基于CUDA流技术的深度学习系统优化  

Optimization of deep learning system based on CUDA stream technology

作　　者：蒋文斌[1,2,3] 刘湃 陈雨浩 张杨松 JIANG Wenbin;LIU Pai;CHEN Yuhao;ZHANG Yangsong(School of Computer of Science and Technology,Huazhong University of Science and Technology,Wuhan 430074,China;Natio nal Engineering Research Center for Big Data Technology and System,Huazhong University of Science and Technology,Wuhan 430074,China;Services Computing Technology and System Laboratory,Huazhong University of Science and Technology,Wuhan 430074,China)

机构地区：[1]华中科技大学计算机科学与技术学院,湖北式汉430074 [2]华中科技大学大数据技术与系统国家工程研究中心,湖北式汉430074 [3]华中科技大学服务计算技术与系统教育部重点实验室,湖北式汉430074

出　　处：《华中科技大学学报（自然科学版）》2020年第7期107-111,共5页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61672250)。

摘　　要：为了在不损失模型准确率的同时优化Caffe深度学习框架的训练速度,提出了一种面向Caffe并基于计算统一设备架构(CUDA)流技术的深度学习系统优化方法,以便充分利用GPU资源,提高计算的并行度.在Caffe网络的各层使用异步CUDA流,使其运行在独立线程以并行执行GPU计算任务;同时将批处理块划分成多个数据片,使用调度算法在前向传播和反向传播过程中以流水线形式进行处理.在数据集MNIST和CIFAR-10上的实验结果表明:优化后的系统在训练速度上有明显提升,同时准确率基本无损失.To optimize the training speed of the deep learning framework of Caffe without losing the accuracy of the model,an optimization method based on the compute unified device architecture(CUDA)stream technology for deep learning system was introduced in the training process of Caffe to make full use of GPU resources and improve the degree of computational parallelism.By applying asynchronous CUDA stream,each layer of the Caffe network ran in a separate thread with CUDA stream to implement the GPU computing task in parallel.Meanwhile,the data batch was divided into multiple data pieces,and the scheduling algorithm was used to process them in the form of pipeline during forward and backward computations.Experiments on dataset MNIST and CIFAR-10 show that the optimized framework has a significant improvement in training speed without obvious accuracy loss.

关 键 词：深度学习 计算统一设备架构(CUDA)流 训练速度 调度算法 准确率 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]