Log function causes floating point underflow/ gradient overflow
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 9k
- Forks
- 1.5k
- Avg merge
- 2d 4h
- Merged PRs (30d)
- 3
Description
we use softmax->log->nll_loss instead of nn.CrossEntropyLoss() to calculate the cross entropy loss. But when training with fp16_optimizer, we found sudden gradient overflows and skip all steps after training several batches. We debug for several hours and read the source code, we found the log function could cause the floating point underflow and gradient overflow as the apex could caluculate the norm of the gradient to determine weather the gradient overflow.
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
The report names softmax, log, nll_loss, nn.CrossEntropyLoss(), and fp16_optimizer, but no file or test. Start by reproducing the gradient overflow in the mixed-precision training path and comparing the two loss calculations; done means the reported underflow and overflow no longer occur under the affected training conditions.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- machine-learning, performance
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100