MultiLayerNetwork Class Reference

Abstraction of a multi-layer perceptron network. More...

#include <MultiLayerNetwork.h>

Inheritance diagram for MultiLayerNetwork:

[legend]Collaboration diagram for MultiLayerNetwork:

[legend]List of all members.

Public Member Functions
	MultiLayerNetwork (uint inputs, uint neuronLabelOffset=0, PublicValues &pv=defaultControl)
	Create a multi-layer network.
	MultiLayerNetwork (uint inputs, PublicValues &pv)
	MultiLayerNetwork (MultiLayerNetwork &srcNet)
	Copy constructor, NOT YET IMPLEMENTED.
	MultiLayerNetwork (const std::string &filename)
	Loads a multi-layer network from the given filename.
virtual	~MultiLayerNetwork ()
virtual void	addLayer (int size)
	Adds a layer of the given size to the network. This new layer becomes the output layer.
virtual void	connectLayer (uint srcLlayer)
	Completely connects the given layer with the layer below it, i.e., all neurons in the given layer will give their output to all the neurons in the layer below.
virtual void	connect (uint srcLlayer, int srcNrn, int destNrn)
	Connects two neurons, with random weight.
virtual void	connect (uint srcLlayer, int srcNrn, int destNrn, real weight)
	Connects two neurons, with the weight provided.
virtual void	setBias (uint layer, int nrn, real bias)
	Sets the bias of the given neuron.
virtual uint	getLayerCount () const
	The number of layers in the network (does not count the input layer as a layer).
virtual Vector	getOutput (const Vector &input)
	Returns the output of the network for the given input.
virtual Vector	getOutput (real *input)
	Wrapper function to allow getOutput() to work for an array of real as input as well.
virtual void	train (TrainingSet &T, uint epochs, real learningRate=DEFAULT_LEARNINGRATE, real momentum=DEFAULT_MOMENTUM)
	Trains the network with data from the given TrainingSet using the backpropagation algorithm.
virtual void	train (TrainingSet &T, Creal epochs, Creal learningRate=CDEFAULT_LEARNINGRATE, Creal momentum=CDEFAULT_MOMENTUM)
virtual void	train (TrainingSet &T, PublicValues &parameters)
	get epochs, learningRate and momentum from the supplied PublicValues
void	getError (TrainingSet &ts)
	compute error for the given traning set. The "epoch error" and "normalized epoch error" are then stored in the _control
void	getErrorGREN (TrainingSet &ts)
	compute error for the given traning set. The "epoch error" and "normalized epoch error" are then stored in the _control
Error	getErrorGREN (const Vector &input, const Vector &desired)
	real error for one example
virtual void	trainExample (const Vector &input, const Vector &desiredOutput, real learningRate=DEFAULT_LEARNINGRATE, real momentum=DEFAULT_MOMENTUM)
	Trains one example only.
virtual void	save (const std::string &filename)
	Saves the network to the given filename.
virtual void	setActivationFunction (uint layer, ActivationFunction f, ActivationFunction df)
	Sets the activation function used by the neurons in the provided layer.
virtual const char *	getClassName () const
	Returns "MultiLayerNetwork".
void	resetWeights ()
	initialize weights to small values
const Layer &	getLayer (uint layer) const
	Get a layer of the network (0=input) Exceptions: Exception  if an invalid layer is given.
uint	getNeuronsCount () const
	Get the total count of neurons in all layers (excluding input "layer").
uint	getLinksCount () const
	Get count of all weights of all neurons.
Layer &	getLayer (uint layer)
	Warning: using this non-const version, you can change the network's behaviour.
	operator std::string () const
	get brief info about the topology, etc.
PublicValues &	getControl ()
void	setControl (PublicValues &ctrl)
void	setLabelOffset (uint firstLabel)
virtual void	trainGREN (MultiLayerNetwork &errorNetwork, TrainingSet &ts, uint epochs, real learningRate=DEFAULT_LEARNINGRATE, real momentum=DEFAULT_MOMENTUM)
	Train the network using the error network.
Static Public Attributes
const real	DEFAULT_MOMENTUM
const real	DEFAULT_LEARNINGRATE
const Creal	CDEFAULT_MOMENTUM
const Creal	CDEFAULT_LEARNINGRATE
Protected Member Functions
InputLayer *	_inputLayer ()
	The input layer.
Layer *	_outputLayer ()
	The output layer.
void	_layerValid (uint layer) const
	impl detail - throws then layer is not valid
void	_connectLayer (Layer &srcLayer, Layer &destLayer)
void	_attachGREN (MultiLayerNetwork &errorNetwork, bool detach)
	attach or detach the gren. TODO: detach doesn't yet restore the input links of GREN, so it's unusable after training standalone
void	_trainExampleGREN (MultiLayerNetwork &errorNetwork, const Vector &input, real learningRate, real momentum)
	unlike _trainExample, _trainExampleGREN assumes that input was already presented
Protected Attributes
uint	_nLayers
	Number of layers in the network.
std::vector< Layer * >	_layers
	The layers.
PublicValues *	_control
Error	_exampleError
uint	_neuronLabelOffset

