> <li><p>5-fold average and 5 times TTA</p></li> </ul> <h3>Others:</h3> <p>The amount of data in this competition is very small and the data distribution is imbalanced. A wrong classification of a sample with multiple disease categories will have significant impact on the final results. So, don’t trust the public leaderboard, trust your CV.</p> <p>Thank you. I have no idea whether it will helps, for there are too many noisy labels.</p> <p>Thank you. Just like <a href="/jinzcoder">@jinzcoder</a> said below.</p> <p>It's right.😃 </p>
Plant Pathology 2020 - FGVC7	20th Place: Bilinear CNN with FL + LS	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Plant Pathology 2020 - FGVC7 <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Welcome!</p> <p><img src="https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1234220%2F9d8ab256053f1510a99a6924c08cca2a%2FBiLinearModel.PNG?generation=1590573890707261&alt=media" alt=""> Kaggle Notebook: <a href="https://www.kaggle.com/jimitshah777/bilinear-efficientnet-focal-loss-label-smoothing">https://www.kaggle.com/jimitshah777/bilinear-efficientnet-focal-loss-label-smoothing</a></p> <p>This was one of the ideas I have used during my experimentation.</p> <p><strong>Apart from this I have tried various other things like:</strong> Different CNN Models(B0,....B7), Different image sizes, different Focal Loss parameters different Label Smoothing parameters Combination of Focal + LS Various Augmentations Early Stopping Late stopping Combination of various CNNs as backbone for BiLinear CNN And a combination of everything that worked.</p> <p>Finally I submitted an ensemble of all my solutions which had a public score of greater than 0.970.</p> <p><strong>Final Note:</strong> Congratulations to all the Winners! It was a great experience for me. I felt like I was learning something or the other day after day, submissions after submissions. Best of luck for future competitions...!</p>
Plant Pathology 2020 - FGVC7	33rd->27th place solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Plant Pathology 2020 - FGVC7 <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>haha</p> <p>Congratulations to everyone. Many thanks to Kaggle and CVPR for holding this nice competition. we are very lucky to win the 33th place. we were concerned about data balance and overfitting,so share our solution: Agumentation: A.IAAFliplr(p=0.5), A.IAAFlipud(p=0.5), A.FancyPCA(alpha=0.1, always_apply=False, p=0.5), A.OneOf([ A.IAAEmboss(p=1.0), A.IAASharpen(p=1.0), A.Blur(p=1.0), ], p=0.5), A.OneOf([ A.ElasticTransform(p=1.0), A.IAAPiecewiseAffine(p=1.0) ], p=0.5), A.Normalize(mean=RGB_MEAN, std=RGB_STD, p=1.0), ToTensorV2(p=1.0), we used so many agumentation because we don't have time to experiment with what works Backbone:efficientnet-b6,efficientnet-b5,We changed the model: model1:cbam_EfficientNet_B6 model2:cbam_EfficientNet_B5 features=model1._fc.in_features + model2._fc.in_feature we also add 'senet' block but it is not works Image sizes : 350,350 loss : CBLoss(not works),Categorical Cross Entropy+label smooth other:smote(Oversampling),5-folds,4*TTA final:ensemble other model Public Leaderboard 0.979->0.981,Private Leaderboard 0.975->0.976 one more thing: Do not trust the LB, trust your CV. You can ensemble many many models，but it is not necessarily.</p>
Plant Pathology 2020 - FGVC7	3rd Place Score Model Summary	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Plant Pathology 2020 - FGVC7 <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>This was my first competition outside of Titanic and was a fun intro to Kaggle competitions. </p> <p>My submission ended up being fairly simple; it was an average of two ensembles, the first with three averaged models the other with two.</p> <p>For all of the models I trained EfficientNet B7 models with Tensorflow on TPU. I used 768x768 resolution for all runs and slightly varying data augmentations. The augmentations I used were: rotation, shear, zoom, shift, flip left/right, flip up/down, random brightness. I also upsampled “multiple diseases” data in my training set by 5x. Thanks to <a href="https://www.kaggle.com/cdeotte">@ChrisDeotte</a> for some helpful data augmentation guides.</p> <p>Ensemble A: Scores: 3 models, all 0.977 public score, averaged public score: 0.979 Model: EfficientNet B7 Data Balancing: Multiple diseases by 5x Data Augmentation: rotation, shear, zoom, shift, flip left/right, flip up/down, random brightness</p> <p>Ensemble B: Scores: 2 models, both 0.979 public score, averaged public score 0.980 Model: EfficientNet B7 Data Balancing: Multiple diseases by 5x Data Augmentation: >35 degree random rotations, >64 pixel shift, same other augmentations as above</p> <p>Ensemble A + B: Average public score 0.981</p>
