weka.classifiers.functions

Class Dl4jMlpClassifier

java.lang.Object

weka.classifiers.AbstractClassifier

weka.classifiers.RandomizableClassifier

weka.classifiers.functions.Dl4jMlpClassifier

All Implemented Interfaces:

java.io.Serializable, java.lang.Cloneable, weka.classifiers.Classifier, weka.classifiers.IterativeClassifier, weka.core.BatchPredictor, weka.core.CapabilitiesHandler, weka.core.CapabilitiesIgnorer, weka.core.CommandlineRunnable, weka.core.OptionHandler, weka.core.Randomizable, weka.core.RevisionHandler

Direct Known Subclasses:

RnnSequenceClassifier

public class Dl4jMlpClassifier
extends weka.classifiers.RandomizableClassifier
implements weka.core.BatchPredictor, weka.core.CapabilitiesHandler, weka.classifiers.IterativeClassifier

A wrapper for DeepLearning4j that can be used to train a multi-layer perceptron.

See Also:

Serialized Form

Field Summary

Fields

Modifier and Type	Field and Description
static int	FILTER_NONE filter: No normalization/standardization.
static int	FILTER_NORMALIZE filter: Normalize training data.
static int	FILTER_STANDARDIZE filter: Standardize training data.
static weka.core.Tag[]	TAGS_FILTER The filter to apply to the training data.

Fields inherited from class weka.classifiers.AbstractClassifier

BATCH_SIZE_DEFAULT, NUM_DECIMAL_PLACES_DEFAULT

Constructor Summary

Constructors

Constructor and Description
Dl4jMlpClassifier() Instantiate the model.