---

Detailed Description

Abstraction of a multi-layer perceptron network.

Basically, layers of SimpleNeurons constitute this network. Training is done using the backpropagation technique which uses the gradient descent method.

The labels of the layers start from 0 (for the input layer) and then keep moving on. The labels of neurons in the layers is = *Layer::MAX_LAYER_SIZE +

All neurons in the layers are allowed to have a bias.

Todo:

The copy constructor

---

Constructor & Destructor Documentation

MultiLayerNetwork (  uint  inputs, uint  neuronLabelOffset = 0, PublicValues &  pv = defaultControl )

Create a multi-layer network. Parameters: inputs  The number of inputs taken in by the network

MultiLayerNetwork (  uint  inputs, PublicValues &  pv )

MultiLayerNetwork (  MultiLayerNetwork &  srcNet  )

Copy constructor, NOT YET IMPLEMENTED.

MultiLayerNetwork (  const std::string &  filename  )

Loads a multi-layer network from the given filename. Parameters: filename  The name of the file from which to load the network Exceptions: Exception  On any error

virtual ~MultiLayerNetwork (   )  [virtual]

---

Member Function Documentation

void _attachGREN (  MultiLayerNetwork &  errorNetwork, bool  detach )  [protected]

attach or detach the gren. TODO: detach doesn't yet restore the input links of GREN, so it's unusable after training standalone

void _connectLayer (  Layer &  srcLayer, Layer &  destLayer )  [protected]

InputLayer* _inputLayer (   )  [protected]

The input layer.

void _layerValid (  uint  layer  )  const [protected]

impl detail - throws then layer is not valid

Layer* _outputLayer (   )  [protected]

The output layer.

void _trainExampleGREN (  MultiLayerNetwork &  errorNetwork, const Vector &  input, real  learningRate, real  momentum )  [protected]

unlike _trainExample, _trainExampleGREN assumes that input was already presented

virtual void addLayer (  int  size  )  [virtual]

Adds a layer of the given size to the network. This new layer becomes the output layer. Reimplemented in TwoLayerNetwork.

virtual void connect (  uint  srcLlayer, int  srcNrn, int  destNrn, real  weight )  [virtual]

Connects two neurons, with the weight provided. Parameters: srcLayer  The index of the layer in which the source neuron lies srcNrn  The index of the source neuron in the source layer destNrn  The index of the destination neuron (will be in the layer with index srcLayer+1) weight  The weight of the link Exceptions: Exception  On any invalid argument values

virtual void connect (  uint  srcLlayer, int  srcNrn, int  destNrn )  [virtual]

Connects two neurons, with random weight. Parameters: srcLayer  The index of the layer in which the source neuron lies srcNrn  The index of the source neuron in the source layer destNrn  The index of the destination neuron (will be in the layer with index srcLayer+1) Exceptions: Exception  On any invalid argument values

virtual void connectLayer (  uint  srcLlayer  )  [virtual]

Completely connects the given layer with the layer below it, i.e., all neurons in the given layer will give their output to all the neurons in the layer below. Weights are random. Parameters: srcLayer  The index of the source layer Exceptions: Exception  If the layer index provided is invalid

virtual const char* getClassName (   )  const [virtual]

Returns "MultiLayerNetwork". Implements Network.

