TDA4VM: product board output is different from X86 simulation

Hi,

Please expect a delay in responses due to US holidays, thank you for your patience!

Best Regards,
Sudheer

Hi zhenzhou,

Can you try sdk 11.0? If sdk 11.0 can solve this problem, we can get the fix faster. 

Regards,

Adam

My colleague is tring sdk 11.0，it need some time

Hi zhenzhou,

Do you see any progress in this issue?

Regards,

Adam

we found sdk 11.0 has has lower diff comapred with 9.2, but how can we fix this in version 9.2, can you share us a patch？ 

How to train our model to reduce quantization loss and improve board performance？

Hi Zhenzhou,

You can try different quantization options: https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_fsg_quantization.md#b2-advanced-calibration It has different effect on different models.

You can also try setting certain layers to 16bit if you can found oud which layer causes the major loss. 

If all these steps does not solve your problem. Please try QAT: https://github.com/TexasInstruments/edgeai-tensorlab/blob/3de61dfa503c408346c3bcd029f49a25e42a8a73/edgeai-modeloptimization/torchmodelopt/edgeai_torchmodelopt/xmodelopt/quantization/README.md#option-2-quantization-aware-training-qat--post-training-calibration-ptc-using-this-repository 

Regards,

Adam

zhenzhou zhenzhou said:

but how can we fix this in version 9.2, can you share us a patch

Hi 

As for this issue, we have a patch now but please wait me to verify this.

Regards,

Adam

hello， Adam Hua

 now we change sdk version to 11.0,  but we found the diff between X86 simulation and product board is still existing. can you check it?

Sure, still the onnx model in this zip file?

zhenzhou zhenzhou said:

shareti.zip

May I also know what tidl version are you using? 11_00_06_00 or 11_00_07_00?

Hello zhenzhou

I compared layer traces on pc an evm on sdk 11.0. I don't see mismatch:

Can you try this as well on your side?

Regards,

Adam

ok, when our board software is ready, i will try.