基于中间图特征提取的卷积网络双标准剪枝

doi:10.19678/j.issn.1000-3428.0064206

摘要/Abstract

摘要： 卷积神经网络（CNN）在计算和存储上存在大量开销，为了使CNN能够在算力和存储能力较弱的嵌入式等端设备上进行部署和运行，提出一种基于中间图特征提取的卷积核双标准剪枝方法。在卷积层后插入中间图互信息特征提取框架，分析卷积核的特征提取能力，结合批量归一化层的缩放因子对卷积核的综合重要性进行评估，获取更为稀疏的CNN模型。针对全连接层存在大量冗余节点的问题，提出一种基于节点相似度与K-means++聚类的全连接层剪枝方法，聚类相似度较高的节点，并对剪枝后的连接层权重进行融合，在一定程度上弥补因剪枝所造成的精度损失。在CIFAR10和CIFAR100数据集上的实验结果表明，使用该剪枝方法对ResNet56网络进行剪枝，在损失0.19%分类精度的情况下能够剪掉48.2%的参数量以及46.7%的浮点运算量，对于VGG16网络，能够剪掉94.5%的参数量以及64.4%的浮点运算量，分类精度仅下降0.01%。与VCNNP、PF等剪枝方法相比，所提剪枝方法能够在保持模型准确率几乎不变的情况下，对CNN的参数量和计算量进行更大比例的裁剪。

关键词: 深度学习, 模型剪枝, 卷积神经网络, 互信息, 节点相似度, K-means++聚类, 中间特征

Abstract: Convolutional Neural Network(CNN) require a considerable amount of overhead in terms of computation and storage.To deploy and run a CNN on embedded devices with a poor computing power and storage capacity, this study proposes a convolution kernel double standard pruning method based on the feature extraction of an intermediate graph.In addition, to obtain a sparser CNN model, this study inserted the mutual information feature extraction framework of the intermediate graph after the convolution layer, analyzed the feature extraction ability of the convolution kernel, and evaluated the comprehensive importance of the convolution kernel combined with the scaling factor of the Batch Normalization(BN) layer.Addressing the problem the fully connected layer containing numerous redundant nodes, a fully connected layer pruning method based on node similarity and K-means++ clustering is proposed.Nodes with high similarity are clustered, and the weight of the connection layer after pruning is fused to compensate for the accuracy loss caused by pruning to a certain extent.Experiments on the CIFAR10 and CIFAR100 datasets show that this pruning method can prune 48.2% of the parameters and 46.7% of the floating-point operations of the ResNet56 network at a classification accuracy loss of 0.19%.For the VGG16 network, 94.5% of the parameters and 64.4% of the floating-point operations can be pruned, and the classification accuracy only decreases by 0.01%.Compared with pruning methods such as VCNNP and PF, the proposed pruning method can prune the parameters and computation of a CNN in larger proportions while maintaining the same model accuracy.

Key words: deep learning, model pruning, Convolutional Neural Network(CNN), mutual information, node similarity, K-means++ clustering, intermediate features

中图分类号:

TP391

程小辉, 李钰, 康燕萍. 基于中间图特征提取的卷积网络双标准剪枝[J]. 计算机工程, 2023, 49(3): 105-112.

CHENG Xiaohui, LI Yu, KANG Yanping. Double Standard Pruning of Convolution Network Based on Feature Extraction of Intermediate Graph[J]. Computer Engineering, 2023, 49(3): 105-112.