PublicValues& getControl (   )  [inline]

void getError (  TrainingSet &  ts  )

compute error for the given traning set. The "epoch error" and "normalized epoch error" are then stored in the _control

Error getErrorGREN (  const Vector &  input, const Vector &  desired )

real error for one example

void getErrorGREN (  TrainingSet &  ts  )

compute error for the given traning set. The "epoch error" and "normalized epoch error" are then stored in the _control

Layer& getLayer (  uint  layer  )

Warning: using this non-const version, you can change the network's behaviour. It's here because Neuron::getOutput() is not (yet) const

const Layer& getLayer (  uint  layer  )  const

Get a layer of the network (0=input) Exceptions: Exception  if an invalid layer is given.

virtual uint getLayerCount (   )  const [virtual]

The number of layers in the network (does not count the input layer as a layer).

uint getLinksCount (   )  const

Get count of all weights of all neurons.

uint getNeuronsCount (   )  const

Get the total count of neurons in all layers (excluding input "layer").

virtual Vector getOutput (  real *  input  )  [virtual]

Wrapper function to allow getOutput() to work for an array of real as input as well. Does exactly the same thing as Network::getOutput(real*). Reimplemented from Network.

virtual Vector getOutput (  const Vector &  input  )  [virtual]

Returns the output of the network for the given input. Parameters: input  A vector of getDimension() reals Returns: The corresponding output of the network Implements Network.

operator std::string (   )  const

get brief info about the topology, etc.

void resetWeights (   )

initialize weights to small values

virtual void save (  const std::string &  filename  )  [virtual]

Saves the network to the given filename. The file is a simple text file, open it up in a text editor to see the format, quite simple Parameters: filename  The name of the file to save the network in. Implements Network.

virtual void setActivationFunction (  uint  layer, ActivationFunction  f, ActivationFunction  df )  [virtual]

Sets the activation function used by the neurons in the provided layer. Parameters: layer  The layer whose activation function is to be changed. layer>0 (as input neurons don't have any activation function) and layer<getLayerCount() f  The activation function to be used df  The derivative of the activation function to be used. Required for training. Exceptions: Exception  if an invalid layer is given

virtual void setBias (  uint  layer, int  nrn, real  bias )  [virtual]

Sets the bias of the given neuron. Parameters: layer  The index of the layer in which the neuron lies nrn  The index of the neuron in the layer bias  The bias to be given to the neuron

void setControl (  PublicValues &  ctrl  )  [inline]

void setLabelOffset (  uint  firstLabel  )  [inline]

virtual void train (  TrainingSet &  T, PublicValues &  parameters )  [virtual]

get epochs, learningRate and momentum from the supplied PublicValues

virtual void train (  TrainingSet &  T, Creal  epochs, Creal  learningRate = CDEFAULT_LEARNINGRATE, Creal  momentum = CDEFAULT_MOMENTUM )  [virtual]

virtual void train (  TrainingSet &  T, uint  epochs, real  learningRate = DEFAULT_LEARNINGRATE, real  momentum = DEFAULT_MOMENTUM )  [virtual]

Trains the network with data from the given TrainingSet using the backpropagation algorithm. Parameters: T  The TrainingSet containing input/desired-output vector pairs epochs  The number of epochs to train the network. An epoch is a single iteration through all input/desired-output vector pairs in T. learningRate  The learning rate to be used for weight updation momentum  The momentum factor to be used during weight updation. 0 by default.

virtual void trainExample (  const Vector &  input, const Vector &  desiredOutput, real  learningRate = DEFAULT_LEARNINGRATE, real  momentum = DEFAULT_MOMENTUM )  [virtual]

Trains one example only. Returns: error for this example

virtual void trainGREN (  MultiLayerNetwork &  errorNetwork, TrainingSet &  ts, uint  epochs, real  learningRate = DEFAULT_LEARNINGRATE, real  momentum = DEFAULT_MOMENTUM )  [virtual]

Train the network using the error network. Generalized Relief Error Network - included in the MultiLayerNetwork, because it only provides new functions and can (better: must) be used together with MultiLayerNetwork The error network has getInputCount() + getOutputCount() inputs (in that order) Parameters: ts  contains input patterns to be trained. May also contain outputs - in this case, "real output error" will be computed

---

Member Data Documentation

PublicValues* _control [protected]

Error _exampleError [protected]

std::vector<Layer *> _layers [protected]

The layers.

uint _neuronLabelOffset [protected]

uint _nLayers [protected]

Number of layers in the network. If you derive your own network from this class, the onus of keeping _nLayers consistent lies on you! Doesn't count the input "layer"

const Creal CDEFAULT_LEARNINGRATE [static]

const Creal CDEFAULT_MOMENTUM [static]

const real DEFAULT_LEARNINGRATE [static]

const real DEFAULT_MOMENTUM [static]

---

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

• MultiLayerNetwork.h

---