org.deeplearning4j.graph.models.embeddings

Class InMemoryGraphLookupTable

java.lang.Object

org.deeplearning4j.graph.models.embeddings.InMemoryGraphLookupTable

All Implemented Interfaces:

GraphVectorLookupTable

public class InMemoryGraphLookupTable
extends java.lang.Object
implements GraphVectorLookupTable

A standard in-memory implementation of a lookup table for vector representations of the vertices in a graph

Field Summary

Fields

Modifier and Type	Field and Description
protected double[]	expTable
protected double	learningRate
protected static double	MAX_EXP
protected int	nVertices
protected org.nd4j.linalg.api.ndarray.INDArray	outWeights
protected BinaryTree	tree
protected int	vectorSize
protected org.nd4j.linalg.api.ndarray.INDArray	vertexVectors

Constructor Summary

Constructors

Constructor and Description
InMemoryGraphLookupTable(int nVertices, int vectorSize, BinaryTree tree, double learningRate)

Method Summary

Modifier and Type	Method and Description
double	calculateProb(int first, int second) Calculate the probability of the second vertex given the first vertex i.e., P(v_second | v_first)
double	calculateScore(int first, int second) Calculate score.
org.nd4j.linalg.api.ndarray.INDArray	getInnerNodeVector(int innerNode)
int	getNumVertices() Returns the number of vertices in the graph
org.nd4j.linalg.api.ndarray.INDArray	getOutWeights()
BinaryTree	getTree()
org.nd4j.linalg.api.ndarray.INDArray	getVector(int idx) Get the vector for the vertex with index idx
org.nd4j.linalg.api.ndarray.INDArray	getVertexVectors()
void	iterate(int first, int second) Conduct learning given a pair of vertices (in and out)
void	resetWeights() Reset (randomize) the weights.
void	setLearningRate(double learningRate) Set the learning rate
void	setVertexVectors(org.nd4j.linalg.api.ndarray.INDArray vertexVectors)
org.nd4j.linalg.api.ndarray.INDArray[][]	vectorsAndGradients(int first, int second) Returns vertex vector and vector gradients, plus inner node vectors and inner node gradients Specifically, out[0] are vectors, out[1] are gradients for the corresponding vectors out[0][0] is vector for first vertex; out[0][1] is gradient for this vertex vector out[0][i] (i>0) is the inner node vector along path to second vertex; out[1][i] is gradient for inner node vertex This design is used primarily to aid in testing (numerical gradient checks)
int	vectorSize() The size of the vector representations

Methods inherited from class java.lang.Object

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

Field Detail

nVertices

protected int nVertices

vectorSize

protected int vectorSize

tree

protected BinaryTree tree

vertexVectors

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

outWeights

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

learningRate

protected double learningRate

expTable

protected double[] expTable

MAX_EXP

protected static double MAX_EXP

Constructor Detail

InMemoryGraphLookupTable

public InMemoryGraphLookupTable(int nVertices,
                                int vectorSize,
                                BinaryTree tree,
                                double learningRate)

Method Detail

getVertexVectors

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

getOutWeights

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

vectorSize

public int vectorSize()

Description copied from interface: GraphVectorLookupTable

The size of the vector representations

Specified by:

vectorSize in interface GraphVectorLookupTable

resetWeights

public void resetWeights()

Description copied from interface: GraphVectorLookupTable

Reset (randomize) the weights.

Specified by:

resetWeights in interface GraphVectorLookupTable

iterate

public void iterate(int first,
                    int second)

Description copied from interface: GraphVectorLookupTable

Conduct learning given a pair of vertices (in and out)

Specified by:

iterate in interface GraphVectorLookupTable

vectorsAndGradients

public org.nd4j.linalg.api.ndarray.INDArray[][] vectorsAndGradients(int first,
                                                                    int second)

Returns vertex vector and vector gradients, plus inner node vectors and inner node gradients
Specifically, out[0] are vectors, out[1] are gradients for the corresponding vectors
out[0][0] is vector for first vertex; out[0][1] is gradient for this vertex vector
out[0][i] (i>0) is the inner node vector along path to second vertex; out[1][i] is gradient for inner node vertex
This design is used primarily to aid in testing (numerical gradient checks)

Parameters:

first - first (input) vertex index

second - second (output) vertex index

calculateProb

public double calculateProb(int first,
                            int second)

Calculate the probability of the second vertex given the first vertex i.e., P(v_second | v_first)

Parameters:

first - index of the first vertex

second - index of the second vertex

Returns:

probability, P(v_second | v_first)

calculateScore

public double calculateScore(int first,
                             int second)

Calculate score. -log P(v_second | v_first)

getTree

public BinaryTree getTree()

getInnerNodeVector

public org.nd4j.linalg.api.ndarray.INDArray getInnerNodeVector(int innerNode)

getVector

public org.nd4j.linalg.api.ndarray.INDArray getVector(int idx)

Description copied from interface: GraphVectorLookupTable

Get the vector for the vertex with index idx

Specified by:

getVector in interface GraphVectorLookupTable

setLearningRate

public void setLearningRate(double learningRate)

Description copied from interface: GraphVectorLookupTable

Set the learning rate

Specified by:

setLearningRate in interface GraphVectorLookupTable

getNumVertices

public int getNumVertices()

Description copied from interface: GraphVectorLookupTable

Returns the number of vertices in the graph

Specified by:

getNumVertices in interface GraphVectorLookupTable

setVertexVectors

public void setVertexVectors(org.nd4j.linalg.api.ndarray.INDArray vertexVectors)