基于损失变化的CNN混合精度量化方法  

Mixed precision quantization method of CNN based on loss variation

作　　者：何益智 李钊[1] 李鉴柏 张少爽 刘文龙 HE Yi-zhi;LI Zhao;LI Jian-bo;ZHANG Shao-shuang;LIU Wen-long(College of Computer Science and Technology,Shandong University of Technology,Zibo 255000,China;School of Electronic and Electrical Engineering,Zibo Vocational Institute,Zibo 255000,China)

机构地区：[1]山东理工大学计算机科学与技术学院,山东淄博255000 [2]淄博职业学院电子电气工程学院,山东淄博255000

出　　处：《计算机工程与设计》2024年第2期571-577,共7页Computer Engineering and Design

基　　金：国家重点研发计划基金项目(2022YFE0107300)。

摘　　要：针对卷积神经网络(convolutional neural network,CNN)在存储和计算资源有限的边缘设备中难以部署应用的问题,提出一种基于损失变化的混合精度量化方法,以低位宽定点数代替全精度浮点数进行运算,降低网络所需资源。根据每个量化层的一阶和二阶信息指导位宽分配,采用K-means方法将量化层聚类成块,降低位宽策略搜索空间。提出一种自适应搜索方式,根据历史策略训练结果自行调整搜索状态。重新整合量化训练过程,减少传统量化训练中计算量。实验结果表明,采用所提方法可在CNN模型推理损失精度较小的前提下,有效压缩模型。Aiming at the problem that convolutional neural networks are difficult to deploy and apply in edge devices with limited memory and computing resources,a mixed precision quantization method based on loss variation was proposed.Resources were reduced using a lower bit fixed-point number instead of float precision.The bit widths were adjusted based on the first and second order information of each quantization layer,and the layers were clustered into blocks using K-means to reduce the search space.An adaptive strategy search method was proposed,the search state was adjusted according to the historical strategy results.The process of quantization training was reintegrated to reduce computation.Experimental results show that the proposed method can effectively compress the CNN with a small inference loss accuracy.

关 键 词：卷积神经网络 混合精度量化 损失变化 K-MEANS聚类 敏感度分析 自适应搜索 模型压缩 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]