AI model optimization

When trying to run the Body Tracking code for the first time, it’s trying to optimize the model but it never finishes. I am stuck at 90.1%. I left the code run for day but it remained stuck with the same completion %.
I then tried to use the ZED Diagnostic tool to manually optimize the models, but as soon as I start the optimization process for body tracking, the tools closes completely.
It works fine for multi class detection though…

Latest SDK version, CUDA 12.1 running on Windows 10.
Is there a log file or something to see why it is failing?

Thanks

Hi @Aleblanc

Can you try cleaning the models and optimizing them via command line?
Intructions are here: https://support.stereolabs.com/hc/en-us/articles/9747407795223-How-can-I-optimize-the-ZED-SDK-AI-models-manually

I’m aware it’s just a workaround, but this could unlock you.

Hi,

I get the following error : Incompatible CUDNN version, found 8.8, min 8.907.

I just installed version 9.2 but it still gives me the same message, even after a reboot.

Thanks

Hi @Aleblanc

There is probably something wrong with either the SDK or CUDNN install.
Can you uninstall the previous CUDNN version and the one you just installed, clean the $path from them, uninstall the ZED SDK, reboot, reinstall CUDNN, reboot, reinstall the ZED SDK, and check if that works?

Hi,
I uninstalled and reinstalled everything but it still doesn’t work.
Whenever I try to optimize the body tracking model, the program crashes.
For some reason, when running the Diagnostics tools I get an error saying that the object detection is corrupted, even after a fresh install…
Diagnostic log attached.

Thanks
ZED_Diagnostic_Results.json (9.2 KB)

Hi,

I think there is a misunderstanding here. You don’t need to install CUDNN yourself. The correct version is already packaged with the ZED SDK.
Maybe that’s what is causing the issue. Can you try to remove your install of CUDNN, clean the optimized models and try again?

Sorry for the trouble.

Best,

Hi,

It doesn’t work.
When trying to clean the models, I get incompatible Cudnn version.