HuBMAP-kaggle

Model Weight dataset Link

• https://www.kaggle.com/datasets/soumya9977/hubmap-coat-512-weights [coatnet-small-aux5-512-fold3]

TODO [today: 11-08-22]

• EXHAUSTE ONE MODEL
  • try pretrained
  • try not pretrained
  • use LR schedule
  • bigger image
  • try multiclass training
  • try effnet-b4
  • change loss
  • add lable smoothing
  • we can change backbone and stuff, but what can we do technique wise.

Ideas:

1. Try Monai
2. Try tiling image segmentation
3. Integrate SMP
4. Train SegFormer with MMseg
5. Try Cellpose
6. Try Unet++ w/ different backbones.
7. Look into stain transforms
8. create a nb to plot all images at the same place
9. create NB to inference one or many images with all the model weights.
10. Apply TTA

TODOs:

• Fix Training loop [deadline: 22/07/22] [done: 23/07]
• Train on image patches [deadline: 24/07]
• Check Dice score implimentation from here
• Train MMseg on HuBMAP

Random:

• virtual env at /home/lakshita/somusan/hubmap_kaggle of python 3.7.0