Fix freezing modules in Ghost Clipping #729
Open
+21
−2
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
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D68656459 |
EnayatUllah
added a commit
to EnayatUllah/opacus
that referenced
this pull request
Feb 11, 2025
Summary: Freezing modules with ghost clipping throws an error as corresponding per-sample norms are (not) calculated. Fix: keep in memory the list of all parameters and checking if corresponding requires_grad is True when calculating norms. Further, unfreezing modules (with and without ghost clipping) wasn't supported because the hooks aren't present for the corresponding modules. Fix: rewrite `requires_grad_' to add the hook. Facebook We initially used a `trainable_parameters(module)` to traverse the list of trainable modules upon norm computation. It was slow because `trainable_parameters(module)` is a generator and it traverses the neural network graph overtime. We replaced it with a list of trainable parameters fixed during model creation time. This is what lead to issues with freezing modules as this list is not updated. Fix: Use **all parameters** **list** -- not a generator, so no traversal happens. Further, we check `requires_grad` when calculating per-sample norm to ascertain whether to compute it or not. This is how this check is done in (non-private) [optimizer](https://github.com/pytorch/pytorch/blob/5725462cd8679dd1dea8a469b1bf2e71f226b664/torch/optim/optimizer.py#L963) to determine which parameters are frozen or not. Differential Revision: D68656459
EnayatUllah
force-pushed
the
export-D68656459
branch
from
6c3ee58 to
60b7522
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68656459 |
Summary: Freezing modules with ghost clipping throws an error as corresponding per-sample norms are (not) calculated. Fix: keep in memory the list of all parameters and checking if corresponding requires_grad is True when calculating norms. Further, unfreezing modules (with and without ghost clipping) wasn't supported because the hooks aren't present for the corresponding modules. Fix: rewrite `requires_grad_' to add the hook. Facebook We initially used a `trainable_parameters(module)` to traverse the list of trainable modules upon norm computation. It was slow because `trainable_parameters(module)` is a generator and it traverses the neural network graph overtime. We replaced it with a list of trainable parameters fixed during model creation time. This is what lead to issues with freezing modules as this list is not updated. Fix: Use **all parameters** **list** -- not a generator, so no traversal happens. Further, we check `requires_grad` when calculating per-sample norm to ascertain whether to compute it or not. This is how this check is done in (non-private) [optimizer](https://github.com/pytorch/pytorch/blob/5725462cd8679dd1dea8a469b1bf2e71f226b664/torch/optim/optimizer.py#L963) to determine which parameters are frozen or not. Differential Revision: D68656459
EnayatUllah
force-pushed
the
export-D68656459
branch
from
60b7522 to
9eb7875
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68656459 |
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.
Summary:
Freezing modules with ghost clipping throws an error as corresponding per-sample norms are (not) calculated. Fix: keep in memory the list of all parameters and checking if corresponding requires_grad is True when calculating norms.
Further, unfreezing modules (with and without ghost clipping) wasn't supported because the hooks aren't present for the corresponding modules. Fix: rewrite
requires_grad_to add the hook.