其他分享
首页 > 其他分享> > INT16乘累加防溢出量化训练——Overflow-aware Quantization

INT16乘累加防溢出量化训练——Overflow-aware Quantization

作者:互联网

简介

image.png

 

基本原理

MNN OAQ

量化方式

小米6

小米9SE

K20Pro

ARMv7

(32位)

ARMv8.1

(64位)

ARMv7

(32位)

ARMv8.1

(64位)

ARMv7

(32位)

ARMv8.1

(64位)

FP32

74.89ms

66.56ms

120.4ms

90.33ms

31.20ms

29.25ms

Normal INT8

66.73ms

51.08ms

155.8ms

114.4ms

36.08ms

35.03ms

OAQ INT8

54.58ms

(1.22倍)

43.87ms

(1.16倍)

95.71ms

(1.63倍)

77.06ms

(1.48倍)

35.39ms

34.71ms

 

PAI量化训练

标签:INT16,ms,Quantization,OAQ,量化,Overflow,累加,溢出
来源: https://blog.csdn.net/nature553863/article/details/112479364