OpenNN::ConjugateGradient Class Reference

#include <conjugate_gradient.h>

Inheritance diagram for OpenNN::ConjugateGradient:

Classes
struct	ConjugateGradientResults

Public Types
enum	TrainingDirectionMethod { PR, FR }

Public Member Functions
	ConjugateGradient (void)
	ConjugateGradient (PerformanceFunctional *)
	ConjugateGradient (const tinyxml2::XMLDocument &)
virtual	~ConjugateGradient (void)
const TrainingRateAlgorithm &	get_training_rate_algorithm (void) const
TrainingRateAlgorithm *	get_training_rate_algorithm_pointer (void)
const TrainingDirectionMethod &	get_training_direction_method (void) const
std::string	write_training_direction_method (void) const
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
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 size_t &	get_maximum_generalization_performance_decreases (void) const
const double &	get_gradient_norm_goal (void) const
const size_t &	get_maximum_iterations_number (void) const
const double &	get_maximum_time (void) const
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_generalization_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_training_direction_history (void) const
const bool &	get_reserve_training_rate_history (void) const
const bool &	get_reserve_elapsed_time_history (void) const
void	set_default (void)
void	set_performance_functional_pointer (PerformanceFunctional *)
void	set_training_direction_method (const TrainingDirectionMethod &)
void	set_training_direction_method (const std::string &)
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 &)
void	set_minimum_parameters_increment_norm (const double &)
void	set_performance_goal (const double &)
void	set_minimum_performance_increase (const double &)
void	set_maximum_generalization_performance_decreases (const size_t &)
void	set_gradient_norm_goal (const double &)
void	set_maximum_iterations_number (const size_t &)
void	set_maximum_time (const double &)
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_generalization_performance_history (const bool &)
void	set_reserve_gradient_history (const bool &)
void	set_reserve_gradient_norm_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 &)
void	set_reserve_all_training_history (const bool &)
void	set_display_period (const size_t &)
void	set_save_period (const size_t &)
double	calculate_PR_parameter (const Vector< double > &, const Vector< double > &) const
double	calculate_FR_parameter (const Vector< double > &, const Vector< double > &) const
Vector< double >	calculate_PR_training_direction (const Vector< double > &, const Vector< double > &, const Vector< double > &) const
Vector< double >	calculate_FR_training_direction (const Vector< double > &, const Vector< double > &, const Vector< double > &) const
Vector< double >	calculate_gradient_descent_training_direction (const Vector< double > &) const
Vector< double >	calculate_training_direction (const Vector< double > &, const Vector< double > &, const Vector< double > &) const
ConjugateGradientResults *	perform_training (void)
std::string	write_training_algorithm_type (void) const
Matrix< std::string >	to_string_matrix (void) const
tinyxml2::XMLDocument *	to_XML (void) const
void	from_XML (const tinyxml2::XMLDocument &)
Public Member Functions inherited from OpenNN::TrainingAlgorithm
	TrainingAlgorithm (void)
	TrainingAlgorithm (PerformanceFunctional *)
	TrainingAlgorithm (const tinyxml2::XMLDocument &)
virtual	~TrainingAlgorithm (void)
virtual TrainingAlgorithm &	operator= (const TrainingAlgorithm &)
virtual bool	operator== (const TrainingAlgorithm &) const
PerformanceFunctional *	get_performance_functional_pointer (void) const
bool	has_performance_functional (void) const
const bool &	get_display (void) const
const size_t &	get_display_period (void) const
const size_t &	get_save_period (void) const
const std::string &	get_neural_network_file_name (void) const
void	set (void)
void	set (PerformanceFunctional *)
void	set_display (const bool &)
void	set_display_period (const size_t &)
void	set_save_period (const size_t &)
void	set_neural_network_file_name (const std::string &)
virtual void	check (void) const
virtual std::string	to_string (void) const
void	print (void) const
void	save (const std::string &) const
void	load (const std::string &)
virtual void	initialize_random (void)

Private Attributes
TrainingDirectionMethod	training_direction_method
TrainingRateAlgorithm	training_rate_algorithm
double	warning_parameters_norm
double	warning_gradient_norm
double	warning_training_rate
double	error_parameters_norm
double	error_gradient_norm
double	error_training_rate
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
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

Additional Inherited Members
Protected Attributes inherited from OpenNN::TrainingAlgorithm
PerformanceFunctional *	performance_functional_pointer
size_t	display_period
size_t	save_period
std::string	neural_network_file_name
bool	display

Detailed Description

This concrete class represents a conjugate gradient training algorithm for a performance functional of a neural network.

Definition at line 47 of file conjugate_gradient.h.

Constructor & Destructor Documentation

OpenNN::ConjugateGradient::ConjugateGradient ( void  )

explicit

Default constructor. It creates a conjugate gradient training algorithm object not associated to any performance functional object. It also initializes the class members to their default values.

