MaskAttentionLoss in DiceCEEdgeLoss doesn't handle images without any edges about super-gradients HOT 3 CLOSED

@davidtvs thanks for reporting this bug, great catch!
Thanks for setting up the unit test for this use case, one thing I would comment on, is instead of checking whether the reduced mask_loss is nan, which can results from other unexpected reasons, it's better IMO to check before reduction whether the tensor is an empty tensor which indicate directly that the mask is empty.

mask_loss = mask_loss[mask == 1]  # consider only mask samples for mask loss computing

        if mask_loss.numel() == 0:
            mask_loss = torch.tensor(0.)

We highly encourage new contributors, feel free to open a PR if you find the time to do so. If not we'll push the fix ASAP.

from super-gradients.

This issue is fixed in master branch and will be available in the next version release. #982

Thanks again @davidtvs for reporting the issue.

from super-gradients.

Join the discussion on DagsHub!

from super-gradients.