-
Notifications
You must be signed in to change notification settings - Fork 69
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
BN Fixes #783
BN Fixes #783
Conversation
Dev -> Main
Dev -> main
MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅ |
Thanks for spotting all these issues. I agree, we should incorporate these fixes.
This prevents the use case where you (don't) want to use the running average in train mode and (don't) want update the batch_norm statistics. Maybe we need an extra arg in the call functions to distinguish between train and eval mode (or just whether or not to use_running_average) instead of inferring from the update_batch_norm arg? |
recheck |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM. JAX changes are in follow up PR.
Nit: Can you run `yapf -i -r -vv -p on the files that are failing the Linting and yapf tests'?
There are some subtle issues with how BatchNorm is handled in the PyTorch version of the code. Currently,
workload.model_fn
has anupdate_batch_norm
parameter, which in theory should allow the submission to control whether the batch-norm statistics are updated during a forward pass. The issues are the following:update_batch_norm_fn
function stores the old momentum parameter for each batchnorm layer in amomentum_backup
variable, so it can be restored later, before zeroing the parameter. However, if it is called withupdate_batch_norm=False
twice in a row, it overwrites themomentum_backup
with 0 on the second call, so momentum then remains zero for the remainder of training.0
indicates that the momentum buffer shouldn't be updated. This is the opposite of how EMA momentum is usually done (i.e. in Adam), where1
would indicate that it shouldn't be updated, and 0 means it's set to the latest value at every step. The custom BatchNorm modules used in the two librispeech workloads follows this second, more standard convention instead. However, theupdate_batch_norm_fn
sets the momentum to zero for all three layer types, resulting in incorrect behavior for the librispeech workloads.update_batch_norm_fn
sets the BN layers to eval mode. This doesn't make sense as it prevents the use-case where you use batch-computed statistics (train mode) without also updating the running statistics. The BN layers can bet set to eval mode separately by passing inForwardPassMode.EVAL
to the forward pass, so removing this.eval()
call doesn't prevent the submission from using eval mode during a forward pass.This PR changes switch the custom BN code to follow the BN convention so that momentum=0 doesn't update the running buffers. It also fixes the issues in the update_batch_norm_fn function mentioned above.