guofengw / deeplearning

Neural network base on c++14, support any number of layers 基于C++14元编程的深度学习神经网络模板类，支持任意层数

Meta-programming DeepLearning

Meta-programming neural network 是一个基于C++14实现的元编程神经网络库 Compile-time matrix constructions, headonly, no dependency, limitless layers, limitless nodes

Feature

• 支持任意深度和超大结点数
• 矩阵运算（CNN采用张量运算）
• 循环类网络输入输出支持多对单、单对多、多对多
• 源码Head-only并且无依赖
• 使用方法极其简单，适合程序局部应用ANN以及用来学习研究
• GPU支持还在计划中

Sample

1) BPNN

#include "BPNN.hpp"
int main()
{
    /// 1. Create a 4 layers NN each layer nodes are 20, 30, 20 and 2
    ///    The first 20 is input layer and the last 2 is output
    typedef mtl::BPNN<20, 30, 20, 2> MyNN;
    MyNN bpnn;
    
    /// 2. Initialize, setup parameters and activate functions
    bpnn.init()
        .set_aberration(0.0001)
        .set_learnrate(0.8)
        .set_sigfunc(mtl::logsig)
        .set_dsigfunc(mtl::dlogsig);

    /// 3. Create input output matrixs, and then enter matrix datas your self
    MyNN::InMatrix inMx;
    MyNN::OutMatrix outMx;
    MyNN::OutMatrix expectMx;
    ///    enter matrix datas ...
    
    /// 4. Training, call train in your own way
    bpnn.train(inMx, outMx, 100);
    
    /// 5. Simulate
    bpnn.simulate(inMx, outMx, expectMx);
}

2) RNN

#include "RNN.hpp"
int main()
{
    /// 1. Create a 4 layers NN each layer nodes are 20, 30, 20 and 2
    ///    The first 20 is input layer and the last 2 is output
    typedef mtl::RNN<20, 30, 20, 2> MyRnn;
    MyRnn rnn;
    
    /// 2. Initialize, setup parameters and activate functions
    rnn.init()
       .set_aberration(0.0001)
       .set_learnrate(0.8)
       .set_sigfunc(mtl::logsig)
       .set_dsigfunc(mtl::dlogsig);

    /// 3. Create input output matrixs, and then enter matrix datas your self
    ///    RNN suport multi-in-out like M:1, 1:M and M:M also 1:1 which is meaningless
    MyRnn::InMatrix<10> inMx; /// 10 input a group, you can change it each training
    MyRnn::OutMatrix<2> outMx; /// 2 ouput a group
    MyRnn::OutMatrix<2> expectMx;
    ///    enter matrix datas ...
    
    /// 4. Training, call train in your own way
    rnn.train(inMx, outMx, 100);
    
    /// 5. Simulate
    rnn.simulate(inMx, outMx，expectMx);
}

3) LSTM

#include "LSTM.hpp"
int main()
{
    /// 1. Create a 4 layers NN each layer nodes are 20, 30, 20 and 2
    ///    The first 20 is input layer and the last 2 is output
    typedef mtl::LSTM<20, 30, 20, 2> MyLSTM;
    MyLSTM lstm;
    
    /// 2. Initialize, setup parameters, LSTM wouldn't setup activate functions
    lstm.init()
        .set_aberration(0.0001)
        .set_learnrate(0.8);

    /// 3. Create input output matrixs, and then enter matrix datas your self
    ///    RNN suport multi-in-out like M:1, 1:M and M:M also 1:1 which is meaningless
    MyLSTM::InMatrix<10> inMx;
    MyLSTM::OutMatrix<2> outMx;
    MyLSTM::OutMatrix<2> expectMx;
    ///    enter matrix datas ...
    
    /// 4. Training, call train in your own way
    lstm.train(inMx, outMx, 100);
    
    /// 5. Simulate
    lstm.simulate(inMx, outMx，expectMx);
}

4) CNN

coding ...

5) instance of MNIST

coming soon ...

6) GPU support

planning ...

7) Future

RCNN, GAN and custom cnn structure like VGG, ResNet, DenseNet ...