摘要
为保证YOLO网络在嵌入式设备上正常运行,需采用剪枝算法精简滤波器以减小网络存储空间和计算量,而现有剪枝算法耗时较长且剪枝精度较低。提出一种基于参数子空间和批量归一化(BN)层缩放因子的双准则剪枝算法。将卷积层滤波器通过k均值聚类得到不同参数子空间,在子空间内使滤波器按权重排序并去除权重较低的滤波器,同时采用BN层缩放因子剪枝算法避免剪枝精度下降。实验结果表明,采用该算法剪枝后的YOLOv3网络在精度不变的情况下,占用的内存减少5/6且计算时间缩短1/3,与PF、CP等剪枝算法相比,该算法在保持较高网络精度的情况下计算量更少。
In order to ensure the normal operation of YOLO network on embedded devices,it is necessary to use pruning algorithm to simplify the filter to reduce the network storage space and the amount of calculation. However,the existing pruning algorithms are time-consuming and have low pruning precision. This paper proposes a bi-criteria pruning algorithm based on parameter subspace and Batch Normalization(BN)layer scaling factor. The convolution layer filter is clustered into different parameter subspaces by k-means clustering.In the subspace,the filters are sorted by weight,and the filters with lower weight are removed.The BN layer scaling factor pruning algorithm is used to avoid the degradation of pruning precision.Experimental results show that the memory occupied by the network which is pruned by the proposed algorithm is reduced by5/6 while the precision remains unchanged and the computing time is reduced by 1/3. Compared with PF,CP and other pruning algorithms,the proposed algorithm requires less computation while maintaining high network precision.
作者
杨民杰
梁亚玲
杜明辉
YANG Minjie;LIANG Yaling;DU Minghui(School of Electronic and Information Engineering,South China University of Technology,Guangzhou 510641,China)
出处
《计算机工程》
CAS
CSCD
北大核心
2021年第2期111-117,共7页
Computer Engineering
基金
国家自然科学基金(61701181)
广东省自然科学基金(2017A030325430)
广州市科技计划项目(201707010070)。
关键词
模型压缩
剪枝
目标检测
均值聚类
缩放因子
model compression
pruning
object detection
mean clustering
scaling factor
