Open W-QY opened 3 months ago
What is certain is that the current onnx inference results are basically consistent with the pth inference results in most cases, but exceptions will occur, and I want to know how to avoid these exceptions.
I want to know if there are inevitable errors between onnx and pth? Or can we definitely get exactly the same results? Is there some inevitable quantization process when generating onnx?
This issue has been automatically marked as stale due to inactivity and will be closed in 30 days if no further activity occurs. If further support is needed, please provide an update and/or more details.
Describe the issue
I designed and trained a 6D pose estimation algorithm model using pytorch. After that I use torch.onnx.export to convert the pth format parameter file into an onnx inference file. Through comparison, it was found that in some input cases (for example, the target in the image is small and the target background is pure black), the inference results using onnxruntime and pytorch are obviously inconsistent, resulting in a large difference in the results of the two (in this case The error results of both inference results are very large compared with the true value).
I want to know how to reduce or completely avoid the inference differences between onnxruntime and pytorch?
To reproduce
Our algorithm can be found at this link: https://github.com/YangHai-1218/PseudoFlow/blob/69e8e7ad11a2a58f06532cc5b89b76300d83613b/models/estimator/wdr_pose.py
Urgency
No response
Platform
Linux
OS Version
86~20.04.2-Ubuntu
ONNX Runtime Installation
Built from Source
ONNX Runtime Version or Commit ID
1.15.1
ONNX Runtime API
Python
Architecture
X64
Execution Provider
CUDA
Execution Provider Library Version
CUDA 11.3