Using opencap bench with mmdeploy ?

[MaximeSabbah]

Hello,

Thank you for your outstanding work !

I was wondering if the SynthPose models you are providing could be interfaced with mmdeploy framework for rapid inferences ?
Are the models available anywhere in .pth format that we can convert .onnx or .engine with the dedicated config files ? or directly .onnx or . engine ?
Do you have any ways to improve the inference time of the model ? For now we quickly tried with transformers and it was very slow.

Thank you in advance for your answer,

Best

Maxime

[yonigozlan]

I recommend using the Transformers version as the mmpose ones are more difficult to maintain, since they were made with an older version of mmpose.
You can have a look at this great post by Pavel Iakubovskii to get better inference speed with Transformers and torch compile: link
You should also be able to export the transformers models to onnx (I think).

Here are the links to the Transformers models: synthpose-vitpose-base-hf, synthpose-vitpose-huge-hf