ONNX face detection model has high GPU usage

[SullyNiu]

Dear ST team,

 

I trained a face detection model.
On https://stedgeai-dc.st.com/, Select a platform->Quantize->Optimize->Benchmark
I found that GPU occupies 93.69% and NPU occupies 6.31%. How to improve npu utilization?

According to the official website documentation, it should be per-channel now. How to convert it to per-tensor?

 

The attachment below is my model file.

Thanks.