Sensitivity-based Integrated Pruning Algorithm for YOLO Network

doi:10.19678/j.issn.1000-3428.0058784

Abstract

Abstract: Deep Convolutional Neural Network(CNN) require considerable storage space and operation times, which hampers their application and deployment on platforms with limited resources.The pruning algorithms can alleviate this problem, but the ones based on a single method for parameter importance evaluation or feature reconstruction have poor generalization performance.To address the problem, an integrated pruning algorithm based on sensitivity is proposed.The algorithm employs sparse scaling factor of BN layer to reduce the density of the layers with many convolutional kernels in YOLO.Then three methods for parameter importance evaluation are used to sort the convolutional kernels by importance.The ratio of to-be-pruned parts of each layer is determined according to the sensitivity.Experimental results show that the proposed algorithm reduces the parameter number of YOLOv3 by 80.5% and YOLOv3-tiny by 92.6%.Compared with the pruning algorithm based on network lightweight method, the proposed algorithm can better improve the detection accuracy and the generalization performance of the pruned model.

Key words: Convolutional Neural Network(CNN), sensitivity, integrated pruning algorithm, YOLO network, importance evaluation

摘要： 深层卷积神经网络所需的计算量和存储空间严重制约了其在资源有限平台上的应用与部署。针对基于单一参数重要性评价或者特征重建的剪枝算法泛化能力较差的问题，提出基于敏感度的集成剪枝算法，利用BN层的缩放因子稀疏YOLO网络中卷积核个数较多的冗余层，结合3种参数重要性评价方法对卷积核做重要性排序，并根据敏感度确定每一层的剪枝比率。实验结果表明，该剪枝算法对于YOLOv3和YOLOv3-tiny网络分别缩减80.5%和92.6%的参数量，并且相比基于网络轻量化方法的剪枝算法提升了网络模型压缩后的检测精度和泛化能力。

关键词: 卷积神经网络, 敏感度, 集成剪枝算法, YOLO网络, 重要性评价

CLC Number:

TP391

ZHANG Jiangyong, XU Zhiyong, ZHANG Jianlin, XU Tao. Sensitivity-based Integrated Pruning Algorithm for YOLO Network[J]. Computer Engineering, 2021, 47(9): 59-68.

张江永, 徐智勇, 张建林, 许涛. 基于敏感度的YOLO网络集成剪枝算法[J]. 计算机工程, 2021, 47(9): 59-68.

/ / Recommend / Download Citations

URL: http://www.ecice06.com/EN/10.19678/j.issn.1000-3428.0058784

http://www.ecice06.com/EN/Y2021/V47/I9/59

Figures/Tables 13

References

[1] ZHANG X Y, ZHOU X Y, LIN M X, et al.ShuffleNet:an extremely efficient convolutional neural network for mobile devices[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2018:6848-6856.
[2] 杨民杰, 梁亚玲, 杜明辉.基于参数子空间和缩放因子的YOLO剪枝算法[J].计算机工程, 2021, 47(2):111-117. YANG M J, LIANG Y L, DU M H.YOLO pruning algorithm based on parameter subspace and scaling factor[J].Computer Engineering, 2021, 47(2):111-117.(in Chinese)
[3] JADERBERG M, VEDALDI A, ZISSERMAN A.Speeding up convolutional neural networks with low rank expansions[EB/OL].[2020-05-11].https://arxiv.org/abs/1405.3866v1.
[4] ZHU C Z, HAN S, MAO H Z, et al.Trained ternary quantization[EB/OL].[2020-05-11].https://arxiv.org/abs/1612.01064v3.
[5] HINTON G, VINYALS O, DEAN J.Distilling the knowledge in a neural network[J].Computer Science, 2015, 14(7):38-39.
[6] NAKKIRAN P, KAPLUN G, BANSAL Y, et al.Deep double descent:where bigger models and more data hurt[EB/OL].[2020-05-11].https://arxiv.org/abs/1912.02292.
[7] LECUN Y, DENKER J S, SOLLA S A.Optimal brain damage[C]//Proceedings of 1990 International Conference on Neural Information Processing.Berlin, Germany:Springer, 1990:598-605.
[8] HAN S, POOL J, TRAN J, et al.Learning both weights and connections for efficient neural networks[C]//Proceedings of 2015 International Conference on Neural Information Processing.Berlin, Germany:Springer, 2015:1135-1143.
[9] GUO Y, YAO A, CHEN Y, et al.Dynamic network surgery for efficient DNNs[C]//Proceedings of 2016 International Conference on Neural Information Processing.Berlin, Germany:Springer, 2016:1387-1395.
[10] LI H, KADAV A, DURDANOVIC I, et al.Pruning filters for efficient ConvNets[EB/OL].[2020-05-11].https://arxiv.org/pdf/1608.08710.pdf.
[11] HU H Y, PENG R, TAI Y W, et al.Network trimming:a data-driven neuron pruning approach towards efficient deep architectures[EB/OL].[2020-05-11].https://arxiv.org/pdf/1607.03250.pdf.
[12] WANG H, ZHANG W H, WONG K Y M, et al.Encoding multisensory information in modular neural networks[C]//Proceedings of 2017 International Conference on Neural Information Processing.Berlin, Germany:Springer, 2017:658-665.
[13] YE J B, LU X, LIN Z, et al.Rethinking the smaller-norm-less-informative assumption in channel pruning of convolution layers[EB/OL].[2020-05-11].https://arxiv.org/abs/1802.00124.
[14] LUO J H, WU J X, LIN W Y.ThiNet:a filter level pruning method for deep neural network compression[C]//Proceedings of 2017 IEEE International Conference on Computer Vision.Washington D.C., USA:IEEE Press, 2017:5068-5076.
[15] ZHUANG Z, TAN M, ZHUANG B, et al.Discrimination-aware channel pruning for deep neural networks[C]//Proceedings of International Conference on Neural Information Processing.Berlin, Germany:Springer, 2018:883-894.
[16] HE Y, LIU P, WANG Z W, et al.Filter pruning via geometric median for deep convolutional neural networks acceleration[C]//Proceedings of 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2019:4335-4344.
[17] LIU Z, LI J G, SHEN Z Q, et al.Learning efficient convolutional networks through network slimming[C]//Proceedings of 2017 IEEE International Conference on Computer Vision.Washington D.C., USA:IEEE Press, 2017:2755-2763.
[18] REDMON J, FARHADI A.YOLOv3:an incremental improvement[EB/OL].[2020-05-11].https://arxiv.org/pdf/1804.02767.pdf.
[19] HE K M, ZHANG X Y, REN S Q, et al.Deep residual learning for image recognition[C]//Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2016:770-778.
[20] HUANG Z Z, WANG X J, LUO P.Convolution-weight-distribution assumption:rethinking the criteria of channel pruning[EB/OL].[2020-05-11].https://arxiv.org/abs/2004.11627.
[21] IOFFE S, SZEGEDY C.Batch normalization:accelerating deep network training by reducing internal covariate shift[C]//Proceedings of the 32nd International Conference on International Conference on Machine Learning.New York, USA:ACM Press, 2015:448-456.
[22] SUN X, REN X C, MA S M, et al.meProp:sparsified back propagation for accelerated deep learning with reduced overfitting[C]//Proceedings of the 34th International Conference on Machine Learning.Washington D.C., USA:IEEE Press, 2017:3299-3308.

Please choose a citation manager

Content to export