Closed JdbermeoUZH closed 5 months ago
DoD:
Carry the following experiments for the brain and wmh datasets to find appropriate cond+uncond checkpoints:
Check no meaningful shift between train and validation for small DDPMs with no uncond training (latest and earliest checkpoint):
[x] 256x256 imgs, Small Unet (7.5M), large batch size, up to 22.5k steps (overfit to train, too big a loss in val in SD) [x] 256x256 imgs, Large Unet (140M), small batch size, up to 100k steps [x] 128x128 imgs, Small Unet (7.5M), large batch size, up to 50k steps
Check no meaningful shift between train and validation for small DDPMs with uncond training (latest and earliest checkpoint):
[x] 256x256 imgs, Small Unet (7.5M), large batch size [x] 256x256 imgs, Large Unet (140M), small batch size, up to 50k steps [x] 256x256 imgs, Large Unet (140M), large batch size, up to 22.5k steps [ ] 128x128 imgs, Small Unet (7.5M), large batch size
DoD:
Carry the following experiments for the brain and wmh datasets to find appropriate cond+uncond checkpoints:
Check no meaningful shift between train and validation for small DDPMs with no uncond training (latest and earliest checkpoint):
Check no meaningful shift between train and validation for small DDPMs with uncond training (latest and earliest checkpoint):