Offers a basic structure for recurrent networks. More...

#include <shark/Models/RecurrentStructure.h>

Inheritance diagram for shark::RecurrentStructure:

Public Types
enum	SigmoidType { Linear, Logistic, Tanh, FastSigmoid }
	type enum for the different variants of sigmoids More...

Public Member Functions
SHARK_EXPORT_SYMBOL	RecurrentStructure ()

const IntMatrix &	connections () const

bool	connection (std::size_t i, std::size_t j) const
	returns whether the connection between neuron i and j exists More...

const RealMatrix &	weights () const
	returns the current weight matrix More...

double	weight (std::size_t i, std::size_t j) const
	returns the weight of the connection between neuron i and j More...

SigmoidType	sigmoidType () const
	returns the type of sigmoid used in this network More...

void	setSigmoidType (SigmoidType sigmoidType)

SHARK_EXPORT_SYMBOL void	setWeights (const RealMatrix &weights)

SHARK_EXPORT_SYMBOL void	setStructure (std::size_t inputs, std::size_t outputs, const IntMatrix &connections, SigmoidType sigmoidType=Logistic)
	Based on a given connection matrix a network is created. More...

SHARK_EXPORT_SYMBOL void	setStructure (std::size_t in, std::size_t hidden, std::size_t out, bool bias=true, SigmoidType sigmoidType=Logistic)
	Creates a fully connected topology for the network with optional bias. More...

SHARK_EXPORT_SYMBOL RealVector	parameterVector () const
	get internal parameters of the model More...

SHARK_EXPORT_SYMBOL void	setParameterVector (RealVector const &newParameters)
	set internal parameters of the model More...

SHARK_EXPORT_SYMBOL void	read (InArchive &archive)
	From ISerializable, reads the Network from an archive. More...

SHARK_EXPORT_SYMBOL void	write (OutArchive &archive) const
	From ISerializable, writes the Network to an archive. More...

std::size_t	inputs () const
	The number of input neurons of the network. More...

std::size_t	outputs () const
	The number of output neurons of the network. More...

std::size_t	numberOfNeurons () const

std::size_t	numberOfUnits () const

std::size_t	bias () const
	The index of the bias unit. More...

std::size_t	parameters () const
	number of parameters of the network More...

SHARK_EXPORT_SYMBOL double	neuron (double activation)
	Activation function for a neuron. More...

SHARK_EXPORT_SYMBOL double	neuronDerivative (double activation)
	Computes the derivative of the neuron. More...

Public Member Functions inherited from shark::ISerializable
virtual	~ISerializable ()
	Virtual d'tor. More...

void	load (InArchive &archive, unsigned int version)
	Versioned loading of components, calls read(...). More...

void	save (OutArchive &archive, unsigned int version) const
	Versioned storing of components, calls write(...). More...

	BOOST_SERIALIZATION_SPLIT_MEMBER ()

Protected Attributes
std::size_t	m_numberOfNeurons
	The total number of neurons of the network (input, output and hidden). More...

std::size_t	m_numberOfUnits
	total number units of the network (input, output, hidden and bias) More...

std::size_t	m_inputNeurons
	The number of input neurons of the network. More...

std::size_t	m_outputNeurons
	The number of output neurons of the network. More...

std::size_t	m_hidden
	The number of hidden neurons of the network. More...

std::size_t	m_bias
	index of the bias unit More...

SigmoidType	m_sigmoidType
	type of Sigmoid used by the network More...

std::size_t	m_numberOfParameters
	The absolute number of parameters of the network. More...

IntMatrix	m_connectionMatrix

RealMatrix	m_weights

Detailed Description

Offers a basic structure for recurrent networks.

it is possible to define the tzpe of sigmoids and the form of the connection matrix. this structure can be shared between different types of ents like the RNNet and the OnlineRNNet

Definition at line 38 of file RecurrentStructure.h.

