ST Edge AI Core - Conversion error for STM32MP2

[NilsW]

Hello, 

I am trying to convert my Onnx model with the ST Edge AI Core tool to the .nb format for the STM32MP25 target.
While converting the model i get the following error:

ST Edge AI Core v2.2.0-20266
 
INTERNAL ERROR: unsupported operand type(s) for *: 'int' and 'shape_struct'

I could track down the error to a resize operation in the Onnx model. Here is a short snippet how the model looks like:

When i remove the resize operations and downscale the 28 x 28 patches to 14 x 14 patches. Then the error is gone.
So i think that the resize operation is the problem. 

This happens for quantized and non quantized models. 

Like stated here (STM32MP2 NPU description - stm32mpu), there should be some resize operations implemented at least on the GPU. 

Did anyone had similar issues and know how to fix them?
Thank you in advance for your help.

Best regards
Nils

[Julian E.]

Hi @NilsW,

 

could you please share the model?

Maybe a untrained version of it if you don't want to disclose everything.

 

Without it, it is difficult to analyze it.

 

Have a good day,

Julian

[NilsW]

Hi @Julian E., 

thanks for your reply.

Sure, i have attatched the model files. It's basically an MobilenetV2 but for the outputs i am using the blocks 2 and 4. I wanted to resize those patches to a common size and concat them.

Best regards
Nils

[Julian E.]

Hi @NilsW,

 

We confirmed the bug. This comes from the NPU.

We opened a ticket at Verisilicon, our NPU supplier.

I'll update you if we get an answer.

 

Have a good day,

Julian

[NilsW]

Hi  @Julian E. 

Are there any updates regarding the bug with the conversion since the last communication? Please let me know if there have been any changes.

Best regards
Nils

[Julian E.]

Hi @NilsW,

 

We received an update from the VSI ticket. We are testing it and looking at how we can deliver it to users.

 

Have a good day,

Julian

 

[NilsW]

Hi @Julian E. 

Thank you for the update!

Best regards
Nils