org.deeplearning4j.models.embeddings.inmemory

Class InMemoryLookupTable<T extends SequenceElement>

java.lang.Object

org.deeplearning4j.models.embeddings.inmemory.InMemoryLookupTable<T>

All Implemented Interfaces:

Serializable, WeightLookupTable<T>

Direct Known Subclasses:

GloveWeightLookupTable

public class InMemoryLookupTable<T extends SequenceElement>
extends Object
implements WeightLookupTable<T>

Default word lookup table

Author:

Adam Gibson

See Also:

Serialized Form

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	InMemoryLookupTable.Builder<T extends SequenceElement>
protected class	InMemoryLookupTable.WeightIterator

Field Summary

Fields

Modifier and Type	Field and Description
protected org.nd4j.linalg.learning.legacy.AdaGrad	adaGrad
protected Map<Integer,org.nd4j.linalg.api.ndarray.INDArray>	codes
protected double[]	expTable
protected com.google.common.util.concurrent.AtomicDouble	lr
protected static double	MAX_EXP
protected double	negative
protected org.nd4j.linalg.api.rng.Random	rng
protected long	seed
protected org.nd4j.linalg.api.ndarray.INDArray	syn0
protected org.nd4j.linalg.api.ndarray.INDArray	syn1
protected org.nd4j.linalg.api.ndarray.INDArray	syn1Neg
protected org.nd4j.linalg.api.ndarray.INDArray	table
protected Long	tableId
protected boolean	useAdaGrad
protected boolean	useHS
protected int	vectorLength
protected VocabCache<T>	vocab

Constructor Summary

Constructors

Constructor and Description
InMemoryLookupTable()
InMemoryLookupTable(VocabCache<T> vocab, int vectorLength, boolean useAdaGrad, double lr, org.nd4j.linalg.api.rng.Random gen, double negative)
InMemoryLookupTable(VocabCache<T> vocab, int vectorLength, boolean useAdaGrad, double lr, org.nd4j.linalg.api.rng.Random gen, double negative, boolean useHS)

Method Summary

Modifier and Type	Method and Description
void	consume(InMemoryLookupTable<T> srcTable) This method consumes weights of a given InMemoryLookupTable PLEASE NOTE: this method explicitly resets current weights
Map<Integer,org.nd4j.linalg.api.ndarray.INDArray>	getCodes()
double[]	getExpTable()
double	getGradient(int column, double gradient) Returns gradient for specified word
com.google.common.util.concurrent.AtomicDouble	getLr() Deprecated.
double	getNegative()
org.nd4j.linalg.api.ndarray.INDArray	getSyn0()
org.nd4j.linalg.api.ndarray.INDArray	getSyn1()
org.nd4j.linalg.api.ndarray.INDArray	getSyn1Neg()
org.nd4j.linalg.api.ndarray.INDArray	getTable()
VocabCache	getVocab()
VocabCache<T>	getVocabCache() Returns corresponding vocabulary
org.nd4j.linalg.api.ndarray.INDArray	getWeights()
protected void	initAdaGrad()
protected void	initExpTable()
void	initNegative()
boolean	isUseAdaGrad()
void	iterate(T w1, T w2) Deprecated.
void	iterateSample(T w1, T w2, AtomicLong nextRandom, double alpha) Deprecated.
int	layerSize() The layer size for the lookup table
org.nd4j.linalg.api.ndarray.INDArray	loadCodes(int[] codes) Loads the co-occurrences for the given codes
protected void	makeTable(int tableSize, double power)
void	plotVocab(org.deeplearning4j.plot.BarnesHutTsne tsne, int numWords, File file) Render the words via TSNE
void	plotVocab(org.deeplearning4j.plot.BarnesHutTsne tsne, int numWords, UiConnectionInfo connectionInfo) Render the words via TSNE
void	plotVocab(int numWords, File file) Render the words via tsne
void	plotVocab(int numWords, UiConnectionInfo connectionInfo) Render the words via tsne
void	putCode(int codeIndex, org.nd4j.linalg.api.ndarray.INDArray code)
void	putVector(String word, org.nd4j.linalg.api.ndarray.INDArray vector) Inserts a word vector
void	resetWeights() Reset the weights of the cache
void	resetWeights(boolean reset) Clear out all weights regardless
void	setCodes(Map<Integer,org.nd4j.linalg.api.ndarray.INDArray> codes)
void	setExpTable(double[] expTable)
void	setLearningRate(double lr) Sets the learning rate
void	setLr(com.google.common.util.concurrent.AtomicDouble lr)
void	setNegative(double negative)
void	setSyn0(org.nd4j.linalg.api.ndarray.INDArray syn0)
void	setSyn1(org.nd4j.linalg.api.ndarray.INDArray syn1)
void	setSyn1Neg(org.nd4j.linalg.api.ndarray.INDArray syn1Neg)
void	setTable(org.nd4j.linalg.api.ndarray.INDArray table)
void	setUseAdaGrad(boolean useAdaGrad)
void	setUseHS(boolean useHS)
void	setVectorLength(int vectorLength)
void	setVocab(VocabCache vocab)
String	toString()
org.nd4j.linalg.api.ndarray.INDArray	vector(String word)
Iterator<org.nd4j.linalg.api.ndarray.INDArray>	vectors() Iterates through all of the vectors in the cache

