模拟神经网络传输的职业教育视频资源开发
|
摘要
面对职业教育视频资源在传输过程中网络结构和硬件资源占用关系这一问题,论文利用矩阵乘法改进了全连接深度神经网络(DNN)的矩阵计算形式,以此动态模拟职业教育视频资源传输状态。将矩阵乘法引入DNN前向传播过程使计算简化,以探究职业教育视频资源库作为硬件实现平台,基于乘累加器IP核与乘加器IP核设计了两种矩阵乘法计算架构,实现了模拟全连接DNN前向传播的职业教育视频资源传输计算过程,并对两种方案在实现不同结构的前向传播计算时的硬件资源占用情况进行对比,得出结论:在实现相同网络的视频资源传播计算情况下,乘累加器方案比乘加器方案消耗更少的硬件资源。
In the face of the problem of network structure and hardware resource occupancy in video transmission in vocational education,this paper uses matrix multiplication to improve the matrix calculation form of full connection depth neural network(DNN) to dynamically simulate the video resources of vocational education Transmission status. The matrix multiplication is introduced into the DNN forward propagation process to simplify the calculation to explore the vocational education video resource libraryas the hardware implementation platform. Based on the accumulator IP core and the multiplier IP core,two matrix multiplication algorithms are designed to realize the simulation connecting the DNN forward communication of the vocational education video resource transmission calculation process,and the two programs in the realization of the different structure of the forward communication calculation of hardware resource occupancy compared to the conclusion:in the realization of the same network of video resource propagation calculation in the case of a multiplier,the multiplier scheme consumes less hardware resources than the multiplier scheme.
In the face of the problem of network structure and hardware resource occupancy in video transmission in vocational education,this paper uses matrix multiplication to improve the matrix calculation form of full connection depth neural network(DNN) to dynamically simulate the video resources of vocational education Transmission status. The matrix multiplication is introduced into the DNN forward propagation process to simplify the calculation to explore the vocational education video resource libraryas the hardware implementation platform. Based on the accumulator IP core and the multiplier IP core,two matrix multiplication algorithms are designed to realize the simulation connecting the DNN forward communication of the vocational education video resource transmission calculation process,and the two programs in the realization of the different structure of the forward communication calculation of hardware resource occupancy compared to the conclusion:in the realization of the same network of video resource propagation calculation in the case of a multiplier,the multiplier scheme consumes less hardware resources than the multiplier scheme.
引文
[1]胡芬婷.远程职业教育微视频资源制作的研究现状及启示[J].新课程研究(中旬刊),2016(8):45-48.HU Fenting.Research status and Enlightenment of microvideo resources production in Distance Vocational Educa-tion[J].The new curriculum research(a),2016(8):45-48.
[2]陈亮,汪景福,王娜,等.基于DNN算法的移动视频推荐策略[J].计算机学报,2016,39(8):1626-1638.CHEN Liang,WANG Jingfu,WANG Na,et,al.Recom-mendation strategy of mobile video based on DNN algo-rithm[J].Journal of Computer Science,2016,39(8):1626-1638.
[3]Wen Qu,Kai-Song Song,Yi-Fei Zhang,Shi et al.A NovelApproach Based on Multi-View Content Analysis andSemi-Supervised Enrichment for Movie Recommendation[J].Journal of Computer Science and Technology,2013(5):1221-1229.
[4]张宝军,卢梦怡,陈治清,等.基于径向基函数神经网络的波束形成算法[J].西安邮电大学学报,2015,20(6):33-36.ZHANG Baojun,LU Mengyi,CHEN Zhiqing,et al.Beamforming algorithm based on radial basis function neuralnetwork[J].Journal of Xi'an University of Posts and Tele-communications,2015,20(6):33-36.
[5]闫明,王秀芬,李强.基于数据对称打包的云计算并行核心失败校验缓解[J].微电子学与计算机,2017,34(5):73-78.YAN Ming,WANG Xiufen,LI Qiang.Parallel core fail-ure verification mitigation for Cloud Computing Based ondata symmetric packing[J].Microellectronics&Comput-er,2017,34(5):73-78.
[6]侯韩旭,李挥,张华宇,等.二元再生码在分布式存储系统的应用[J].计算机研究与发展,2013,50(2):45-53.HOU Hanxu,LI Hui,ZHANG Huayu,et al Applicationof two yuan regeneration code in distributed storage system[J].Research and development on Computer,2013,50(2):45-53.
