de.datexis.sector.tagger

Class SectorTagger

java.lang.Object

de.datexis.annotator.AnnotatorComponent

de.datexis.tagger.Tagger

de.datexis.sector.tagger.SectorTagger

All Implemented Interfaces:

IComponent

public class SectorTagger
extends Tagger

SECTOR Recurrent Network with separated FW/BW layers. Implementation of: Sebastian Arnold, Rudolf Schneider, Philippe Cudré-Mauroux, Felix A. Gers and Alexander Löser: "SECTOR: A Neural Model for Coherent Topic Segmentation and Classification." Transactions of the Association for Computational Linguistics (2019).

Author:

Sebastian Arnold

Field Summary

Fields

Modifier and Type	Field and Description
protected Encoder	bagEncoder
protected Encoder	embEncoder
protected ModelEvaluation	eval
protected org.deeplearning4j.nn.conf.preprocessor.FeedForwardToRnnPreProcessor	ff2rnn
protected Encoder	flagEncoder
protected static org.slf4j.Logger	log
protected boolean	requireSubsampling
protected Encoder	targetEncoder
protected int	workers

Fields inherited from class de.datexis.tagger.Tagger

batchSize, embeddingLayerSize, embeddingVectorSize, inputVectorSize, maxTimeSeriesLength, net, numEpochs, numExamples, outputVectorSize, randomize

Fields inherited from class de.datexis.annotator.AnnotatorComponent

id, model, modelAvailable, name, timer

Constructor Summary

Constructors

Constructor and Description
SectorTagger() used by XML deserializer
SectorTagger(Resource modelPath)
SectorTagger(String id)

Method Summary

Modifier and Type	Method and Description
void	attachVectors(Collection<Document> docs, AbstractMultiDataSetIterator.Stage stage, Class<? extends Encoder> targetClass)
protected void	attachVectors(DocumentSentenceIterator.DocumentBatch batch, Class<? extends Encoder> targetClass)
SectorTagger	buildSECTORModel(int ffwLayerSize, int lstmLayerSize, int embeddingLayerSize, int iterations, double learningRate, double dropout, org.nd4j.linalg.lossfunctions.ILossFunction lossFunc, org.nd4j.linalg.activations.Activation activation)
protected static void	clearLayerStates(org.deeplearning4j.nn.graph.ComputationGraph net) clear layer states to avoid leaks
void	enableTrainingUI()
Map<String,org.nd4j.linalg.api.ndarray.INDArray>	encodeMatrix(DocumentSentenceIterator.DocumentBatch batch)
static Map<String,org.nd4j.linalg.api.ndarray.INDArray>	feedForward(org.deeplearning4j.nn.graph.ComputationGraph net, org.nd4j.linalg.dataset.api.MultiDataSet next)
List<Encoder>	getEncoders()
org.deeplearning4j.nn.conf.ComputationGraphConfiguration	getGraphConfiguration()
org.deeplearning4j.nn.graph.ComputationGraph	getNN()
Encoder	getTargetEncoder()
boolean	isRequireSubsampling()
void	loadModel(Resource modelFile)
void	saveModel(Resource modelPath, String name) Saves the model to .bin.gz
void	setEncoders(List<Encoder> encoders)
void	setGraphConfiguration(org.nd4j.shade.jackson.databind.JsonNode conf)
void	setInputEncoders(Encoder bagEncoder, Encoder embEncoder, Encoder flagEncoder)
void	setRequireSubsampling(boolean requireSubsampling)
void	setTargetEncoder(Encoder targetEncoder)
SectorTagger	setWorkspaceParams(int workers)
void	tag(Collection<Document> docs)
void	testModel(Dataset dataset)
void	trainModel(Dataset dataset)
org.deeplearning4j.earlystopping.EarlyStoppingResult<org.deeplearning4j.nn.graph.ComputationGraph>	trainModel(Dataset train, Dataset validation, org.deeplearning4j.earlystopping.EarlyStoppingConfiguration conf)
void	trainModel(Dataset dataset, int numEpochs)
protected void	triggerEpochListeners(boolean epochStart, int epochNum)

Methods inherited from class de.datexis.tagger.Tagger

getBatchSize, getEmbeddingLayerSize, getLayerConfiguration, getNumEpochs, isModelAvailableInChildren, isRandomize, loadConf, saveUpdater, setBatchSize, setEmbeddingLayerSize, setLayerConfiguration, setListeners, setNumEpochs, setRandomize, setTrainingParams, tag

