[X-CUBE-AI] Error while analyzing model

[thetectivestm]

Hello everyone,

when I try to analyze my model (mynetwork.onnx) clicking on Analyze in the target selection window an error message occurs without any details. 

After the selection of the target in the configuration area, if I analyze the same model I receive this error: "NOT IMPLEMENTED: Convolution with multiple inputs not supported".

The model mynetwork.onnx is obtained by converting a pytorch model obtained by running an algorithm downloaded from github. 

I use Windows 10 Home 64 bit (10.0, build 19045), STM32CubeIDE 1.13.1 and X-Cube-AI 8.1.0 (I've also tried 8.0.1).

Could anyone please help me?

[fauvarque.daniel]

Can you share the model so we can reproduce the issue.

Thanks in advance.

Regards

Daniel

[thetectivestm]

I converted the pytorch model on Colab in this way:

example_input = torch.randn(1, 4, 256)

torch.onnx.export(my_model, example_input, onnx_path, verbose=True)

because my input has width=1, height=256, channels=4 and I know the example_input must be in this format (N=1, channels, width*height). 

[fauvarque.daniel]

Thanks for the sharing. 

could you share the onnx file or does it contains information that you can't share ?

It will speed up the process for us.

Regards

[thetectivestm]

I can't attach that file because the file type .onnx is not supported and the content is not perfectly readable from a .txt.

[fauvarque.daniel]

You should be able to zip it before and attach the zip

Thanks

[thetectivestm]

Yes, right.

[fauvarque.daniel · Accepted solution]

Thanks a lot I've reproduced the limitation, as you can see you have in the model convolution layers with 2 inputs. 

We'll see if this can be supported

[thetectivestm]

My input is just a sequence (width=1, height=256, channels=4).

What do you mean by "Convolution layers with 2 inputs"?

Thanks in advance.

[fauvarque.daniel]

If you open the onnx file with Netron you'll see that you have some Conv layers with 2 inputs

Regards