Add focal loss option for classification #9
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Note I explicitly keep the current implementation
F.cross_entropy
ifself.options.classification_focal_gamma == 0
out of an abundance of caution even though I checked that, numerically, the more complex formula gives the same answer.The difference is that
F.cross_entropy
fusesF.log_softmax
andF.nll_loss
, so it is supposed to be more numerically stable, which I wanted to keep. But if you think it's unnecessary, we can remove this if-then clause.@mstamenk