PROCESSOR-SDK-J721E: Problem with converting tidl model.

Please follow the steps outlined in below doc to identify the functional mismatch issue

https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_osr_debug.md

Also refer below for quantization related parameters

https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_fsg_quantization.md

 Once you identify the mis matching layers, Please the details of the layer to help on this

Hi Kumar,

What do you mean by your last sentence?

Regards,

Josel

The steps in the document helps in identifying the layer which is cause the result mismatch.

Once we know the layer configuration and steps to reproduce the issue, we can suggest the fix or follow up

Hi Kumar,

thanks for your attention. I have compared feature maps with different psdk versions and quantitative scales (int8 or int16). For the int16 model, I can get the almost same tensor result, only the diff in the int8 model, all the comparisons show in 4 figures.Both the int8 model look terrible( Maybe, the psdk71-int8 model gets well target is a coincidence?), in addition, I use the same tidl_import_*.txt file to the ./out/tidl_model_import.out and I have QAT.Do you have any other suggestions, thank you so much.

Linxin





Regards,

Josel

Can you please try with calibrationOption = 64 in import config file?

Hi Kumar,

Thank you very much for your suggestion.

I have tried to set calibrationOption=64, it seemed not effective when I execute ./out/tidl_model_import.out tidl_import_model.txt


I found the details of calibrationOption in psdk Official documents. It seems like only 0,1,7,13,16,32 could be set. Here is my tidl_import_model file.


Please advise

Regards,

Josel

Could you please share the complete model import log with debugTraceLevel=3

Hi Kumar,

I have attached the relevant documents.
convert_log_20230415.txt
1200.model.zip

Regards,

Josel

Below prints are not expected with calibrationOption = 64 in SDK 8.6. Could you please try sing the SDK 8.6?

Warning :: Tensor Scale for layer 72 is reduced by 1 Bits
Warning :: Tensor Scale for layer 120 is reduced by 1 Bits
Warning :: Tensor Scale for layer 121 is reduced by 1 Bits
Warning :: Tensor Scale for layer 123 is reduced by 3 Bits
Warning :: Tensor Scale for layer 161 is reduced by 1 Bits
Warning :: Tensor Scale for layer 162 is reduced by 1 Bits
Warning :: Tensor Scale for layer 164 is reduced by 4 Bits
Warning :: Tensor Scale for layer 195 is reduced by 1 Bits
Warning :: Tensor Scale for layer 196 is reduced by 1 Bits
Warning :: Tensor Scale for layer 199 is reduced by 2 Bits
Warning :: Tensor Scale for layer 200 is reduced by 1 Bits
Warning :: Tensor Scale for layer 201 is reduced by 2 Bits

Hi Kumar,

I am required to use SDK8.2. Should I try to convert the 8.2 model again, or can I use the model converted by psdk8.6 for version psdk8.2?

Regards,

Josel

You can convert the model with 8.6 validate the accuracy in Host emulation mode, if the issue that you are observing is solved, then you can update your PSDK 8.2 to use the TIDL version 8.6