ViTPose

[ariG23498]

This PR will:

1. Add the dummy inputs
2. Build a small model
3. Add test

CC: @xenova

I am following this guide. While @xenova has already taken care of the configurations and registry, I am trying to push this PR to the end line.

Currently the following works

from pathlib import Path
from optimum.exporters import TasksManager
from optimum.exporters.onnx import export
from transformers import VitPoseForPoseEstimation

base_model = VitPoseForPoseEstimation.from_pretrained("usyd-community/vitpose-base-simple")

onnx_path = Path("model.onnx")
onnx_config_constructor = TasksManager.get_exporter_config_constructor("onnx", base_model, library_name="transformers", task="keypoint-detection")
onnx_config = onnx_config_constructor(base_model.config)

print(onnx_config)

onnx_inputs, onnx_outputs = export(base_model, onnx_config, onnx_path, onnx_config.DEFAULT_ONNX_OPSET)

Are there any other immediate check that I need to perform?

[xenova]

Thanks @ariG23498! Could you merge the commits from #2098 into this PR too so we only have one PR to review/merge? Thanks!

[ariG23498]

@xenova I have made the adjustments. Is it okay now?

[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.

[xenova]

I added support for exporting models with num_experts>1 and things seem to work well. I've also created a tiny-random model which can be used for testing. Other than that, LGTM ✅

cc @echarlaix for final review :)

[echarlaix]

Looks good, thanks a lot !