Cannot analyze ONNX model: INTERNAL ERROR: 'numpy.ndarray' object has no attribute 'get_shape_map'

[GfbGoesBrr]

Hi all, I am trying to benchmark a TCN (causal cnn)-model on an STM32u5, however, when using any of the AI-Tools (online, CubeIDE or CLI), the analysis step fails with the output:

ST Edge AI Core v1.0.0-19894
INTERNAL ERROR: 'numpy.ndarray' object has no attribute 'get_shape_map'

Which is sadly not helpful. The model is attached, it was converted to onnx from pytorch (or rather fastai) using opset 13. It should take an input of 1x500x1 (bs x 1DSample x channels) and give an output of the same shape.

I would be happy to get any pointers on what I can do to fix this, thank you!

[Julian E. · Accepted solution]

Hello @GfbGoesBrr ,

I have confirmed it is an error with the current public version of ST Edge AI Core. It is fixed in the upcoming version that should come out mid-December. 

Unfortunately, you can't do anything until the new version of the ST Edge AI Core comes out...

Have a good day,

Julian

[GfbGoesBrr]

Thank you for the effort!

I removed a somewhat custom part of the model definition, now it works already.

Side note: Giving the option to have more usefull logs would be appreciated, I tried setting the verbosity to different values but it does not change the behavior. Online and in Cube IDE the verbosity is always set to a specific level too, even if I try to overwrite it using the user flag field.