[Clean-up] Planned removal of the max_size argument

[HichTala]

What does this PR do?

This PR aims to serve as a progress tracker for updating the size configurations of detection models on the Hugging Face Hub. The updates focus on aligning the configs with the deprecation of the max_size argument while ensuring backward compatibility where needed. Each update will be submitted as an individual PR directly to the model repositories on the Hub, following the script and guidance provided by @qubvel.

All individual pull requests made directly to the model repositories on the Hub will be tracked here via links, providing a clear overview of progress and ensuring comprehensive updates across relevant models.

Here's a brief summary of what's been done so far regarding the planned deletion of the max_size argument:

This pull request includes multiple changes across several image processing files in the transformers library. The primary focus is on deprecating the max_size parameter and replacing it with size['longest_edge']. Additionally, backwards compatibility adjustments have been made to handle the size parameter more flexibly.

Deprecation of max_size parameter:

• src/transformers/models/conditional_detr/image_processing_conditional_detr.py: Replaced logger.warning_once with logger.error for the max_size parameter deprecation and removed handling of max_size in __init__, resize, and preprocess methods. [1] [2] [3] [4]
• src/transformers/models/deformable_detr/image_processing_deformable_detr.py: Similar changes to deprecate max_size and update methods accordingly. [1] [2] [3] [4]
• src/transformers/models/detr/image_processing_detr.py: Updated to use logger.error for deprecation and removed max_size handling. [1] [2] [3] [4]
• src/transformers/models/grounding_dino/image_processing_grounding_dino.py: Changed logger.warning_once to raise ValueError for max_size deprecation and removed related code. [1] [2] [3] [4]
• src/transformers/models/rt_detr/image_processing_rt_detr.py: Updated to handle max_size deprecation similarly.

Backwards compatibility adjustments:

• src/transformers/models/conditional_detr/image_processing_conditional_detr.py: Added handling for size when it is an integer.
• src/transformers/models/deformable_detr/image_processing_deformable_detr.py: Included similar handling for integer size.
• src/transformers/models/detr/image_processing_detr.py: Adjusted size handling for backwards compatibility.
• src/transformers/models/grounding_dino/image_processing_grounding_dino.py: Added compatibility handling for size.
• src/transformers/models/rt_detr/image_processing_rt_detr.py: Included backwards compatibility for size.

It follows an issue and a pull request linked bellow:

• Deprecation Warning for max_size in DetrImageProcessor.preprocess #34977
• Fix max size deprecated warning #34998

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@qubvel

[HichTala]

I started by testing on my own repository to make sure everything works good and avoid spamming all models.

Here is the link to commit that is returned by push_to_hub method: https://huggingface.co/HichTala/test_pull_requests/commit/bb6b52f2f2039b5f4b1a27c33ebc42af072a9206

The pushed config file adds optional arguments and rearranges the arguments in an order sometimes different from that of the original file. If this is ok, we can start pushing the first 5 models.

[qubvel]

Hi @HichTala! Thanks for continuing with this issue 🤗

It's great that you tried using your own model! I suggest:

1. Replacing "feature_extractor_type": "DetrFeatureExtractor" with "image_processor_type": "DetrImageProcessor" because the feature extractor is deprecated.
2. Removing max_size and replacing size with a dictionary.

This way, it will be easier to merge those pull requests, and the difference will be smaller! I suppose it will require modifying the raw JSON file, instead of saving with ImageProcessor.

[HichTala]

Hi @qubvel,
I followed your suggestion, here is the PR in my own model, I think we're ready to start PRing the first 5 models!
I'll probably launch it in the next few days.

[qubvel]

Hi @HichTala! Looks good to me, let's proceed with the first five 👍

[HichTala]

Hi @qubvel, here are the 5 firsts!