Member Enumeration Documentation

◆ SigmoidType

enum shark::RecurrentStructure::SigmoidType

type enum for the different variants of sigmoids

Enumerator
Linear	f(x) = x
Logistic	f(x) = 1/(1+exp(-x))
Tanh	f(x) = tanh(x)
FastSigmoid	f(x) = x/(1+\|x\|)

Definition at line 47 of file RecurrentStructure.h.

Constructor & Destructor Documentation

◆ RecurrentStructure()

SHARK_EXPORT_SYMBOL shark::RecurrentStructure::RecurrentStructure ( )

Creates an empty recurrent neural network. A call to setStructure is needed afterwards to configure the topology of the network

Member Function Documentation

◆ bias()

std::size_t shark::RecurrentStructure::bias ( ) const

inline

The index of the bias unit.

Definition at line 175 of file RecurrentStructure.h.

References m_bias.

Referenced by setSigmoidType().

◆ connection()

bool shark::RecurrentStructure::connection	(	std::size_t	i,
		std::size_t	j
	)		const

inline

returns whether the connection between neuron i and j exists

Definition at line 85 of file RecurrentStructure.h.

References m_connectionMatrix.

◆ connections()

const IntMatrix& shark::RecurrentStructure::connections ( ) const

inline

returns the connection Matrix of the network.

The format is best described with an example: given a Network with 2 inputs, 2 outputs and 2 hidden and no bias unit, where the inputs are only connected to the hidden units. The corresponding matrix looks like this:

1 2 3 4 5 6 7 1 1 0 1 1 1 1 first hidden 1 1 0 1 1 1 1 second hidden 0 0 0 1 1 1 1 first output 0 0 0 1 1 1 1 second output

The ith row stands for the ith neuron of the network and when an element (i,j) is 1, the ith unit will receive input from unit j. the first =0,..,inputs-1 columns are the input neurons followd by the column of the bias, which is completely zero in this example if j is a hidden or output neuron, the activation from the PREVIOUS time step is used. if j is an input neuron, the current input is used. input neurons can't receive activation. This is no limitation, since the hidden layer can be subdivided in arbitrary sublayers when the right topology is used. the last column of the matrix is reserved for the bias neuron. So the matrix has size NxN+1

Definition at line 81 of file RecurrentStructure.h.

References m_connectionMatrix.

Referenced by setSigmoidType().

◆ inputs()

std::size_t shark::RecurrentStructure::inputs ( ) const

inline

The number of input neurons of the network.

Definition at line 160 of file RecurrentStructure.h.

References m_inputNeurons.

Referenced by shark::OnlineRNNet::inputSize(), shark::RNNet::inputSize(), and setSigmoidType().

◆ neuron()

SHARK_EXPORT_SYMBOL double shark::RecurrentStructure::neuron ( double activation )

Activation function for a neuron.

Referenced by parameters().

◆ neuronDerivative()

SHARK_EXPORT_SYMBOL double shark::RecurrentStructure::neuronDerivative ( double activation )

Computes the derivative of the neuron.

Referenced by parameters().

◆ numberOfNeurons()

std::size_t shark::RecurrentStructure::numberOfNeurons ( ) const

inline

Definition at line 167 of file RecurrentStructure.h.

References m_numberOfNeurons.

◆ numberOfUnits()

std::size_t shark::RecurrentStructure::numberOfUnits ( ) const

inline

Definition at line 170 of file RecurrentStructure.h.

References m_numberOfUnits.

Referenced by shark::OnlineRNNet::setOutputActivation().

◆ outputs()

std::size_t shark::RecurrentStructure::outputs ( ) const

inline

The number of output neurons of the network.

Definition at line 164 of file RecurrentStructure.h.

References m_outputNeurons.

Referenced by shark::OnlineRNNet::outputSize(), shark::RNNet::outputSize(), and setSigmoidType().

◆ parameters()

std::size_t shark::RecurrentStructure::parameters ( ) const

inline

number of parameters of the network

Definition at line 180 of file RecurrentStructure.h.

