To run this solution:

0. put train/test files in the same file structure as required in kaggle kernels
1. change variable MODEL_DIR to be the directory where you put se-resnext50 weights file
2. change variable MODEL_NAME to be the name of se-resnext50 weights file
3. run :)

LB journey:

What worked for me:

• a good combination of learning rate and batch size.
• larger image size
• average prediction from multiple folds
• test time augmentation

What didn't work for me:

• Se-resnext101 + 288 + (bs==40): kind of interesting, my guess is that the batch size is too low.
• focal loss: somehow my validation f2 score became unstable and could not converge better than that of BCE loss.
• default Fastai Se-resnext50 structure.