融合权重与卷积核删减的SSD网络压缩

doi:10.11896/jsjkx.180901630

摘要/Abstract

摘要： 目标检测是计算机视觉领域中重要的研究方向。近几年,深度学习在基于视频的目标检测领域取得了突破性研究进展。深度学习强大的特征学习和特征表达能力,使其能够自动学习和提取相关特征并加以利用。然而,复杂的网络结构使得深度学习模型具有参数规模大、计算需求高、占用存储空间大等问题。基于深度神经网络的单发多框检测器(Single-shot Multi-box Detector 300,SSD300)能够对视频中的目标进行实时检测,但无法移植到嵌入式设备或移动终端以满足实际应用中的需求。为了解决该问题,文中提出了一种权重删减和卷积核删减融合的方法。首先,针对深度卷积神经网络模型权重参数过多导致模型过大的问题,采用权重删减的方法移除各卷积层中的冗余权重,确定各层权重的稀疏度;然后,针对卷积层计算量大的问题,根据各卷积层中的权重稀疏度对冗余卷积核进行删减,以减少冗余参数和计算量;最后,对删减后的神经网络进行训练,以恢复其检测精度。为验证该方法的有效性,在卷积神经网络框架caffe平台上对SSD网络模型进行验证。结果表明,压缩加速后的SSD300网络模型的大小为12.5MB,检测速度最高可达50FPS (frames per second)。实验实现了在网络检测准确率下降尽量小的前提下,将SSD300网络压缩了8.4×,加速了2×。权重删减和卷积核删减融合的方法为SSD300网络在视频检测中的智能化应用提供了可行性方案。

关键词: 深度神经网络, 单发多框检测器, 网络压缩与加速, 权重删减, 卷积核删减

Abstract: Object detection is an important research direction in the field of computer vision.In recent years,deep lear-ning has achieved great breakthroughs in object detection which is based on the video.Deep learning has powerful ability of feature learning and feature representation.The ability enables it to automatically learn,extract and utilize relevant features.However,complex network structure makes the deep learning model have a large scale of parameter.The deep neural network is both computationally intensive and memory intensive.Single Shot MultiBox Detector300 (SSD300),a single-shot detector,produces markedly superior detection accuracy and speed by using a single deep neural network.But it is difficult to deploy it on object detection systems with limited hardware resources.To address this limitation,the fusing method of weight pruning and filter pruning was proposed to reduce the storage requirement and inference time required by neural networks without affecting its accuracy.Firstly,in order to reduce the number of excessive weight parameters in the model of deep neural network,the weight pruning method is proposed.Network connections is pruned,in which weight is unimportant.Then,to reduce the large computation in convolution layer,the redundant filters are pruned according to the percentage of effective weights in each layer.Finally,the pruned neural network is trained to restore its detection accuracy.To verify the effectiveness of the method,the SSD300 was validated on caffe which is the convolutional neural network framework.After compression and acceleration,the storage of SSD300 neural network required is 12.5MB and the detection speed is 50FPS.The fusion of weight and filter pruning achieves the result by 2× speed-up,which reduces the storage required by SSD300 by 8.4×,as little increase of error as possible.The fusing method of weight and filter pruning makes it possible for SSD300 to be embedded in intelligent systems to detect and track objects.

Key words: Deep neural networks, Single-shot multi-box detector (SSD), Network compression and acceleration, Weight pruning, Filter pruning

中图分类号:

TP183

韩佳林, 王琦琦, 杨国威, 陈隽, 王以忠. 融合权重与卷积核删减的SSD网络压缩[J]. 计算机科学, 2019, 46(11): 272-276.

HAN Jia-lin, WANG Qi-qi, YANG Guo-wei, CHEN Jun, WANG Yi-zhong. SSD Network Compression Fusing Weight and Filter Pruning[J]. Computer Science, 2019, 46(11): 272-276.

