gradient_descent.h

/****************************************************************************************************************/
/*                                                                                                              */
/*   OpenNN: Open Neural Networks Library                                                                       */
/*   www.artelnics.com/opennn                                                                                   */
/*                                                                                                              */
/*   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                                                  */
/*                                                                                                              */
/*   Roberto Lopez                                                                                              */
/*   Artelnics - Making intelligent use of data                                                                 */
/*   [email protected]                                                                                 */
/*                                                                                                              */
/****************************************************************************************************************/

#ifndef __GRADIENTDESCENT_H__
#define __GRADIENTDESCENT_H__

// System includes

#include <string>
#include <sstream>
#include <iostream>
#include <fstream>
#include <algorithm>
#include <functional>
#include <limits>
#include <cmath>
#include <ctime>

// OpenNN includes

#include "performance_functional.h"

#include "training_algorithm.h"
#include "training_rate_algorithm.h"


namespace OpenNN
{


class GradientDescent : public TrainingAlgorithm
{

public:

   // DEFAULT CONSTRUCTOR

   explicit GradientDescent(void); 

   // PERFORMANCE FUNCTIONAL CONSTRUCTOR

   explicit GradientDescent(PerformanceFunctional*);

   // XML CONSTRUCTOR

   explicit GradientDescent(const tinyxml2::XMLDocument&); 


   // DESTRUCTOR

   virtual ~GradientDescent(void);

   // STRUCTURES


   struct GradientDescentResults : public TrainingAlgorithm::TrainingAlgorithmResults
   {

       GradientDescentResults(void)
       {
           gradient_descent_pointer = NULL;
       }


       GradientDescentResults(GradientDescent* new_gradient_descent_pointer)
       {
           gradient_descent_pointer = new_gradient_descent_pointer;
       }


       virtual ~GradientDescentResults(void)
       {
       }


      GradientDescent* gradient_descent_pointer;

      // Training history


      Vector< Vector<double> > parameters_history;


      Vector<double> parameters_norm_history;


      Vector<double> performance_history;


      Vector<double> generalization_performance_history;


      Vector< Vector<double> > gradient_history;


      Vector<double> gradient_norm_history;


      Vector< Vector<double> >  training_direction_history;


      Vector<double> training_rate_history;


      Vector<double> elapsed_time_history;

      // Final values


      Vector<double> final_parameters;


      double final_parameters_norm;


      double final_performance;


      double final_generalization_performance;


      Vector<double> final_gradient;


      double final_gradient_norm;


      Vector<double> final_training_direction;


      double final_training_rate;


      double elapsed_time;


      size_t iterations_number;


      std::string stopping_criterion;

      void resize_training_history(const size_t&);

      std::string to_string(void) const;

      Matrix<std::string> write_final_results(const size_t& precision = 3) const;
   };

   // METHODS

   const TrainingRateAlgorithm& get_training_rate_algorithm(void) const;
   TrainingRateAlgorithm* get_training_rate_algorithm_pointer(void);

   // Training parameters

   const double& get_warning_parameters_norm(void) const;
   const double& get_warning_gradient_norm(void) const;
   const double& get_warning_training_rate(void) const;

   const double& get_error_parameters_norm(void) const;
   const double& get_error_gradient_norm(void) const;
   const double& get_error_training_rate(void) const;

   // Stopping criteria

   const double& get_minimum_parameters_increment_norm(void) const;

   const double& get_minimum_performance_increase(void) const;
   const double& get_performance_goal(void) const;
   const double& get_gradient_norm_goal(void) const;
   const size_t& get_maximum_generalization_performance_decreases(void) const;

   const size_t& get_maximum_iterations_number(void) const;
   const double& get_maximum_time(void) const;

   // Reserve training history

   const bool& get_reserve_parameters_history(void) const;
   const bool& get_reserve_parameters_norm_history(void) const;

   const bool& get_reserve_performance_history(void) const;
   const bool& get_reserve_gradient_history(void) const;
   const bool& get_reserve_gradient_norm_history(void) const;
   const bool& get_reserve_generalization_performance_history(void) const;

   const bool& get_reserve_training_direction_history(void) const;
   const bool& get_reserve_training_rate_history(void) const;
   const bool& get_reserve_elapsed_time_history(void) const;

   // Set methods

   void set_performance_functional_pointer(PerformanceFunctional*);

   void set_training_rate_algorithm(const TrainingRateAlgorithm&);


   void set_default(void);

   void set_reserve_all_training_history(const bool&);


   // Training parameters

   void set_warning_parameters_norm(const double&);
   void set_warning_gradient_norm(const double&);
   void set_warning_training_rate(const double&);

   void set_error_parameters_norm(const double&);
   void set_error_gradient_norm(const double&);
   void set_error_training_rate(const double&);

   // Stopping criteria

   void set_minimum_parameters_increment_norm(const double&);

   void set_minimum_performance_increase(const double&);
   void set_performance_goal(const double&);
   void set_gradient_norm_goal(const double&);
   void set_maximum_generalization_performance_decreases(const size_t&);

   void set_maximum_iterations_number(const size_t&);
   void set_maximum_time(const double&);

   // Reserve training history

   void set_reserve_parameters_history(const bool&);
   void set_reserve_parameters_norm_history(const bool&);

   void set_reserve_performance_history(const bool&);
   void set_reserve_gradient_history(const bool&);
   void set_reserve_gradient_norm_history(const bool&);
   void set_reserve_generalization_performance_history(const bool&);

   void set_reserve_training_direction_history(const bool&);
   void set_reserve_training_rate_history(const bool&);
   void set_reserve_elapsed_time_history(const bool&);

   // Utilities

   void set_display_period(const size_t&);

   // Training methods

   Vector<double> calculate_training_direction(const Vector<double>&) const;

   GradientDescentResults* perform_training(void);

   std::string write_training_algorithm_type(void) const;

   // Serialization methods

   Matrix<std::string> to_string_matrix(void) const;

   tinyxml2::XMLDocument* to_XML(void) const;
   void from_XML(const tinyxml2::XMLDocument&);

private:

   // TRAINING OPERATORS


   TrainingRateAlgorithm training_rate_algorithm;

   // TRAINING PARAMETERS


   double warning_parameters_norm;


   double warning_gradient_norm;   


   double warning_training_rate;

   
   double error_parameters_norm;


   double error_gradient_norm;


   double error_training_rate;


   // STOPPING CRITERIA


   double minimum_parameters_increment_norm;


   double minimum_performance_increase;


   double performance_goal;


   double gradient_norm_goal;


   size_t maximum_generalization_performance_decreases;


   size_t maximum_iterations_number;


   double maximum_time;

   // TRAINING HISTORY


   bool reserve_parameters_history;


   bool reserve_parameters_norm_history;


   bool reserve_performance_history;


   bool reserve_gradient_history;


   bool reserve_gradient_norm_history;

   
   bool reserve_training_direction_history;


   bool reserve_training_rate_history;


   bool reserve_elapsed_time_history;


   bool reserve_generalization_performance_history;
};

}

#endif