This repository has been archived by the owner on Nov 3, 2023. It is now read-only.
Allow varying number of student model heads during distillation #3447
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Outdated suggestions cannot be applied.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.
Patch description
When doing knowledge distillation, allow for the number of attention heads in the student model to differ from that in the teacher model, provided that:
(1) We are doing TinyBERT-style distillation, in which the weights of the teacher model are not copied to the student
(2) The coefficients on the loss terms of the MSE of the attention matrices between the two models are both exactly 0, meaning that we do not have to minimize the differences between them, which would be impossible with a different number of attention heads in each model
Testing steps
Existing distillation CI checks