###
计算机系统应用:2018,27(8):49-55
本文二维码信息
码上扫一扫!
基于统计分析的卷积神经网络模型压缩方法
杨扬, 蓝章礼, 陈巍
(重庆交通大学 信息科学与工程学院, 重庆 400074)
Convolution Neural Network Model Compression Method Based on Statistical Analysis
YANG Yang, LAN Zhang-Li, CHEN Wei
(School of Information Science & Engineering, Chongqing Jiaotong University, Chongqing 400074, China)
摘要
图/表
参考文献
相似文献
本文已被:浏览 27次   下载 27
投稿时间:2017-12-10    修订日期:2018-01-04
中文摘要: 针对卷积神经网络中卷积层参数冗余,运算效率低的问题,从卷积神经网络训练过程中参数的统计特性出发,提出了一种基于统计分析裁剪卷积核的卷积神经网络模型压缩方法,在保证卷积神经网络处理信息能力的前提下,通过裁剪卷积层中对整个模型影响较小的卷积核对已训练好的卷积神经网络模型进行压缩,在尽可能不损失模型准确率的情况下减少卷积神经网络的参数,降低运算量.通过实验,证明了本文提出的方法能够有效地对卷积神经网络模型进行压缩.
Abstract:Aiming at the problem of convolutional layer parameter redundancy and low operation efficiency in convolutional neural network, a convolution neural network (CNN) model compression method based on statistical analysis is proposed in this paper. On the premise of ensuring a good ability of convolutional neural network to process information, the well-trained convolution neural network model is compressed by pruning the convolution kernels which have less influence on the whole model in the convolution layer, meanwhile, reducing the parameters of CNN without losing the model accuracy so as to reduce the amount of computation. Experiments show that the proposed method can effectively compress the convolution neural network model while maintaining a good performance.
文章编号:     中图分类号:    文献标志码:
基金项目:重庆市基础科学与前沿技术研究专项项目(cstc2016jcyjA1953)
引用文本:
杨扬,蓝章礼,陈巍.基于统计分析的卷积神经网络模型压缩方法.计算机系统应用,2018,27(8):49-55
YANG Yang,LAN Zhang-Li,CHEN Wei.Convolution Neural Network Model Compression Method Based on Statistical Analysis.COMPUTER SYSTEMS APPLICATIONS,2018,27(8):49-55

用微信扫一扫

用微信扫一扫