KhuDaNet(Perceptron, MLP, CNN)

• KhuDaNet: Open source platform for neural networks (Kyung Hee University, Data Analysis & Vision Intelligence)
• C++, (Perceptron, MLP(DNN), CNN)

Source file

• KhuDaNetLayer.cpp / KhuDaNetLayer.h
  • Fully connected layer
• KhuDaNet.cpp / KhuDaNet.h
  • Sequential network

MNIST test

• Download train and test data from http://yann.lecun.com/exdb/mnist/

• Move to decompressed files to [Run] folder

• Run the following files

  • KhuDaNetApp.exe: Windows application using MFC
  • KhuDaNetConsole.exe: console application

• Network I - 1.KhuDaNet-Perceptron

  • Input: 28*28 pixel values ranging [0, 1]
  • Output: 1 node for binary decision (class 0: digits 0, 1, 2, 3, 4, class 1: digits 5, 6, 7, 8, 9)
    
    

• Network II - 2.KhuDaNet-MLP-DNN

  • Input layer: 28*28 pixel values ranging [0, 1]
  • Hidden layers: 2 hidden layers
  • Output layer: 10 node for 10 digits
    
    

• Network III (LeNet5-like CNN) - 3.KhuDaNet-CNN

  • Input layer: 32*32 pixel values ranging [0, 1]
  • Hidden layers: CNN-POOL-CNN-POOL-CNN-FC
  • Output layer: 10 node for 10 digits
    
    

• Solution files for Visual Studio

  • KhuDaNetApp.sln: Windows application using MFC
  • KhuDaNetConsole.sln: console application