ML-KWS-for-MCU
ML-KWS-for-MCU copied to clipboard
Quantization. How to decide the best settings for act_max parameters?
Hi All, I use quant.py to scan the CONV1, CONV2....FC 's input/output. Can anyone tell me how to choose the best setting and why? Thank you.
My example is -- act_max's one content with differnt settings get different "Train/Validation/Test accuracy rates" as below,
128: 86.17%, 83.73%, 83.54% 64: 95.62%, 93.30%, 93.33% 32: 95.82%, 93.41%, 93.62% 16: 95.79%, 93.41%, 93.89% 8: 95.77%, 93.45%, 93.64% 4: 95.89%, 93.27%, 93.62% 2: 94.01%, 91.32%, 91.35%