Fix loss computation in TFBertForPreTraining#17898
Merged
Rocketknight1 merged 1 commit intomainfrom Jun 28, 2022
Merged
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
sgugger
approved these changes
Jun 27, 2022
Collaborator
sgugger
left a comment
sgugger
left a comment
There was a problem hiding this comment.
Thanks for fixing. Is this copied somewhere else too or just in BERT?
Member
Author
|
@sgugger I believe it's unique to BERT, because I tried searching the codebase for any similar lines and it couldn't find any. I suspect this is how it stayed undetected for so long - it uses the NSP loss and people generally don't train with that anymore. |
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
4 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
With thanks to @Sreyan88 for writing up a clean bug report and reproducer, and to @ydshieh for locating the problematic code!
Our
hf_compute_loss()function forTFBertForPreTrainingwas incorrect. However, it still appeared to work when the number of masked positions was evenly divisible by the batch size. Other, more commonly-used models likeTFBertForMaskedLMdo not have this issue.The problem was incorrect handling of the reduction for the masked loss, so I took the opportunity to rewrite the function in modern TF. All shapes are now static in the rewritten function as well, which means it should now compile with XLA.
Fixes #17883