This project is my attempt at the Human Protein Atlas Classification competition on Kaggle. To make the model work better, I've used multiple approaches and optimizations such as:
- Multilabel stratification
- 4 channel Resnets
- Making transfer learning work for 4 channels
- Data Augmentation
- One Cycle Learning rate scheduling (Cosine Annealing)
- Progressive resizing
- Discriminative layer training
At the time of taking part in this competition, I had to suddenly stop working on this project due to some unforeseen circumstances. As a result I couldn't train the model for a significant amount of time. I managed to get an accuracy of about 88% on the multilabel classification task invloving 29 classes. The best fbeta score was 0.412.
This is an incomplete / dropped project