org.deeplearning4j.nn.graph.vertex

Class BaseGraphVertex

java.lang.Object

org.deeplearning4j.nn.graph.vertex.BaseGraphVertex

All Implemented Interfaces:

Serializable, Trainable, GraphVertex

Direct Known Subclasses:

DuplicateToTimeSeriesVertex, ElementWiseVertex, InputVertex, L2NormalizeVertex, L2Vertex, LastTimeStepVertex, LayerVertex, MergeVertex, PoolHelperVertex, PreprocessorVertex, ReshapeVertex, ReverseTimeSeriesVertex, SameDiffGraphVertex, ScaleVertex, ShiftVertex, StackVertex, SubsetVertex, UnstackVertex

public abstract class BaseGraphVertex
extends Object
implements GraphVertex

See Also:

Serialized Form

Field Summary

Fields

Modifier and Type	Field and Description
protected DataType	dataType
protected INDArray	epsilon
protected ComputationGraph	graph
protected INDArray[]	inputs
protected VertexIndices[]	inputVertices A representation of the vertices that are inputs to this vertex (inputs during forward pass) Specifically, if inputVertices[X].getVertexIndex() = Y, and inputVertices[X].getVertexEdgeNumber() = Z then the Zth output of vertex Y is the Xth input to this vertex
protected boolean	outputVertex
protected VertexIndices[]	outputVertices A representation of the vertices that this vertex is connected to (outputs duing forward pass) Specifically, if outputVertices[X].getVertexIndex() = Y, and outputVertices[X].getVertexEdgeNumber() = Z then the output of this vertex (there is only one output) is connected to the Zth input of vertex Y
protected int	vertexIndex The index of this vertex
protected String	vertexName

Constructor Summary

Constructors

Modifier	Constructor and Description
protected	BaseGraphVertex(ComputationGraph graph, String name, int vertexIndex, VertexIndices[] inputVertices, VertexIndices[] outputVertices, DataType dataType)

Method Summary

Modifier and Type	Method and Description
boolean	canDoBackward() Whether the GraphVertex can do backward pass.
boolean	canDoForward() Whether the GraphVertex can do forward pass.
void	clear() Clear the internal state (if any) of the GraphVertex.
void	clearVertex() This method clears inpjut for this vertex
TrainingConfig	getConfig()
INDArray	getEpsilon() Get the epsilon/error (i.e., dL/dOutput) array previously set for this GraphVertex
INDArray	getGradientsViewArray()
VertexIndices[]	getInputVertices() A representation of the vertices that are inputs to this vertex (inputs duing forward pass) Specifically, if inputVertices[X].getVertexIndex() = Y, and inputVertices[X].getVertexEdgeNumber() = Z then the Zth output of vertex Y is the Xth input to this vertex
int	getNumInputArrays() Get the number of input arrays.
int	getNumOutputConnections() Get the number of outgoing connections from this GraphVertex.
VertexIndices[]	getOutputVertices() A representation of the vertices that this vertex is connected to (outputs duing forward pass) Specifically, if outputVertices[X].getVertexIndex() = Y, and outputVertices[X].getVertexEdgeNumber() = Z then the Xth output of this vertex is connected to the Zth input of vertex Y
int	getVertexIndex() Get the index of the GraphVertex
String	getVertexName() Get the name/label of the GraphVertex
boolean	isInputVertex() Whether the GraphVertex is an input vertex
long	numParams()
INDArray	params()
Map<String,INDArray>	paramTable(boolean backpropOnly) Get the parameter table for the vertex
void	setEpsilon(INDArray epsilon) Set the errors (epsilon - aka dL/dActivation) for this GraphVertex
void	setInput(int inputNumber, INDArray input, LayerWorkspaceMgr workspaceMgr) Set the input activations.
void	setInputVertices(VertexIndices[] inputVertices) Sets the input vertices.
void	setLayerAsFrozen() Only applies to layer vertices.
void	setOutputVertices(VertexIndices[] outputVertices) set the output vertices.
abstract String	toString()
boolean	updaterDivideByMinibatch(String paramName) DL4J layers typically produce the sum of the gradients during the backward pass for each layer, and if required (if minibatch=true) then divide by the minibatch size. However, there are some exceptions, such as the batch norm mean/variance estimate parameters: these "gradients" are actually not gradients, but are updates to be applied directly to the parameter vector.

