[RT-Detr] - Add RT-Detr Onnx Config

[YHallouard]

What does this PR do?

• Add RT-Detr Onnx config to optimum from this PR in transformers : [RT-DETR] Fix onnx inference bug for Optype (Where) transformers#33877

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

[YHallouard]

Hi @fxmarty, @echarlaix, @JingyaHuang, @michaelbenayoun,

I'm creating this PR following @qubvel and @ArthurZucker advices on that PR: huggingface/transformers#33877

[echarlaix]

Thanks a lot for the addition @YHallouard

[YHallouard]

Hi @echarlaix, thanks for the review!

I added "PekingU/rtdetr_r18vd" for testing, but I saw that you used hf-internal-testing. Should I create a random initiated model their ?

[YHallouard]

Hi @echarlaix, what are your insights ? I made the changes but should I create a random initiated model in hf-internal-testing ?

[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.

[echarlaix]

Apologies for the delay @YHallouard, just fixed the conflicts so that we can soon merge. Two things must be fixed before we can merge :

• it seems that the height / width used for export needs to have a minimum size (won't work with default value 64) failing at https://github.com/huggingface/transformers/blob/deac971c469bcbb182c2e52da0b82fb3bf54cccf/src/transformers/models/rt_detr/modeling_rt_detr.py#L1900 as we need enc_outputs_class.shape[1] >= config.num_queries : can be fixed by overwriting _create_dummy_input_generator_classes and give an appropriate default value for height and width (+ raising an error if value not correctly provided)
• it seems pixel_mask is considered as a constant during export and thus removed from the graph inputs

[echarlaix]

why is it needed ? (not needed for ViT models but might be needed for RT-DETR)

[echarlaix]

also if needed shouldn't the height / width dimension set to dynamic as well?