ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2016, Vol. 53 ›› Issue (6): 1202-1210.doi: 10.7544/issn1000-1239.2016.20150147

• 系统结构 • 上一篇    下一篇



  1. 1(国防科学技术大学并行与分布处理国防重点实验室 长沙 410073);2(国防科学技术大学计算机学院 长沙 410073) (
  • 出版日期: 2016-06-01
  • 基金资助: 

DLPF: A Parallel Deep Learning Programming Framework Based on Heterogeneous Architecture

Wang Yueqing1, Dou Yong1, Lü Qi1, Li Baofeng2, Li Teng1   

  1. 1(Science and Technology on Parallel and Distributed Processing Laboratory, National University of Defense Technology, Changsha 410073);2(College of Computer, National University of Defense Technology, Changsha 410073)
  • Online: 2016-06-01

摘要: 深度学习在机器学习领域扮演着十分重要的角色,已被广泛应用于各种领域,具有十分巨大的研究和应用前景.然而,深度学习也面临3方面的挑战:1)现有深度学习工具使用便捷性不高,尽管深度学习领域工具越来越多,然而大多使用过程过于繁杂,不便使用;2)深度学习模型灵活性不高,限制了深度学习模型发展的多样性;3)深度学习训练时间较长,超参数搜索空间大,从而导致超参数寻优比较困难.针对这些挑战,设计了一种基于深度学习的并行编程框架,该框架设计了统一的模块库,能可视化地进行深度学习模型构建,提高了编程便捷性;同时在异构平台对算法模块进行加速优化,较大程度减少训练时间,进而提高超参数寻优效率.实验结果表明,该编程框架可以灵活构建多种模型,并且对多种应用取得了较高的分类精度.通过超参数寻优实验,可以便捷地获得最优超参数组合,从而推断各种超参数与不同应用的联系.

关键词: 深度学习, 编程框架, 可视化, 异构平台, 加速

Abstract: Deep learning plays an important role in machine learning field, and it has been widely used in various applications. The prospect of research and applications of deep learning are huge. However, deep learning also faces several challenges. Firstly, there are many tools in deep learning field, but these tools are not convenient to use for non-expert users because the installation and usage of them are really complex. Secondly, the diversity of deep learning is limited because the flexibility of existing deep learning models is not enough. Furthermore, the training time of deep learning is so long that the optimal hyper-parameters combination cannot be found in a short time. To solve these problems, we design a deep learning programming framework based on heterogeneous architecture in this paper. The programming framework establishes a unified module library which can be used to build a deep model through the visual interface conveniently. Besides, the framework also accelerates the basic modules on heterogeneous platform, and makes the speed of searching optimal hyper-parameters combination be faster. Experimental results show that the programming framework can construct deep models flexibly, and more importantly, it can achieve comparative classification results and better timing performance for a variety of applications. In addition, the framework can search optimal hyper-parameters efficiently and make us infer the relationship of all hyper-parameters.

Key words: deep learning, programming framework, visualization, heterogeneous architecture, accelerate