References m_numberOfParameters, neuron(), neuronDerivative(), and SHARK_EXPORT_SYMBOL.

Referenced by shark::OnlineRNNet::numberOfParameters(), and shark::RNNet::numberOfParameters().

◆ parameterVector()

SHARK_EXPORT_SYMBOL RealVector shark::RecurrentStructure::parameterVector ( ) const

get internal parameters of the model

Referenced by shark::OnlineRNNet::parameterVector(), shark::RNNet::parameterVector(), and setSigmoidType().

◆ read()

SHARK_EXPORT_SYMBOL void shark::RecurrentStructure::read ( InArchive & archive )

virtual

From ISerializable, reads the Network from an archive.

Reimplemented from shark::ISerializable.

Referenced by setSigmoidType().

◆ setParameterVector()

SHARK_EXPORT_SYMBOL void shark::RecurrentStructure::setParameterVector ( RealVector const & newParameters )

set internal parameters of the model

Referenced by shark::OnlineRNNet::setParameterVector(), shark::RNNet::setParameterVector(), and setSigmoidType().

◆ setSigmoidType()

void shark::RecurrentStructure::setSigmoidType ( SigmoidType sigmoidType )

inline

sets the type of sigmoid used in this network

Parameters

sigmoidType the type of sigmoid

Definition at line 105 of file RecurrentStructure.h.

References bias(), connections(), inputs(), Logistic, m_sigmoidType, outputs(), parameterVector(), read(), setParameterVector(), setStructure(), setWeights(), SHARK_EXPORT_SYMBOL, sigmoidType(), weights(), and write().

◆ setStructure() [1/2]

SHARK_EXPORT_SYMBOL void shark::RecurrentStructure::setStructure	(	std::size_t	inputs,
		std::size_t	outputs,
		const IntMatrix &	connections,
		SigmoidType	sigmoidType = `Logistic`
	)

Based on a given connection matrix a network is created.

This method needs to know how many inputs and outputs the network has and how the units are connected.

If a standard structure is needed, see the other version of this method. Also see connections for a quick explanation of the matrix format

The same mechanic applies alo to recurrentConnections but every element can be set, not only the lower triangular part.

After this operation, all weights are initialized to 0.

Parameters

inputs	number of input neurons of the network
outputs	number of output neurons of the network
connections	feed-forward connections. default is true
sigmoidType	the type of the sigmoid to be used. the default is the Logistic function

Referenced by setSigmoidType().

◆ setStructure() [2/2]

SHARK_EXPORT_SYMBOL void shark::RecurrentStructure::setStructure	(	std::size_t	in,
		std::size_t	hidden,
		std::size_t	out,
		bool	bias = `true`,
		SigmoidType	sigmoidType = `Logistic`
	)

Creates a fully connected topology for the network with optional bias.

After a call, the network will have hidden+out units.

Parameters

in	number of input neurons
hidden	number of output neurons
out	number of input neurons
bias	enables bias neuron, default is true
sigmoidType	the type of the sigmoid to be used. the default is the Logistic function

◆ setWeights()

SHARK_EXPORT_SYMBOL void shark::RecurrentStructure::setWeights ( const RealMatrix & weights )

Sets the weight matrix. It is not allowed that elements are non-zero when the element in the connection matrix is 0!

Parameters

weights the new weight matrix

Referenced by setSigmoidType().

◆ sigmoidType()

SigmoidType shark::RecurrentStructure::sigmoidType ( ) const

inline

returns the type of sigmoid used in this network

Definition at line 99 of file RecurrentStructure.h.

References m_sigmoidType.

Referenced by setSigmoidType().

◆ weight()

double shark::RecurrentStructure::weight	(	std::size_t	i,
		std::size_t	j
	)		const

inline

returns the weight of the connection between neuron i and j

Definition at line 94 of file RecurrentStructure.h.

References m_weights.

