Fix ISIC dataloader, shape error in UNet & questions about paper

[halleewong]

Hi,

Thanks for sharing this code! Posting here because it seems like there's no "issues" tab in your repo.

I am very interested in comparing to your model. Would you be able to release the pre-trained weights? It would make it much easier to compare and the pre-trained weights are necessary to run the val.py examples in your readme.

I tried running the train.py example using the provided ISIC dataloader (fixed some bugs in 1cb9712 to get it working) but I encountered a shape error with the recommended unet backbone. The image embeddings output by the unet encoder are [image_size/16, image_size/16] instead of [image_size/4, image_size/4] expected here, leading to a shape mismatch here.

Which backbone did you use in the paper? I wasn't able to find these architectural details in the main paper or supplement.

Thanks!
Hallee

[j-Halloran]

yes seconded, either pre-trained weights or a better description of training data would be much appreciated