cancel
Showing results for 
Search instead for 
Did you mean: 

Error with ST Developper Cloud AI

R.SOLER
Associate

Hello,

The ST developper cloud website returns an error message when downloading the CubeIde project, and only on large neural networks. The error seems to be due to the api (https://stm32ai-cs.st.com/api/8.0.1/project/) returning a Gateway timeout 504 due to an internal server problem. The error is found on ST developper cloud but also via the api directly. It seems to work with the 7.3.0 version but I never succeeded with the 8.0.1 version.

Attached is a screenshot of the error and the tflite that generated it. The tflite is not corrupted and can be validated on a computer.

What can be the issue ?

R.SOLER

ALTEN


_legacyfs_online_stmicro_images_0693W00000dDVBgQAO.png

3 REPLIES 3
hamitiya
ST Employee

Hello,

A fix has been deployed today in order to fix your issue. You should now be able to generate a project for your larger models (with a run up to 5min)

Best regards,

Yanis

Hello, 

Thank you for the fix regarding the timeout.

An other issues also happens when using the ST developper cloud. When unselecting the option " Use activation buffer for input buffer (--allocate-inputs) ", the benchmark can be launched but no inferences times are prompted, and while generating the cubeide project an error message is prompted. Here is a tflite if you want to try it out.

Alex_0-1687512226050.png

 

hamitiya
ST Employee

Hello Alex,

Thanks for your request.

Without "allocate-inputs" or "allocate-outputs", your model requires way more RAM to fit on STM32 targets. When not using these options, we try to fit inputs and outputs in default internal memory. For your model it needs:

- Input/Output: 768 KiB / 1.04 MiB

As we show in Dev. Cloud benchmark panel, we don't have a board with this kind of internal memory, with the exception of STM32H7B3I-DK which can run your model (fyi: version 8.1.0, inference time: 3124ms).
It will not run if you uncheck allocate-outputs since you will need 1MB of contiguous memory.

Best regards,

Yanis