Inference model with BatchNorm2D generates different output via PyTorch and OnnxRuntime

My setup is:
OS - Microsoft Windows 10 Enterprise 2016 LTSB
GPU - Quadro M2000M
CUDA vers - 9.0, 10.0 & 11.0
Visual Studio - 2017, 2019
Python - 3.6.8
Torch 1.1.0\1.6.0
onnxruntme - 1.5.1
onnxruntime-gpu - 1.4.0
onnx version - 1.7.0

Please refer to the following link which show that torch Onnx exporter maybe has a problem with BatchNorm2D operation:
BatchNorm2D Export problem

You will be able to find there all required information which describe the problem.

I thought that it will push forward the investigation if I will open an issue here also about the referenced Onnxruntime issue.


Hello, is there any update regarding this topic?

Any kind of support will be much appreciated.

It is a major concern that we cannot include BatchNorm2D operator in our models due to *.onnx export problem.

In order to be able using NVIDIA TensorRT sdk, we shall understand what is the root cause of the wrong BatchNorm2D behavior after converting the TorchScript *.pt to *.onnx format.