[7]刘仲,田希.面向多核向量处理器的矩阵乘法向量化方法[J].计算机学报,2017(7):1-13.LIU Zhong,TIAN Xi.Matrix multiplication oriented vec-torization method for multi-core vector processors[J].Journal of Computer Science,2017(7):1-13.
[8]杨娟,曾苗祥,徐晶,等.基于众核处理器和GPU的视频快速检测方案[J].计算机科学,2015,42(3):266-270,295.YANG Juan,Zeng Miaoxiang,XU Jing,et al.Video fastdetection scheme based on multi core processor and GPU[J].computer science,2015,42(3):266-270,295.
[9]姜宏旭,李辉勇,刘亭杉,等.基于FPGA的高效机载视频采集及预处理方法[J].北京航空航天大学学报,2015,41(1):14-22.JIANG Hongxu,LI Huiyong,LIU Tingbing,et al.Efficientairborne video acquisition and preprocessing methodbased on FPGA[J].Journal of Beihang University.2015,41(1):14-22.
[10]易盟,隋立春.基于改进全卷积神经网络的航拍图像语义分类方法[J].计算机工程,2017,43(10):216-221.YI Meng,SUI Lichun.Aerial image semantic classifica-tion method based on improved total convolution neuralnetwork[J].Computer Engineering,2017,43(10):216-221.
[11]张蕾,章毅.大数据分析的无限深度神经网络方法[J].计算机研究与发展,2016,53(1):68-79.ZHANG Lei,ZHANG Yi.Infinite depth neural networkmethod for large data analysis[J].Research and develop-ment of Computer,2016,53(1):68-79.
[12]潘炜深,金连文,冯子勇.基于多尺度梯度及深度神经网络的汉字识别[J].北京航空航天大学学报,2015,41(4):751-756.PAN Weishen,JIN Lianwen,FENG Ziyong.Chinesecharacter recognition based on multi-scale gradient anddepth neural network[J].Journal of Beihang University.2015,41(4):751-756.
[13]赵宏伟,李圣普.基于粒子群算法和RBF神经网络的云计算资源调度方法研究[J].计算机科学,2016,43(3):113-117,150.ZHAO Hongwei,LI Shengpu.Cloud computing resourcescheduling method based on particle swarm optimizationand RBF neural network[J].Computer Science,2016,43(3):113-117,150.
[14]陈智,王贵锋,柳莺.一种基于IP Core实现FFT变换的新方法[J].自动化与仪器仪表,2012(2):163-164.CHEN Zhi,WANG Guifeng,LIU Ying.A new method forimplementing FFT transformation based on IP Core[J].Automation and instrumentation,2012(2):163-164.
[15]姚若河,徐新才.基于冗余符号数的定点乘法器的设计[J].华南理工大学学报(自然科学版),2014,42(3):27-34.YAO Ruohe,XU Xincai.Design of fixed point multiplierbased on redundant symbol number[J].Journal of SouthChina University of Technology(Natural Science Edi-tion),2014,42(3):27-34.
[16]周怡,李树国.一种改进的基4-Booth编码流水线大数乘法器设计[J].微电子学与计算机,2014,31(1):60-63,67.ZHOU Yi,LI Shuguo.Design of a modified large num-ber multiplier based on 4-Booth coding[J].Microellec-tronics&Computer,2014,31(1):60-63,67.
[2]陈亮,汪景福,王娜,等.基于DNN算法的移动视频推荐策略[J].计算机学报,2016,39(8):1626-1638.CHEN Liang,WANG Jingfu,WANG Na,et,al.Recom-mendation strategy of mobile video based on DNN algo-rithm[J].Journal of Computer Science,2016,39(8):1626-1638.
[3]Wen Qu,Kai-Song Song,Yi-Fei Zhang,Shi et al.A NovelApproach Based on Multi-View Content Analysis andSemi-Supervised Enrichment for Movie Recommendation[J].Journal of Computer Science and Technology,2013(5):1221-1229.