Method Summary

Modifier and Type	Method and Description
boolean	arithmeticUnderflow(org.nd4j.linalg.api.ndarray.INDArray array) Checks the array (as output from ComputationGraph.outputSingle()) for arithmetic underflow
void	buildClassifier(weka.core.Instances data) The method used to train the classifier.
double[]	distributionForInstance(weka.core.Instance inst) The method to use when making a prediction for a test instance.
double[][]	distributionsForInstances(weka.core.Instances insts) The method to use when making predictions for test instances.
void	done() Clean up after learning.
org.nd4j.linalg.api.ndarray.INDArray	featurizeForLayer(java.lang.String layerName, org.nd4j.linalg.dataset.api.iterator.DataSetIterator iter, PoolingType poolingType) Uses the DL4J TransferLearningHelper to featurize the instances using activations from the given layer
weka.core.Instances	getActivationsAtLayers(java.lang.String[] layerNames, weka.core.Instances input) Overridden method - if no pooling type is given then set it to NONE Uses the given set of layers to extract features for the given dataset
weka.core.Instances	getActivationsAtLayers(java.lang.String[] layerNames, weka.core.Instances input, PoolingType poolingType) Uses the given set of layers to extract features for the given dataset
CacheMode	getCacheMode()
weka.core.Capabilities	getCapabilities() Returns default capabilities of the classifier.
boolean	getDoNotClearFilesystemCache()
EarlyStopping	getEarlyStopping()
weka.core.SelectedTag	getFilterType()
org.deeplearning4j.nn.conf.inputs.InputType.InputTypeConvolutional	getInputShape(CustomModelSetup customModelSetup)
AbstractInstanceIterator	getInstanceIterator()
TrainingListener	getIterationListener()
Layer[]	getLayers()
boolean	getLoadLayerSpecification()
LogConfiguration	getLogConfig() Get the log configuration.
org.deeplearning4j.nn.graph.ComputationGraph	getModel() Get the ComputationGraph model
java.lang.String	getModelName() Get the name of the loaded model
NeuralNetConfiguration	getNeuralNetConfiguration()
int	getNumEpochs()
int	getNumGPUs()
int	getParameterAveragingFrequency()
int	getPrefetchBufferSize()
int	getQueueSize()
boolean	getResume() Returns true if the model is to be finalized (or has been finalized) after training.
AbstractZooModel	getZooModel() Get the modelzoo model
java.lang.String	globalInfo() .
boolean	implementsMoreEfficientBatchPrediction() Performs efficient batch prediction
void	initializeClassifier(weka.core.Instances data) The method used to initialize the classifier.
boolean	isFilterMode()
static boolean	isValidOutputLayer(boolean filterMode, org.deeplearning4j.nn.conf.layers.Layer layer) Checks if the layer is a valid output layer
static Dl4jMlpClassifier	loadInferenceModel(java.io.File serializedModelFile, AbstractZooModel zooModelType) Load a Dl4jMlpClassifier for use in the Inference Panel - no need to supply Instances or InstanceIterators
static Dl4jMlpClassifier	loadModel(weka.core.Instances data, java.io.File serializedModelFile, AbstractZooModel zooModelType, AbstractInstanceIterator instanceIterator) Load a Dl4jMlpClassifier for use with the given instances and iterator
void	loadZooModelNoData(int numClasses, long seed, int[] newShape) Load a ComputationGraph without any data - used in the Dl4j Inference panel
static void	main(java.lang.String[] argv) The main method for running this class.
boolean	next() Perform another epoch.
org.nd4j.linalg.api.ndarray.INDArray	outputSingle(org.nd4j.linalg.api.ndarray.INDArray image) Convenience method to allow the Dl4j Inference panel to call outputSingle for a single image
void	setCacheMode(CacheMode cm)
void	setDoNotClearFilesystemCache(boolean clear)
void	setEarlyStopping(EarlyStopping config)
void	setFilterMode(boolean filterMode)
void	setFilterType(weka.core.SelectedTag newType)
void	setInstanceIterator(AbstractInstanceIterator iterator)
void	setIterationListener(TrainingListener l)
void	setLayers(Layer... layers)
void	setLoadLayerSpecification(boolean loadLayerSpecification)
void	setLogConfig(LogConfiguration logConfig) Set the log configuration.
void	setNeuralNetConfiguration(NeuralNetConfiguration config)
void	setNumEpochs(int numEpochs)
void	setNumGPUs(int numGPUs)
void	setParameterAveragingFrequency(int frequency)
void	setPrefetchBufferSize(int prefetchBufferSize)
void	setQueueSize(int QueueSize)
void	setResume(boolean resume) If called with argument true, then the next time done() is called the model is effectively "frozen" and no further iterations can be performed
void	setZooModel(AbstractZooModel zooModel) Set the modelzoo zooModel
static weka.core.Instances[]	splitTrainVal(weka.core.Instances data, double p) Split the dataset into p% train an (100-p)% test set.
java.lang.String	toString() Returns a string describing the model.
static Dl4jMlpClassifier	tryLoadFromFile(java.io.File serializedModelFile, AbstractZooModel zooModelType) Tries to load from a saved model file (if it exists), otherwise loads the given zoo model
boolean	useEarlyStopping() Use early stopping only if valid split percentage
boolean	useZooModel() Check if the user has selected to use a zoomodel

Methods inherited from class weka.classifiers.RandomizableClassifier

getOptions, getSeed, listOptions, seedTipText, setOptions, setSeed

Methods inherited from class weka.classifiers.AbstractClassifier

batchSizeTipText, classifyInstance, debugTipText, doNotCheckCapabilitiesTipText, forName, getBatchSize, getDebug, getDoNotCheckCapabilities, getNumDecimalPlaces, getRevision, makeCopies, makeCopy, numDecimalPlacesTipText, postExecution, preExecution, run, runClassifier, setBatchSize, setDebug, setDoNotCheckCapabilities, setNumDecimalPlaces

Methods inherited from class java.lang.Object

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

Methods inherited from interface weka.core.BatchPredictor

getBatchSize, setBatchSize

Methods inherited from interface weka.classifiers.Classifier

classifyInstance

Methods inherited from interface weka.core.OptionHandler

makeCopy

Field Detail

FILTER_NORMALIZE

public static final int FILTER_NORMALIZE

filter: Normalize training data.

See Also:

Constant Field Values

FILTER_STANDARDIZE

public static final int FILTER_STANDARDIZE

filter: Standardize training data.

