This project contains code for multi-label classification on the kaggle subset of the Human Protein Atlas dataset.
Implemented features:
- loss functions
- binary cross entropy
- focal loss
- soft F-beta loss
- CNN architectures
- ResNet (50, 152)
- Senet154
- InceptionV4
- learning rate scheduling
- loss weighting
- sample weighting
- stratified split for train/val
- extensive data augmentation
- test-time augmentation
- rule-based post-processing
- adaptive f1 thresholding