- Implement packed padded sequence for LSTM
- Implement save interval features
- Reduce saving, don't save 6 times per epoch...
- Maybe try lowering the minlen for segments?
- Try starting from lower learning rate for pre-trained models? Or have higher decay? % didn't work
- Try doing a MIL experiment on short segments only, to make training faster
- Try running median filters with size 1
- Add an ablation study with BCE
- Contextual information?