Channel wise cross entropy issue

vision
21

Your target is probably all zeros.
I would recommend to avoid narmalizing the target and rescaling it again. Instead just keep the class indices and remove the transformation for your target.
If you call target.long() on small numbers, they will be all zeros. Scaling with 255 won’t change that.

A fast fix would be to first multiply it with 255 and then convert it to a long tensor or as I described avoid normalizing it in the first place:

F.cross_entropy(pred, (target.squeeze(1) * 255.).long())
22

Thanks for the tip. I guess that’s the problem. As you mentioned, I can see that most of the values in the target are in the order of 1e-3.