切比雪夫逼近的深度神经网络并行加速  

Parallel Acceleration for Deep Neural Network of Chebyshev Approximation

作　　者：李方舒 钱慧[1] 陈晓旭 LI Fang-shu;QIAN Hui;CHEN Xiao-xu(School of Physics and Information Engineering,Fuzhou University,Fuzhou 350116,China)

机构地区：[1]福州大学物理与信息工程学院,福州350116

出　　处：《小型微型计算机系统》2020年第10期2206-2211,共6页Journal of Chinese Computer Systems

基　　金：数字福建物联网工程应用实验室建设项目(82917002)资助。

摘　　要：深度神经网络(Deep Neural Network,DNN)中数据量巨大,且卷积层计算复杂度高,使得其难以在资源有限的嵌入式GPU上进行部署,因此需要对其进行并行加速设计.本文提出采用切比雪夫多项式对卷积核进行逼近,并将该优化方案应用在面向图像重构的DNN中以实现卷积操作的并行化处理,降低计算复杂度.然后为优化后的网络卷积层进行基于GPU的并行加速设计,最后将网络整体移植到NVIDIA AGX Xavier嵌入式开发板上来实现图像的重构推理过程.实验结果表明,并行加速后的网络重构推理的速度是原始网络的2.2倍.Due to the large amount of data of Deep Neural Netw ork(DNN)and the high computational complexity of convolutional layer,it is difficult for DNN to be deployed on the embedded GPU w ith limited resources,therefore,parallel acceleration design is needed.This paper proposes to use Chebyshev polynomial to approximate the convolution kernel,and applies this optimization scheme to image reconstruction DNN to realize parallel processing of convolution operation and reduce computational complexity.Then the GPUbased parallel acceleration designs are carried out for the convolutional layers of the optimized netw ork.Finally,the netw ork is transplanted to NVIDIA AGX Xavier embedded development board to realize the reasoning process of image reconstruction.The experimental results show that the reconstruction reasoning speed of the parallel accelerated netw ork is 2.2 times faster than that of the original netw ork.

关 键 词：深度神经网络 图像重构 并行计算 嵌入式GPU 切比雪夫逼近 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]