de.jungblut.classification.nn

Class RBM

java.lang.Object

de.jungblut.classification.nn.RBM

public final class RBM
extends Object

Class for training and stacking Restricted Boltzmann Machines (RBMs). Stacked RBMs are called DBN (deep belief net). Usually every layer of a deep belief net is training greedily with the contrastive divergence algorithm implemented in RBMCostFunction. Create new instances with the RBM.RBMBuilder or with the static factory methods.

Author:

thomas.jungblut

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	RBM.RBMBuilder

Method Summary

Modifier and Type	Method and Description
static RBM	deserialize(DataInputStream in) Deserializes the RBM back from the binary stream input.
WeightMatrix[]	getNeuralNetworkWeights(int outputLayerSize) Creates a weight matrix that can be used for unsupervised weight initialization in the MultilayerPerceptron.
de.jungblut.math.DoubleMatrix[]	getWeights()
de.jungblut.math.DoubleVector	predict(de.jungblut.math.DoubleVector input) Returns the hidden activations of the last RBM.
de.jungblut.math.DoubleVector	reconstructInput(de.jungblut.math.DoubleVector hiddenActivations) Creates a reconstruction of the input using the given hidden activations.
static void	serialize(RBM model, DataOutput out) Serializes this RBM model into the given output stream.
void	setSeed(long seed) Sets the internally used rng seed.
static RBM	single(int numHiddenNodes)
static RBM	single(int numHiddenNodes, ActivationFunction func)
static RBM	singleGPU(int numHiddenNodes, ActivationFunction func)
static RBM	stacked(ActivationFunction func, int... numHiddenNodes) Creates a new stacked RBM with sigmoid activation and with the given number of hidden nodes in each stacked layer.
static RBM	stacked(int... numHiddenNodes) Creates a new stacked RBM with sigmoid activation and with the given number of hidden nodes in each stacked layer.
static RBM	stackedGPU(ActivationFunction func, int... numHiddenNodes) Creates a new stacked RBM with sigmoid activation and with the given number of hidden nodes in each stacked layer.
void	train(de.jungblut.math.DoubleVector[] trainingSet, double alpha, int numIterations) Trains the RBM on the given training set.
void	train(de.jungblut.math.DoubleVector[] trainingSet, Minimizer minimizer, int numIterations) Trains the RBM on the given training set.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Detail

train

public void train(de.jungblut.math.DoubleVector[] trainingSet,
                  double alpha,
                  int numIterations)

Trains the RBM on the given training set.

Parameters:

trainingSet - the training set to train on.

alpha - the learning rate for gradient descent.

numIterations - how many iterations of training have to be done. (if converged before, it will stop training)

train

public void train(de.jungblut.math.DoubleVector[] trainingSet,
                  Minimizer minimizer,
                  int numIterations)

Trains the RBM on the given training set.

Parameters:

trainingSet - the training set to train on.

minimizer - the minimizer to use. Note that the costfunction's gradient isn't the real gradient and thus can't be optimized by line searching minimizers like Fmincg.

numIterations - how many iterations of training have to be done. (if converged before, it will stop training)

predict

public de.jungblut.math.DoubleVector predict(de.jungblut.math.DoubleVector input)

Returns the hidden activations of the last RBM.

Parameters:

input - the input of the first RBM.

Returns:

a vector that contains the values of the hidden activations on the last layer.

reconstructInput

public de.jungblut.math.DoubleVector reconstructInput(de.jungblut.math.DoubleVector hiddenActivations)

Creates a reconstruction of the input using the given hidden activations. (That, what is returned by predict(DoubleVector)).

Parameters:

hiddenActivations - the activations of the predict method.

Returns:

the reconstructed input vector.

getWeights

public de.jungblut.math.DoubleMatrix[] getWeights()

Returns:

the weight matrices.

getNeuralNetworkWeights

public WeightMatrix[] getNeuralNetworkWeights(int outputLayerSize)

Creates a weight matrix that can be used for unsupervised weight initialization in the MultilayerPerceptron.

Parameters:

outputLayerSize - the size of the classification layer on top of this RBM.

Returns:

the WeightMatrix that maps layers to the weights.

setSeed

public void setSeed(long seed)

Sets the internally used rng seed.

serialize

public static void serialize(RBM model,
                             DataOutput out)
                      throws IOException

Serializes this RBM model into the given output stream.

Throws:

IOException

deserialize

public static RBM deserialize(DataInputStream in)
                       throws IOException

Deserializes the RBM back from the binary stream input.

Throws:

IOException

single

public static RBM single(int numHiddenNodes,
                         ActivationFunction func)

Returns:

a single RBM which isn't stacked and emits to the given number of hidden nodes.

stacked

public static RBM stacked(ActivationFunction func,
                          int... numHiddenNodes)

Creates a new stacked RBM with sigmoid activation and with the given number of hidden nodes in each stacked layer. For example: 4,3,2 will create the first RBM with 4 hidden nodes, the second layer will operate on the 4 hidden node outputs of the RBM before and emit to 3 hidden nodes. Similarly the last layer will receive three inputs and emit 2 output's, which state you receive in the predict method.

single

public static RBM single(int numHiddenNodes)

Returns:

a single RBM with sigmoid activation which isn't stacked and emits to the given number of hidden nodes.

stacked

public static RBM stacked(int... numHiddenNodes)

Creates a new stacked RBM with sigmoid activation and with the given number of hidden nodes in each stacked layer. For example: 4,3,2 will create the first RBM with 4 hidden nodes, the second layer will operate on the 4 hidden node outputs of the RBM before and emit to 3 hidden nodes. Similarly the last layer will receive three inputs and emit 2 output's, which state you receive in the predict method.

singleGPU

public static RBM singleGPU(int numHiddenNodes,
                            ActivationFunction func)

Returns:

a single RBM which isn't stacked and emits to the given number of hidden nodes.

stackedGPU

public static RBM stackedGPU(ActivationFunction func,
                             int... numHiddenNodes)

Creates a new stacked RBM with sigmoid activation and with the given number of hidden nodes in each stacked layer. For example: 4,3,2 will create the first RBM with 4 hidden nodes, the second layer will operate on the 4 hidden node outputs of the RBM before and emit to 3 hidden nodes. Similarly the last layer will receive three inputs and emit 2 output's, which state you receive in the predict method.