This repository was archived by the owner on Apr 8, 2025. It is now read-only.
Fix random behavior when loading ELECTRA models #599
Merged
Conversation
This file contains hidden or 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
tholor
approved these changes
Oct 21, 2020
Member
tholor
left a comment
tholor
left a comment
There was a problem hiding this comment.
Nice finding! This was a nasty one :)
Timoeller
approved these changes
Oct 21, 2020
Contributor
Timoeller
left a comment
Timoeller
left a comment
There was a problem hiding this comment.
Looking good. Thanks for the detailed description @bogdankostic
What a big headache this behaviour created and what small changes needed to fix that headache.
Closed
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
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Outdated suggestions cannot be applied.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.
This PR should fix #519. @brandenchan @stefan-it @PhilipMay
Summary of the bug
When training an ELECTRA model, saving it and loading it later on, the loaded model produced different (much worse) predictions than the trained model.
Root of the bug
In this transformers PR,
summary_use_projwas set by default toTrue, which means that a linear layer is stacked on top of pooling. Therefore, each time an ELECTRA model is loaded in FARM, the weights of this linear layer are randomly initialized.Fix
Set
summary_use_projtoFalsein order to not add that linear layer.