Issue with TensorRT model export: ONNX model from PyTorch 1.14 showing PyTorch 2.0.0 in TensorRT logs

[ridabenhassi2002]

Hi,

I’m encountering an issue when exporting a YOLOv8 model to ONNX using PyTorch 1.14, but when I run the model through TensorRT, the logs indicate that it’s using PyTorch 2.0.0 as the producer version. Here are the details:

1. Export Process:

   • I used PyTorch version 1.14 for exporting the model to ONNX:

     torch.onnx.export(model, dummy_input, "yolov8s.onnx", opset_version=11)
     

   • The ONNX export completed without errors, and the model file yolov8s.onnx was generated successfully.

2. TensorRT Conversion:

   • When I use TensorRT’s trtexec to convert the ONNX model, I notice that the logs show that the model’s producer is listed as PyTorch 2.0.0:

     [12/06/2024-12:32:39] [I] [TRT] Producer name:    pytorch
     [12/06/2024-12:32:39] [I] [TRT] Producer version: 2.0.0
     

   • The export was done with PyTorch 1.14, not 2.0.0.

3. Expected Behavior:

   • I expected the producer version to show 1.14, consistent with the PyTorch version used during export.

4. What I’ve Tried:

   • I’ve double-checked my PyTorch version with torch.__version__ and confirmed it’s 1.14.
   • I’ve tried re-exporting the model and rebuilding the environment but still get the same result.

5. Environment:

   • PyTorch version: 1.14
   • TensorRT version: 8.5.2
   • CUDA version: 11.4
     Has anyone encountered this issue before? Any suggestions on how to resolve this discrepancy?

Thanks!

[AastaLLL]

Hi,

Could you share the exported file so we can check it?
Thanks.

[ridabenhassi2002]

Thanks for your reply, here’s the file onnx and the engine.

yolov8s_.zip (56.5 MB)

[AastaLLL]

Hi,

This is what we get from your ONNX model:

$ head -5 yolov8s_opset\ 11.onnx.txt 
pytorch2.0.0:���
�

input
0.model.0.conv.weight
0.model.0.conv.bias/0/model.0/conv/Conv_output_0/0/model.0/conv/Conv"Conv*
...

The version is really written as 2.0.0.
Maybe you can check this issue with the PyTorch team.

Thanks.