Fix computation in normalization layers #876
Merged
+36
−14
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
|
Looks like the current implementation doesn't pass under strict-dtype-promotion (see failing tests), but other than that this sounds good to me! |
|
Thanks @patrick-kidger I have fixed it. Please let me know if you have any more suggestions on this one |
|
Alright, this LGTM! Final question before I merge this, do you have a reference for needing higher precision for these operations? (As much for me to read up on as anything else :D ) |
|
I think this is pretty much enough: https://pytorch.org/docs/stable/amp.html#autocast-op-reference |
|
Alright, looks good to me! Thanks for spotting this, and for the reference. Merged :) |
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Outdated suggestions cannot be applied.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.
Irrespective of the dtype passed to the normalization layers, the calculations should be done with fp32 or higher dtype. Not doing that can cause instabilities in training which is too common when someone is training models in half-precision. Now, we can delegate this to the end user to do this explicitly, but it is an easy thing to miss. This PR takes care of that.