Repository navigation
Conversation
before_backward checked loss availability with `if not strategy.loss:`, but an
all-zero loss tensor (e.g. a numerically perfect batch from CrossEntropyLoss)
evaluates to False as a bool, so MASPlugin raised ValueError("Loss is not
available") even though the SGD update had already run. Switch the check to
`requires_grad`, which is True for a real training loss and False only when no
autograd graph was built.
Adds regression tests covering both directions: a grad-bearing zero loss is
accepted and the penalty applied, while a loss with no autograd graph (the
BaseSGDTemplate initial torch.zeros(1)) still raises ValueError.
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
When a training batch is numerically perfect, nn.CrossEntropyLoss can return an all-zero loss tensor, e.g. tensor(0., grad_fn=). In MASPlugin.before_backward the availability of the loss was checked with
if not strategy.loss:, which treats an all-zero tensor as falsy, so MASPlugin raised ValueError("Loss is not available") even though the SGD update had already run correctly.This changes the check to
if not strategy.loss.requires_grad:, which is true for a real training loss and false only when no autograd graph was built. This matches the fix suggested in the issue and avoids introducing None handling elsewhere in the training loop. A regression test covers both directions: a grad-bearing zero loss is accepted and the penalty is still applied, while a loss with no autograd graph (the BaseSGDTemplate initial torch.zeros(1)) still raises ValueError.Fixes #1676.