Methods inherited from class de.datexis.annotator.AnnotatorComponent

appendTestLog, appendTestLog, appendTrainLog, appendTrainLog, clearTestLog, clearTrainLog, getConf, getId, getModel, getName, getTestLog, getTrainLog, isModelAvailable, setConf, setId, setModel, setModelAvailable, setModelFilename, setName

Methods inherited from class java.lang.Object

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

Field Detail

log

protected static final org.slf4j.Logger log

bagEncoder

protected Encoder bagEncoder

embEncoder

protected Encoder embEncoder

flagEncoder

protected Encoder flagEncoder

targetEncoder

protected Encoder targetEncoder

workers

protected int workers

requireSubsampling

protected boolean requireSubsampling

eval

protected ModelEvaluation eval

ff2rnn

protected final org.deeplearning4j.nn.conf.preprocessor.FeedForwardToRnnPreProcessor ff2rnn

Constructor Detail

SectorTagger

public SectorTagger()

used by XML deserializer

SectorTagger

public SectorTagger(String id)

SectorTagger

public SectorTagger(Resource modelPath)

Method Detail

getNN

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

Overrides:

getNN in class Tagger

isRequireSubsampling

public boolean isRequireSubsampling()

setRequireSubsampling

public void setRequireSubsampling(boolean requireSubsampling)

setInputEncoders

public void setInputEncoders(Encoder bagEncoder,
                             Encoder embEncoder,
                             Encoder flagEncoder)

setTargetEncoder

public void setTargetEncoder(Encoder targetEncoder)

setWorkspaceParams

public SectorTagger setWorkspaceParams(int workers)

getEncoders

public List<Encoder> getEncoders()

Specified by:

getEncoders in interface IComponent

Overrides:

getEncoders in class Tagger

getTargetEncoder

public Encoder getTargetEncoder()

setEncoders

public void setEncoders(List<Encoder> encoders)

Specified by:

setEncoders in interface IComponent

Overrides:

setEncoders in class Tagger

buildSECTORModel

public SectorTagger buildSECTORModel(int ffwLayerSize,
                                     int lstmLayerSize,
                                     int embeddingLayerSize,
                                     int iterations,
                                     double learningRate,
                                     double dropout,
                                     org.nd4j.linalg.lossfunctions.ILossFunction lossFunc,
                                     org.nd4j.linalg.activations.Activation activation)

trainModel

public void trainModel(Dataset dataset)

Overrides:

trainModel in class Tagger

trainModel

public void trainModel(Dataset dataset,
                       int numEpochs)

trainModel

public org.deeplearning4j.earlystopping.EarlyStoppingResult<org.deeplearning4j.nn.graph.ComputationGraph> trainModel(Dataset train,
                                                                                                                     Dataset validation,
                                                                                                                     org.deeplearning4j.earlystopping.EarlyStoppingConfiguration conf)

testModel

public void testModel(Dataset dataset)

Overrides:

testModel in class Tagger

tag

public void tag(Collection<Document> docs)

Specified by:

tag in class Tagger

encodeMatrix

public Map<String,org.nd4j.linalg.api.ndarray.INDArray> encodeMatrix(DocumentSentenceIterator.DocumentBatch batch)

feedForward

public static Map<String,org.nd4j.linalg.api.ndarray.INDArray> feedForward(org.deeplearning4j.nn.graph.ComputationGraph net,
                                                                           org.nd4j.linalg.dataset.api.MultiDataSet next)

triggerEpochListeners

protected void triggerEpochListeners(boolean epochStart,
                                     int epochNum)

attachVectors

public void attachVectors(Collection<Document> docs,
                          AbstractMultiDataSetIterator.Stage stage,
                          Class<? extends Encoder> targetClass)

attachVectors

protected void attachVectors(DocumentSentenceIterator.DocumentBatch batch,
                             Class<? extends Encoder> targetClass)

clearLayerStates

protected static void clearLayerStates(org.deeplearning4j.nn.graph.ComputationGraph net)

clear layer states to avoid leaks

enableTrainingUI

public void enableTrainingUI()

saveModel

public void saveModel(Resource modelPath,
                      String name)

Saves the model to .bin.gz

Specified by:

saveModel in interface IComponent

Overrides:

saveModel in class Tagger

Parameters:

modelPath -

name -

loadModel

public void loadModel(Resource modelFile)

Specified by:

loadModel in interface IComponent

Overrides:

loadModel in class Tagger

getGraphConfiguration

public org.deeplearning4j.nn.conf.ComputationGraphConfiguration getGraphConfiguration()

Overrides:

getGraphConfiguration in class Tagger

setGraphConfiguration

public void setGraphConfiguration(org.nd4j.shade.jackson.databind.JsonNode conf)

Overrides:

setGraphConfiguration in class Tagger