Definition at line 27 of file conjugate_gradient.cpp.

OpenNN::ConjugateGradient::ConjugateGradient ( PerformanceFunctional *  new_performance_functional_pointer )

explicit

General constructor. It creates a conjugate gradient training algorithm associated to a performance functional object. It also initializes the rest of class members to their default values.

Parameters

new_performance_functional_pointer

Pointer to a performance functional object.

Definition at line 40 of file conjugate_gradient.cpp.

OpenNN::ConjugateGradient::ConjugateGradient ( const tinyxml2::XMLDocument &  conjugate_gradient_document )

explicit

XML constructor. It creates a conjugate gradient training algorithm not associated to any performance functional object. It also loads the class members from a XML document.

Parameters

conjugate_gradient_document

TinyXML document with the members of a conjugate gradient object.

Definition at line 56 of file conjugate_gradient.cpp.

Member Function Documentation

double OpenNN::ConjugateGradient::calculate_FR_parameter	(	const Vector< double > &	old_gradient,
		const Vector< double > &	gradient
	)		const

Returns the Fletcher-Reeves parameter used to calculate the training direction.

Parameters

old_gradient	Previous objective function gradient.
gradient	Current objective function gradient.

Definition at line 1059 of file conjugate_gradient.cpp.

Vector< double > OpenNN::ConjugateGradient::calculate_FR_training_direction	(	const Vector< double > &	old_gradient,
		const Vector< double > &	gradient,
		const Vector< double > &	old_training_direction
	)		const

Returns the training direction using the Fletcher-Reeves update.

Parameters

old_gradient	Previous objective function gradient.
gradient	Current objective function gradient.
old_training_direction	Previous training direction vector.

Definition at line 1302 of file conjugate_gradient.cpp.

Vector< double > OpenNN::ConjugateGradient::calculate_gradient_descent_training_direction

(

const Vector< double > &

gradient

)

const

Returns the gradient descent training direction, which is the negative of the normalized gradient.

Parameters

gradient

Gradient vector.

Definition at line 1476 of file conjugate_gradient.cpp.

double OpenNN::ConjugateGradient::calculate_PR_parameter	(	const Vector< double > &	old_gradient,
		const Vector< double > &	gradient
	)		const

Returns the Polak-Ribiere parameter used to calculate the training direction.

Parameters

old_gradient	Previous objective function gradient.
gradient	Current objective function gradient.

Definition at line 1139 of file conjugate_gradient.cpp.

Vector< double > OpenNN::ConjugateGradient::calculate_PR_training_direction	(	const Vector< double > &	old_gradient,
		const Vector< double > &	gradient,
		const Vector< double > &	old_training_direction
	)		const

Returns the training direction using the Polak-Ribiere update.

Parameters

old_gradient	Previous objective function gradient.
gradient	Current objective function gradient.
old_training_direction	Previous training direction vector.

Definition at line 1225 of file conjugate_gradient.cpp.

Vector< double > OpenNN::ConjugateGradient::calculate_training_direction	(	const Vector< double > &	old_gradient,
		const Vector< double > &	gradient,
		const Vector< double > &	old_training_direction
	)		const

Returns the conjugate gradient training direction, which has been previously normalized.

Parameters

old_gradient	Gradient vector in the previous iteration.
gradient	Current gradient vector.
old_training_direction	Training direction in the previous iteration.

Definition at line 1379 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::from_XML ( const tinyxml2::XMLDocument &  document )

virtual

Deserializes the conjugate gradient object from a XML document of the TinyXML library.

Parameters

document

TinyXML document containing the member data.

Reimplemented from OpenNN::TrainingAlgorithm.

Definition at line 2729 of file conjugate_gradient.cpp.

const double & OpenNN::ConjugateGradient::get_error_gradient_norm

(

void

)

const

Returns the value for the norm of the gradient vector at wich an error message is written to the screen and the program exits.

Definition at line 186 of file conjugate_gradient.cpp.

const double & OpenNN::ConjugateGradient::get_error_training_rate

(

void

)

const

Returns the training rate value at wich the line minimization algorithm is assumed to fail when bracketing a minimum.

Definition at line 197 of file conjugate_gradient.cpp.

const double & OpenNN::ConjugateGradient::get_gradient_norm_goal

(

void

)

const

Returns the goal value for the norm of the objective function gradient. This is used as a stopping criterion when training a multilayer perceptron

Definition at line 239 of file conjugate_gradient.cpp.

const double & OpenNN::ConjugateGradient::get_performance_goal

(

void

)

const

Returns the goal value for the performance. This is used as a stopping criterion when training a multilayer perceptron

Definition at line 228 of file conjugate_gradient.cpp.

