gradient_descent.h

//   OpenNN: Open Neural Networks Library
//   www.opennn.net
//
//   G R A D I E N T   D E S C E N T   C L A S S   H E A D E R
//
//   Artificial Intelligence Techniques SL
//   artelnics@artelnics.com
 
#ifndef GRADIENTDESCENT_H
#define GRADIENTDESCENT_H
 
// System includes
 
#include <string>
#include <sstream>
#include <iostream>
#include <fstream>
#include <algorithm>
#include <functional>
#include <limits>
#include <limits.h>
#include <cmath>
#include <ctime>
 
// OpenNN includes
 
#include "loss_index.h"
 
#include "optimization_algorithm.h"
#include "learning_rate_algorithm.h"
#include "config.h"
 
namespace OpenNN
{
 
struct GradientDescentData;
 
 
 
class GradientDescent : public OptimizationAlgorithm
{
 
public:
 
   // Constructors
 
   explicit GradientDescent(); 
 
   explicit GradientDescent(LossIndex*);
 
   // Destructor
 
   virtual ~GradientDescent();   
 
   const LearningRateAlgorithm& get_learning_rate_algorithm() const;
   LearningRateAlgorithm* get_learning_rate_algorithm_pointer();
 
   string get_hardware_use() const;
 
   // Stopping criteria   
 
   const type& get_minimum_loss_decrease() const;
   const type& get_loss_goal() const;
 
   const Index& get_maximum_selection_failures() const;
 
   const Index& get_maximum_epochs_number() const;
   const type& get_maximum_time() const;
 
   // Set methods
 
   void set_loss_index_pointer(LossIndex*);
 
   void set_learning_rate_algorithm(const LearningRateAlgorithm&);
 
   void set_default();
 
   // Stopping criteria
 
   void set_maximum_epochs_number(const Index&);
 
   
 
   void set_minimum_loss_decrease(const type&);
   void set_loss_goal(const type&);
 
   void set_maximum_selection_failures(const Index&);
 
   void set_maximum_time(const type&);
 
   // Training methods
 
   void calculate_training_direction(const Tensor<type, 1>&, Tensor<type, 1>&) const;
 
   void update_parameters(
           const DataSetBatch& batch,
           NeuralNetworkForwardPropagation& forward_propagation,
           LossIndexBackPropagation& back_propagation,
           GradientDescentData& optimization_data);
 
   TrainingResults perform_training();   
 
   string write_optimization_algorithm_type() const;
 
   // Serialization methods
 
   Tensor<string, 2> to_string_matrix() const;
 
   void from_XML(const tinyxml2::XMLDocument&);
 
   void write_XML(tinyxml2::XMLPrinter&) const;
 
private:
 
   // TRAINING OPERATORS
 
 
   LearningRateAlgorithm learning_rate_algorithm;
 
   const type first_learning_rate = static_cast<type>(0.01);
 
   // Stopping criteria 
 
 
   type minimum_loss_decrease;
 
 
   type training_loss_goal;
 
 
   Index maximum_selection_failures;
 
 
   Index maximum_epochs_number;
 
 
   type maximum_time;
 
};
 
 
struct GradientDescentData : public OptimizationAlgorithmData
{
 
    explicit GradientDescentData()
    {
    }
 
 
    explicit GradientDescentData(GradientDescent* new_gradient_descent_pointer)
    {
        set(new_gradient_descent_pointer);
    }
 
 
    virtual ~GradientDescentData() {}
 
    void set(GradientDescent* new_gradient_descent_pointer)
    {
        gradient_descent_pointer = new_gradient_descent_pointer;
 
        LossIndex* loss_index_pointer = gradient_descent_pointer->get_loss_index_pointer();
 
        NeuralNetwork* neural_network_pointer = loss_index_pointer->get_neural_network_pointer();
 
        const Index parameters_number = neural_network_pointer->get_parameters_number();
 
        // Neural network data
 
        potential_parameters.resize(parameters_number);
 
        parameters_increment.resize(parameters_number);
 
        // Optimization algorithm data
 
        training_direction.resize(parameters_number);
    }
 
 
    void print() const
    {
        cout << "Training direction:" << endl;
        cout << training_direction << endl;
 
        cout << "Learning rate:" << endl;
        cout << learning_rate << endl;
    }
 
    GradientDescent* gradient_descent_pointer = nullptr;
 
    // Neural network data
 
    Tensor<type, 1> parameters_increment;
 
    // Optimization algorithm data
 
    Index epoch = 0;
 
    Tensor<type, 0> training_slope;
 
    type learning_rate = type(0);
    type old_learning_rate = type(0);
};
 
}
 
#endif