Package org.nd4j.autodiff.functions

Class DifferentialFunction

java.lang.Object

org.nd4j.autodiff.functions.DifferentialFunction

Direct Known Subclasses:

BaseOp, DynamicCustomOp

public abstract class DifferentialFunction
extends Object

Field Summary

Fields

Modifier and Type	Field	Description
protected int[]	dimensions
protected Object[]	extraArgs
protected boolean	inPlace
protected String	ownName
protected boolean	ownNameSetWithDefault
protected SameDiff	sameDiff
protected INDArray	scalarValue

Constructor Summary

Constructors

Constructor	Description
DifferentialFunction()
DifferentialFunction(boolean sameDiff)
DifferentialFunction(SameDiff sameDiff, boolean inPlace, Object[] extraArgs)
DifferentialFunction(SameDiff sameDiff, boolean inPlace, SDVariable[] args)	Add the various arguments for this function
DifferentialFunction(SameDiff sameDiff, Object[] extraArgs)
DifferentialFunction(SameDiff sameDiff, Onnx.NodeProto node, Map<String,Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph)	Initialize the function from the given Onnx.NodeProto
DifferentialFunction(SameDiff sameDiff, SDVariable[] args)
DifferentialFunction(SameDiff sameDiff, NodeDef nodeDef, Map<String,AttrValue> attributesForNode, GraphDef graph)	Initialize the function from the given NodeDef

Method Summary

Modifier and Type	Method	Description
SDVariable	arg()	Return the first argument
SDVariable	arg(int num)	Return the specified argument for this function
String[]	argNames()
SDVariable[]	args()	Return the arguments for a given function
Map<String,Map<String,AttributeAdapter>>	attributeAdaptersForFunction()	Returns the AttributeAdapter s for each of the possible ops for import (typically tensorflow and onnx) See AttributeAdapter for more information on what the adapter does.
List<DataType>	calculateOutputDataTypes(List<DataType> dataTypes)	Calculate the data types for the output arrays.
List<LongShapeDescriptor>	calculateOutputShape()	Calculate the output shape for this op
List<LongShapeDescriptor>	calculateOutputShape(OpContext oc)
abstract void	clearArrays()	Clear the input and output INDArrays, if any are set
String	configFieldName()	Returns the name of the field to be used for looking up field names.
void	configureWithSameDiff(SameDiff sameDiff)
List<SDVariable>	diff(List<SDVariable> i_v1)	Perform automatic differentiation wrt the input variables
abstract List<SDVariable>	doDiff(List<SDVariable> f1)	The actual implementation for automatic differentiation.
DifferentialFunction	dup()	Duplicate this function
boolean	equals(Object o)
protected Boolean	getBooleanFromProperty(String propertyName, Map<String,Object> properties)
protected Double	getDoubleValueFromProperty(String propertyName, Map<String,Object> properties)
INDArray	getInputArgument(int index)
protected Integer	getIntValueFromProperty(String propertyName, Map<String,Object> properties)
protected Long	getLongValueFromProperty(String propertyName, Map<String,Object> properties)
int	getNumOutputs()
protected String	getStringFromProperty(String propertyName, Map<String,Object> properties)
Object	getValue(Field property)	Get the value for a given property for this function
int	hashCode()
abstract void	initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String,Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph)	Iniitialize the function from the given Onnx.NodeProto
abstract void	initFromTensorFlow(NodeDef nodeDef, SameDiff initWith, Map<String,AttrValue> attributesForNode, GraphDef graph)	Initialize the function from the given NodeDef
boolean	isConfigProperties()	Returns true if the fields for this class should be looked up from a configuration class.
SDVariable	larg()	The left argument for this function
Map<String,Map<String,PropertyMapping>>	mappingsForFunction()	Returns the mappings for a given function ( for tensorflow and onnx import mapping properties of this function).
abstract String	onnxName()	The opName of this function in onnx
String[]	onnxNames()	The opName of this function in onnx
String	opName()	The name of the op
int	opNum()	The number of the op (mainly for old legacy XYZ ops like Op)
Op.Type	opType()	The type of the op
List<SDVariable>	outputs()
SDVariable	outputVariable()
SDVariable[]	outputVariables()	Return the output variables for this differential function.
abstract SDVariable[]	outputVariables(String baseName)	Return the output functions for this differential function.
String[]	outputVariablesNames()
Map<String,Object>	propertiesForFunction()	Returns the properties for a given function
SDVariable	rarg()	The right argument for this function.
void	replaceArg(int i, SDVariable newArg)	Replace argument at the specified index
protected void	setInstanceId()
void	setPropertiesForFunction(Map<String,Object> properties)
void	setValueFor(Field target, Object value)	Set the value for this function.
abstract String	tensorflowName()	The opName of this function tensorflow
String[]	tensorflowNames()	The opName of this function tensorflow

