Class representing a multi-layer neural network based on backpropagation/gradient descent. More...

#include <BackpropagationNeuralNetwork.hpp>

Inheritance diagram for mic::mlnn::BackpropagationNeuralNetwork< eT >:

Collaboration diagram for mic::mlnn::BackpropagationNeuralNetwork< eT >:

Public Member Functions
	BackpropagationNeuralNetwork (std::string name_="bp_net")

virtual	~BackpropagationNeuralNetwork ()
	Virtual descriptor - empty. More...

template<typename LossFunction >
void	setLoss ()

void	forward (mic::types::MatrixPtr< eT > input_data, bool skip_dropout=false)

bool	verify ()

void	backward (mic::types::MatrixPtr< eT > gradients_)

eT	train (mic::types::MatrixPtr< eT > encoded_batch_, mic::types::MatrixPtr< eT > encoded_targets_, eT learning_rate_, eT decay_=0.0f)

eT	test (mic::types::MatrixPtr< eT > encoded_batch_, mic::types::MatrixPtr< eT > encoded_targets_)

eT	calculateMeanLoss (mic::types::MatrixPtr< eT > encoded_targets_, mic::types::MatrixPtr< eT > encoded_predictions_)

Public Member Functions inherited from mic::mlnn::MultiLayerNeuralNetwork< eT >
	MultiLayerNeuralNetwork (std::string name_="mlnn")

virtual	~MultiLayerNeuralNetwork ()

template<typename LayerType >
void	pushLayer (LayerType *layer_ptr_)

template<typename LayerType >
std::shared_ptr< LayerType >	getLayer (size_t index_)

std::shared_ptr< Layer< eT > >	getLayer (size_t index_)

void	popLayer (size_t number_of_layers_=1)

size_t	layerInputsSize (size_t layer_number_=-1)

size_t	lastLayerOutputsSize (size_t layer_number_=-1)

size_t	lastLayerBatchSize (size_t layer_number_=-1)

template<typename omT >
void	setOptimization ()

void	update (eT alpha_, eT decay_=0.0f)

void	resetGrads ()

void	resizeBatch (size_t batch_size_)

mic::types::MatrixPtr< eT >	getPredictions ()

mic::types::MatrixPtr< eT >	getPredictions (size_t layer_nr_)

size_t	countCorrectPredictions (mic::types::MatrixPtr< eT > targets_, mic::types::MatrixPtr< eT > predictions_)

bool	save (std::string filename_)

bool	load (std::string filename_)

Protected Attributes
std::shared_ptr < mic::neural_nets::loss::Loss < eT > >	loss

Protected Attributes inherited from mic::mlnn::MultiLayerNeuralNetwork< eT >
std::vector< std::shared_ptr < mic::mlnn::Layer< eT > > >	layers

std::string	name

bool	connected
	Flag denoting whether the layers are interconnected, thus no copying between inputs and outputs of the neighboring layers will be required. More...

Detailed Description

template<typename eT>
class mic::mlnn::BackpropagationNeuralNetwork< eT >

Class representing a multi-layer neural network based on backpropagation/gradient descent.

Author: tkornuta

Template Parameters

eT	Template parameter denoting precision of variables (float for calculations/double for testing).

Definition at line 40 of file BackpropagationNeuralNetwork.hpp.

Constructor & Destructor Documentation

template<typename eT>

mic::mlnn::BackpropagationNeuralNetwork< eT >::BackpropagationNeuralNetwork ( std::string name_ = "bp_net" )

inline

Constructor. Sets the neural network name.

Parameters

name_ Name of the network.

Definition at line 47 of file BackpropagationNeuralNetwork.hpp.

template<typename eT>

virtual mic::mlnn::BackpropagationNeuralNetwork< eT >::~BackpropagationNeuralNetwork ( )

inlinevirtual

Virtual descriptor - empty.

Definition at line 58 of file BackpropagationNeuralNetwork.hpp.

Member Function Documentation

template<typename eT>

void mic::mlnn::BackpropagationNeuralNetwork< eT >::backward ( mic::types::MatrixPtr< eT > gradients_ )

inline

Performs the back propagation algorithm i.e. propagates the gradients from the last layer to the first.

