摘 要:深度学习框架是实现机器学习的关键工具,合适的深度学习框架可以达到事半功倍的效果。为助力研究者选择合适的框架,在回顾近十种常见框架的基础上,聚焦当前受众最广、热度最高的两种深度学习框架 TensorFlow 和 PyTorch,从历程、现状、机制、训练模式、可视化、工业部署等角度对两者进行比对分析,并归类对应适用场景的建议,为框架选择提供思路参考。
关键词:深度学习;TensorFlow;PyTorch;适用场景
中图分类号:TP181;TP391.41 文献标识码:A 文章编号:2096-4706(2020)04-0080-04
Comparative Analysis of Deep Learning Frameworks Based on TensorFlow and PyTorch
HUANG Yuping,LIANG Weixuan,XIAO Zuhuan
(Guangdong Communication Polytechnic,Guangzhou 510650,China)
Abstract:A deep learning framework is a key tool for implementing machine learning. A suitable deep learning framework can achieve more results with less effort. In order to help researchers choose the appropriate framework,based on reviewing nearly ten common frameworks,focus on the two deep learning frameworks (TensorFlow and PyTorch) that are currently the most widely used and most popular. From history,status,mechanism,training mode,visualization,industrial deployment and other perspectives compare and analyze the two,and categorize suggestions corresponding to applicable scenarios,providing a reference for framework selection.
Keywords:deep learning framework;TensorFlow;PyTorch;application scene
基金项目:广东交通职业技术学院大学生科技创新项目(GDCP-ZX-2019-003-N2);广东交通职业技术学院一院一坊项目(ZC-AB-05-10 01-51);2019 年广东交通职业技术学院大学生创新创业训练计划项目(ZC-AB-05-1003-59)
参考文献:
[1] HE H. The State of Machine Learning Frameworks in 2019 [EB/OL].(2019-10-10).https://thegradient.pub/state-of-ml-frameworks-2019-pytorch-dominates-researchtensorflow-dominates-industry/.
[2] HALE J. Which Deep Learning Framework is Growing Fastest [EB/OL].(2019-04-01).https://towardsdatascience.com/which-deeplearning-framework-is-growing-fastest-3f77f14aa318.
[3] Sharma P. 5 Amazing Deep Learning Frameworks Every Data Scientist Must Know [EB/OL].(2019-03-14).https://www.analyticsvidhya.com/blog/2019/03/deep-learning-frameworks-comparison/.
[4] DUBOVIKOV K. PyTorch vs TensorFlow — Spotting the Difference [EB/OL].(2017-06-21).https://medium.com/@dubovikov.kirill/pytorch-vs-tensorflow-spotting-the-difference-25c75777377b.
[5] 段仲静,李少波,胡建军,等 . 深度学习目标检测方法及主流框架综述 [J/OL]. 激光与光电子学进展:1-18(2019-12-11).http://kns.cnki.net/kcms/detail/31.1690.TN.20191210.1 209.022.html.
[6] 许淑扬.深度学习框架Caffe在图像分类中的应用探析 [J].信息与电脑(理论版),2019,31(23):95-96.
[7] 加日拉 • 买买提热衣木,常富蓉,刘晨,等 . 主流深度学习框架对比 [J]. 电子技术与软件工程,2018(7):74.
[8] 庞涛 . 开源深度学习框架发展现状与趋势研究 [J]. 互联网天地,2018(4):46-54.
[9] 李西 . 人工智能背景下的各大深度学习框架展望 [J]. 计算机产品与流通,2018(8):145.
[10] 51CTO.2019 深度学习框架排行榜 [EB/OL].(2019-08-02).https://ai.51cto.com/art/201908/600692.htm.
[11] 阿里云云栖号 . 关于 TensorFlow 你需要了解的 9 件事 [EB/OL].(2018-08-27).https://www.jianshu.com/p/435f9 f109b22.
[12] 燚智能物联网 .Chainer 时代即将结束 [EB/OL].(2019-12-13).http://www.openpcba.com/web/contents/get?id= 5845&tid=15.
作者简介:
黄玉萍(1989-),女,汉族,江西赣州人,讲 师,硕士研究生,研究方向:轨道交通控制,弓网建模;
梁炜萱(2000-),女,汉族,广东清远人,研究方向:城市轨道交通通信信号处理;
肖祖环(1998-),男,汉族,四川达州人,研究方向:城市轨道交通通信信号处理。