Methods inherited from class java.lang.Object

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

Field Detail

sameDiff

protected SameDiff sameDiff

inPlace

protected boolean inPlace

scalarValue

protected INDArray scalarValue

dimensions

protected int[] dimensions

extraArgs

protected Object[] extraArgs

ownName

protected String ownName

ownNameSetWithDefault

protected boolean ownNameSetWithDefault

Constructor Detail

DifferentialFunction

public DifferentialFunction()

DifferentialFunction

public DifferentialFunction(boolean sameDiff)

DifferentialFunction

public DifferentialFunction(SameDiff sameDiff,
                            NodeDef nodeDef,
                            Map<String,AttrValue> attributesForNode,
                            GraphDef graph)

Initialize the function from the given NodeDef

Parameters:

nodeDef -

DifferentialFunction

public DifferentialFunction(SameDiff sameDiff,
                            Onnx.NodeProto node,
                            Map<String,Onnx.AttributeProto> attributesForNode,
                            Onnx.GraphProto graph)

Initialize the function from the given Onnx.NodeProto

Parameters:

node -

DifferentialFunction

public DifferentialFunction(SameDiff sameDiff,
                            boolean inPlace,
                            Object[] extraArgs)

Parameters:

sameDiff -

extraArgs -

DifferentialFunction

public DifferentialFunction(SameDiff sameDiff,
                            Object[] extraArgs)

Parameters:

sameDiff -

extraArgs -

DifferentialFunction

public DifferentialFunction(SameDiff sameDiff,
                            SDVariable[] args)

Parameters:

sameDiff -

args -

DifferentialFunction

public DifferentialFunction(SameDiff sameDiff,
                            boolean inPlace,
                            SDVariable[] args)

Add the various arguments for this function

Parameters:

sameDiff -

inPlace -

args -

Method Detail

attributeAdaptersForFunction

public Map<String,Map<String,AttributeAdapter>> attributeAdaptersForFunction()

Returns the AttributeAdapter s for each of the possible ops for import (typically tensorflow and onnx) See AttributeAdapter for more information on what the adapter does. Similar to mappingsForFunction(), the returned map contains a AttributeAdapter for each field name when one is present. (It is optional for one to exist)_

Returns:

mappingsForFunction

public Map<String,Map<String,PropertyMapping>> mappingsForFunction()

Returns the mappings for a given function ( for tensorflow and onnx import mapping properties of this function). The mapping is indexed by field name. If the function has no properties, this returned map will be empty. Note that some functions have multiple names. This function returns a map indexed by each alias it has for a given name. These names include both onnx and tensorflow names (which might be 1 or more)

Returns:

propertiesForFunction

public Map<String,Object> propertiesForFunction()

Returns the properties for a given function

Returns:

configureWithSameDiff

public void configureWithSameDiff(SameDiff sameDiff)

setPropertiesForFunction

public void setPropertiesForFunction(Map<String,Object> properties)

getBooleanFromProperty

protected Boolean getBooleanFromProperty(String propertyName,
                                         Map<String,Object> properties)

getStringFromProperty

protected String getStringFromProperty(String propertyName,
                                       Map<String,Object> properties)

getIntValueFromProperty

protected Integer getIntValueFromProperty(String propertyName,
                                          Map<String,Object> properties)

getLongValueFromProperty

protected Long getLongValueFromProperty(String propertyName,
                                        Map<String,Object> properties)

getDoubleValueFromProperty

protected Double getDoubleValueFromProperty(String propertyName,
                                            Map<String,Object> properties)

getValue

public Object getValue(Field property)

Get the value for a given property for this function

Parameters:

property - the property to get

Returns:

the value for the function if it exists

setValueFor

public void setValueFor(Field target,
                        Object value)

Set the value for this function. Note that if value is null an ND4JIllegalStateException will be thrown.