Parameters

gradients_ The input gradient (i.e. result of the derivative of the loss function).

Definition at line 161 of file BackpropagationNeuralNetwork.hpp.

Referenced by mic::mlnn::BackpropagationNeuralNetwork< float >::train().

template<typename eT>

eT mic::mlnn::BackpropagationNeuralNetwork< eT >::calculateMeanLoss	(	mic::types::MatrixPtr< eT >	encoded_targets_,
		mic::types::MatrixPtr< eT >	encoded_predictions_
	)

inline

Calculates the loss function according to the selected loss function.

Parameters

encoded_targets_	Targets (labels) encoded in the form of pointer to matrix of size [label_size x batch_size].
encoded_predictions_	Predicted outputs of the network encoded in the form of pointer to matrix of size [label_size x batch_size].

Returns: Loss computed according to the selected loss function.

Definition at line 242 of file BackpropagationNeuralNetwork.hpp.

Referenced by main().

template<typename eT>

void mic::mlnn::BackpropagationNeuralNetwork< eT >::forward	(	mic::types::MatrixPtr< eT >	input_data,
		bool	skip_dropout = `false`
	)

inline

Passes the data in a feed-forward manner through all consecutive layers, from the input to the output layer.

Parameters

input_data	Input data - a matrix containing [sample_size x batch_size].
skip_dropout	Flag for skipping dropouts - which should be set to true during testing.

Definition at line 76 of file BackpropagationNeuralNetwork.hpp.

Referenced by main(), mic::mlnn::BackpropagationNeuralNetwork< float >::test(), and mic::mlnn::BackpropagationNeuralNetwork< float >::train().

template<typename eT>

template<typename LossFunction >

void mic::mlnn::BackpropagationNeuralNetwork< eT >::setLoss ( )

inline

Sets the loss function.

Template Parameters

LossFunction Template parameter denoting the loss function type (e.g. mic::neural_nets::loss::CrossEntropyLoss<eT>).

Definition at line 66 of file BackpropagationNeuralNetwork.hpp.

Referenced by batch_function(), mic::applications::MNISTPatchReconstructionApplication::initializePropertyDependentVariables(), mic::applications::MNISTPatchSoftmaxApplication::initializePropertyDependentVariables(), mic::neural_nets::unit_tests::Simple2LayerRegressionNN::Simple2LayerRegressionNN(), and mic::neural_nets::unit_tests::Tutorial2LayerNN::Tutorial2LayerNN().

template<typename eT>

eT mic::mlnn::BackpropagationNeuralNetwork< eT >::test	(	mic::types::MatrixPtr< eT >	encoded_batch_,
		mic::types::MatrixPtr< eT >	encoded_targets_
	)

inline

Tests the neural network with a given batch.

Parameters

encoded_batch_	Batch encoded in the form of matrix of size [sample_size x batch_size].
encoded_targets_	Targets (labels) encoded in the form of matrix of size [label_size x batch_size].

Returns: Loss computed according to the selected loss function. If function not set - returns INF.

Definition at line 222 of file BackpropagationNeuralNetwork.hpp.

Referenced by mic::applications::MNISTPatchReconstructionApplication::collectTestStatistics(), mic::applications::MNISTPatchSoftmaxApplication::collectTestStatistics(), and main().

template<typename eT>

eT mic::mlnn::BackpropagationNeuralNetwork< eT >::train	(	mic::types::MatrixPtr< eT >	encoded_batch_,
		mic::types::MatrixPtr< eT >	encoded_targets_,
		eT	learning_rate_,
		eT	decay_ = `0.0f`
	)

inline

Trains the neural network with a given batch.

Parameters

encoded_batch_	Batch encoded in the form of matrix of size [sample_size x batch_size].
encoded_targets_	Targets (labels) encoded in the form of matrix of size [label_size x batch_size].
learning_rate_	The learning rate.
decay_	Weight decay rate (determining that the "unused/unupdated" weights will decay to 0) (DEFAULT=0.0 - no decay).