Semantic segmentation loss function

Note that (some) torchvision segmentation models will use a dict as the output.
Could you check that?