基于聚类和生成对抗学习模型的滤波器剪枝  

FPCC-GAN: CLUSTER CENTER AND GENERATIVE ADVERSARIAL LEARNINGIN FILTER LEVEL PRUNING

作　　者：冯叶棋 张俊三[1] 邵明文[1] 张世栋 Feng Yeqi;Zhang Junsan;Shao Mingwen;Zhang Shidong(College of Computer Science and Technology,China University of Petroleum,Qingdao 266580,Shandong,China;State Grid Shandong Electric Power Research Institute,Jinan 250003,Shandong,China)

机构地区：[1]中国石油大学(华东)计算机科学与技术学院,山东青岛266580 [2]国网山东电科院,山东济南250003

出　　处：《计算机应用与软件》2024年第1期253-260,共8页Computer Applications and Software

基　　金：国家自然科学基金项目(61673396);中央高校基本科研业务费专项资金项目(20CX05019A);中石油重大科技项目(ZD2019-183-004)。

摘　　要：深度神经网络过深的网络架构和冗余的参数会导致昂贵的计算成本,近年来深度神经网络的压缩与加速已成为研究热点。针对现有方法的范数准则局限性以及标签依赖问题,提出一种基于聚类中心和生成对抗学习的结构化滤波器剪枝方法(FPCC-GAN):使用K-means聚类算法按卷积层将滤波器逐层聚类;比例化修剪各簇内离聚类中心较近的提取冗余特征的滤波器;使用生成对抗学习迭代训练。实验结果分析表明,与当前主流方法相比,该方法具有更高的准确率。The deep architecture and parameter redundancy of deep neural network will lead to high computational cost.Deep neural network compression and acceleration has become an important issue in recent years.To address the norm-criterion limitation and label dependence of current methods,we propose a structured filter pruning method based on cluster center and generative adversarial learning(FPCC-GAN).(1)The filters were clustered by K-means clustering algorithm for every convolution layer.(2)Filters closer to the cluster center were pruned proportionally,which extracted redundant features.(3)Generative adversarial learning was used for iteratively training.The experimental results show that compared with current mainstream methods,the proposed method has higher accuracy.

关 键 词：网络压缩 深度神经网络加速 参数剪枝 聚类 生成对抗学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]