ConjugateGradient::ConjugateGradientResults * OpenNN::ConjugateGradient::perform_training ( void  )

virtual

Trains a neural network with an associated performance functional according to the conjugate gradient algorithm. Training occurs according to the training operators, training parameters and stopping criteria.

Implements OpenNN::TrainingAlgorithm.

Definition at line 1770 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_default ( void  )

virtual

Sets the default values into a conjugate gradient object. Training operators:

• Training direction method = Polak-Ribiere;
• Training rate method = Brent;

Training parameters:

• First training rate: 1.0.
• Bracketing factor: 2.0.
• Training rate tolerance: 1.0e-3.

Stopping criteria:

• Performance goal: -1.0e99.
• Gradient norm goal: 0.0.
• Maximum training time: 1.0e6.
• Maximum number of iterations: 100.

User stuff:

• Warning training rate: 1.0e6.
• Error training rate: 1.0e12.
• Display: true.
• Display period: 10.
• Save period: 0.

Reserve:

• Reserve training direction history: false.
• Reserve training direction norm history: false.
• Reserve training rate history: false.

Reimplemented from OpenNN::TrainingAlgorithm.

Definition at line 504 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_display_period

(

const size_t &

new_display_period

)

Sets a new number of iterations between the training showing progress.

Parameters

new_display_period

Number of iterations between the training showing progress.

Definition at line 1000 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_error_gradient_norm

(

const double &

new_error_gradient_norm

)

Sets a new value for the gradient vector norm at which an error message is written to the screen and the program exits.

Parameters

new_error_gradient_norm

Error norm of gradient vector value.

Definition at line 679 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_error_parameters_norm

(

const double &

new_error_parameters_norm

)

Sets a new value for the parameters vector norm at which an error message is written to the screen and the program exits.

Parameters

new_error_parameters_norm

Error norm of parameters vector value.

Definition at line 648 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_error_training_rate

(

const double &

new_error_training_rate

)

Sets a new training rate value at wich a the line minimization algorithm is assumed to fail when bracketing a minimum.

Parameters

new_error_training_rate

Error training rate value.

Definition at line 710 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_gradient_norm_goal

(

const double &

new_gradient_norm_goal

)

Sets a new the goal value for the norm of the objective function gradient. This is used as a stopping criterion when training a multilayer perceptron

Parameters

new_gradient_norm_goal

Goal value for the norm of the objective function gradient.

Definition at line 813 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_maximum_generalization_performance_decreases

(

const size_t &

new_maximum_generalization_performance_decreases

)

Sets a new maximum number of generalization failures.

Parameters

new_maximum_generalization_performance_decreases

Maximum number of iterations in which the generalization evalutation decreases.

Definition at line 843 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_maximum_iterations_number

(

const size_t &

new_maximum_iterations_number

)

Sets a maximum number of iterations for training.

Parameters

new_maximum_iterations_number

Maximum number of iterations for training.

Definition at line 854 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_maximum_time

(

const double &

new_maximum_time

)

Sets a new maximum training time.

Parameters

new_maximum_time

Maximum training time.

Definition at line 865 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_minimum_parameters_increment_norm

(

const double &

new_minimum_parameters_increment_norm

)

Sets a new value for the minimum parameters increment norm stopping criterion.

Parameters

new_minimum_parameters_increment_norm

Value of norm of parameters increment norm used to stop training.

Definition at line 740 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_minimum_performance_increase

(

const double &

new_minimum_performance_increase

)

Sets a new minimum performance improvement during training.

Parameters

new_minimum_performance_increase

Minimum improvement in the performance between two iterations.

Definition at line 770 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_performance_functional_pointer ( PerformanceFunctional *  new_performance_functional_pointer )

virtual

Sets a pointer to a performance functional object to be associated to the conjugate gradient object. It also sets that performance functional to the training rate algorithm.

Parameters

new_performance_functional_pointer

Pointer to a performance functional object.

Reimplemented from OpenNN::TrainingAlgorithm.

Definition at line 372 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_performance_goal

(

const double &

new_performance_goal

)

Sets a new goal value for the performance. This is used as a stopping criterion when training a multilayer perceptron

Parameters

new_performance_goal

Goal value for the performance.

Definition at line 801 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_reserve_all_training_history

(

const bool &

new_reserve_all_training_history

)

Makes the training history of all variables to reseved or not in memory when training.

• Parameters.
• Parameters norm.
• performance.
• Gradient.
• Gradient norm.
• Generalization performance.
• Training direction.
• Training direction norm.
• Training rate.

Parameters

new_reserve_all_training_history

True if all training history variables are to be reserved, false otherwise.

Definition at line 443 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_reserve_elapsed_time_history

(

const bool &

new_reserve_elapsed_time_history

)

Makes the elapsed time over the iterations to be reseved or not in memory. This is a vector.

Parameters

new_reserve_elapsed_time_history

True if the elapsed time history vector is to be reserved, false otherwise.

Definition at line 976 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_reserve_generalization_performance_history

(

const bool &

new_reserve_generalization_performance_history

)

Makes the Generalization performance history to be reserved or not in memory. This is a vector.

Parameters

new_reserve_generalization_performance_history

True if the Generalization performance history is to be reserved, false otherwise.

Definition at line 988 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_reserve_gradient_history

(

const bool &

new_reserve_gradient_history

)

Makes the gradient history vector of vectors to be reseved or not in memory.

Parameters

new_reserve_gradient_history

True if the gradient history matrix is to be reserved, false otherwise.

Definition at line 928 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_reserve_gradient_norm_history

(

const bool &

new_reserve_gradient_norm_history

)

Makes the gradient norm history vector to be reseved or not in memory.

Parameters

new_reserve_gradient_norm_history

True if the gradient norm history matrix is to be reserved, false otherwise.

Definition at line 940 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_reserve_parameters_history

(

const bool &

new_reserve_parameters_history

)

Makes the parameters history vector of vectors to be reseved or not in memory.

Parameters

new_reserve_parameters_history

True if the parameters history vector of vectors is to be reserved, false otherwise.

Definition at line 895 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_reserve_parameters_norm_history

(

const bool &

new_reserve_parameters_norm_history

)

Makes the parameters norm history vector to be reseved or not in memory.

Parameters

new_reserve_parameters_norm_history

True if the parameters norm history vector is to be reserved, false otherwise.

Definition at line 906 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_reserve_performance_history

(

const bool &

new_reserve_performance_history

)

Makes the performance history vector to be reseved or not in memory.

Parameters

new_reserve_performance_history

True if the performance history vector is to be reserved, false otherwise.

Definition at line 917 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_reserve_training_direction_history

(

const bool &

new_reserve_training_direction_history

)

Makes the training direction history vector of vectors to be reseved or not in memory.

Parameters

new_reserve_training_direction_history

True if the training direction history matrix is to be reserved, false otherwise.

Definition at line 952 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_reserve_training_rate_history

(

const bool &

new_reserve_training_rate_history

)

Makes the training rate history vector to be reseved or not in memory.

Parameters

new_reserve_training_rate_history

True if the training rate history vector is to be reserved, false otherwise.

Definition at line 964 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_save_period

(

const size_t &

new_save_period

)

Sets a new number of iterations between the training saving progress.

Parameters

new_save_period

Number of iterations between the training saving progress.

Definition at line 1029 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_training_direction_method

(

const TrainingDirectionMethod &

new_training_direction_method

)

Sets a new training direction method to be used for training.

Parameters

new_training_direction_method

Conjugate gradient training direction method.

Definition at line 386 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_training_direction_method

(

const std::string &

new_training_direction_method_name

)

Sets a new conjugate gradient training direction from a string representation. Possible values are:

• "PR"
• "FR"

Parameters

new_training_direction_method_name

String with the name of the training direction method.

Definition at line 402 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_warning_gradient_norm

(

const double &

new_warning_gradient_norm

)

Sets a new value for the gradient vector norm at which a warning message is written to the screen.

Parameters

new_warning_gradient_norm

Warning norm of gradient vector value.

Definition at line 588 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_warning_parameters_norm

(

const double &

new_warning_parameters_norm

)

Sets a new value for the parameters vector norm at which a warning message is written to the screen.

Parameters

new_warning_parameters_norm

Warning norm of parameters vector value.

Definition at line 557 of file conjugate_gradient.cpp.

void OpenNN::ConjugateGradient::set_warning_training_rate

(

const double &

new_warning_training_rate

)

Sets a new training rate value at wich a warning message is written to the screen during line minimization.

Parameters

new_warning_training_rate

Warning training rate value.

Definition at line 619 of file conjugate_gradient.cpp.

Matrix< std::string > OpenNN::ConjugateGradient::to_string_matrix ( void  ) const

virtual

Returns a default (empty) string matrix containing the members of the training algorithm object.

Reimplemented from OpenNN::TrainingAlgorithm.

Definition at line 2187 of file conjugate_gradient.cpp.

tinyxml2::XMLDocument * OpenNN::ConjugateGradient::to_XML ( void  ) const

virtual

Serializes the conjugate gradient object into a XML document of the TinyXML library. See the OpenNN manual for more information about the format of this element.

Reimplemented from OpenNN::TrainingAlgorithm.

Definition at line 2360 of file conjugate_gradient.cpp.

Member Data Documentation

size_t OpenNN::ConjugateGradient::maximum_generalization_performance_decreases

private

Maximum number of iterations at which the generalization performance decreases. This is an early stopping method for improving generalization.

Definition at line 380 of file conjugate_gradient.h.

---

The documentation for this class was generated from the following files:

• conjugate_gradient.h
• conjugate_gradient.cpp