Plant Pathology 2020 - FGVC7	4th place solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Plant Pathology 2020 - FGVC7 <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Congrats to everyone. I learned a lot and had fun. Thank you so much to Kaggle and the competition organizers for a really interesting and relevant problem. I attempted this Kaggle competition after titanic and mnist (though I read a lot of code from Flowers TPU).</p> <p>I think my solution is simple, much of the interesting stuff I tried didn't work sadly.</p> <p>Also would like to mention some Kagglers for helping me, my success is their success: <a href="https://www.kaggle.com/anyexiezouqu" target="_blank">@anyexiezouqu</a> for efficientnet b7 and incepresnetv2 ensemble at 800 image size. <a href="https://www.kaggle.com/ateplyuk" target="_blank">@ateplyuk</a> for efficientnet b7 on TPU. <a href="https://www.kaggle.com/mgornergoogle" target="_blank">@mgornergoogle</a> for multiple TPU tutorials and TPU troubleshooting posts I found during debugging.</p> <p>Submission 1: private score .978, public score .982 (placed 4th)<br> Submission 2: private score .980, public score .978 (hypothetically ~20th)</p> <p>I'll focus on the first submission here.<br> model1 = b7<br> model2 = (b6+incepresnetv2)/2<br> Final model = (model1 + model2) / 2</p> <h2><strong>What worked:</strong></h2> <ul> <li>Increasing the image size (I used 896 or 800 before memory limitations). If you use 896 you need to drop the batch size.</li> <li>Using noisystudent weights instead of imagenet</li> <li>drop val to train on more data (fairly standard practice)</li> <li>random flip LR augmentation</li> <li>random flip up/down augmentation</li> </ul> <h2><strong>No change or slight improvement:</strong></h2> <ul> <li>Use b6 instead of b7. With (b7 + incepresnetv2)/2 I got .975 LB, and also many LB scores lower than that. With (b6 + incepresnetv2)/2 I got .975 LB on the first try. So needs more data but maybe it improves. Though CV was not representative of LB for me, It was clear from model's confidence I was still overfitting. Also others used b5 and b6 successfully on LB so these 3 factors contributed to me bumping down to b6. I also wanted to try a b5 but not enough time.</li> </ul> <h2><strong>No change or slightly worse:</strong></h2> <ul> <li>Random cropping size 896 down to size 800. Not sure why this didn't improve the score, it's possible the pretrained networks are already quite translationally invariant.</li> <li>random rotation</li> <li>TTA with same, flip lr, flip down, flip lr flip down.</li> <li>Saturation (I might have done too much, but it didn't help)</li> </ul> <h2><strong>Implementation did not work:</strong></h2> <ul> <li>Gaussian blur on images (too slow), though I suspect it might have been useful if done carefully because of blurry/out of focus images.</li> <li>L2 regularization (seemed to make CV better but LB worse). After discovering this <a href="https://www.hds.utc.fr/~fdavoine/dokuwiki/_media/en/pr2020.pdf" target="_blank">I found a paper</a> saying, for transfer learning, to make weights approach imagenet initialization versus 1 (L2) or 0 (L1). Didn't try this yet.</li> </ul> <p>I think the theme of this competition for me was trying many techniques I thought would help, a few did but many didn't. So my model was quite simple in the end.</p> <p>Thanks, hope to see you all on Kaggle in future.</p>
Plant Pathology 2020 - FGVC7	7th place solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Plant Pathology 2020 - FGVC7 <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>my architecture for this competition was <a href="https://github.com/mrT23/TResNet">TResNet</a>, that provides a very good speed-accuracy-batch_size tradeoff for GPUs. for example, i was able to train on input resolution 600x600 with batch size of 64 and training rate of 90 img/sec, which enabled me to experiment fast and efficient.</p> <ul> <li>as a loss function, i used <a href="https://arxiv.org/pdf/1910.07038.pdf">soft triplet loss</a> + cross entropy</li> <li>i used squish cropping with <a href="https://docs.fast.ai/vision.transform.html#_dihedral">8-rotation</a> augmentation, both for training and for inference.</li> <li>i used class balancing to give more weight to class1. without class-balancing, i could not learn that specific class properly, and the competition loss metric would punish me greatyl.</li> <li>i used pseudo-label to enrich a bit the ultra-small train set (added about 60% of the test set to train set)</li> <li>i used EMA to shorten the training even further and provide better results.</li> <li>my ensemble was composed of different input resolutions and different augmentations regimes.</li> </ul> <p>p.s. 1 "Do not trust the LB, trust your CV" ? everyone on every kaggle forums say that, but in hindsight there was a strong correlation for me between private and public scores. if i had chosen the top public score as submission,i would have ended at 2nd place :-)</p> <p>p.s. 2 test set was way too small, and just injected randomness and luck to the final positions</p> <p>6th place is infinitesimal close :-)</p> <p><img src="https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F658430%2Fbd834381a11701e06ca75cadf4906639%2Faaa.png?generation=1590644297590525&alt=media" alt=""></p> <p>7th (or 6th) in a competition with 1317 teams is not that bad :-)</p> <p>my model (<a href="https://github.com/mrT23/TResNet">TResNet</a>) is fully available on github. regarding training code and tricks, they dont contain ideas very different from other public kernals already published, so i dont plan to release the code.</p>