search

ORB-HD / deface

Video anonymization by face detection
MIT License
684 stars 95 forks source link

Fix ort explicit providers required #22

Closed Aziks0 closed 2 years ago

Aziks0 commented 2 years ago

Fix for #21

I simply added the providers CUDAExecutionProvider and CPUExecutionProvider as arguments for the onnxruntime.InferenceSession class instantiation.

mdraw commented 2 years ago

Thanks! I am not entirely sure how to best handle this issue. On Linux this explicit setting is not necessary (yet?). Your fixed version emits a

UserWarning: Specified provider 'CUDAExecutionProvider' is not in available provider names.Available providers: 'CPUExecutionProvider'

on my machine, which does not have CUDA installed. It does run fine aside from this warning but it would in general be better to make the provider user-configurable, since automatically choosing the "best" one is apparently deliberately not supported by onnxruntime. CUDAExecutionProvider and CPUExecutionProvider are not always the best choice. For example on Intel CPUs the OneDNN provider could be better or on CUDA-capable systems the TensorRT provider could be better.

I'll merge this now but in the long term we should introduce this choice as an additional CLI arg and/or envvar or find another solution.