- https://www.kaggle.com/datasets/soumya9977/hubmap-coat-512-weights [coatnet-small-aux5-512-fold3]
- 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.
- Try
Monai
- Try tiling image segmentation
- Integrate
SMP
- Train SegFormer with MMseg
- Try Cellpose
- Try Unet++ w/ different backbones.
- Look into stain transforms
- create a nb to plot all images at the same place
- create NB to inference one or many images with all the model weights.
- Apply
TTA
- 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
- virtual env at
/home/lakshita/somusan/hubmap_kaggle
of python 3.7.0