Modifying DPLossFastGradientClipping to add support for generative tasks with ghost clipping (#716) #722
+20
−6
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: D68047256 |
|
This pull request was exported from Phabricator. Differential Revision: D68047256 |
aparna-aketi
pushed a commit
to aparna-aketi/opacus
that referenced
this pull request
Jan 25, 2025
…sks with ghost clipping (pytorch#722) Summary: Pull Request resolved: pytorch#722 Generative tasks for NLP output predictions of shape (B,T,C) i.e., (batch_size, sequence_length, vocab_size). To compute the cross-entropy loss in this case, usually the predictions are reshaped to (BxT, C) and targets to (BxT). This creates an issue with Ghost Clipping per sample loss computation as BxT is seen as the batch_size. In particular, the current implementation of Ghost Clipping results in loss_per_sample, coeff variables to have a shape of BxT and B respectively. This causes a shape mismatch error. This diff fixes that error by collapsing the loss_per_sample variable to shape B i.e., the loss across the sequence_length dim is averaged/summed. Differential Revision: D68047256
aparna-aketi
force-pushed
the
export-D68047256
branch
from
3c32510 to
d701dc3
Compare
…sks with ghost clipping (pytorch#722) Summary: Generative tasks for NLP output predictions of shape (B,T,C) i.e., (batch_size, sequence_length, vocab_size). To compute the cross-entropy loss in this case, usually the predictions are reshaped to (BxT, C) and targets to (BxT). This creates an issue with Ghost Clipping per sample loss computation as BxT is seen as the batch_size. In particular, the current implementation of Ghost Clipping results in loss_per_sample, coeff variables to have a shape of BxT and B respectively. This causes a shape mismatch error. This diff fixes that error by collapsing the loss_per_sample variable to shape B i.e., the loss across the sequence_length dim is averaged/summed. Differential Revision: D68047256
aparna-aketi
force-pushed
the
export-D68047256
branch
from
d701dc3 to
8a6c564
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68047256 |
|
This pull request has been merged in c7d6144. |
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:
Generative tasks for NLP output predictions of shape (B,T,C) i.e., (batch_size, sequence_length, vocab_size). To compute the cross-entropy loss in this case, usually the predictions are reshaped to (BxT, C) and targets to (BxT). This creates an issue with Ghost Clipping per sample loss computation as BxT is seen as the batch_size. In particular, the current implementation of Ghost Clipping results in loss_per_sample, coeff variables to have a shape of BxT and B respectively. This causes a shape mismatch error. This diff fixes that error by collapsing the loss_per_sample variable to shape B i.e., the loss across the sequence_length dim is averaged/summed.
Differential Revision: D68047256