See Also:

Constant Field Values

FILTER_NONE

public static final int FILTER_NONE

filter: No normalization/standardization.

See Also:

Constant Field Values

TAGS_FILTER

public static final weka.core.Tag[] TAGS_FILTER

The filter to apply to the training data.

Constructor Detail

Dl4jMlpClassifier

public Dl4jMlpClassifier()

Instantiate the model.

Method Detail

main

public static void main(java.lang.String[] argv)

The main method for running this class.

Parameters:

argv - the command-line arguments

splitTrainVal

public static weka.core.Instances[] splitTrainVal(weka.core.Instances data,
                                                  double p)
                                           throws java.lang.Exception

Split the dataset into p% train an (100-p)% test set.

Parameters:

data - Input data

p - train percentage

Returns:

Array of instances: (0) Train, (1) Test

Throws:

java.lang.Exception - Filterapplication went wrong

getLogConfig

public LogConfiguration getLogConfig()

Get the log configuration.

Returns:

Log configuration

setLogConfig

@OptionMetadata(displayName="log config",
                description="The log configuration.",
                commandLineParamName="logConfig",
                commandLineParamSynopsis="-logConfig <LogConfiguration>",
                displayOrder=1)
public void setLogConfig(LogConfiguration logConfig)

Set the log configuration.

Parameters:

logConfig - Log configuration

globalInfo

public java.lang.String globalInfo()

.

Returns:

Global info to return.

getCapabilities

public weka.core.Capabilities getCapabilities()

Returns default capabilities of the classifier.

Specified by:

getCapabilities in interface weka.classifiers.Classifier

Specified by:

getCapabilities in interface weka.core.CapabilitiesHandler

Overrides:

getCapabilities in class weka.classifiers.AbstractClassifier

Returns:

the capabilities of this classifier

getLayers

public Layer[] getLayers()

setLayers

@OptionMetadata(displayName="layer specification.",
                description="The specification of a layer. This option can be used multiple times.",
                commandLineParamName="layer",
                commandLineParamSynopsis="-layer <string>",
                displayOrder=2)
public void setLayers(Layer... layers)

getNumEpochs

public int getNumEpochs()

setNumEpochs

@OptionMetadata(description="The number of epochs to perform.",
                displayName="number of epochs",
                commandLineParamName="numEpochs",
                commandLineParamSynopsis="-numEpochs <int>",
                displayOrder=4)
public void setNumEpochs(int numEpochs)

getInstanceIterator

@OptionMetadata(description="The instance trainIterator to use.",
                displayName="instance iterator",
                commandLineParamName="iterator",
                commandLineParamSynopsis="-iterator <string>",
                displayOrder=6)
public AbstractInstanceIterator getInstanceIterator()

setInstanceIterator

public void setInstanceIterator(AbstractInstanceIterator iterator)

getNeuralNetConfiguration

@OptionMetadata(description="The neural network configuration to use.",
                displayName="network configuration",
                commandLineParamName="config",
                commandLineParamSynopsis="-config <string>",
                displayOrder=7)
public NeuralNetConfiguration getNeuralNetConfiguration()

setNeuralNetConfiguration

public void setNeuralNetConfiguration(NeuralNetConfiguration config)

getEarlyStopping

@OptionMetadata(description="The early stopping configuration to use.",
                displayName="early stopping",
                commandLineParamName="early-stopping",
                commandLineParamSynopsis="-early-stopping <string>",
                displayOrder=7)
public EarlyStopping getEarlyStopping()

setEarlyStopping

public void setEarlyStopping(EarlyStopping config)

getFilterType

@OptionMetadata(description="The type of normalization to perform.",
                displayName="attribute normalization",
                commandLineParamName="normalization",
                commandLineParamSynopsis="-normalization <int>",
                displayOrder=12)
public weka.core.SelectedTag getFilterType()

setFilterType

public void setFilterType(weka.core.SelectedTag newType)

getQueueSize

public int getQueueSize()

setQueueSize

@OptionMetadata(description="The queue size for asynchronous data transfer (default: 0, synchronous transfer).",
                displayName="data queue size",
                commandLineParamName="queueSize",
                commandLineParamSynopsis="-queueSize <int>",
                displayOrder=30)
public void setQueueSize(int QueueSize)

getResume

public boolean getResume()

Returns true if the model is to be finalized (or has been finalized) after training.

