While I am coding in fastai, I found a potential bug. I think I know the problem. It related to loss function and ambiguity in pytorch 1.3 documentation (in pytorch 1.4 under development, I don’t see they make modification even though some thread online talked about this issue), which results in wrongly calculation of certain loss of models. I have some examples to reproduce such error.
I wonder if I can make a pull request? Thanks.