Use canonical get_size_with_aspect_ratio (with max_size) from transformers.image_transforms to fix #37939

[sonianuj287]

Removed max_size parameter from ConditionalDetrImageProcessor
to simplify image resizing logic and fix #37939.

Used canonical get_size_with_aspect_ratio (with max_size) from transformers.image_transforms

Tested with pytest suite and all tests are passed.

Happy to be guided further for improvement.
Thanks

[Rocketknight1]

cc @molbap @yonigozlan

[molbap]

Thanks for the contribution! I'm all for simplifying the image processing logic wherever we can, however this will break Backwards Compatibility on a few models (from what I checked, not many). cc @yonigozlan wdyt?

@sonianuj287 , Can you please edit the PR's name to add "🚨🚨🚨" in front so we know it's breaking? Thank you!

[sonianuj287]

Hi @molbap , Thanks for replying. Okay will change the PR's name.
For the issue, the other way to resolve is to fix all function calls for this function - "get_size_with_aspect_ratio" by removing max_size parameter.

[yonigozlan]

Hey @sonianuj287 ! Thanks for working on this. However I'm not sure this fixes #37939 as max_size is still accepted as an arg?
To deprecate it fully we need to modify the resize function of conditional detr. However I agree with removing get_size_with_aspect_ratio from the image processor and importing it from image_transforms as they seem to be equivalent.

We can first start with merging this PR to deprecate max_size and import get_size_with_aspect_ratio from image_transforms, and I think we can then extend this to other models which have an equivalent get_size_with_aspect_ratio to the image_transforms one plainly written in their image processing file, along with support for max_size :)

[yonigozlan]

However as @qubvel noted in #37939, we still have a lot of popular models using max_size on the hub, so we must be careful how we approach this. We can first remove all get_size_with_aspect_ratio in imaging files if they are equivalent to the image_transforms one, as this shouldn't cause any issues with BC

[sonianuj287]

HI @yonigozlan, Thanks for confirming to merge this PR.

Let me get the task to change in other models which have an equivalent get_size_with_aspect_ratio to the image_transforms one plainly written in their image processing file, along with support for max_size.

Delighted to help you. :)

[sonianuj287]

I can see some conflict arouse due to some new changes, I'll fix them in couple of minutes.

[sonianuj287]

Hi @yonigozlan , I resolved the conflict, and the workflows also passed. Let me know for any other changes, and we good to close. :)

[sonianuj287]

@yonigozlan @molbap

I can see files like -
image_transformers.py
image_processing_deformable_detr.py
image_processing_data.py
image_processing_detr.py
image_processing_eomt.py
image_processing_grounding_dino.py
image_processing_masking2former.py
image_processing_maskformer.py
image_processing_rt_detr.py

having this function - def get_size_with_aspect_ratio with max_size attribute, I can change it to avoid errors

These files also have this function but max_size is an optional attribute there, so we can skip them. right?
image_processing_yolos_past.py
image_processing_yolos.py
modular_yolos.py

should I go further and make changes?

[yonigozlan]

Hey @sonianuj287 , if the get_size_with_aspect_ratio defined is these files are the exact same than in image_transforms, let's remove them form these files and import them from image_transforms, even for image_processing_yolos etc.

[yonigozlan]

Hey @sonianuj287, be careful not to remove get_size_with_aspect_ratio if it's not the same as in image_transforms, for the yolos image processors, they are not the same! Also it looks like you haven't

[yonigozlan]

Not the same as in image_transforms! (mode_size)

[yonigozlan]

Same here

[sonianuj287]

Hi @yonigozlan @molbap

This PR updates all relevant image processors to use the canonical get_size_with_aspect_ratio function from transformers.image_transforms.
The YOLOS model is temporarily excluded to avoid altering its output dimensions — it will be handled in a dedicated follow-up PR once backward compatibility is ensured.

[sonianuj287]

Hi @yonigozlan , just a follow up. Am I going correct?
shall we go and merge and proceed with yolo changes next..

[yonigozlan]

Thanks for working on this @sonianuj287 ! Looks good to me, let's wait for the CI to pass then I'll merge :)

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: conditional_detr, deformable_detr, detr, eomt, grounding_dino, mask2former, maskformer, rt_detr

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.