Merged
Conversation
pengwa
added
the
training
justinchuby
reviewed
May 13, 2022
justinchuby
reviewed
May 13, 2022
orttraining/orttraining/test/training_ops/cuda/optimizer/adam_test.cc
Outdated
Show resolved
Hide resolved
askhade
reviewed
May 16, 2022
askhade
reviewed
May 16, 2022
askhade
reviewed
May 16, 2022
orttraining/orttraining/training_ops/cuda/optimizer/adam/adam.h
Outdated
Show resolved
Hide resolved
askhade
reviewed
May 16, 2022
orttraining/orttraining/training_ops/cuda/optimizer/adam/adam.cc
Outdated
Show resolved
Hide resolved
baijumeswani
previously approved these changes
May 25, 2022
ashbhandare
reviewed
May 25, 2022
onnxruntime/test/testdata/test_data_generator/adamw_test_data_generator.py
Outdated
Show resolved
Hide resolved
ashbhandare
reviewed
May 25, 2022
onnxruntime/test/testdata/test_data_generator/adamw_test_data_generator.py
Outdated
Show resolved
Hide resolved
ashbhandare
reviewed
May 25, 2022
onnxruntime/test/testdata/test_data_generator/adamw_test_data_generator.py
Outdated
Show resolved
Hide resolved
ashbhandare
reviewed
May 25, 2022
orttraining/orttraining/test/training_ops/cuda/optimizer/adamw_test.cc
Outdated
Show resolved
Hide resolved
ashbhandare
reviewed
May 25, 2022
ashbhandare
reviewed
May 25, 2022
orttraining/orttraining/test/training_ops/cuda/optimizer/adamw_test.cc
Outdated
Show resolved
Hide resolved
ashbhandare
reviewed
May 25, 2022
ashbhandare
reviewed
May 25, 2022
orttraining/orttraining/training_ops/cuda/optimizer/adamw/adamw.cc
Outdated
Show resolved
Hide resolved
ashbhandare
reviewed
May 25, 2022
ashbhandare
reviewed
May 25, 2022
ashbhandare
reviewed
May 25, 2022
| // > there is a minor difference compared with Apex's implementation, | ||
| // which uses double storing corrections before casting to float passing to kernels. | ||
| // > std::pow(float, int) return double since C++11, so we cast back to float. | ||
| alpha_correction = 1.f - static_cast<float>(std::pow(alpha, update_count)); |
Contributor
There was a problem hiding this comment.
will we get better precision by casting to float after the subtraction?
std::pow(alpha, update_count) is a small number < 1, so a precision loss will affect it much more than (1-<a small number)>) will, same for beta correction. Please correct me if I'm wrong
Contributor
Author
There was a problem hiding this comment.
Yeah, it might be. While as the comment says, it is to match what we see from other frameworks.
I original use double to calculate them, then changed this from double to float, to avoid for some rare case, that might bring differences between our runs and torch/HF runs. Make sense?
Contributor
There was a problem hiding this comment.
nit: The comment is not clear if we are matching apex or diverging from it, so if you have other things to fix, you can make the comment a bit clearer too.
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
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.
Description: MTA (Multiple Tensor Apply) Adam Optimizer Implementation.
Motivation and Context