基于Caffe的嵌入式多核处理器深度学习框架并行实现

高榕; 张良; 梅魁志

doi:10.7652/xjtuxb201806006

您当前的位置：

首页 >

文章列表页 >

基于Caffe的嵌入式多核处理器深度学习框架并行实现

更新时间：2025-07-09

- 基于Caffe的嵌入式多核处理器深度学习框架并行实现
- A Deep Learning Frame on Embedded Multicore Processors Based on Caffe and Its Parallel Implementation
- 西安交通大学学报 2018年52卷第6期页码：36-41+113
- 作者机构：
  
  西安交通大学电子与信息工程学院,西安,710049
- 作者简介：
- 基金信息：
  
  : 2018-01-16。作者简介: 高榕(1993—),男,硕士生;梅魁志(通信作者),男,教授,博士生导师。基金项目: 国家重点研发计划资助项目(2017YFB1301100)
- DOI：10.7652/xjtuxb201806006
  中图分类号： TP183
- 网络首发：2018-06-10，
  
  纸质出版：2018
- 稿件说明：
移动端阅览
高榕, 张良, 梅魁志. 基于Caffe的嵌入式多核处理器深度学习框架并行实现[J]. 西安交通大学学报, 2018,52(6):36-41+113.

A Deep Learning Frame on Embedded Multicore Processors Based on Caffe and Its Parallel Implementation[J]. 2018, 52(6): 36-41+113.
高榕, 张良, 梅魁志. 基于Caffe的嵌入式多核处理器深度学习框架并行实现[J]. 西安交通大学学报, 2018,52(6):36-41+113. DOI： 10.7652/xjtuxb201806006.

A Deep Learning Frame on Embedded Multicore Processors Based on Caffe and Its Parallel Implementation[J]. 2018, 52(6): 36-41+113. DOI： 10.7652/xjtuxb201806006.

摘要

针对开源深度学习快速特征嵌入的卷积框架(Caffe)在Android移动端进行前向计算时存在的兼容性和时间性能差的问题

提出了基于Caffe的嵌入式同构、异构并行化改进设计方法。该方法将Caffe及其第三方库通过交叉编译移植到嵌入式移动平台后

利用同构的多核多线程方法分别对卷积层、输入帧之间的部分前向计算过程进行了并行化; 实现了采用开放运算语言(OpenCL)的异构图形处理器(GPU)卷积计算

进一步提升了框架的处理速度。对3种经典的深度神经网络模型MNIST、Cifar-10和CaffeNet进行了测试对比

测试结果表明:在没有任何模型精度损失的条件下

并行后的前向计算耗时明显低于并行前

时间性能提升最高达到2倍。所提方法能够将深度学习框架Caffe高效地、并行地部署和应用于嵌入式移动多核芯片上。

Abstract

An effective embedded homogeneous and heterogeneous parallel improvement design on the basis of Caffe is proposed to solve the poor compatibility and low efficiency of forward inference in Android mobile terminals by using the open-sourced deep learning frame named Caffe(Convolutional architecture for fast feature embedding). The scheme transplants Caffe and its third-party library to arm architecture using a cross compiler

and then the multi-core and multi-thread technology is used to parallelize partial forward inference between convolution layer and input frame group. An heterogeneous parallel convolutional implementation based on OpenCL is also presented to further improve the time performance of the scheme. Comparison tests with three classic deep learning neural networks MNIST

Cifar-10 and CaffeNet show that in the absence of any model precision loss

the time consuming after parallelization is far less than that before parallel

and time performance increases up to 2 times. It is concluded that the proposal can make the deep learning frame Caffe effectively deploy and work in parallel on portable embedded multicore devices.

关键词

Keywords

references

JIA Yangqing, SHELHAMER E, DONAHUE J, et al. Caffe: convolutional architecture for fast feature embedding [C]∥Proceedings of the 22nd ACM International Conference on Multimedia. New York, USA: ACM, 2014: 675-678.

Google Inc. Tensorflow [EB/OL].(2017-09-23)[2017-10-10]. https: ∥github.com/tensorflow/tenso-rflow.

Facebook Inc. caffe2 [EB/OL].(2017-05-22)[2017-06-24]. https: ∥github.com/caffe2/caffe2.

Tencent Inc. ncnn [EB/OL].(2017-09-23)[2017-11-04]. https: ∥github.com/Tencent/ncnn.

Baidu Inc. Mobile-deep-learning [EB/OL].(2017-09-03)[2017-10-02]. https: ∥github.com/baidu/mobi-le-deep-learning.

HAN Song, MAO Huizi, DALLY W J. Deep compression: compressing deep neural networks with pruning, trained quantization and Huffman coding [J]. Fiber, 2015, 56(4): 3-7.

LIN D D, TALATHI S S, ANNAPUREDDY V S. Fixed point quantization of deep convolutional networks [J]. Computer Science, 2016: arXiv: 1511. 06393.

ZHANG Xianyi. OpenBLAS [EB/OL].(2017-02-03)[2017-02-24]. https: ∥github.com/xianyi/OpenBLAS.

MARAT D. NNPACK [EB/OL].(2017-07-21)[2017-09-07]. https: ∥github.com/Maratyszcza/NNPACK.

OSKOUEI S S L, GOLESTANI H, KACHUEE M, et al. GPU-based acceleration of deep convolutional neural networks on mobile platforms [J]. Computer Science, 2015: arXiv: 1511.07376v1.

Apple Inc. OpenCL reference guide [EB/OL].(2017-08-17)[2017-08-22]. https: ∥www.khronos.org/files/opencl22-reference-guide.pdf.

SONG I, KIM H J, JEON P B. Deep learning for real-time robust facial expression recognition on a smartphone [C]∥Proceedings of the IEEE International Conference on Consumer Electronics. Piscataway, NJ, USA: IEEE, 2014: 564-567.

FARABET C, MARTINI B, AKSELROD P, et al. Hardware accelerated convolutional neural networks for synthetic vision systems [J]. IEEE International Symposium on Circuits and Systems, 2010, 54(3): 257-260.

崔继岳, 梅魁志, 刘冬冬, 等. 面向OpenCL的MaliGPU仿真器构建研究 [J]. 西安交通大学学报, 2015, 49(2): 20-24.

CUI Jiyue, MEI Kuizhi, LIU Dongdong, et al. Construction of embedded Mali GPU simulator for OpenCL [J]. Journal of Xi'an Jiaotong University, 2015, 49(2): 20-24.

SH1R0. Caffe-android-lib [EB/OL].(2017-01-01)[2017-01-15]. https: ∥github.com/sh1r0/caffe-and-roid-lib.

浏览量

下载量

CSCD

文章被引用时，请邮件提醒。

提交

工具集

关联资源

深度学习驱动的碳纤维管壁缺陷镜面红外检测无损定量评估方法

面向模型预测控制的风电机组深度Koopman全局线性建模方法

主蒸汽管道破裂事故下蒸汽发生器的动态液位预测

融合U-net网络的纯卷积视频预测模型

一种双阻尼小波赋能的可解释卷积神经网络在轴承故障诊断中的应用