Support optional flag to clamp gradient in 'backward' to prevent crash #45
Conversation
This file contains 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
This commit addresses a very intermittent, but deadly crash bug that is destroying my training runs - a very occasional infinite gradient in the 'backward' function. In this commit, functionality remains unchanged by default. However, an optional flag has been added that allows clamping the gradient in the 'backward' function. The flag takes the form of an int or sequence giving the max value (or min/max if sequence). An optional third value in the passed sequence is interpreted as a Boolean that indicates whether to print a warning to the console whenever an infinite gradient is clamped. The default is False. Support for PyTorch only.
…tions Also, add dummy argument in 'backward' to match new backward_clamp_gradient_mag argument
|
This looks great. Any chance you can help writing a small test for this to make sure it is indeed clamping the gradients correctly? |
|
I'd be happy to if you can give just a hint or two about how/where it would go... It's a bit tricky for me to understand how to extract the logic in this case from the actual runtime location in the 'backward' function... |
|
See new PR that is identical to this but uses a different branch from my fork: |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
This commit addresses a very intermittent, but deadly crash bug that is destroying my training runs - a very occasional infinite gradient in the 'backward' function.
In this commit, functionality remains unchanged by default.
However, an optional flag has been added that allows clamping the gradient in the 'backward' function. The flag takes the form of an int or sequence giving the min/max value.
An optional third value in the passed sequence is interpreted as a Boolean that indicates whether to print a warning to the console whenever an infinite gradient is clamped. The default is False.
Support for PyTorch only.