GPU异构计算环境中长短时记忆网络模型的应用及优化  

Application and optimization of long short-term memory network model in GPU heterogeneous computing environment

作　　者：梁桂才[1] 梁思成 陆莹[1] LIANG Guicai;LIANG Sicheng;LU Ying(Information Management Center,Guangxi Vocational College of Mechanical and Electrical Engineering,Nanning 530007,China)

机构地区：[1]广西机电职业技术学院信息管理中心,南宁530007

出　　处：《计算机应用文摘》2024年第10期37-41,共5页Chinese Journal of Computer Application

基　　金：2023年广西科技厅广西重点研发计划项目:基于GPU的高性能AI算力一体化资源池的构建(2023AB01399)。

摘　　要：随着深度学习的广泛应用及算力资源的异构化,在GPU异构计算环境下的深度学习加速成为又一研究热点。文章探讨了在GPU异构计算环境中如何应用长短时记忆网络模型,并通过优化策略提高其性能。首先,介绍了长短时记忆网络模型的基本结构(包括门控循环单元、丢弃法、Adam与双向长短时记忆网络等);其次,提出了在GPU上执行的一系列优化方法,如CuDNN库的应用及并行计算的设计等。最终,通过实验分析了以上优化方法在训练时间、验证集性能、测试集性能、超参数和硬件资源使用等方面的差异。With the widespread application of deep learning and the isomerization of computing resources,the acceleration of deep learning in GPU heterogeneous computing environments has become another research hotspot.This article explores how to apply long short-term memory network models in GPU heterogeneous computing environments and improve their performance through optimization strategies.Firstly,the basic structure of the long short-term memory network model was introduced,including gate recurrent unit,Dropout,Adam,and BiLSTM.Secondly,a series of optimization methods were proposed for execution on GPUs,such as the application of CuDNN library and the design of parallel computing.Finally,the differences in training time,validation set performance,test set performance,hyperparameters,and hardware resource utilization among the above optimization methods were analyzed through experiments.

关 键 词：GPU异构 长短时记忆网络 门控循环单元 ADAM DROPOUT CuDNN 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]