transformers
Enable users to use their own loss functions + deal with prefetching for grad accum
#34198
Merged
Merged
Enable users to use their own loss functions + deal with prefetching for grad accum #34198
muellerzr
merged
20 commits into
main
from
muellerzr-fix-loss-calc
bookmark
57c698fe
Bookmark
3c579479
Bookmark
1d57bd8f
Actually implement
15e61f12
Pass in kwarg explicitly
928e9271
Adjust for if we do or don't have labels
b59c8f16
Bookmark fix for od
79f9479e
bookmark
13f33692
Fin
8080f28a
closer
13160e08
Negate accelerate grad accum div
6fa155a8
Fixup not training long enough
c2a705fd
BenjaminBossan
commented
on 2024-10-17
Add in compute_loss to take full model output
ac04e610
winglian
commented
on 2024-10-17
Document
af8411b8
compute_loss -> compute_loss_fn
a5fac5a0
Add a test
39d8f28c
muellerzr
changed the title
Refactor
42849302
Refactor
932a4910
Uncomment tests
2a6b0383
danielhanchen
requested changes
on 2024-10-17
ArthurZucker
approved these changes
on 2024-10-17
Update tests/trainer/test_trainer.py
54d10ded
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub