J721EXSOMXEVM: J721EXSOMXEVM

Hi,

Since this thread was inactive, it got locked.

I have checked the thread, and as i see we have requested logs/observations on 9.1 sdk.

Could you please revert the same ? This will help us to cross out sdk version related possibility. 

Hi, Pratik,

I tested the same model and configuration with SDK 9.1, but results are the same.

I send you config file and logs.

config_and_logs_sdk_9_1.zip

Best Regards!

Thanks for sharing the logs.

I will verify the findings at my end and get back to you in 1 week.

Hi, Pratik,

Any update? Our customer is pushing us a lot!

BR

Hi Lachezar,

Apologies for the delay in response.

Due to bandwidth I was not able to look into above issue.

I have checked the logs, since you are pointing to accuracy related issue here are few things you can try.

Please refer to the vast documentation on accuracy debug here : https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_osr_debug.md

I have created jupyter notebooks that would be helpful for layer level accuracy debug (Float 32 Vs Quantized) : https://github.com/kumardesappan/colab-notebooks/blob/main/tidlrt_tensor_debug.ipynb

If you observe any issue in accuracy wrt particular layer, please share the trimmed model(having only that layer) to verify at our end.

Thanks for your patience again.

Hi, Pratik,

Any update?

Hi Lachezar,

Can you please refer my above response.

Thank you