◆ weights()

const RealMatrix& shark::RecurrentStructure::weights ( ) const

inline

returns the current weight matrix

Definition at line 90 of file RecurrentStructure.h.

References m_weights.

Referenced by setSigmoidType().

◆ write()

SHARK_EXPORT_SYMBOL void shark::RecurrentStructure::write ( OutArchive & archive ) const

virtual

From ISerializable, writes the Network to an archive.

Reimplemented from shark::ISerializable.

Referenced by setSigmoidType().

Member Data Documentation

◆ m_bias

std::size_t shark::RecurrentStructure::m_bias

protected

index of the bias unit

Definition at line 207 of file RecurrentStructure.h.

Referenced by bias().

◆ m_connectionMatrix

IntMatrix shark::RecurrentStructure::m_connectionMatrix

protected

stores the topology of the network. Element (i,j) is 1 if the ith neuron receives input from neuron j. The data for neuron i is stored in the ith row of the matrix.

Definition at line 219 of file RecurrentStructure.h.

Referenced by connection(), and connections().

◆ m_hidden

std::size_t shark::RecurrentStructure::m_hidden

protected

The number of hidden neurons of the network.

Definition at line 204 of file RecurrentStructure.h.

◆ m_inputNeurons

std::size_t shark::RecurrentStructure::m_inputNeurons

protected

The number of input neurons of the network.

Definition at line 200 of file RecurrentStructure.h.

Referenced by inputs().

◆ m_numberOfNeurons

std::size_t shark::RecurrentStructure::m_numberOfNeurons

protected

The total number of neurons of the network (input, output and hidden).

Definition at line 194 of file RecurrentStructure.h.

Referenced by numberOfNeurons().

◆ m_numberOfParameters

std::size_t shark::RecurrentStructure::m_numberOfParameters

protected

The absolute number of parameters of the network.

Definition at line 214 of file RecurrentStructure.h.

Referenced by parameters().

◆ m_numberOfUnits

std::size_t shark::RecurrentStructure::m_numberOfUnits

protected

total number units of the network (input, output, hidden and bias)

Definition at line 197 of file RecurrentStructure.h.

Referenced by numberOfUnits().

◆ m_outputNeurons

std::size_t shark::RecurrentStructure::m_outputNeurons

protected

The number of output neurons of the network.

Definition at line 202 of file RecurrentStructure.h.

Referenced by outputs().

◆ m_sigmoidType

SigmoidType shark::RecurrentStructure::m_sigmoidType

protected

type of Sigmoid used by the network

Definition at line 210 of file RecurrentStructure.h.

Referenced by setSigmoidType(), and sigmoidType().

◆ m_weights

RealMatrix shark::RecurrentStructure::m_weights

protected

stores the feed-forward part of the weights. the recurrent part is added via m_recurrentWeights. The weights for neuron i are stored in the ith row of the matrix

Definition at line 223 of file RecurrentStructure.h.

Referenced by weight(), and weights().

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

include/shark/Models/RecurrentStructure.h

Public Types

Public Member Functions

Protected Attributes

Detailed Description

Member Enumeration Documentation

◆ SigmoidType

Constructor & Destructor Documentation

◆ RecurrentStructure()

Member Function Documentation

◆ bias()

◆ connection()

◆ connections()

◆ inputs()

◆ neuron()

◆ neuronDerivative()

◆ numberOfNeurons()

◆ numberOfUnits()

◆ outputs()

◆ parameters()

◆ parameterVector()

◆ read()

◆ setParameterVector()

◆ setSigmoidType()

◆ setStructure() [1/2]

◆ setStructure() [2/2]

◆ setWeights()

◆ sigmoidType()

◆ weight()

◆ weights()

◆ write()

Member Data Documentation

◆ m_bias

◆ m_connectionMatrix

◆ m_hidden

◆ m_inputNeurons

◆ m_numberOfNeurons

◆ m_numberOfParameters

◆ m_numberOfUnits

◆ m_outputNeurons

◆ m_sigmoidType

◆ m_weights