transformers
Reset loss to zero on logging in Trainer to avoid bfloat16 issues
#8561
Merged

Reset loss to zero on logging in Trainer to avoid bfloat16 issues #8561

bminixhofer
bminixhofer make tr_loss regular float
c9d7ccfa
sgugger
bminixhofer
bminixhofer
sgugger
bminixhofer
sgugger
bminixhofer
bminixhofer
sgugger
bminixhofer
bminixhofer Revert "make tr_loss regular float"
d3e8b825
bminixhofer reset loss at each logging step
142d34db
bminixhofer
sgugger
bminixhofer keep track of total loss with _total_loss_scalar
1f40edb0
bminixhofer
bminixhofer add remaining tr_loss at the end
ce16f5c7
sgugger
sgugger approved these changes on 2020-11-17
sgugger sgugger requested a review from LysandreJik LysandreJik 5 years ago
bminixhofer bminixhofer changed the title make tr_loss in Trainer regular float to avoid overflow Reset loss to zero on logging in Trainer to avoid bfloat16 issues 5 years ago
LysandreJik
LysandreJik approved these changes on 2020-11-18
LysandreJik LysandreJik merged f6fe41c9 into master 5 years ago
bminixhofer bminixhofer deleted the tr_loss_fix branch 5 years ago

Login to write a write a comment.

Login via GitHub

Reviewers
Assignees
No one assigned
Labels
Milestone