convert mmpose model to ONNX for TensorRT

[khanhha]

Hello,
I would like to ask if it's possible to convert MMPose models to ONNX format for using with TensorRT for better real-time performance.
Thanks

[innerlee]

this will be added to the next iteration, with great possibility

[khanhha]

That would be fantastic. I can help implement C++ inference code using TensorRT, for the bottom-up approach.

[innerlee]

maybe this pr open-mmlab/mmaction2#160 is helpful

[khanhha]

Thanks

…

On Tue, Sep 15, 2020 at 5:30 PM lizz ***@***.***> wrote: maybe this pr open-mmlab/mmaction2#160 <open-mmlab/mmaction2#160> is helpful — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#131 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AARLXKGR4NU7QKHWMOEEJO3SF6CCBANCNFSM4RNH4YGA> .

-- Ha Huy Khanh Student School of Information & Communication Technology - Hanoi University of Science & Technology Mobile: 01686499853 Y!M: khanhaaaa Skype: khanhhh89

[innerlee]

Hi @khanhha , you can try this pull request #305, any feedback would be great

[alicera]

I find the bottom-up approach have the matrix algorithm.
Do you have any idea for implement C++ inference code

[innerlee]

np & torch can be very fast for matrix things. Mostly there is no need for c++ reimplementation. Could you please open a new issue giving details on where is the slowness? We will try to optimize it once the bottleneck is clear

[alicera]

Hi @khanhha ,
Do you have any idea for implement C++ inference code for buttom-up?
This project only have top-down
https://github.com/linghu8812/tensorrt_inference/tree/master/mmpose