Fix batchnorm in testmode without track stats #576
Merged
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
In test mode, the CUDA cuDNN implementation of batchnorm was not matching the CPU batchnorm in FLUX. In FLUX, with track_stats=False, the mean and variance of the current batch are used. Here, mean and variance were initialized to 0 and 1, respectively, and passed to cudnnBatchNormalizationForwardInference. To fix this, we need to calculate the mean and variance over the current batch to match the CPU implementation. Unfortunately, cudnnBatchNormalizationForwardInference requires a trained running mean and variance. However, batchnorm train and test should be identical without tracked stats since they both normalize over the current batch. As a result we can use cudnnBatchNormalizationForwardTraining in test mode as well, which works without a running mean and variance.
3 tasks
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Outdated suggestions cannot be applied.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.
In test mode, the CUDA cuDNN implementation of batchnorm was not matching the CPU batchnorm in FLUX. In FLUX, with track_stats=False, the mean and variance of the current batch are used. Here, mean and variance were initialized to 0 and 1, respectively, and passed to cudnnBatchNormalizationForwardInference.
To fix this, we need to calculate the mean and variance over the current batch to match the CPU implementation. Unfortunately, cudnnBatchNormalizationForwardInference requires a trained running mean and variance. However, batchnorm train and test should be identical without tracked stats since they both normalize over the current batch. As a result we can use cudnnBatchNormalizationForwardTraining in test mode as well, which works without a running mean and variance.
This is needed to help address FluxML/Flux.jl#1606 along with Flux.jl PR 2427.
PR Checklist