Metaheuristics-Template-Library

Metaheuristics Template Library(MTL) is supported Genetic Algorithm, Simulated Annealing and Neural Network.

Attention: It is under development.

##Overview MTL is a header-only library for the meta-heuristics programming.

When you want to use the meta-heuristics, or when you need more abstract and highly reusable architecture, this library would be helpful.

Description

MTL is supporting the following algorithm.

1. Genetic Algorithm

2. Simulated Annealing

3. Neural Network

MTL is a static implementation and works faster than the other library. Instead, the compilation will take more time.

Anyone who is using recently C++ would be able to handle this MTL easily.

Requirement

C++14 Compiler

OpenCV 2.49 (TSP of GA/SA sample)

Install

Please place the MTL repository in the same directory with your project and compile it.

compile option: -std=c++1y

Usage

NeuralNetwork

#include"NNSolver.hpp"
#include<fstream>
#include<vector>
#include<utility>

int main() {
/* Neural Network Learning Solver(NNSolver) requires template arguments            */
/* First template arg: network structure                  */
/* Second template arg: activation function            */

/* sample: static network                              */
/* mtl::FeedForward constructs multilayer perceptron network as statically       */
/* mtl::FeedForward requires variadic template arguments */
/* variadic template arguments :  the number of units on each layer. order: input-output-hide(multi) */
/* below sample constructs 2 unit input layers / 1 unit output layers */
 mtl::NNSolver< mtl::FeedForward<2, 1>, mtl::tanh_af > solver(0.05);

/* training samples */
/* pair.first : input, pair.second : output */
 std::vector< std::pair< std::array<double,2>, std::array<double,1> > > list;
 list.push_back(std::make_pair( std::array<double,2>{1,1}, std::array<double,1>{1} ));
 list.push_back(std::make_pair( std::array<double,2>{1,-1}, std::array<double,1>{-1} ));
 list.push_back(std::make_pair( std::array<double,2>{-1,1}, std::array<double,1>{-1} ));
 list.push_back(std::make_pair( std::array<double,2>{-1,-1}, std::array<double,1>{-1} ));
 
 /* Exec learning */
 /* training function requires learning algorithm */
 solver.training<mtl::Backpropagation>(list);
 
 /* if training succeeded, solveAnswer function returns correctly value...  */
 auto ans = solver.solveAnswer({0, 0});
 
 /* dynamic network version */
 std::vector<mtl::FeedForward_Dy::size_t> network_struct = {2, 4, 1}; // input: 2 hide: 4 output: 1
 mtl::NNSolver< mtl::FeedForward_Dy, mtl::tanh_af > solver(network_struct);
 ...
 solver.training<mtl::Backpropagation>(0.15, list);
 
 /* gpu-accelerated version */
std::vector<mtl::FeedForward_Dy::size_t> network_struct = {2, 4, 1};
mtl::FeedForward_Amp<4> network; //template argument requires max unit size
   network.setStruct(network_struct);

   mtl::NNSolver< mtl::FeedForward_Amp_View<4>, mtl::tanh_af_gpu_accel > solver(network);

   std::vector< std::pair< std::vector<float>, std::vector<float> > > list;
   list.push_back(std::make_pair(std::vector<float>{1, 1}, std::vector<float>{-1}));
   ...
   
   solver.training<mtl::Backpropagation_Gpu_Accel>(0.15, list); //gpe acceleration algorithm
 
 return 0;
}

Licence

MIT

Link

My thesis

Development of OCR system by Gpu-Accelerated Deep-Learning
Original(ja): GPUによって加速化されたディープラーニングによる光学文字認識システムの開発

Author

りやさん(@Riyaaaa_a)