meka.classifiers.multilabel

Class BPNN

java.lang.Object

weka.classifiers.AbstractClassifier

weka.classifiers.SingleClassifierEnhancer

meka.classifiers.multilabel.MultilabelClassifier

meka.classifiers.multilabel.NN.AbstractNeuralNet

meka.classifiers.multilabel.BPNN

All Implemented Interfaces:

java.io.Serializable, java.lang.Cloneable, weka.classifiers.Classifier, weka.core.CapabilitiesHandler, weka.core.OptionHandler, weka.core.RevisionHandler

public class BPNN
extends AbstractNeuralNet

BPNN.java - Back Propagation Neural Network. This is a standard back-propagated Neural Network with multiple outputs that correspond to multiple labels.
If trained 'from scratch' only 1 layer is possible, but if you initialise it (from another method) with pre-trained weight matrices, the number of layers is inferred from that.

Version:

March 2013

Author:

Jesse Read

See Also:

Serialized Form

Field Summary

Fields

Modifier and Type	Field and Description
Jama.Matrix[]	W Weight Matrix

Constructor Summary

Constructors

Constructor and Description
BPNN()

Method Summary

Methods

Modifier and Type	Method and Description
double	backPropagate(double[][] X_, double[][] Y_) Back Propagate - Do one round of Back Propagation on batch X_,Y_.
void	buildClassifier(weka.core.Instances D)
double[]	distributionForInstance(weka.core.Instance xy)
Jama.Matrix[]	forwardPass(double[][] X_) Forward Pass - Given input X_, get output of all layers Z[0]...
void	initWeights(int d, int L, int[] H) InitWeights - Initialize a BPNN of H.length hidden layers with H[0], H[1], etc hidden units in each layer (W will be random, and of the corresponding dimensions).
static void	main(java.lang.String[] args)
double[]	popy(double[] x_) Forward Pass - Given input x_, get output y_.
double[][]	popY(double[][] X_) Forward Pass - Given input X_, get output Y_.
void	setWeights(Jama.Matrix[] W, int L) SetWeights - Initialize a BPNN with (pre-trained) weight matrices W (which also determines X dimensions).
double	train(double[][] X_, double[][] Y_)
double	train(double[][] X_, double[][] Y_, int I) Train - Train for I iterations.
double	update(double[][] X_, double[][] Y_) Update - A single training epoch.

Methods inherited from class meka.classifiers.multilabel.NN.AbstractNeuralNet

getE, getH, getOptions, listOptions, setE, setH, setOptions, toString

Methods inherited from class meka.classifiers.multilabel.MultilabelClassifier

evaluation, getCapabilities, getRevision, getTemplate, globalInfo, makeCopies, runClassifier, testCapabilities

Methods inherited from class weka.classifiers.SingleClassifierEnhancer

classifierTipText, getClassifier, setClassifier

Methods inherited from class weka.classifiers.AbstractClassifier

classifyInstance, debugTipText, forName, getDebug, makeCopies, makeCopy, runClassifier, setDebug

Methods inherited from class java.lang.Object

equals, getClass, hashCode, notify, notifyAll, wait, wait, wait

Field Detail

W

public Jama.Matrix[] W

Weight Matrix

Constructor Detail

BPNN

public BPNN()

Method Detail

buildClassifier

public void buildClassifier(weka.core.Instances D)
                     throws java.lang.Exception

Specified by:

buildClassifier in interface weka.classifiers.Classifier

Specified by:

buildClassifier in class MultilabelClassifier

Throws:

java.lang.Exception

distributionForInstance

public double[] distributionForInstance(weka.core.Instance xy)
                                 throws java.lang.Exception

Specified by:

distributionForInstance in interface weka.classifiers.Classifier

Specified by:

distributionForInstance in class MultilabelClassifier

Throws:

java.lang.Exception

setWeights

public void setWeights(Jama.Matrix[] W,
              int L)
                throws java.lang.Exception

SetWeights - Initialize a BPNN with (pre-trained) weight matrices W (which also determines X dimensions).

Parameters:

W - pre-trained weight matrix (should include bias weights, assume W[-1]-1 hidden units in penultimate layer not including bias])

L - the number of labels (for making the final matrix)

Throws:

java.lang.Exception

initWeights

public void initWeights(int d,
               int L,
               int[] H)
                 throws java.lang.Exception

InitWeights - Initialize a BPNN of H.length hidden layers with H[0], H[1], etc hidden units in each layer (W will be random, and of the corresponding dimensions).

Parameters:

d - number of visible units

L - number of labels (output units)

H - number of units in hidden layers, H.length = number of hidden layers. CURRENTLY LIMITED TO 1.

Throws:

java.lang.Exception

train

public double train(double[][] X_,
           double[][] Y_)
             throws java.lang.Exception

Throws:

java.lang.Exception

train

public double train(double[][] X_,
           double[][] Y_,
           int I)
             throws java.lang.Exception

Train - Train for I iterations. I is not necessarily m_E (yet)!

Throws:

java.lang.Exception

update

public double update(double[][] X_,
            double[][] Y_)
              throws java.lang.Exception

Update - A single training epoch.

Throws:

java.lang.Exception

popy

public double[] popy(double[] x_)

Forward Pass - Given input x_, get output y_.

Parameters:

x_ - input

Returns:

y_ output

popY

public double[][] popY(double[][] X_)

Forward Pass - Given input X_, get output Y_.

Parameters:

X_ - input

Returns:

Y_ output

forwardPass

public Jama.Matrix[] forwardPass(double[][] X_)

Forward Pass - Given input X_, get output of all layers Z[0]...

Parameters:

X_ - input (no bias included)

Returns:

output Z[] = {X,Z1,Z2,...,Y}

backPropagate

public double backPropagate(double[][] X_,
                   double[][] Y_)
                     throws java.lang.Exception

Back Propagate - Do one round of Back Propagation on batch X_,Y_.

Parameters:

X_ - input

Y_ - teacher values

Throws:

java.lang.Exception

main

public static void main(java.lang.String[] args)
                 throws java.lang.Exception

Throws:

java.lang.Exception