基于均方误差的8位深度神经网络量化  

8-BIT deep neural network quantization based on mean square error

作　　者：冯鹏程 禹龙[2,4] 田生伟[3,4] 耿俊 龚国良[5] FENG Peng-cheng;YU Long;TIAN Sheng-wei;GENG Jun;GONG Guo-liang(College of Information Science and Engineering,Xinjiang University,Urumqi 830046,China;Network Center,Xinjiang University,Urumqi 830046,China;School of Software,Xinjiang University,Urumqi 830008,China;Key Laboratory of Software Engineering Technology,Xinjiang University,Urumqi 830008,China;Laboratory of High-Speed Circuit and Artificial Neural Networks,Institute of Semiconductors,Chinese Academy of Sciences,Beijing 100083,China)

机构地区：[1]新疆大学信息科学与工程学院,新疆乌鲁木齐830046 [2]新疆大学网络中心,新疆乌鲁木齐830046 [3]新疆大学软件学院,新疆乌鲁木齐830008 [4]新疆大学软件工程技术重点实验室,新疆乌鲁木齐830008 [5]中国科学院半导体研究所高速电路与神经网络实验室,北京100083

出　　处：《计算机工程与设计》2022年第5期1258-1264,共7页Computer Engineering and Design

基　　金：新疆维吾尔自治区科技援疆基金项目(2020E0234);国家自然科学基金项目(U2003208、61962057、U19A2080、61563051、61662074);北京市科技计划基金项目(Z181100001518006);中国科学院STS计划基金项目(KFJ-STS-ZDTP-070);新疆自治区科技人才培养基金项目(QN2016YX0051)。

摘　　要：为模型量化后具有更高的准确度,提出以量化均方误差(QMSE)为指标的确定量化系数的方法,针对量化后性能损失严重的小型网络,进一步提出更新统计参数(USP)的方法。QMSE将量化过程中的舍入和截断操作产生的噪声相结合,以此作为选取合适量化系数的指标;USP通过更新批次归一化层中的均值和方差,矫正模型量化产生的均值和方差偏移。实验结果表明,在不进行重训练的情况下,使用QMSE+USP对常见的深度神经网络量化,模型性能优于其它算法。To achieve higher accuracy after the model is quantized,a method to determine the quantization scale with the quantized mean square error(QMSE)as an indicator was proposed.For small models with severe performance loss after quantization,the update of statistical parameters(USP)was further proposed.QMSE combined the noise generated by the rounding and truncation operations in the quantization process as an indicator for selecting appropriate quantization scale.USP corrected the mean and variance generated by model quantization by updating the mean and variance in the batch normalization layer.Experimental results show that using QMSE+USP to quantify common deep neural networks without retraining,the model perfor-mance is better than using other algorithms.

关 键 词：深度神经网络 模型压缩 量化 卷积神经网络 均方误差 

分 类 号：TP305[自动化与计算机技术—计算机系统结构]