Problem when moving a github project model to stm32cube AI

[uveuvenouve]

First , the github project is this
https://github.com/Xiaobin-Rong/gtcrn/tree/main

It has converted to onnx and tensorRT for ppl to use
but when i load the onnx model into stm32 for optimize/import (setting , compression : none , optimization : balance)
it just jumps out an error

X-cube-AI version is 9.1.0

ST Edge AI Core v1.0.0-19894 
INTERNAL ERROR: Trying to remove node_339 which has multiple inputs

is it because the multi input and output network limitation or something else?

and if i success to import this model
how to deal with the multi dimensional  tensor arrays that the github author demo ?

if you want to test on the model
there is the link for the onnx and tensorRT model
https://github.com/Xiaobin-Rong/TRT-SE/tree/main/online/models

Thanks in advance

Uveuvenouve

[Julian E. · Accepted solution]

Hello @uveuvenouve,

The issue here is that bidirectional LSTM are not supported in ST Edge AI

You can see below that your node_339 is the bidirectional LSTM:

You can find the list of supported ONNX operator in the ONNX toolboox page in the CubeAI Documentation:

https://wiki.st.com/stm32mcu/wiki/AI:X-CUBE-AI_documentation#X-CUBE-AI_embedded_documentation_access 

Concerning the LSTM:

Have a good day,

Julian

[uveuvenouve]

Thank you

[uveuvenouve]

Hi Julian,

Do you know any way to use this model on stm32 btw?

Uveuvenouve

[Julian E.]

Hello @uveuvenouve,

Sadly, with the current state of the ST Edge AI Core, you can't.

If you are able to edit the model and only use supported operator, then it would work. But it probably requires a lot of work, and it will not be the same model anymore...

In the future, the people working on the ST Edge AI Core might add it, but it won't be soon sadly.

Best Regards,

Julian