fix(multimodal): handle BaseModelOutputWithPooling return type in CLIPScore for transformers >= 5.0 - #3428
Open
rityagodala wants to merge 2 commits into
Open
Conversation
…>= 5.0 in CLIPScore
In transformers >= 5.0, `CLIPModel.get_image_features` and `get_text_features`
return `BaseModelOutputWithPooling` instead of a raw tensor. This caused
`CLIPScore` to crash with:
AttributeError: 'BaseModelOutputWithPooling' object has no attribute 'norm'
Fix: unwrap the output in `_get_features` when it is not already a `Tensor`,
extracting `.pooler_output` which contains the projected embedding.
Fixes Lightning-AI#3407
for more information, see https://pre-commit.ci
Contributor
There was a problem hiding this comment.
Pull request overview
Fixes CLIPScore compatibility with transformers >= 5.0, where CLIPModel.get_image_features() / get_text_features() can return a BaseModelOutputWithPooling object instead of a raw embedding tensor, leading to runtime failures during feature normalization.
Changes:
- Unwrap
get_image_featuresreturn values by extracting.pooler_outputwhen the returned value is not aTensor. - Unwrap
get_text_featuresreturn values by extracting.pooler_outputwhen the returned value is not aTensor.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+134
to
+136
| # transformers >= 5.0 changed get_image_features / get_text_features to return | ||
| # BaseModelOutputWithPooling instead of a raw tensor; unwrap when needed. | ||
| return image_features if isinstance(image_features, Tensor) else image_features.pooler_output |
Comment on lines
+133
to
+136
| image_features = model.get_image_features(processed["pixel_values"].to(device)) | ||
| # transformers >= 5.0 changed get_image_features / get_text_features to return | ||
| # BaseModelOutputWithPooling instead of a raw tensor; unwrap when needed. | ||
| return image_features if isinstance(image_features, Tensor) else image_features.pooler_output |
Comment on lines
+150
to
+155
| text_features = model.get_text_features( | ||
| processed["input_ids"].to(device), processed["attention_mask"].to(device) | ||
| ) | ||
| # transformers >= 5.0 changed get_image_features / get_text_features to return | ||
| # BaseModelOutputWithPooling instead of a raw tensor; unwrap when needed. | ||
| return text_features if isinstance(text_features, Tensor) else text_features.pooler_output |
|
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #3428 +/- ##
========================================
- Coverage 37% 31% -5%
========================================
Files 349 349
Lines 19901 19907 +6
========================================
- Hits 7264 6250 -1014
- Misses 12637 13657 +1020 🚀 New features to boost your workflow:
|
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
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
CLIPScorecrashes withAttributeError: 'BaseModelOutputWithPooling' object has no attribute 'norm'when used withtransformers >= 5.0.Root cause: In transformers >= 5.0,
CLIPModel.get_image_featuresandget_text_featuresnow returnBaseModelOutputWithPoolingrather than a raw tensor. The code in_get_featurespassed this object directly to the caller, where.norm(p=2, dim=-1, keepdim=True)was called on it — causing the crash.Changes
src/torchmetrics/functional/multimodal/clip_score.py: In_get_features, after callingget_image_features/get_text_features, checkisinstance(features, Tensor). If the return value is aBaseModelOutputWithPooling(transformers ≥ 5.0), extract.pooler_output— the projected embedding tensor — before returning. Both old and new transformers versions are handled transparently.Repro (before fix)
Fixes #3407