PyTorch implementation for Kaggle Human Protein Atlas Image Classification Challenge

• Model
  • ResNet18 backbone (pretrained on ImageNet)
  • Small decoder for image reconstruction
• Input size: (512, 512, 4) (RGBY)
• Augmentation
  • Random horizontal flip
  • Random vertical flip
  • Random affine transformation (rotation, translation, scale)
• Batch size: 40
• Optimizer: SGD
• Learning rate: cosine annealing from 5e-2 to 4e-4 (1 cycle)
• Weight decay: 1e-4
• Loss functions
  • Binary Cross Entropy (BCE)
    • Log-damped class frequency pos_weights, without oversampling rare data
  • Mean Square Error (MSE) for image reconstruction (unsupervised)
    • To learn a good feature representation as an auxiliary signal
  • Prediction for # of class co-occurrence (COOC)
    • Get higher confidence of predictions with higher threshold
• 5 folds (random split), 20 epochs for each fold

• The mean of the predictions of 5-fold model with 8 TTA (transpose + flip)
• Threshold: 0.4
• Score (with data leak): public LB 0.568 / private LB 0.516 (138th / 2236)

• pytorch 0.4.1
• torchvision 0.2.1
• numpy
• opencv
• scikit-learn
• pandas
• tqdm

pip install -r requirements.txt

• Download data from Kaggle HPA competition
• Extract train/test.zip
• Download HPAv18 external data and HPAv18RGBY_WithoutUncertain_wodpl.csv
• Put the external data into the corresponding external directory
• Modify the path appropriately in config.json
• Run PYTHONPATH=. python loaders/hpa_loader.py first to generate training/validation data lists

python [train, test, merge].py -h for more details