摘要
在模型压缩中,单独使用权重剪枝或卷积核剪枝对卷积神经网络进行压缩,压缩后的模型中仍然存在较多冗余参数.针对这一问题,提出了一种结合权重剪枝和卷积核剪枝的混合剪枝方法.首先,剪除对卷积神经网络整体精度贡献较小的卷积核;其次,对剪枝过的模型再进行权重剪枝实现进一步的模型压缩.在剪枝过程中通过重新训练来恢复模型精度.在MNIST和CIFAR-10数据集上的实验结果表明,提出的混合剪枝方法在几乎不降低模型精度的前提下,将LeNet-5和VGG-16分别压缩了13. 01倍和19. 20倍.
The compressed convolutional neural network that only using weight or filter pruning still exists redundant parameters. A method of combining weight pruning and filter pruning is proposed. Firstly,the filters with small effect on the output accuracy are removed.By removing the whole channels in the network together with their connecting filters,the computational costs are reduced significantly.After obtaining the pruned model,weight pruning is performed for further compressing. Experiments on the MNIST and the CIFAR-10 datasets indicate that the mixed pruning is effective and feasible. The proposed method achieves 13. 01 x compression on LeNet-5 and 9. 20 x compression on VGG-16 without obviously accuracy decline.
作者
靳丽蕾
杨文柱
王思乐
崔振超
陈向阳
陈丽萍
JIN Li-lei;YANG Wen-zhu;WANG Si-le;CUI Zhen-chao;CHEN Xiang-yang;CHEN Li-ping(School of Cyber Security and Computer,Hebei University,Baoding 071002,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2018年第12期2596-2601,共6页
Journal of Chinese Computer Systems
基金
河北省自然科学基金项目(F2015201033F201701069)资助
"云数融合
科教创新"基金课题项目(2017A20004)资助
关键词
卷积神经网络
模型压缩
网络剪枝
混合剪枝
convolutional neural network
model compression
network pruning
mixed pruning