Parameters:

target - the target field

value - the value to set

isConfigProperties

public boolean isConfigProperties()

Returns true if the fields for this class should be looked up from a configuration class.

Returns:

configFieldName

public String configFieldName()

Returns the name of the field to be used for looking up field names. This should be used in conjunction with isConfigProperties() to facilitate mapping fields for model import.

Returns:

replaceArg

public void replaceArg(int i,
                       SDVariable newArg)

Replace argument at the specified index

Parameters:

i - the index

newArg - the new argument

outputVariables

public SDVariable[] outputVariables()

Return the output variables for this differential function. Note that this op *may* dynamically generate variable outputs.

Returns:

outputVariable

public SDVariable outputVariable()

Returns:

The output variable, or the first output variable, if multiple outputs exist

outputs

public List<SDVariable> outputs()

outputVariablesNames

public String[] outputVariablesNames()

outputVariables

public abstract SDVariable[] outputVariables(String baseName)

Return the output functions for this differential function.

Returns:

doDiff

public abstract List<SDVariable> doDiff(List<SDVariable> f1)

The actual implementation for automatic differentiation.

Parameters:

f1 -

Returns:

args

public SDVariable[] args()

Return the arguments for a given function

Returns:

the arguments for a given function

arg

public SDVariable arg(int num)

Return the specified argument for this function

Parameters:

num - Number of the argument. Must be in range 0 to numArgs - 1 inclusive

Returns:

Specified argument

argNames

public String[] argNames()

arg

public SDVariable arg()

Return the first argument

Returns:

diff

public List<SDVariable> diff(List<SDVariable> i_v1)

Perform automatic differentiation wrt the input variables

Parameters:

i_v1 - the input variables

Returns:

the differentiated output wrt each input variable

setInstanceId

protected void setInstanceId()

opName

public String opName()

The name of the op

Returns:

opType

public Op.Type opType()

The type of the op

Returns:

opNum

public int opNum()

The number of the op (mainly for old legacy XYZ ops like Op)

Returns:

getInputArgument

public INDArray getInputArgument(int index)

initFromTensorFlow

public abstract void initFromTensorFlow(NodeDef nodeDef,
                                        SameDiff initWith,
                                        Map<String,AttrValue> attributesForNode,
                                        GraphDef graph)

Initialize the function from the given NodeDef

Parameters:

nodeDef -

initWith -

attributesForNode -

graph -

initFromOnnx

public abstract void initFromOnnx(Onnx.NodeProto node,
                                  SameDiff initWith,
                                  Map<String,Onnx.AttributeProto> attributesForNode,
                                  Onnx.GraphProto graph)

Iniitialize the function from the given Onnx.NodeProto

Parameters:

node -

initWith -

attributesForNode -

graph -

larg

public SDVariable larg()

The left argument for this function

Returns:

rarg

public SDVariable rarg()

The right argument for this function. Note that this assumes that there are 2 args for this function, if 2 are not set, it throws an ND4JIllegalStateException

Returns:

dup

public DifferentialFunction dup()

Duplicate this function

Returns:

calculateOutputShape

public List<LongShapeDescriptor> calculateOutputShape()

Calculate the output shape for this op

Returns:

List of output shape descriptors

calculateOutputShape

public List<LongShapeDescriptor> calculateOutputShape(OpContext oc)

calculateOutputDataTypes

public List<DataType> calculateOutputDataTypes(List<DataType> dataTypes)

Calculate the data types for the output arrays. Though datatypes can also be inferred from calculateOutputShape(), this method differs in that it does not require the input arrays to be populated. This is important as it allows us to do greedy datatype inference for the entire net - even if arrays are not available.

Parameters:

dataTypes - The data types of the inputs

Returns:

The data types of the outputs

equals

public boolean equals(Object o)

Overrides:

equals in class Object

hashCode

public int hashCode()

Overrides:

hashCode in class Object

onnxNames

public String[] onnxNames()

The opName of this function in onnx

Returns:

tensorflowNames

public String[] tensorflowNames()

The opName of this function tensorflow

Returns:

onnxName

public abstract String onnxName()

The opName of this function in onnx

Returns:

tensorflowName

public abstract String tensorflowName()

The opName of this function tensorflow

Returns:

getNumOutputs

public int getNumOutputs()

clearArrays

public abstract void clearArrays()

Clear the input and output INDArrays, if any are set