参考文献 25

[1]	GIRSHICK R,DONAHUE J,DARRELL T,et al.Rich feature hierarchies for accurate object detection and semantic segmentation[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Columbus:IEEE,2014:580-587.
[2]	GIRSHICK R.Fast R-CNN[C]∥Proceedings of the IEEE Conference on International Conference on Computer Vision.Boston:IEEE,2015:1440-1448.
[3]	REN S,HE K,GIRSHICK R,et al.Faster R-CNN:towards real-time object detection with region proposal networks [J].IEEE Transactions on Pattern Analysis & Machine Intelligence,2017,39(6):1137-1149.
[4]	REDMON J,DIVVALA S,GIRSHICK R,et al.You only look once:unified,real-time object detection[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.New York:IEEE Press,2016:779-788.
[5]	LIU W,ANGUELOV D,ERHAN D,et al.SSD:single shotmultibox detector[C]∥Proceedings of European Conference on Computer Vision.Amsterdam:Springer International Publishing,2016:21-37.
[6]	HAN S,POOL J,TRAN J,et al.Learning both weights andconnections for efficient neural networks[M]∥Neural Information Processing Systems.Morgan Kaufmann Publishers Inc,2015:1135-1143.
[7]	HAN S,POOL J,DALLY W J,et al.Deep Compression:compressing deep neuralnetworks with pruning,trained quantization and huffman coding[C]∥Proceedings of Conference on Learning Representations.San Juan:IEEE,2016:233-242.
[8]	MOHAMMAD R,VICENTE O,JOSEPH R,et al.XOR-Net:ImageNet Classification Using Binary Convolutional Neural Networks[C]∥Proceedings of European Conference on Computer Vision.Amsterdam:ECCV,2016:525-542.
[9]	MATTHIEU C,ITAY H,DANIEL S,et al.Binarized NeuralNetworks:Training Neural Networks with Weights and Activations Constrained to +1 or －1[EB/OL].https://arxiv.org/abs/1704.04861.pdf.
[10]	GEOFFREY H,ORIOL V,JEFF D,et al.Distilling the knowledge in a Neural Network[C]∥Proceedings of Conference on Advances in Neural Infermation Processing Systems.Montreal:IEEE,2014:2644-2652.
[11]	BHARAT BHUSAN S,VINEETH N.B.Deep Model Compression:Distilling Knowledge from Noisy Teachers[EB/OL].https://arxiv.org/abs/1610.09650.pdf.
[12]	MAX J,ANDREA V,ANDREW Z,et al.Speeding up Convolutional Neural Networks with Low Rank Expansions[J].Computer Science,2014,4(4):1-7.
[13]	VIKAS S,TARA N S,SANJIV K,et al.Structured Transforms for Small-Footprint Deep Learning[EB/OL].https://arxiv.org/abs/1510.01722.pdf.
[14]	WEN W,WU C,WANG Y,et al.Learning structured sparsity in deep neural networks[M]∥Advances in Neural Information Processing Systems.Berlin:Springer,2016:2074-2082.
[15]	LIU Z,SHEN Z,HUANG G,et al.Learning efficient convolutional networks through network slimming[C]∥Proceedings of the IEEE International Conference on Computer Vision(ICCV).IEEE,2017:2755-2763.
[16]	HE Y,ZHANG X,SUN J,et al.Channel pruning for accelerating very deep neural networks [EB/OL].https://arxiv.org/abs/1707.06168.pdf.
[17]	IANDOLA F N,HAN S,MOSKEWICZ M W,et al.SqueezeNet:AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size[C]∥Proceedings of International Conference on Learning Representations.San Juan:ICLR,2016.
[18]	HOWARD A G,ZHU M,CHEN B,et al.MobileNets:efficient convolutional neural networks for mobile vision applications [EB/OL].https://arxiv.org/abs/1704.04861.pdf.
[19]	ZHANG X,ZHOU X,LIN M,et al.ShuffleNet:an extremely efficient convolutional neural network for mobile devices [EB/OL].https:// arxiv.org/abs/1707.01083.pdf.
[20]	EVERINGHAM M,VAN G L,WILLIAMS C K I,et al.Thepascal visual object classes (voc) challenge[J].International journal of computer vision,2010,88(2):303-338.
[21]	HANSON S J,PRATT L Y.Comparing biases for minimal network construction with back-propagation[M]∥Neural Information Processing Systems.Morgan Kaufmann Publishers Inc,1989:177-185.
[22]	CUN Y L,DENKER J S,SOLLA S A,et al.Optimal brain damage[C]∥Neural Information Processing Systems.Morgan Kaufmann Publishers Inc,1990:598-605.
[23]	HASSIBI B,STORK D G.Second Order derivatives for network pruning:optimal brain surgeon[C]∥Neural Information Processing Systems.Morgan Kaufmann Publishers Inc,1992:164-171.
[24]	HAN Y F,JIANG T H,MA Y P,et al.Compression of deep neural networks [J].Application Research of Computers,2018,35(10):2894-2897.(in Chinese)韩云飞,蒋同海,马玉鹏,等.深度神经网络的压缩研究[J].计算机应用研究,2018,35(10):2894-2897.
[25]	焦李成.深度学习、优化与识别[M].北京:清华大学出版社,2017:104.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed