A classic backpropagation SGD Trainer. More...

#include <SGDTrainer.h>

Inherited by net::Adadelta, and net::Backpropagation.

Public Member Functions
	SGDTrainer ()
	Initialize empty Backpropagation object. More...

	SGDTrainer (double targetErrorLevel_, int maximumEpochs_)
	Initialize the object with necessary constants. More...

double	train (net::NeuralNet *network, const std::vector< std::vector< double > > &input, const std::vector< std::vector< double > > &correctOutput)
	Trains a neural network on a training set until the target error level is reached. More...

double	trainEpocs (double numberOfEpochs, net::NeuralNet *network, const std::vector< std::vector< double > > &input, const std::vector< std::vector< double > > &correctOutput)
	Trains a neural network on a training set for a specified number of epochs. More...

void	store (std::ofstream *out)
	Stores a Trainer object using specified stream. More...

bool	initFromStream (std::ifstream *in)

Public Member Functions inherited from net::Trainer
std::vector< std::vector < std::vector< std::vector < double > > > >	getGradients ()

std::vector< std::vector < std::vector< std::vector < double > > > >	getWeightChanges ()

std::vector< std::vector < std::vector< double > > >	getInitialWeights ()

std::vector< std::vector < std::vector< double > > >	getFinalWeights ()

Public Attributes
double	targetErrorLevel
	The target error level, set by constructor. More...

int	maximumEpochs
	The maximum number of iterations, set by constructor. More...

Protected Member Functions
double	trainOnDataPoint (net::NeuralNet *network, const std::vector< double > &input, const std::vector< double > &correctOutput)
	Gets the output of the neural network, calculates the error of each neuron, and edits the weights of the neurons to reduce error. More...

virtual void	resetNetworkVectors (net::NeuralNet *network)
	Resets the Backpropagation object's neural network specific vectors using a neural network (NN is needed because the number of layers, neurons, and weights are needed). More...

virtual double	getChangeInWeight (double weight, int layerIndex, int neuronIndex, int weightIndex)=0

Additional Inherited Members
Protected Attributes inherited from net::Trainer
std::vector< std::vector < std::vector< std::vector < double > > > >	gradients

std::vector< std::vector < std::vector< std::vector < double > > > >	weightChanges

std::vector< std::vector < std::vector< double > > >	initialWeights

std::vector< std::vector < std::vector< double > > >	finalWeights

Detailed Description

A classic backpropagation SGD Trainer.

Constructor & Destructor Documentation

SGDTrainer::SGDTrainer ( )

Initialize empty Backpropagation object.

SGDTrainer::SGDTrainer	(	double	targetErrorLevel_,
		int	maximumEpochs_
	)

Initialize the object with necessary constants.

Parameters

targetErrorLevel_	at this error level, a net will be considered trained
maximumEpochs_	after this number of training iterations (one pass through all of the data points), a net will stop being trained no matter what

Member Function Documentation

virtual double net::SGDTrainer::getChangeInWeight	(	double	weight,
		int	layerIndex,
		int	neuronIndex,
		int	weightIndex
	)

protectedpure virtual

Implemented in net::Backpropagation, and net::Adadelta.

bool SGDTrainer::initFromStream ( std::ifstream * in )

void SGDTrainer::resetNetworkVectors ( net::NeuralNet * network )

protectedvirtual

Resets the Backpropagation object's neural network specific vectors using a neural network (NN is needed because the number of layers, neurons, and weights are needed).

Reimplemented in net::Adadelta.

void SGDTrainer::store ( std::ofstream * output )

virtual

Stores a Trainer object using specified stream.

Parameters

output pointer to the output stream which the neural network will be written to

Implements net::Trainer.

double SGDTrainer::train	(	net::NeuralNet *	network,
		const std::vector< std::vector< double > > &	input,
		const std::vector< std::vector< double > > &	correctOutput
	)

virtual

Trains a neural network on a training set until the target error level is reached.

Edits the weights of the neural network until its error in predicting the correctOutput of each input reaches the value of targetErrorLevel or the number of training cycles reaches the value of maximumIterations. NOTE: If learning rate is not low enough, the weights of the neural network may got to infinity due to the nature of backpropagation.

Parameters

network	the neural network to be trained
input	a vector of neural network inputs; each element in input, should have a corresponding output in correctOutput
correctOutput	network is trained to output an element of correctOutput when fed a corresponding element of the input vector

Implements net::Trainer.

double SGDTrainer::trainEpocs	(	double	numberOfEpochs,
		net::NeuralNet *	network,
		const std::vector< std::vector< double > > &	input,
		const std::vector< std::vector< double > > &	correctOutput
	)

Trains a neural network on a training set for a specified number of epochs.

Edits the weights of the neural network until its error in predicting the correctOutput of each input reaches the value of targetErrorLevel or the number of training cycles reaches the value of maximumIterations. NOTE: If learning rate is not low enough, the weights of the neural network may got to infinity due to the nature of backpropagation.

Parameters

numberOfEpochs	the number of training passes that will be made through the data
network	the neural network to be trained
input	a vector of neural network inputs; each element in input, should have a corresponding output in correctOutput
correctOutput	network is trained to output an element of correctOutput when fed a corresponding element of the input vector

double SGDTrainer::trainOnDataPoint	(	net::NeuralNet *	network,
		const std::vector< double > &	input,
		const std::vector< double > &	correctOutput
	)

protected

Gets the output of the neural network, calculates the error of each neuron, and edits the weights of the neurons to reduce error.

Parameters

network	the neural network to be trained
input	the input fed to the neural network
correctOutput	network is trained to output this when fed the input vector

Member Data Documentation

int net::SGDTrainer::maximumEpochs

The maximum number of iterations, set by constructor.

double net::SGDTrainer::targetErrorLevel

The target error level, set by constructor.

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

/home/truell20/Documents/Fido/Fido/include/SGDTrainer.h
/home/truell20/Documents/Fido/Fido/src/SGDTrainer.cpp

Public Member Functions

Public Attributes

Protected Member Functions

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

Member Function Documentation

Member Data Documentation