Zero Zeng

It should not return error, emmm... what kind of error did you get?

No, openpose in tensorrt-zoo is ported from the official openpose repo. which is caffe based.

Sry I don't familiar with windows platform so can not help you with this :(

@kevinch-nv ^ ^

Did you do the int8 calibration with real data? and how many images you are using for calibration?

Resnet50 accuracy is in our L1 and QA test, I will be surprised if we had accuracy issue for it. Have you checked the code that do the calibration and...

Just don't install the python whl? I believe they need extra steps to build(https://github.com/NVIDIA/TensorRT/tree/main/python).

I think it's feasible, see https://docs.nvidia.com/deeplearning/tensorrt/api/c_api/classnvinfer1_1_1_i_gpu_allocator.html @nvpohanh correct me if I'm wrong.

From what I see TF-TRT uses a custom GPU allocator and there may be multiple streams running in parallel.

Looks like a problem of mmdeploy, I would suggest ask in mmdeploy repo.