Showing results for 
Search instead for 
Did you mean: 

How to deploy a quantized model on an embedded platform

Associate III

Hello everyone, 

I have a pytorch model (mymodel.pth) obtained by running the mixed-precision quantization of this algorithm: * I followed the instructions in the readme section.

Python models are not natively supported by X-CUBE-AI so I convert mymodel.pth to ONNX.

But if I open mymodel.onnx with Netron, I can see that each convolutional layer has 2 inputs (see image) and STM32CubeIDE doesnt' support this. 


* Reading the paper, I know that the quantized model obtained by Q-PPG has been deployed on STM32WB55. So how can I deploy the model on the embedded platform?

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?

ST Employee

in STM32Cube.AI for quantized ONNX networks we support the QDQ format.

The current recommended flow is to export the pytorch float model to ONNX and then use the ONNX Runtime post training quantization to quantize the model.

You can refer to the embedded documentation for details on how to perform quantization and the supported parameters

Generally the packs are located in your home and this chapter is located here: