Which Intermediate Representation Version needed - ONNX & X-Cube-AI/ST Edge AI Cloud

[MoCo]

Dear Sir or Madam,

I'm trying to import a ONNX network with X-Cube-AI(9.1)/ST Edge AI Cloud which was generated and exported using Matlab/Simulink. When trying to work with the model I always receive the following error:

[... ST Edge AI Core v1.0.0-19894 
INTERNAL ERROR: Your model ir_version is higher than the checker's.

...] - same error in X-Cube-AI(9.1) and ST Edge AI Cloud.

The intermediate representation version of the Matlab/Simulink generated ONNX Model is 9 and I choose OpsetVersion to be 13 (recommendation somewhere in the docs). Attached is an example of the file I would like to import.

Now it would be interesting for me to know which intermediate representation version is required by X-Cube-AI(9.1)/ST Edge AI Cloud? Then I can convert the onnx file to the correct ir version.

Which ONNX version is the basis for X-Cube-AI(9.1)/ST Edge AI Cloud?

Have a nice day and best regards,

MoCo

STM32CubeAI  ST Edge AI Developer Cloud 

[hamitiya · Accepted solution]

Hello @MoCo ,

You can find all the embedded versions with this command:

$ ./stedgeai.exe --tools-version
stedgeai - ST Edge AI Core v1.0.0-19894
- Python version : 3.9.13
- Numpy version : 1.23.5
- TF version : 2.15.1
- TF Keras version : 2.15.0
- ONNX version : 1.10.2
- ONNX RT version : 1.17.3

I tried your model with our development version and it is working as expected.

Could you please try to downgrade your op_set to see if the behavior change ?

For your information, with our development version:

Best regards,

Yanis

[MoCo]

Dear Yanis,

thanks a lot for your great help!

I was able to retrieve the version of ONNX which is important to allow conversion.

stedgeai - ST Edge AI Core v1.0.0-19894
- Python version : 3.9.13
- Numpy version : 1.23.5
- TF version : 2.15.1
- TF Keras version : 2.15.0
- ONNX version : 1.10.2
- ONNX RT version : 1.17.3

I tried to change the opset down to 6 (in steps of one) from the current 13 - but that changes nothing.

Under https://onnxruntime.ai/docs/reference/compatibility.html you can see that  IR version 9 was introduced with ONNX version 1.14. This means that with the current ONNX 1.10.2 we have in tIR version 8 you require with stedgeai - is that true? 

Then my path would be to install ONNX version 1.14/1.15 an try to downgrade the IR version from 9 to 8.

I have only one remaining question - can you give me an rough estimate when the release of the new version will happen (is a beta test possible)?

Many thanks and best regards

MoCo

[hamitiya]

Hello @MoCo ,

If op_set does not correct your issue, then you may try to downgrade your ONNX install. As I can see, IR_VERSION is set to 8 in our installation:

From onnx_ml_pb2.py embedded in ONNX package:

IR_VERSION = 8

 For the next release, I cannot say much but I can share you a hint, so stay tuned !

https://www.st.com/content/st_com/en/events/stm32-summit.html

Best regards,

Yanis

[MoCo]

Hello Yanis,

thanks a lot - you and the team developing this software are doing are great job.

I'm looking forward to the release.

I found a simple fix - I installed Matlab R2024a and did the export with it (ir version = 7) - that works great.

Best regards and have a nice rest of the week,

MoCo