search

ARM-software / ML-KWS-for-MCU

Keyword spotting on Arm Cortex-M Microcontrollers
Apache License 2.0
1.13k stars 417 forks source link

quant_test with different act_max parameters but produces same content weights.h ? #122

Closed jessonwtp closed 4 years ago

jessonwtp commented 4 years ago

Hi All,

When I use quant_test.py to tune act_max parameters (one by one with value 8, 16, 32), I can see the training accuracy, validation accuracy and test accuracy are changing according to different act_max parameters. But, why I got the same weights.h ? Is it reasonable ? I can get good accuracy by test.py and label_wav.py. But when I use this quanted weights file to my embedded device, the test result is great different. that's why I think maybe caused by quant function.

It's appreciate if anyone knows the reason and share your experience. Thanks.

ShintaZachariah commented 4 years ago

Hi All, what is act_max parameter?

jessonwtp commented 4 years ago

Hi Shintzchariah, act_max is followed by 12 numbers when execute quant_test on ds_cnn model. If DNN mode, act_max is followed by 5 numbers. You can refer below linking. https://github.com/ARM-software/ML-KWS-for-MCU/blob/master/Deployment/Quant_guide.md

jessonwtp commented 4 years ago

I'm trying the quant bias shift and out shift which set in the embeded Device side.