Forums - Model quantification

2 posts / 0 new
or Register
Last post
Model quantification
MarkWu
Join Date: 8 May 22
Posts: 37
Posted: Wed, 2022-08-17 04:11
Top

Hi,

I encountered some questions and wanted to ask. I encountered a headache.
 
After I quantized the dlc model to int8, I ran it on the dsp. The accuracy of the int8 model is relatively low, so it cannot be used,
But the non-quantized model is used on the dsp. The accuracy of running it is very high, but the time comparison of loading the non-quantized model on the dsp is about 3s, how can I make the loading time of the non-quantized model on the dsp shorter?
Or how can I quantize the model so that the int8 model can achieve the accuracy of the non-quantized model?
 
My device is Snapdragon 8Gen,snpe version  is 1.64, looking forward to your answer
  • Up0
  • Down0
Re: Model quantification
SahilBandar
Join Date: 23 May 18
Posts: 37
Posted: Fri, 2022-08-19 09:12
Top

Hi Smith,

If you pass on the non-quantized model to DSP, SNPE takes care of converting it to quantized internally. 
This is the something, which may cause the delay in loading the model.

For pre-quantized model, make sure that the input data you passed is uint8.

Regards,
Sahil Bandar

  • Up0
  • Down0
or Register

Opinions expressed in the content posted here are the personal opinions of the original authors, and do not necessarily reflect those of Qualcomm Incorporated or its subsidiaries (“Qualcomm”). The content is provided for informational purposes only and is not meant to be an endorsement or representation by Qualcomm or any other party. This site may also provide links or references to non-Qualcomm sites and resources. Qualcomm makes no representations, warranties, or other commitments whatsoever about any non-Qualcomm sites or third-party resources that may be referenced, accessible from, or linked to this site.