ProbSpace Competition: Ukiyo-e Author Prediction

This slide shows the 2nd place overall solution of the ProbSpace Ukiyo-e competition which was held from Nov.2019 to Jan.2020.

Competition URL: https://prob.space/competitions/ukiyoe-author

Updated on Jul. 2022

Transcript

1. Ukiyo-e Author Prediction Open Review Competition Maxwell
2. 0 1 2 3 4 5 6 7 8 9 3158 train images 10 classes 224 x 224 397 test images 10 classes 224 x 224 Remove 10 images But this was less helphul for me... (might be regularization?) 3158 => 3148 images ProbSpace DenseNet 121 (GR=32) DenseNet 169 (GR=32) SE-ResNeXt 50 SE-ResNeXt 101 ResNet 18 Inception - ResNetV2 Common to All Models - Stratified 8 folds - Images divided by 255 - AdamW - 3 Stage Learning CyclicLR (1e-3~1e-4) => LRonPlateau (1e-5) => LRonPlateau (5e-6) - Batch Size: 36 - Augmenatation: RandomSizedCrop, Shift, Scale, Brightness, Rotate HorizontalFlip, GridDistortion - Softmax with 10 classes Predict FC 1024 FC 128 FC 10 DenseNet 121 Psuedo threshold: 0.99 1st stage Geometric Weighted Blending CV: 0.8974, LB: 0.902 2nd stage Model 1 (TTA) CV: 0.8736, LB: 0.914 Model 2 (TTA) CV: 0.8792, LB: 0.914 Model 3 (TTA) CV: 0.8743, LB: 0.899 DenseNet 121 Psuedo threshold: 0.95 DenseNet 169 Psuedo threshold: 0.99 Common to All Models - Stratified 8 folds - Images divided by 255 - AdamW - 3 Stage Learning CyclicLR => LRonPlateau => LRonPlateau - Batch Size: 36 - Augmenatation: RandomSizedCrop, Shift, Scale, Brightness, Rotate, HorizontalFlip, GridDistortion - CutMix: p=0.5, alpha=0.5 - Softmax with 10 classes FC 1024 FC 128 FC 10 397 test images 10 classes 224 x 224 Predict Geometric Weighted Blending Model 1 : Model 2 : Model 3 = 40% : 25% : 35% Prediction with TTA - Augmentation: - Shift, Scale, Rotate, ... - HorizontalFlip - 50-times averaging Post-Processing Image No.150 No.150 is all white. Classified as `0` Model 1 Model 2 Model 3 CV: 0.883 LB: 0.929 Author Class 2 nd Ubuntu 18.04 - Keras - TITAN RTX x 1 - Geforce GTX 1080 Ti x 2 @ Maxwell_110
3. What worked  DenseNet 121, 169 - DenseNet outperformed other model architectures - Color information seemed to be important  Data Augmentation - Crop, Shift, Scale, Rotate, Brightness, HorizontalFlip, GridDistortion  CutMix - alpha=0.5  Psuedo Labeling (Hard Labeling) - Threshold 0.99/0.95 - Threshold ensemble in 2nd stage  3 stage Learning - AdamW - CyclicLR (1e-3-1e-4, 200 epoch) => RLRonP (1e-5, 100 epoch) => RLRonP (5e-6, 100 epoch)  TTA - 50 times, Shift, Scale, Rotate, HorizontalFlip Confusion matrix of DenseNet121 GradCAM (In this case, CNN looks paying attention to the back ground and umbrellas) ProbSpace
4. What did `NOT` worked  Metric Learning - Arcface  Stacking - Arcface features as additional features in 2nd layer - Statistical values (mean, std, CoV, ...) as additional features in 2nd layer  A custom function to improve class imbalance - I have never succeeded to this kind of tricks... you? Arcface with t-SNE ( 4 of 8 folds ) ProbSpace