拆轮子 笔记

拆程序

[ ] 分析一个发展到半程的DL项目CXXNET

• 开始时间
• 源码阅读,code structure
  
  • coding guide
    
    • 格式,注释文档
  • getting started
    
    • src文件夹下的文件夹都有一个.h文件.这个文件是一个模块module,.h文件是模块接口interface.
    • 所有其他的-inl.hpp文件是接口的实现implementations.
      
      • 所有接口都是对其他模块可视的.
      • Templatized class with parameter xpu that can stands for cpu or gpu.
    • The project depends on mshadow for tensor operations.
  • 项目的逻辑布局
    
    • 依赖顺序:nnet->updater->layer
      
      • 所有模块依赖global.h and utils
      • io is an independent module
    • ayer is implementation of neural net layers and defines forward and backward propagation
    • updater is the parameter updating module,it defines update rule of weights.
      
      • AsyncUpdater is a special updater that handles asynchronize communication and update
      • It uses mshadow-ps to do async communication
    • nnet is the neural net structure that combines layers together to form a neural net
      
      • Dependency in nnet: CXXNetThreadTrainer->NeuralNetThread->NeuralNet
    • io is the input module to handle reading various data and preprocessing.
      -( io看来是数据传输模块,涉及底层的操作
      
      • io uses iterator pattern to handle data processing pipeline
      • The pipeline can be mult-threaded using threadbuffer trick
  • 所有的源码怎么组织工作起来.
    
    • 数据从io模块拉入nnet
    • nnet获得已经获取部分数据的#gpu线程，调用层对象来做forwardbackprop(forward and backward propagation)
    • 对于每个权重，更新(updater)被创建
      
      • AsyncUpdater.AfterBackprop is called after backprop of the corresponding layer to push out gradient
      • AsyncUPdater.UpdateWait is called before forward to the layer
      • mshadow-ps does the async trick of parameter communication
    • AsyncUpdater will call IUpdater, which does the updating trick
      
      • If update_on_server is on, IUpdater will be created on server-side instead
  • 文件命名惯例
• .h files are data structures and interface
  
  • In each folder, there is one .h file that have same name as the folder, this file defines everything needed for other module to use this module
  • Interface headers: layer/layer.h, updater/updater.h
• -inl.hpp files are implementations of interface, like cpp file in most project.
  
  • You only need to understand the interface file to understand the usage of that layer
• In each folder, there can be a .cpp file, and .cu file that that compiles the module of that layer
  
  • the .cpp file and .cu file does not contain implementation, but reuse common implementation in file ends with _impl-inl.hpp

分析项目

• automake 工具使用
  
  • 原因是缺少libtool,其实也是缺少aclocal等工具.
  • ?但是安装automake时已经自动安装了m4,autoconf.是否可以通过设置 m4_pattern_allow?

configure.ac:16: error: possibly undefined macro: AC_PROG_LIBTOOL

                     If this token and others are legitimate, please use m4_pattern_allow.
                     See the Autoconf documentation.
                     autoreconf: /usr/bin/autoconf failed with exit status: 1

[ ] 怎么查找错误?

源码阅读

• 源码阅读心得