基于层融合特征系数的动态网络结构化剪枝  

Dynamic Network Structured Pruning via Feature Coefficients of Layer Fusion

作　　者：卢海伟 袁晓彤[1] LU Haiwei;YUAN Xiaotong(School of Automation,Nanjing University of Information Science and Technology,Nanjing 210044)

机构地区：[1]南京信息工程大学自动化学院

出　　处：《模式识别与人工智能》2019年第11期1051-1059,共9页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.61876090,61936005);国家新一代人工智能重大项目(No.2018AAA0100401)资助~~

摘　　要：剪枝是一种减少模型复杂度的有效方式,现有的剪枝方法仅考虑卷积层对特征图的影响,无法准确判断冗余滤波器,文中提出基于层融合特征系数的动态网络结构化剪枝方法.同时考虑卷积层和批规范化(BN)层对输出特征图的影响,利用多动态参数确定滤波器的重要性,动态寻找冗余滤波器,获取最优网络结构.在CIFAR-10、CIFAR-100数据集上的实验表明,无论是残差网络还是轻量型网络,文中方法在采用较大剪枝率时仍能保持较高精度.Pruning is an effective way to reduce the complexity of the model.In the existing pruning methods,only the influence of the convolutional layer on the feature map is taken into account,and therefore the redundant filter cannot be determined accurately.In this paper,a dynamic network structured pruning method based on layer fusion feature coefficients is proposed.Considering the influence of convolutional layer and Batch Norm layer on the feature map,the importance of the filter is determined by multiple dynamic parameters,and the redundant filter is dynamically searched to obtain the optimal network structure.Experiments on the standard datasets of CIFAR-10 and CIFAR-100 shows that both the residual network and the lightweight network maintain high accuracy while using large pruning rates.

关 键 词：模型复杂度 层融合特征系数 结构化剪枝 动态参数 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]