Methods inherited from class java.lang.Object

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

Methods inherited from interface org.deeplearning4j.nn.graph.vertex.GraphVertex

doBackward, doForward, feedForwardMaskArrays, getInputs, getLayer, hasLayer, isOutputVertex, setBackpropGradientsViewArray, setInputs, setOutputVertex

Field Detail

graph

protected ComputationGraph graph

vertexName

protected String vertexName

vertexIndex

protected int vertexIndex

The index of this vertex

inputVertices

protected VertexIndices[] inputVertices

A representation of the vertices that are inputs to this vertex (inputs during forward pass) Specifically, if inputVertices[X].getVertexIndex() = Y, and inputVertices[X].getVertexEdgeNumber() = Z then the Zth output of vertex Y is the Xth input to this vertex

outputVertices

protected VertexIndices[] outputVertices

A representation of the vertices that this vertex is connected to (outputs duing forward pass) Specifically, if outputVertices[X].getVertexIndex() = Y, and outputVertices[X].getVertexEdgeNumber() = Z then the output of this vertex (there is only one output) is connected to the Zth input of vertex Y

inputs

protected INDArray[] inputs

epsilon

protected INDArray epsilon

outputVertex

protected boolean outputVertex

dataType

protected DataType dataType

Constructor Detail

BaseGraphVertex

protected BaseGraphVertex(ComputationGraph graph,
                          String name,
                          int vertexIndex,
                          VertexIndices[] inputVertices,
                          VertexIndices[] outputVertices,
                          DataType dataType)

Method Detail

getVertexName

public String getVertexName()

Description copied from interface: GraphVertex

Get the name/label of the GraphVertex

Specified by:

getVertexName in interface GraphVertex

getVertexIndex

public int getVertexIndex()

Description copied from interface: GraphVertex

Get the index of the GraphVertex

Specified by:

getVertexIndex in interface GraphVertex

getNumInputArrays

public int getNumInputArrays()

Description copied from interface: GraphVertex

Get the number of input arrays. For example, a Layer may have only one input array, but in general a GraphVertex may have an arbtrary (>=1) number of input arrays (for example, from multiple other layers)

Specified by:

getNumInputArrays in interface GraphVertex

getNumOutputConnections

public int getNumOutputConnections()

Description copied from interface: GraphVertex

Get the number of outgoing connections from this GraphVertex. A GraphVertex may only have a single output (for example, the activations out of a layer), but this output may be used as the input to an arbitrary number of other GraphVertex instances. This method returns the number of GraphVertex instances the output of this GraphVertex is input for.

Specified by:

getNumOutputConnections in interface GraphVertex

getInputVertices

public VertexIndices[] getInputVertices()

A representation of the vertices that are inputs to this vertex (inputs duing forward pass)
Specifically, if inputVertices[X].getVertexIndex() = Y, and inputVertices[X].getVertexEdgeNumber() = Z then the Zth output of vertex Y is the Xth input to this vertex

Specified by:

getInputVertices in interface GraphVertex

setInputVertices

public void setInputVertices(VertexIndices[] inputVertices)

Description copied from interface: GraphVertex

Sets the input vertices.

Specified by:

setInputVertices in interface GraphVertex

See Also:

GraphVertex.getInputVertices()

getOutputVertices

public VertexIndices[] getOutputVertices()

A representation of the vertices that this vertex is connected to (outputs duing forward pass) Specifically, if outputVertices[X].getVertexIndex() = Y, and outputVertices[X].getVertexEdgeNumber() = Z then the Xth output of this vertex is connected to the Zth input of vertex Y

Specified by:

getOutputVertices in interface GraphVertex

setOutputVertices

public void setOutputVertices(VertexIndices[] outputVertices)

Description copied from interface: GraphVertex

set the output vertices.