Methods inherited from class java.lang.Object

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

Methods inherited from interface org.deeplearning4j.models.embeddings.WeightLookupTable

getTableId, setTableId

Field Detail

syn0

protected org.nd4j.linalg.api.ndarray.INDArray syn0

syn1

protected org.nd4j.linalg.api.ndarray.INDArray syn1

vectorLength

protected int vectorLength

rng

protected transient org.nd4j.linalg.api.rng.Random rng

lr

protected com.google.common.util.concurrent.AtomicDouble lr

expTable

protected double[] expTable

MAX_EXP

protected static double MAX_EXP

seed

protected long seed

table

protected org.nd4j.linalg.api.ndarray.INDArray table

syn1Neg

protected org.nd4j.linalg.api.ndarray.INDArray syn1Neg

useAdaGrad

protected boolean useAdaGrad

negative

protected double negative

useHS

protected boolean useHS

vocab

protected VocabCache<T extends SequenceElement> vocab

codes

protected Map<Integer,org.nd4j.linalg.api.ndarray.INDArray> codes

adaGrad

protected org.nd4j.linalg.learning.legacy.AdaGrad adaGrad

tableId

protected Long tableId

Constructor Detail

InMemoryLookupTable

public InMemoryLookupTable()

InMemoryLookupTable

public InMemoryLookupTable(VocabCache<T> vocab,
                           int vectorLength,
                           boolean useAdaGrad,
                           double lr,
                           org.nd4j.linalg.api.rng.Random gen,
                           double negative,
                           boolean useHS)

InMemoryLookupTable

public InMemoryLookupTable(VocabCache<T> vocab,
                           int vectorLength,
                           boolean useAdaGrad,
                           double lr,
                           org.nd4j.linalg.api.rng.Random gen,
                           double negative)

Method Detail

initAdaGrad

protected void initAdaGrad()

getExpTable

public double[] getExpTable()

setExpTable

public void setExpTable(double[] expTable)

getGradient

public double getGradient(int column,
                          double gradient)

Description copied from interface: WeightLookupTable

Returns gradient for specified word

Specified by:

getGradient in interface WeightLookupTable<T extends SequenceElement>

Returns:

layerSize

public int layerSize()

Description copied from interface: WeightLookupTable

The layer size for the lookup table

Specified by:

layerSize in interface WeightLookupTable<T extends SequenceElement>

Returns:

the layer size for the lookup table

resetWeights

public void resetWeights(boolean reset)

Description copied from interface: WeightLookupTable

Clear out all weights regardless

Specified by:

resetWeights in interface WeightLookupTable<T extends SequenceElement>

plotVocab

public void plotVocab(org.deeplearning4j.plot.BarnesHutTsne tsne,
                      int numWords,
                      File file)

Description copied from interface: WeightLookupTable

Render the words via TSNE

Specified by:

plotVocab in interface WeightLookupTable<T extends SequenceElement>

Parameters:

tsne - the tsne to use

plotVocab

public void plotVocab(int numWords,
                      File file)

Render the words via tsne

Specified by:

plotVocab in interface WeightLookupTable<T extends SequenceElement>

plotVocab

public void plotVocab(int numWords,
                      UiConnectionInfo connectionInfo)

Render the words via tsne

Specified by:

plotVocab in interface WeightLookupTable<T extends SequenceElement>

plotVocab

public void plotVocab(org.deeplearning4j.plot.BarnesHutTsne tsne,
                      int numWords,
                      UiConnectionInfo connectionInfo)

Render the words via TSNE

Specified by:

plotVocab in interface WeightLookupTable<T extends SequenceElement>

Parameters:

tsne - the tsne to use

numWords -

connectionInfo -

putCode

public void putCode(int codeIndex,
                    org.nd4j.linalg.api.ndarray.INDArray code)

Specified by:

putCode in interface WeightLookupTable<T extends SequenceElement>

Parameters:

codeIndex -

code -

loadCodes

public org.nd4j.linalg.api.ndarray.INDArray loadCodes(int[] codes)

Loads the co-occurrences for the given codes

Specified by:

loadCodes in interface WeightLookupTable<T extends SequenceElement>

Parameters:

codes - the codes to load

Returns:

an ndarray of code.length by layerSize

initNegative

public void initNegative()

initExpTable

protected void initExpTable()

iterateSample

@Deprecated
public void iterateSample(T w1,
                                      T w2,
                                      AtomicLong nextRandom,
                                      double alpha)

Deprecated.