Specified by:

getResume in interface weka.classifiers.IterativeClassifier

Returns:

the current value of finalize

setResume

@OptionMetadata(description="Set whether training can be resumed at a later date",
                displayName="resume",
                commandLineParamName="resume",
                commandLineParamSynopsis="-resume",
                commandLineParamIsFlag=true,
                displayOrder=31)
public void setResume(boolean resume)

If called with argument true, then the next time done() is called the model is effectively "frozen" and no further iterations can be performed

Specified by:

setResume in interface weka.classifiers.IterativeClassifier

Parameters:

resume - true if the model is to be finalized after performing iterations

getLoadLayerSpecification

public boolean getLoadLayerSpecification()

setLoadLayerSpecification

@OptionMetadata(description="Set whether you want the GUI to load the layer specification when selecting a zoo model. This does not affect the model\'s function, but simply allows you to view the model\'s layers. You may need to click \'OK\' and open this window again for this setting to come into effect.",
                displayName="Preview zoo model layer specification in GUI",
                commandLineParamName="load-layer-spec",
                commandLineParamSynopsis="-load-layer-spec",
                commandLineParamIsFlag=true,
                displayOrder=3)
public void setLoadLayerSpecification(boolean loadLayerSpecification)

getDoNotClearFilesystemCache

public boolean getDoNotClearFilesystemCache()

setDoNotClearFilesystemCache

@OptionMetadata(displayName="Preserve filesystem cache",
                description="If true, the filesystem cache will not be cleared when starting or resuming training of a model. This can save time on data preparation for a given problem, but will cause errors if a dataset different to that in the cache is used",
                commandLineParamName="preserve-file-cache",
                commandLineParamSynopsis="-preserve-file-cache",
                commandLineParamIsFlag=true,
                displayOrder=32)
public void setDoNotClearFilesystemCache(boolean clear)

getNumGPUs

public int getNumGPUs()

setNumGPUs

@OptionMetadata(displayName="Number of GPUs",
                description="Number of available GPUs (ignored if no GPU backend available)",
                commandLineParamName="numGPUs",
                commandLineParamSynopsis="-numGPUs <integer>",
                displayOrder=33)
public void setNumGPUs(int numGPUs)

getPrefetchBufferSize

public int getPrefetchBufferSize()

setPrefetchBufferSize

@OptionMetadata(displayName="Size of prefetch buffer for multiple GPUs",
                description="Size of the prefetch buffer that will be used for background data prefetching (0 = disable prefetch). Ignored if there is only one GPU, or no GPU backend available.",
                commandLineParamName="prefetchSize",
                commandLineParamSynopsis="-prefetchSize <integer>",
                displayOrder=34)
public void setPrefetchBufferSize(int prefetchBufferSize)

getParameterAveragingFrequency

public int getParameterAveragingFrequency()

setParameterAveragingFrequency

@OptionMetadata(displayName="Model parameter averaging frequency",
                description="How often (in iterations, not epochs) to average model parameters when leveraging multiple GPUs (ignored if there is only one GPU, or no GPU backend available). Set no greater than num instances/num GPUs or no averaging will occur!",
                commandLineParamName="averagingFrequency",
                commandLineParamSynopsis="-averagingFrequency <integer>",
                displayOrder=35)
public void setParameterAveragingFrequency(int frequency)

getModelName

public java.lang.String getModelName()

Get the name of the loaded model

Returns:

Model name

buildClassifier

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

The method used to train the classifier.

Specified by:

buildClassifier in interface weka.classifiers.Classifier

Parameters:

data - set of instances serving as training data

Throws:

java.lang.Exception - if something goes wrong in the training process

isValidOutputLayer

public static boolean isValidOutputLayer(boolean filterMode,
                                         org.deeplearning4j.nn.conf.layers.Layer layer)

Checks if the layer is a valid output layer

Parameters:

filterMode - true if the model is being used for a filter

layer - output layer of the model

Returns:

true if the supplied layer is a valid output layer

initializeClassifier

public void initializeClassifier(weka.core.Instances data)
                          throws java.lang.Exception

The method used to initialize the classifier.