https://www.ecice06.com/CN/Y2023/V49/I3/105

图/表 12

20230314184959

20230314185003

20230314185006

20230314185010

20230314185013

20230314185016

20230314185019

20230314185023

20230314185027

20230314185030

20230314185044

20230314185047

参考文献

[1] SIMONYAN K, ZISSERMAN A.Very deep convolutional networks for large-scale image recognition[EB/OL].[2022-02-05].https://arxiv.org/abs/1409.1556.
[2] SZEGEDY C, IOFFE S, VANHOUCKE V, et al.Inception-v4, inception-ResNet and the impact of residual connections on learning[J].Proceedings of the AAAI Conference on Artificial Intelligence, 2017, 31(1):15-20.
[3] HE K M, ZHANG X Y, REN S Q, et al.Deep residual learning for image recognition[EB/OL].[2022-02-05].https://arxiv.org/abs/1512.03385.
[4] KRIZHEVSKY A, SUTSKEVER I, HINTON G E.ImageNet classification with deep convolutional neural networks[J].Communications of the ACM, 2017, 60(6):84-90.
[5] LIU N, MA X L, XU Z Y, et al.AutoCompress:an automatic DNN structured pruning framework for ultra-high compression rates[J].Proceedings of the AAAI Conference on Artificial Intelligence, 2020, 34(4):4876-4883.
[6] DENIL M, SHAKIBI B, DINH L, et al.Predicting parameters in deep learning[EB/OL].[2022-02-05].https://arxiv.org/abs/1306.0543.
[7] 曾焕强, 胡浩麟, 林向伟, 等.深度神经网络压缩与加速综述[J].信号处理, 2022, 38(1):183-194. ZENG H Q, HU H L, LIN X W, et al.Deep neural network compression and acceleration:an overview[J].Journal of Signal Processing, 2022, 38(1):183-194.(in Chinese)
[8] 林景栋, 吴欣怡, 柴毅, 等.卷积神经网络结构优化综述[J].自动化学报, 2020, 46(1):24-37. LIN J D, WU X Y, CHAI Y, et al.Structure optimization of convolutional neural networks:a survey[J].Acta Automatica Sinica, 2020, 46(1):24-37.(in Chinese)
[9] 李江昀, 赵义凯, 薛卓尔, 等.深度神经网络模型压缩综述[J].工程科学学报, 2019, 41(10):1229-1239. LI J Y, ZHAO Y K, XUE Z E, et al.A survey of model compression for deep neural networks[J].Chinese Journal of Engineering, 2019, 41(10):1229-1239.(in Chinese)
[10] LI S Y, HANSON E, QIAN X H, et al.ESCALATE:boosting the efficiency of sparse CNN accelerator with kernel decomposition[C]//Proceedings of the 54th Annual IEEE/ACM International Symposium on Microarchitecture.Washington D.C., USA:IEEE Press, 2021:992-1004.
[11] HAN S, MAO H Z, DALLY W J.Deep compression:compressing deep neural networks with pruning, trained quantization and huffman coding[EB/OL].[2022-02-05].https://arxiv.org/abs/1510.00149.
[12] AGHLI N, RIBEIRO E.Combining weight pruning and knowledge distillation for CNN compression[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2021:3185-3192.
[13] LI H, KADAV A, DURDANOVIC I, et al.Pruning filters for efficient ConvNets[EB/OL].[2022-02-05].https://www.cs.umd.edu/~hjs/pubs/nipswemdnn16-header.pdf.
[14] HE Y, LIU P, WANG Z W, et al.Filter pruning via geometric median for deep convolutional neural networks acceleration[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2019:4335-4344.
[15] HE Y H, ZHANG X Y, SUN J.Channel pruning for accelerating very deep neural networks[C]//Proceedings of IEEE International Conference on Computer Vision.Washington D.C., USA:IEEE Press, 2017:1398-1406.
[16] ANWAR S, HWANG K, SUNG W.Structured pruning of deep convolutional neural networks[J].ACM Journal on Emerging Technologies in Computing Systems, 2017, 13(3):1-18.
[17] 卢海伟, 夏海峰, 袁晓彤.基于滤波器注意力机制与特征缩放系数的动态网络剪枝[J].小型微型计算机系统, 2019, 40(9):1832-1838. LU H W, XIA H F, YUAN X T.Dynamic network pruning via filter attention mechanism and feature scaling factor[J].Journal of Chinese Computer Systems, 2019, 40(9):1832-1838.(in Chinese)
[18] LIU Z, LI J G, SHEN Z Q, et al.Learning efficient convolutional networks through network slimming[C]//Proceedings of IEEE International Conference on Computer Vision.Washington D.C., USA:IEEE Press, 2017:2755-2763.
[19] ALQAHTANI A.Pruning CNN filters via quantifying the importance of deep visual representations[J].Computer Vision and Image Understanding, 2021, 208/209:103220.
[20] KUMAR A, SHAIKH A M, LI Y, et al.Pruning filters with L1-norm and capped L1-norm for CNN compression[J].Applied Intelligence, 2021, 51(2):1152-1160.
[21] CHANG X P, PAN H H, LIN W Y, et al.A mixed-pruning based framework for embedded convolutional neural network acceleration[J].IEEE Transactions on Circuits and Systems I:Regular Papers, 2021, 68(4):1706-1715.
[22] DING X H, DING G G, GUO Y C, et al.Approximated oracle filter pruning for destructive CNN width optimization[EB/OL].[2022-02-05].https://arxiv.org/abs/1905.04748.
[23] YEOM S K, SEEGERER P, LAPUSCHKIN S, et al.Pruning by explaining:a novel criterion for deep neural network pruning[J].Pattern Recognition, 2021, 115:107899.
[24] KRIZHEVSKY A.Learning multiple layers of features from Tiny images[EB/OL].[2022-02-05].http://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf.
[25] LIN M, CHEN Q, YAN S C.Network in network[EB/OL].[2022-02-05].https://arxiv.org/abs/1312.4400.
[26] IOFFE S, SZEGEDY C.Batch normalization:accelerating deep network training by reducing internal covariate shift[EB/OL].[2022-02-05].https://arxiv.org/abs/1502.03167.
[27] HE Y, KANG G L, DONG X Y, et al.Soft filter pruning for accelerating deep convolutional neural networks[EB/OL].[2022-02-05].https://arxiv.org/abs/1808.06866.
[28] MOLCHANOV P, TYREE S, KARRAS T, et al.Pruning convolutional neural networks for resource efficient inference[EB/OL].[2022-02-05].https://arxiv.org/abs/1611.06440.
[29] ZHAO C, NI B, ZHANG J, et al.Variational convolutional neural network pruning[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2019:2780-2789.

选择文件类型/文献管理软件名称

选择包含的内容