I am trying to follow the "Getting started with X-CUBE-AI Expansion Package for Artificial Intelligence (AI)" document (User Manual UM2526), but I cannot validate the model on target. Note that I'm using the STM32F746-DISCO instead of the recommended NUCLEO-F746ZG, but that's a minor difference.
I can sucessfully validate the model on desktop (using random numbers and real data), but get an error when trying to validate on target using the "automatic compilation and download" option.
CubeMX reports "Automatic build and run failed".
Consulting the STM32CubeMX.log file I see the following error:
2023-09-19 11:57:49,034 [ERROR] STM32CubeIDEBuilder:207 - !MESSAGE The project description file (.project) for 'mxAI1803873821400' is missing. This file contains important information about the project. The project will not function properly until this file is restored.
My project is named STM32F746-DISCO, not the random string of characters shown. Even when I've used the "Generate Code" option to try to export to the CubeIDE, I don't get a .project file associated with 'mxAI1803873821400'. What am I doing wrong?
Thanks!
1 ACCEPTED SOLUTION
Accepted Solutions
Got it working! (Just not using STM32CubeIDE).
Here's what I had to install on my Windows 10 machine to get it working.
- STM32CubeMx v 6.9.2
- Arm Keil MDK v 5.38 Keil MDK (arm.com)
- Arm Compiler 5.06 update 7 (separate install, per these instructions) Can I add Arm Compiler 5 to Keil MDK 5.37 or newer?
- The necessary CMSIS Packs for my board using the Keil Pack Installer and this link: Arm Keil | CMSIS Packs
Once I had all those prerequisites set up, I could select MDK-ARM V5 from the dropdown menu and the automatic compilation and download worked!
As an aside, I would recommend not validating your on-target model against your entire dataset. I tried validation on target for the MNIST test dataset (10,000 samples), and CubeMX would freeze up after 1000 or so. Much better to use the "validate on desktop" option to compute the accuracy of your deployed model, then use "validate on target" with a few hundred (or fewer) samples to estimate actual on-device inference time.
Got it working! (Just not using STM32CubeIDE).
Here's what I had to install on my Windows 10 machine to get it working.
- STM32CubeMx v 6.9.2
- Arm Keil MDK v 5.38 Keil MDK (arm.com)
- Arm Compiler 5.06 update 7 (separate install, per these instructions) Can I add Arm Compiler 5 to Keil MDK 5.37 or newer?
- The necessary CMSIS Packs for my board using the Keil Pack Installer and this link: Arm Keil | CMSIS Packs
Once I had all those prerequisites set up, I could select MDK-ARM V5 from the dropdown menu and the automatic compilation and download worked!
As an aside, I would recommend not validating your on-target model against your entire dataset. I tried validation on target for the MNIST test dataset (10,000 samples), and CubeMX would freeze up after 1000 or so. Much better to use the "validate on desktop" option to compute the accuracy of your deployed model, then use "validate on target" with a few hundred (or fewer) samples to estimate actual on-device inference time.