Specified by:

initializeClassifier in interface weka.classifiers.IterativeClassifier

Parameters:

data - set of instances serving as training data

Throws:

java.lang.Exception - if something goes wrong in the training process

getInputShape

public org.deeplearning4j.nn.conf.inputs.InputType.InputTypeConvolutional getInputShape(CustomModelSetup customModelSetup)

loadZooModelNoData

public void loadZooModelNoData(int numClasses,
                               long seed,
                               int[] newShape)

Load a ComputationGraph without any data - used in the Dl4j Inference panel

Parameters:

numClasses - Number of classes to load with

seed - Random seed

newShape - Input shape

outputSingle

public org.nd4j.linalg.api.ndarray.INDArray outputSingle(org.nd4j.linalg.api.ndarray.INDArray image)

Convenience method to allow the Dl4j Inference panel to call outputSingle for a single image

Parameters:

image - Image to predict for

Returns:

Model predictions

next

public boolean next()
             throws java.lang.Exception

Perform another epoch.

Specified by:

next in interface weka.classifiers.IterativeClassifier

Throws:

java.lang.Exception

useEarlyStopping

public boolean useEarlyStopping()

Use early stopping only if valid split percentage

Returns:

True if early stopping should be used

done

public void done()

Clean up after learning.

Specified by:

done in interface weka.classifiers.IterativeClassifier

getZooModel

public AbstractZooModel getZooModel()

Get the modelzoo model

Returns:

The modelzoo model object

setZooModel

@OptionMetadata(displayName="zooModel",
                description="The model-architecture to choose from the modelzoo (default = no model).",
                commandLineParamName="zooModel",
                commandLineParamSynopsis="-zooModel <string>",
                displayOrder=11)
public void setZooModel(AbstractZooModel zooModel)

Set the modelzoo zooModel

Parameters:

zooModel - The predefined zooModel

tryLoadFromFile

public static Dl4jMlpClassifier tryLoadFromFile(java.io.File serializedModelFile,
                                                AbstractZooModel zooModelType)
                                         throws weka.core.WekaException

Tries to load from a saved model file (if it exists), otherwise loads the given zoo model

Parameters:

serializedModelFile - Saved model path

zooModelType - Type of Zoo Model

Returns:

Dl4jMlpClassifier with the loaded ComputationGraph

Throws:

weka.core.WekaException - From errors occurring during loading the model file

loadModel

public static Dl4jMlpClassifier loadModel(weka.core.Instances data,
                                          java.io.File serializedModelFile,
                                          AbstractZooModel zooModelType,
                                          AbstractInstanceIterator instanceIterator)
                                   throws java.lang.Exception

Load a Dl4jMlpClassifier for use with the given instances and iterator

Parameters:

data - Instances to prime the model with

serializedModelFile - Saved model file

zooModelType - Type of Zoo Model

instanceIterator - Instance iterator to prime the model with

Returns:

Dl4jMlpClassifier setup with the instances and iterator

Throws:

java.lang.Exception - From errors occurring during loading the model file, or from intializing from the data

loadInferenceModel

public static Dl4jMlpClassifier loadInferenceModel(java.io.File serializedModelFile,
                                                   AbstractZooModel zooModelType)
                                            throws weka.core.WekaException

Load a Dl4jMlpClassifier for use in the Inference Panel - no need to supply Instances or InstanceIterators

Parameters:

serializedModelFile - Saved model file

zooModelType - Type of Zoo Model

Returns:

Dl4jMlpClassifier ready to be used in the Inference Panel

Throws:

weka.core.WekaException - From errors occurring during loading the model file, or from intializing from the data

getIterationListener

public TrainingListener getIterationListener()

setIterationListener

@OptionMetadata(displayName="set the iteration listener",
                description="Set the iteration listener.",
                commandLineParamName="iteration-listener",
                commandLineParamSynopsis="-iteration-listener <string>",
                displayOrder=9)
public void setIterationListener(TrainingListener l)

getCacheMode

public CacheMode getCacheMode()

setCacheMode

@OptionMetadata(displayName="set the cache mode",
                description="Set the cache mode.",
                commandLineParamName="cache-mode",
                commandLineParamSynopsis="-cache-mode <string>",
                displayOrder=13)
public void setCacheMode(CacheMode cm)

isFilterMode

public boolean isFilterMode()

setFilterMode

@ProgrammaticProperty
public void setFilterMode(boolean filterMode)

implementsMoreEfficientBatchPrediction

public boolean implementsMoreEfficientBatchPrediction()

Performs efficient batch prediction

Specified by:

implementsMoreEfficientBatchPrediction in interface weka.core.BatchPredictor

Overrides:

implementsMoreEfficientBatchPrediction in class weka.classifiers.AbstractClassifier

Returns:

true

distributionForInstance

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

The method to use when making a prediction for a test instance. Use distributionsForInstances() instead for speed if possible.

Specified by:

distributionForInstance in interface weka.classifiers.Classifier

Overrides:

distributionForInstance in class weka.classifiers.AbstractClassifier

Parameters:

inst - the instance to get a prediction for

Returns:

the class probability estimates (if the class is nominal) or the numeric prediction (if it is numeric)

Throws:

java.lang.Exception - if something goes wrong at prediction time

arithmeticUnderflow

public boolean arithmeticUnderflow(org.nd4j.linalg.api.ndarray.INDArray array)

Checks the array (as output from ComputationGraph.outputSingle()) for arithmetic underflow

Parameters:

array - Array to check

Returns:

True if the model returned an arithmetic underflow

distributionsForInstances

public double[][] distributionsForInstances(weka.core.Instances insts)
                                     throws java.lang.Exception

The method to use when making predictions for test instances.

Specified by:

distributionsForInstances in interface weka.core.BatchPredictor

Overrides:

distributionsForInstances in class weka.classifiers.AbstractClassifier

Parameters:

insts - the instances to get predictions for

Returns:

the class probability estimates (if the class is nominal) or the numeric predictions (if it is numeric)

Throws:

java.lang.Exception - if something goes wrong at prediction time

getModel

public org.deeplearning4j.nn.graph.ComputationGraph getModel()

Get the ComputationGraph model

Returns:

ComputationGraph model

toString

public java.lang.String toString()

Returns a string describing the model.

Overrides:

toString in class java.lang.Object

Returns:

the model string

useZooModel

public boolean useZooModel()

Check if the user has selected to use a zoomodel

Returns:

True if zoomodel is not CustomNet

getActivationsAtLayers

public weka.core.Instances getActivationsAtLayers(java.lang.String[] layerNames,
                                                  weka.core.Instances input)
                                           throws java.lang.Exception

Overridden method - if no pooling type is given then set it to NONE Uses the given set of layers to extract features for the given dataset

Parameters:

layerNames - Layer

input - data to featurize

Returns:

Instances transformed to the image features

Throws:

java.lang.Exception

featurizeForLayer

public org.nd4j.linalg.api.ndarray.INDArray featurizeForLayer(java.lang.String layerName,
                                                              org.nd4j.linalg.dataset.api.iterator.DataSetIterator iter,
                                                              PoolingType poolingType)
                                                       throws java.lang.Exception

Uses the DL4J TransferLearningHelper to featurize the instances using activations from the given layer

Parameters:

layerName - layer activations to use for instances

iter - iterator for the instances

poolingType - pooling type to be used (only necessary if using intermediary layers with 3D activations)

Returns:

INDArray containing the newly transformed instances

Throws:

java.lang.Exception

getActivationsAtLayers

public weka.core.Instances getActivationsAtLayers(java.lang.String[] layerNames,
                                                  weka.core.Instances input,
                                                  PoolingType poolingType)
                                           throws java.lang.Exception

Uses the given set of layers to extract features for the given dataset

Parameters:

layerNames - Layer

input - data to featurize

poolingType - pooling type to use

Returns:

Instances transformed to the image features

Throws:

java.lang.Exception