• 0 Model: microsoft/table-transformer-detection
  Created: 2022-10-14 09:14:13+00:00
  Downloads: 3207449
  PR: [Clean-up] Planned removal of the max_size argument

• 1 Model: microsoft/table-transformer-structure-recognition
  Created: 2022-10-14 09:19:57+00:00
  Downloads: 1029420
  PR: [Clean-up] Planned removal of the max_size argument

• 2 Model: TahaDouaji/detr-doc-table-detection
  Created: 2022-03-11 15:55:14+00:00
  Downloads: 238448
  PR: [Clean-up] Planned removal of the max_size argument

• 3 Model: Aryn/deformable-detr-DocLayNet
  Created: 2024-03-19 00:46:32+00:00
  Downloads: 25095
  PR: [Clean-up] Planned removal of the max_size argument

• 4 Model: valentinafeve/yolos-fashionpedia
  Created: 2022-11-17 16:04:03+00:00
  Downloads: 16479
  PR: [Clean-up] Planned removal of the max_size argument

[qubvel]

Hi @HichTala! Thanks for creating the PRs and providing links. I just verified them, and everything looks perfect. Special thanks for providing a message with the PR. 🤗 I will ask to merge them and then we can proceed with the next batch.

[qubvel]

Hi @HichTala, hope you are doing great! 2 PRs were merged by owners, however we can't merge another PRs on our side, so we have to wait to see if they will be accepted for a while. Let's open more PRs 🤗 I would say the next 20-30 would be enough. And if you can, please provide links to the PRs directly instead of commits, thank you!

[HichTala]

Hi @qubvel,
Hope you are doing great aswell, here is the next batch ! I opened 10 PRs, the next ones have less than 100 downloads, should I include them ?

• 0 Model: isalia99/detr-resnet-50-sku110k
  Created: 2024-03-14 15:12:44+00:00
  Downloads: 5155
  PR: [Clean-up] Planned removal of the max_size argument

• 1 Model: aParadigmP/table-transformer-detection-custom-ale
  Created: 2023-11-03 12:38:33+00:00
  Downloads: 3906
  PR: [Clean-up] Planned removal of the max_size argument

• 2 Model: nickmuchi/yolos-small-finetuned-license-plate-detection
  Created: 2022-08-18 13:21:39+00:00
  Downloads: 3790
  PR: [Clean-up] Planned removal of the max_size argument

• 3 Model: facebook/detr-resnet-101-dc5
  Created: 2022-03-02 23:29:05+00:00
  Downloads: 3472
  PR: [Clean-up] Planned removal of the max_size argument

• 4 Model: facebook/detr-resnet-50-dc5
  Created: 2022-03-02 23:29:05+00:00
  Downloads: 3191
  PR: [Clean-up] Planned removal of the max_size argument

• 5 Model: nickmuchi/yolos-small-rego-plates-detection
  Created: 2022-07-07 21:31:21+00:00
  Downloads: 2568
  PR: [Clean-up] Planned removal of the max_size argument

• 6 Model: hilmantm/detr-traffic-accident-detection
  Created: 2024-01-22 14:35:40+00:00
  Downloads: 669
  PR: [Clean-up] Planned removal of the max_size argument

• 7 Model: Omnifact/conditional-detr-resnet-101-dc5
  Created: 2024-03-26 22:35:56+00:00
  Downloads: 263
  PR: [Clean-up] Planned removal of the max_size argument

• 8 Model: nickmuchi/yolos-small-plant-disease-detection
  Created: 2022-11-04 12:18:11+00:00
  Downloads: 182
  PR: [Clean-up] Planned removal of the max_size argument

• 9 Model: nickmuchi/detr-resnet50-license-plate-detection
  Created: 2022-08-23 17:44:21+00:00
  Downloads: 129
  PR: [Clean-up] Planned removal of the max_size argument

[qubvel]

Thank you, 10 is enough! We also need to check the other tags, e.g. zero-shot-object-detection, there might be a max_size too

[HichTala]

Hi @qubvel,
That's a good point! I looked at other tags and found that the following tags can have the max_size parameter in the preprocess_config file:

• object-detection
• image segmentation
• text-image

I'll try to look in the code to see exactly where it's used and whether I may need to remove some parts as we did for object detection.

[HichTala]

However, I can confirm that none of the zero-shot-object-detection models or other tags have the max_size parameter in the configurations, but some of them have size as a number and not as a dict. I don't know what we want, do we want to normalize the code or not? I may have to investigate the code a little further to see what's in there and see if it might be good or not.

[qubvel]

Hi @HichTala! Thanks a lot for investigating it. Let's limit the scope for max_size now, integer size is fine 👍