Kaggle Humpback Whale Identification Challenge 2019 2nd place code

This is the source code for my part of the 2nd place solution to the Humpback Whale Identification Challenge hosted by Kaggle.com.

End to End Whale Identification Model

Recent Update

2019.03.13: code upload.

Dependencies

• imgaug == 0.2.8
• opencv-python==3.4.2
• scikit-image==0.14.0
• scikit-learn==0.19.1
• scipy==1.1.0
• torch==1.0.1.
• torchvision==0.2.2 pandas

Solution Development

single model design

• Input: 256x512 or 512*512 cropped images;
• Backbone: resnet101, seresnet101, seresnext101;
• Loss function: arcface loss + triplet loss + focal loss;
• optimizer: adam with warm up lr strategy；
• Augmentation: blur,grayscale,noise,shear,rotate,perspective transform；
• Horizontal flip to create more ids -> 5004*2
• Pseudo Labeling

Single model performace

single model	privare LB
resnet101_fold0_256x512	0.9696
seresnet101_fold0_256x512	0.9691
seresnext101_fold0_256x512	0.9692
resnet101_fold0_512x512	0.9682
seresnet101_fold0_512x512	0.9664
seresnext101_fold0_512x512	-

Single model performace with pseudo labeling

I generate a pseudo label list containing 1.5k samples when I reached 0.940 in public LB, and I kept using this list till the competition ended. I used the bottleneck feature of the arcface model (my baseline model) to calculate cosine distance of train test images. For those few shot classes (less than 2 samples), I choose 0.65 as the threshold to filter high confidence samples. I think it will be better result using 0.970 LB model to find pseudo label.

single model	privare LB
resnet101_fold0_256x512	0.9705
seresnet101_fold0_256x512	0.9704
seresnext101_fold0_256x512	-

Model ensemble performace

single model	privare LB
resnet101_seresnet101_seresnext101_fold0_256x512	0.97113
resnet101_seresnet101_seresnext101_fold0_512x512_pseudo	0.97072
10 models(final submisson)	0.97209

Path Setup

Set the following path to your own in ./process/data_helper.py

PJ_DIR = r'/Kaggle_Whale2019_2nd_place_solution'#project path
train_df = pd.read_csv('train.csv') #train.csv path
TRN_IMGS_DIR = '/train/'#train data path
TST_IMGS_DIR = '/test/' #test data path

Single Model Training

train resnet101 256x512 fold 0：

CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py --mode=train --model=resnet101 --image_h=256 --image_w=512 --fold_index=0 --batch_size=128

train resnet101 512x512 fold 0：

CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py --mode=train --model=resnet101 --image_h=512 --image_w=512 --fold_index=0 --batch_size=128

predict resnet101 256x512 fold 0 model：

CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py --mode=test --model=resnet101 --image_h=256 --image_w=512 --fold_index=0 --batch_size=128 --pretrained_mode=max_valid_model.pth

train resnet101 256x512 fold 0 with pseudo labeling：

CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py --mode=train --model=resnet101 --image_h=256 --image_w=512 --fold_index=0 --batch_size=128 --is_pseudo=True

predict resnet101 256x512 fold 0 model with pseudo labeling：

CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py --mode=test --model=resnet101 --image_h=256 --image_w=512 --fold_index=0 --batch_size=128 --is_pseudo=True --pretrained_mode=max_valid_model.pth

Inferring

For each model you can run command below to infer：

CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py --mode=test --model=resnet101 --image_h=256 --image_w=512 --fold_index=0 --batch_size=128 --pretrained_model=max_valid_model.pth

Thease params will determine the model path, for example: --model = resnet101 --fold_index=0 --image_h=256 --image_w=512 --is_pseudo=False --pretrained_model=max_valid_model.pth

It will find the model './model/resnet101_fold0_256_512/checkpoint/max_valid_model.pth', and run inference. So for the 10 models, you may need to change the params above and run inference for each model.

Final Ensemble

the final submission is the weight average result of 10 ckpts

python ensemble.py

Ken Notes

bboxes

bbox50 entry e.g. (filename, x0, y0, x1, y1)

965de6a7c.jpg,2,1031,51,461 265e16f33.jpg,6,1023,199,445 e01e495c7.jpg,215,859,375,531

bbox101 entry e.g.

965de6a7c.jpg,2,1036,55,461 265e16f33.jpg,2,1007,199,445 e01e495c7.jpg,215,859,375,543

Output files for test images for one infer run are put here. Redirect this to our "work" directory for the batch of images.

./models/resnet101_fold0_pseudo_256_512/checkpoint/max_valid_model/2TTA_0.csv ./models/resnet101_fold0_pseudo_256_512/checkpoint/max_valid_model/2TTA_1.csv

So, we could just run infer one model at a time until we are above threshold so as to save time in case the first model finds the whale? Be curious to see how far down the line it takes to finally get above the threshold. And if it is almost always finds it with first model or doesn't find it with any model. Is it worth, going down the line at all? Have to play with stats to see what is worth it.

Questions

• How are the bboxes created?
• Are the inference results for the training data built into the model? i.e. how are those created and where are they stored?
  • WhaleDataset - val - not needed, just a validate subset. 5 subsets of 600 images each.
  • I think the features are built in. Have to ask them this.