I am now trying to use tensorrt to speed up the detection algorithm. Using FP32, I find that the output is inconsistent between tensorrt and pytorch, and the model’s featmaps are inconsistent.
We transfer pytorch model to onnx model.And
- inputs are consistent.
- Comparing the model parameters of onnx and pytorch, we find that the model parameters of onnx and pytorch are consistent
I don’t know what went wrong.
I hope the authorities can help me out.
the Platform of tensorrt is Win10，the Platform of pytorch is ubuntu,I use the RTX5000 GPU,
the model is ResNext-101