org.deeplearning4j.sda

Class StackedDenoisingAutoEncoder

java.lang.Object

org.deeplearning4j.nn.BaseMultiLayerNetwork

org.deeplearning4j.sda.StackedDenoisingAutoEncoder

All Implemented Interfaces:

Serializable, Persistable

public class StackedDenoisingAutoEncoder
extends BaseMultiLayerNetwork

A JBlas implementation of stacked denoising auto encoders.

Author:

Adam Gibson

See Also:

Serialized Form

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	StackedDenoisingAutoEncoder.Builder

Field Summary

Fields inherited from class org.deeplearning4j.nn.BaseMultiLayerNetwork

errorTolerance, layers, learningRateUpdate

Constructor Summary

Constructors

Constructor and Description
StackedDenoisingAutoEncoder()
StackedDenoisingAutoEncoder(int nIns, int[] hiddenLayerSizes, int nOuts, int n_layers, org.apache.commons.math3.random.RandomGenerator rng)
StackedDenoisingAutoEncoder(int n_ins, int[] hiddenLayerSizes, int nOuts, int nLayers, org.apache.commons.math3.random.RandomGenerator rng, org.jblas.DoubleMatrix input, org.jblas.DoubleMatrix labels)

Method Summary

Methods

Modifier and Type	Method and Description
NeuralNetwork	createLayer(org.jblas.DoubleMatrix input, int nVisible, int nHidden, org.jblas.DoubleMatrix W, org.jblas.DoubleMatrix hbias, org.jblas.DoubleMatrix vBias, org.apache.commons.math3.random.RandomGenerator rng, int index) Creates a layer depending on the index.
NeuralNetwork[]	createNetworkLayers(int numLayers)
void	pretrain(double lr, double corruptionLevel, int epochs)
void	pretrain(org.jblas.DoubleMatrix input, double lr, double corruptionLevel, int epochs) Unsupervised pretraining based on reconstructing the input from a corrupted version
void	trainNetwork(org.jblas.DoubleMatrix input, org.jblas.DoubleMatrix labels, Object[] otherParams) Train the network running some unsupervised pretraining followed by SGD/finetune

Methods inherited from class org.deeplearning4j.nn.BaseMultiLayerNetwork

applyTransforms, asDecoder, backProp, backPropStep, clone, encode, fanIn, feedForward, finetune, finetune, getActivation, getColumnMeans, getColumnStds, getColumnSums, getDist, getErrorTolerance, getFanIn, getHiddenLayerSizes, getInput, getL2, getLabels, getLayers, getLearningRateUpdate, getLogLayer, getMomentum, getnIns, getnLayers, getnOuts, getOptimizer, getRenderWeightsEveryNEpochs, getRng, getSigmoidLayers, getSparsity, getWeightTransforms, initializeLayers, initializeNetwork, isForceNumEpochs, isShouldBackProp, isShouldInit, isToDecode, isUseRegularization, load, loadFromFile, merge, negativeLogLikelihood, predict, reconstruct, reconstruct, setActivation, setColumnMeans, setColumnStds, setColumnSums, setDist, setErrorTolerance, setFanIn, setForceNumEpochs, setHiddenLayerSizes, setInput, setL2, setLabels, setLayers, setLearningRateUpdate, setLogLayer, setMomentum, setnIns, setnLayers, setnOuts, setOptimizer, setRenderWeightsEveryNEpochs, setRng, setShouldBackProp, setShouldInit, setSigmoidLayers, setSparsity, setToDecode, setUseRegularization, setWeightTransforms, update, write

Methods inherited from class java.lang.Object

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

Constructor Detail

StackedDenoisingAutoEncoder

public StackedDenoisingAutoEncoder()

StackedDenoisingAutoEncoder

public StackedDenoisingAutoEncoder(int n_ins,
                           int[] hiddenLayerSizes,
                           int nOuts,
                           int nLayers,
                           org.apache.commons.math3.random.RandomGenerator rng,
                           org.jblas.DoubleMatrix input,
                           org.jblas.DoubleMatrix labels)

StackedDenoisingAutoEncoder

public StackedDenoisingAutoEncoder(int nIns,
                           int[] hiddenLayerSizes,
                           int nOuts,
                           int n_layers,
                           org.apache.commons.math3.random.RandomGenerator rng)

Method Detail

pretrain

public void pretrain(double lr,
            double corruptionLevel,
            int epochs)

pretrain

public void pretrain(org.jblas.DoubleMatrix input,
            double lr,
            double corruptionLevel,
            int epochs)

Unsupervised pretraining based on reconstructing the input from a corrupted version

Parameters:

input - the input to train on

lr - the starting learning rate

corruptionLevel - the corruption level (the smaller number of inputs; the higher the corruption level should be) the percent of inputs to corrupt

epochs - the number of iterations to run

trainNetwork

public void trainNetwork(org.jblas.DoubleMatrix input,
                org.jblas.DoubleMatrix labels,
                Object[] otherParams)

Description copied from class: BaseMultiLayerNetwork

Train the network running some unsupervised pretraining followed by SGD/finetune

Specified by:

trainNetwork in class BaseMultiLayerNetwork

Parameters:

input - input examples

labels - output labels

otherParams - (double) learningRate (double) corruptionLevel (int) epochs Optional: (double) finetune lr (int) finetune epochs

createLayer

public NeuralNetwork createLayer(org.jblas.DoubleMatrix input,
                        int nVisible,
                        int nHidden,
                        org.jblas.DoubleMatrix W,
                        org.jblas.DoubleMatrix hbias,
                        org.jblas.DoubleMatrix vBias,
                        org.apache.commons.math3.random.RandomGenerator rng,
                        int index)

Description copied from class: BaseMultiLayerNetwork

Creates a layer depending on the index. The main reason this matters is for continuous variations such as the CDBN where the first layer needs to be an CRBM for continuous inputs. Please be sure to call super.initializeNetwork to handle the passing of baseline parameters such as fanin and rendering.

Specified by:

createLayer in class BaseMultiLayerNetwork

Parameters:

input - the input to the layer

nVisible - the number of visible inputs

nHidden - the number of hidden units

W - the weight vector

hbias - the hidden bias

vBias - the visible bias

rng - the rng to use (THiS IS IMPORTANT; YOU DO NOT WANT TO HAVE A MIS REFERENCED RNG OTHERWISE NUMBERS WILL BE MEANINGLESS)

index - the index of the layer

Returns:

a neural network layer such as RBM

createNetworkLayers

public NeuralNetwork[] createNetworkLayers(int numLayers)

Specified by:

createNetworkLayers in class BaseMultiLayerNetwork