search

majianjia / nnom

A higher-level Neural Network library for microcontrollers.
Apache License 2.0
820 stars 235 forks source link

RNN量化 #161

Open wenzhu888 opened 2 years ago

wenzhu888 commented 2 years ago

RNN量化损失相对较大，可以增加支持16bit算子吗？

majianjia commented 2 years ago

16bit正在做，现阶段内部和state用的是16bit，输入输出为了兼容还是8bit