Iterate on the given 2 vocab words

Specified by:

iterateSample in interface WeightLookupTable<T extends SequenceElement>

Parameters:

w1 - the first word to iterate on

w2 - the second word to iterate on

nextRandom - next random for sampling

alpha - the alpha to use for learning

isUseAdaGrad

public boolean isUseAdaGrad()

setUseAdaGrad

public void setUseAdaGrad(boolean useAdaGrad)

getNegative

public double getNegative()

setUseHS

public void setUseHS(boolean useHS)

setNegative

public void setNegative(double negative)

iterate

@Deprecated
public void iterate(T w1,
                                T w2)

Deprecated.

Iterate on the given 2 vocab words

Specified by:

iterate in interface WeightLookupTable<T extends SequenceElement>

Parameters:

w1 - the first word to iterate on

w2 - the second word to iterate on

resetWeights

public void resetWeights()

Reset the weights of the cache

Specified by:

resetWeights in interface WeightLookupTable<T extends SequenceElement>

makeTable

protected void makeTable(int tableSize,
                         double power)

putVector

public void putVector(String word,
                      org.nd4j.linalg.api.ndarray.INDArray vector)

Inserts a word vector

Specified by:

putVector in interface WeightLookupTable<T extends SequenceElement>

Parameters:

word - the word to insert

vector - the vector to insert

getTable

public org.nd4j.linalg.api.ndarray.INDArray getTable()

setTable

public void setTable(org.nd4j.linalg.api.ndarray.INDArray table)

getSyn1Neg

public org.nd4j.linalg.api.ndarray.INDArray getSyn1Neg()

setSyn1Neg

public void setSyn1Neg(org.nd4j.linalg.api.ndarray.INDArray syn1Neg)

vector

public org.nd4j.linalg.api.ndarray.INDArray vector(String word)

Specified by:

vector in interface WeightLookupTable<T extends SequenceElement>

Parameters:

word -

Returns:

setLearningRate

public void setLearningRate(double lr)

Description copied from interface: WeightLookupTable

Sets the learning rate

Specified by:

setLearningRate in interface WeightLookupTable<T extends SequenceElement>

vectors

public Iterator<org.nd4j.linalg.api.ndarray.INDArray> vectors()

Description copied from interface: WeightLookupTable

Iterates through all of the vectors in the cache

Specified by:

vectors in interface WeightLookupTable<T extends SequenceElement>

Returns:

an iterator for all vectors in the cache

getWeights

public org.nd4j.linalg.api.ndarray.INDArray getWeights()

Specified by:

getWeights in interface WeightLookupTable<T extends SequenceElement>

getSyn0

public org.nd4j.linalg.api.ndarray.INDArray getSyn0()

setSyn0

public void setSyn0(org.nd4j.linalg.api.ndarray.INDArray syn0)

getSyn1

public org.nd4j.linalg.api.ndarray.INDArray getSyn1()

setSyn1

public void setSyn1(org.nd4j.linalg.api.ndarray.INDArray syn1)

getVocabCache

public VocabCache<T> getVocabCache()

Description copied from interface: WeightLookupTable

Returns corresponding vocabulary

Specified by:

getVocabCache in interface WeightLookupTable<T extends SequenceElement>

setVectorLength

public void setVectorLength(int vectorLength)

getLr

@Deprecated
public com.google.common.util.concurrent.AtomicDouble getLr()

Deprecated.

This method is deprecated, since all logic was pulled out from this class and is not used anymore. However this method will be around for a while, due to backward compatibility issues.

Returns:

initial learning rate

setLr

public void setLr(com.google.common.util.concurrent.AtomicDouble lr)

getVocab

public VocabCache getVocab()

setVocab

public void setVocab(VocabCache vocab)

getCodes

public Map<Integer,org.nd4j.linalg.api.ndarray.INDArray> getCodes()

setCodes

public void setCodes(Map<Integer,org.nd4j.linalg.api.ndarray.INDArray> codes)

toString

public String toString()

Overrides:

toString in class Object

consume

public void consume(InMemoryLookupTable<T> srcTable)

This method consumes weights of a given InMemoryLookupTable PLEASE NOTE: this method explicitly resets current weights

Parameters:

srcTable -