de.jungblut.classification.nn

Class MultilayerPerceptron

java.lang.Object

de.jungblut.classification.AbstractPredictor

de.jungblut.classification.AbstractClassifier

de.jungblut.classification.nn.MultilayerPerceptron

All Implemented Interfaces:

Classifier, Predictor

public final class MultilayerPerceptron
extends AbstractClassifier

Multilayer perceptron implementation that works on GPU via JCuda and CPU. It features l0/1 regularization (lambda) and dropout. A wide variety of activation functions and error functions can be configured. You can set this network up by using the builder given by MultilayerPerceptron. MultilayerPerceptron.MultilayerPerceptronBuilder.

Author:

thomas.jungblut

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	MultilayerPerceptron.MultilayerPerceptronBuilder Configuration for training a neural net through the Classifier

Field Summary

Fields

Modifier and Type	Field and Description
static long	SEED

Method Summary

Modifier and Type	Method and Description
static MultilayerPerceptron	deserialize(DataInput in) Deserializes a new neural network from the given input stream.
ActivationFunction[]	getActivations()
de.jungblut.math.DoubleVector	getFoldedThetaVector()
int[]	getLayers()
WeightMatrix[]	getWeights()
de.jungblut.math.DoubleVector	predict(de.jungblut.math.DoubleVector xi) Predicts the outcome of the given input by doing a forward pass.
de.jungblut.math.DoubleVector	predict(de.jungblut.math.DoubleVector xi, double threshold) Predicts the outcome of the given input by doing a forward pass.
static void	serialize(MultilayerPerceptron model, DataOutput out) Serializes this network at its current state to a binary file.
void	train(de.jungblut.math.DoubleVector[] features, de.jungblut.math.DoubleVector[] outcome) Trains this classifier with the given features and the outcome.
double	train(de.jungblut.math.DoubleVector[] features, de.jungblut.math.DoubleVector[] outcome, Minimizer minimizer, int maxIterations, double lambda, boolean verbose) Full backpropagation training method.

Methods inherited from class de.jungblut.classification.AbstractClassifier

train

Methods inherited from class de.jungblut.classification.AbstractPredictor

extractPredictedClass, extractPredictedClass, predictedClass, predictedClass, predictProbability

Methods inherited from class java.lang.Object

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

Methods inherited from interface de.jungblut.classification.Predictor

asClassifier, extractPredictedClass, extractPredictedClass, predictedClass, predictedClass, predictProbability

Field Detail

SEED

public static long SEED

Method Detail

predict

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

Predicts the outcome of the given input by doing a forward pass.

Returns:

the vector that contains an indicator at the index of the class. Usually zero or 1, in some cases it is a probability or activation value.

predict

public de.jungblut.math.DoubleVector predict(de.jungblut.math.DoubleVector xi,
                                             double threshold)

Predicts the outcome of the given input by doing a forward pass. Used for binary classification by a threshold. Everything above threshold will be considered as 1, the other case as 0.

train

public void train(de.jungblut.math.DoubleVector[] features,
                  de.jungblut.math.DoubleVector[] outcome)

Description copied from interface: Classifier

Trains this classifier with the given features and the outcome.

Specified by:

train in interface Classifier

Overrides:

train in class AbstractClassifier

outcome - the outcome must have classes labeled as doubles. E.G. in the binary case you have a single element and decide between 0d and 1d. In higher dimensional cases you have each of these single elements mapped to a dimension.

train

public final double train(de.jungblut.math.DoubleVector[] features,
                          de.jungblut.math.DoubleVector[] outcome,
                          Minimizer minimizer,
                          int maxIterations,
                          double lambda,
                          boolean verbose)

Full backpropagation training method. It performs weight finding by using a minimizer. Note that it only guarantees to find a global minimum solution in case of linear or convex problems (zero / one hidden layer), of course this is also dependend on the concrete minimizer implementation. If you have more than a single hidden layer, then it will usually trap into a local minimum. It supplies a vector so training can be resumed from a good starting point.

Parameters:

features - the training examples.

outcome - the outcomes for the training examples.

minimizer - the minimizer to use to train the neural network.

maxIterations - the number of maximum iterations to train.

lambda - the given regularization parameter.

verbose - output to console with the last given errors.

theta - initial spot to start the minimizations.

Returns:

the cost of the training.

getFoldedThetaVector

public de.jungblut.math.DoubleVector getFoldedThetaVector()

Returns:

the folded theta vector, seeded by the current weight matrices.

getWeights

public WeightMatrix[] getWeights()

getLayers

public int[] getLayers()

getActivations

public ActivationFunction[] getActivations()

deserialize

public static MultilayerPerceptron deserialize(DataInput in)
                                        throws IOException

Deserializes a new neural network from the given input stream. Note that "in" will not be closed by this method.

Throws:

IOException

serialize

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

Serializes this network at its current state to a binary file. Note that "out" will not be closed in this method.

Throws:

IOException