[4]张宝军,卢梦怡,陈治清,等.基于径向基函数神经网络的波束形成算法[J].西安邮电大学学报,2015,20(6):33-36.ZHANG Baojun,LU Mengyi,CHEN Zhiqing,et al.Beamforming algorithm based on radial basis function neuralnetwork[J].Journal of Xi'an University of Posts and Tele-communications,2015,20(6):33-36.
[5]闫明,王秀芬,李强.基于数据对称打包的云计算并行核心失败校验缓解[J].微电子学与计算机,2017,34(5):73-78.YAN Ming,WANG Xiufen,LI Qiang.Parallel core fail-ure verification mitigation for Cloud Computing Based ondata symmetric packing[J].Microellectronics&Comput-er,2017,34(5):73-78.
[6]侯韩旭,李挥,张华宇,等.二元再生码在分布式存储系统的应用[J].计算机研究与发展,2013,50(2):45-53.HOU Hanxu,LI Hui,ZHANG Huayu,et al Applicationof two yuan regeneration code in distributed storage system[J].Research and development on Computer,2013,50(2):45-53.
[7]刘仲,田希.面向多核向量处理器的矩阵乘法向量化方法[J].计算机学报,2017(7):1-13.LIU Zhong,TIAN Xi.Matrix multiplication oriented vec-torization method for multi-core vector processors[J].Journal of Computer Science,2017(7):1-13.
[8]杨娟,曾苗祥,徐晶,等.基于众核处理器和GPU的视频快速检测方案[J].计算机科学,2015,42(3):266-270,295.YANG Juan,Zeng Miaoxiang,XU Jing,et al.Video fastdetection scheme based on multi core processor and GPU[J].computer science,2015,42(3):266-270,295.
[9]姜宏旭,李辉勇,刘亭杉,等.基于FPGA的高效机载视频采集及预处理方法[J].北京航空航天大学学报,2015,41(1):14-22.JIANG Hongxu,LI Huiyong,LIU Tingbing,et al.Efficientairborne video acquisition and preprocessing methodbased on FPGA[J].Journal of Beihang University.2015,41(1):14-22.
[10]易盟,隋立春.基于改进全卷积神经网络的航拍图像语义分类方法[J].计算机工程,2017,43(10):216-221.YI Meng,SUI Lichun.Aerial image semantic classifica-tion method based on improved total convolution neuralnetwork[J].Computer Engineering,2017,43(10):216-221.
[11]张蕾,章毅.大数据分析的无限深度神经网络方法[J].计算机研究与发展,2016,53(1):68-79.ZHANG Lei,ZHANG Yi.Infinite depth neural networkmethod for large data analysis[J].Research and develop-ment of Computer,2016,53(1):68-79.
[12]潘炜深,金连文,冯子勇.基于多尺度梯度及深度神经网络的汉字识别[J].北京航空航天大学学报,2015,41(4):751-756.PAN Weishen,JIN Lianwen,FENG Ziyong.Chinesecharacter recognition based on multi-scale gradient anddepth neural network[J].Journal of Beihang University.2015,41(4):751-756.
[13]赵宏伟,李圣普.基于粒子群算法和RBF神经网络的云计算资源调度方法研究[J].计算机科学,2016,43(3):113-117,150.ZHAO Hongwei,LI Shengpu.Cloud computing resourcescheduling method based on particle swarm optimizationand RBF neural network[J].Computer Science,2016,43(3):113-117,150.
[14]陈智,王贵锋,柳莺.一种基于IP Core实现FFT变换的新方法[J].自动化与仪器仪表,2012(2):163-164.CHEN Zhi,WANG Guifeng,LIU Ying.A new method forimplementing FFT transformation based on IP Core[J].Automation and instrumentation,2012(2):163-164.
[15]姚若河,徐新才.基于冗余符号数的定点乘法器的设计[J].华南理工大学学报(自然科学版),2014,42(3):27-34.YAO Ruohe,XU Xincai.Design of fixed point multiplierbased on redundant symbol number[J].Journal of SouthChina University of Technology(Natural Science Edi-tion),2014,42(3):27-34.
[16]周怡,李树国.一种改进的基4-Booth编码流水线大数乘法器设计[J].微电子学与计算机,2014,31(1):60-63,67.ZHOU Yi,LI Shuguo.Design of a modified large num-ber multiplier based on 4-Booth coding[J].Microellec-tronics&Computer,2014,31(1):60-63,67.