无人驾驶深度学习模型组合剪枝算法  

Combined pruning algorithm for unmanned deep learning model

作　　者：赵丽君 周永军 汤小红[1] 蒋淑霞[1] 董寅宾 廖慕钦 ZHAO Lijun;ZHOU Yongjun;TANG Xiaohong;JIANG Shuxia;DONG Yinbin;LIAO Muqin(School of Mechanical and Electrical Engineering,Central South University of Forestry and Technology,Changsha 410000,China)

机构地区：[1]中南林业科技大学机电工程学院,湖南长沙410000

出　　处：《传感器与微系统》2021年第3期127-129,共3页Transducer and Microsystem Technologies

基　　金：长沙市科技计划资助项目(kq1701102)。

摘　　要：首先对网络模型进行正则化,提高网络泛化能力及模型稀疏度;接着采用组合剪枝算法,先对卷积层卷积核进行删减,再对全连接层权重进行删减。剪枝过程分多次进行,每次剪枝后通过再训练以恢复模型分类识别性能。实验结果表明:提出的组合剪枝算法能够在保证VGG—16网络对车辆数据集分类识别率保持在92.84%的情况下,对网络模型压缩了79%,加速3.38倍。Firstly,the network model is regularized to improve the generalization ability network and sparsity of model.Then,the combined pruning algorithm is adopted.Convolutional layer convolution kernel is firstly deleted,and then the full connection layer weight is deleted.The pruning process is carried out several times,and each time after pruning,it is retrained to restore the model classification and recognition performance.The experimental results show that the combined pruning algorithm can compress the network model by 79%and accelerate by 3.38 times while ensuring that the VGG-16 network keeps the classification and recognition rate of the vehicle dataset at 92.84%.

关 键 词：无人驾驶 卷积神经网络 权重剪枝 卷积核剪枝 组合剪枝 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]