Does STM32MP257 support MMAction2 with TSN or SlowFast behavior recognition? 3D CONV?

[fanronghua0123456]

Can I use MMAction2 with TSN or SlowFast behavior recognition models and deploy them to the STM32MP257? Does the STM32MP257 support 3D CONV operators?

[Julian E.]

Hello @fanronghua0123456,

Here is the list of supported layers for the MP2:

https://wiki.st.com/stm32mpu/wiki/STM32MP2_NPU_description#Operation_support

There is indeed support for the CONV3D with the operation VSI_NN_OP_CONV3D with NPU in 8bits and on the GPU in 16bits.

For the model MMAction, it would be required to quantize the model in fflite or onnx for it to work on the HW. We don't have the description of the model, but it seems similar to this, if you want to take a look:

https://wiki.st.com/stm32mpu/wiki/How_to_deploy_your_NN_model_on_STM32MPU

Have a good day,

Julian

[fanronghua0123456]

Hello @Julian E.

 Thanks ,

How do I evaluate the operational efficiency of the model on the development stm32mp25x board? Currently, the model trained with yolov11 has a very large inference speed after quantization. I would like to know how long each network layer of my model takes to infer, so that we can optimize the model structure better.

[Julian E.]

Hello @fanronghua0123456,

I think the ST Developer Cloud could help you: https://stedgeai-dc.st.com/session

You can import model, select the MP2 and benchmark its inference time on a real board in our board farm

You can also do it manually with the validate command: https://stedgeai-dc.st.com/assets/embedded-docs/stm32mpu_command_line_interface.html 

Have a good day,

Julian

[fanronghua0123456]

hello,@Julian E.

thanks, so I want to know  details of this duration times . and  how long each network layer of my model takes to infer?

 thanks.

[Julian E.]

Hello @fanronghua0123456,

I asked the dev team, but unfortunately, we don't have such tool for the MP2.

Have a good day,

Julian

[fanronghua0123456]

Because the performance of each GPU and NPU is different, when we want to deploy the model on embedded devices, the inference speed of each device is different. Therefore, I believe you must have a tool to analyze the inference speed of each layer of the model's network.

[Julian E. · Accepted solution]

Hello @fanronghua0123456,

The dev team probably have such tools, but I cannot share them.

Have a good day,

Julian