Specified by:

setOutputVertices in interface GraphVertex

See Also:

GraphVertex.getOutputVertices()

isInputVertex

public boolean isInputVertex()

Description copied from interface: GraphVertex

Whether the GraphVertex is an input vertex

Specified by:

isInputVertex in interface GraphVertex

setInput

public void setInput(int inputNumber,
                     INDArray input,
                     LayerWorkspaceMgr workspaceMgr)

Description copied from interface: GraphVertex

Set the input activations.

Specified by:

setInput in interface GraphVertex

Parameters:

inputNumber - Must be in range 0 to GraphVertex.getNumInputArrays()-1

input - The input array

setEpsilon

public void setEpsilon(INDArray epsilon)

Description copied from interface: GraphVertex

Set the errors (epsilon - aka dL/dActivation) for this GraphVertex

Specified by:

setEpsilon in interface GraphVertex

clear

public void clear()

Description copied from interface: GraphVertex

Clear the internal state (if any) of the GraphVertex. For example, any stored inputs/errors

Specified by:

clear in interface GraphVertex

canDoForward

public boolean canDoForward()

Description copied from interface: GraphVertex

Whether the GraphVertex can do forward pass. Typically, this is just whether all inputs are set.

Specified by:

canDoForward in interface GraphVertex

canDoBackward

public boolean canDoBackward()

Description copied from interface: GraphVertex

Whether the GraphVertex can do backward pass. Typically, this is just whether all errors/epsilons are set

Specified by:

canDoBackward in interface GraphVertex

getEpsilon

public INDArray getEpsilon()

Description copied from interface: GraphVertex

Get the epsilon/error (i.e., dL/dOutput) array previously set for this GraphVertex

Specified by:

getEpsilon in interface GraphVertex

toString

public abstract String toString()

Overrides:

toString in class Object

setLayerAsFrozen

public void setLayerAsFrozen()

Description copied from interface: GraphVertex

Only applies to layer vertices. Will throw exceptions on others. If applied to a layer vertex it will treat the parameters of the layer within it as constant. Activations through these will be calculated as they would as test time regardless of training mode

Specified by:

setLayerAsFrozen in interface GraphVertex

clearVertex

public void clearVertex()

Description copied from interface: GraphVertex

This method clears inpjut for this vertex

Specified by:

clearVertex in interface GraphVertex

paramTable

public Map<String,INDArray> paramTable(boolean backpropOnly)

Description copied from interface: GraphVertex

Get the parameter table for the vertex

Specified by:

paramTable in interface Trainable

Specified by:

paramTable in interface GraphVertex

Parameters:

backpropOnly - If true: exclude unsupervised training parameters

Returns:

Parameter table

numParams

public long numParams()

Specified by:

numParams in interface Trainable

Returns:

Number of parameters

getConfig

public TrainingConfig getConfig()

Specified by:

getConfig in interface Trainable

Returns:

Training configuration

params

public INDArray params()

Specified by:

params in interface Trainable

Returns:

1d parameter vector

getGradientsViewArray

public INDArray getGradientsViewArray()

Specified by:

getGradientsViewArray in interface Trainable

Returns:

1D gradients view array

updaterDivideByMinibatch

public boolean updaterDivideByMinibatch(String paramName)

Description copied from interface: Trainable

DL4J layers typically produce the sum of the gradients during the backward pass for each layer, and if required (if minibatch=true) then divide by the minibatch size.
However, there are some exceptions, such as the batch norm mean/variance estimate parameters: these "gradients" are actually not gradients, but are updates to be applied directly to the parameter vector. Put another way, most gradients should be divided by the minibatch to get the average; some "gradients" are actually final updates already, and should not be divided by the minibatch size.

Specified by:

updaterDivideByMinibatch in interface Trainable

Parameters:

paramName - Name of the parameter

Returns:

True if gradients should be divided by minibatch (most params); false otherwise (edge cases like batch norm mean/variance estimates)