TDA4VM: Prequantized TFLite model on TDA4

Ashay Mulye

Part Number: TDA4VM

Prequantized TFLite model on TDA4

Hi,

I read in the github repo of edge ai tidl that it is possible to import pre-quantized TFLite model directly in edge ai tidl workflow. However, some more description would be nice.

For example:

is the model supposed to be compiled again and generate corresponding artifacts (direct inference on this model does not work, it expects "allowednode.txt")
if we can run direct inference, does it expect some other flag during inference which will indicate that this is a pre-quantized model.

Regards

Ashay

over 2 years ago

0 Luca Vicenzi over 2 years ago

Prodigy 80 points

I think you need to compile it in order to generate the model artifacts

0 Ashay Mulye over 2 years ago in reply to Luca Vicenzi

Prodigy 149 points

Thanks, would you happen to know, if it supports both Post training quantization or just model trained with quantization-aware training?

Best

Ashay

0 Varun Tripathi over 2 years ago in reply to Ashay Mulye

TI__Genius 11355 points

Hi,
You can set the quantization_scale_type option appropriately to run PTQ in TIDL for a TF-Lite model as well

Regards,
Varun

0 Ashay Mulye over 2 years ago in reply to Varun Tripathi

Prodigy 149 points

Hi,

thanks for your response. I tried this option but the model could not be compiled with this option. I get the following error:

Import Error: TFL PTQ Model with all ranges not supplied

Best

Ashay

Processors

Processors forum

TDA4VM: Prequantized TFLite model on TDA4