Package org.deeplearning4j.nn.modelimport.keras

Class KerasLayer

java.lang.Object

org.deeplearning4j.nn.modelimport.keras.KerasLayer

Direct Known Subclasses:

Keras2DEmbedding, KerasActivation, KerasAlphaDropout, KerasBatchNormalization, KerasBidirectional, KerasConvolution, KerasCropping1D, KerasCropping2D, KerasCropping3D, KerasDense, KerasDropout, KerasELU, KerasEmbedding, KerasFlatten, KerasGaussianDropout, KerasGaussianNoise, KerasGlobalPooling, KerasInput, KerasLambda, KerasLeakyReLU, KerasLoss, KerasLRN, KerasLSTM, KerasMasking, KerasMerge, KerasPermute, KerasPoolHelper, KerasPooling1D, KerasPooling2D, KerasPooling3D, KerasPReLU, KerasReLU, KerasRepeatVector, KerasReshape, KerasSimpleRnn, KerasSoftmax, KerasSpaceToDepth, KerasSpatialDropout, KerasTFOpLayer, KerasThresholdedReLU, KerasUpsampling1D, KerasUpsampling2D, KerasUpsampling3D, KerasZeroPadding1D, KerasZeroPadding2D, KerasZeroPadding3D

public class KerasLayer
extends Object

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static class	KerasLayer.DimOrder

Field Summary

Fields

Modifier and Type	Field	Description
protected String	className
protected KerasLayerConfiguration	conf
protected KerasLayer.DimOrder	dimOrder
protected double	dropout
protected List<String>	inboundLayerNames
protected int[]	inputShape
protected Integer	kerasMajorVersion
protected Layer	layer
protected String	layerName
protected Map<String,Object>	originalLayerConfig
protected List<String>	outboundLayerNames
protected GraphVertex	vertex
protected double	weightL1Regularization
protected double	weightL2Regularization
protected Map<String,INDArray>	weights

Constructor Summary

Constructors

Modifier	Constructor	Description
protected	KerasLayer()	Default constructor.
protected	KerasLayer(Integer kerasVersion)	Constructor with Keras version only.
protected	KerasLayer(Map<String,Object> layerConfig)	Constructor.
protected	KerasLayer(Map<String,Object> layerConfig, boolean enforceTrainingConfig)	Constructor.

Method Summary

Modifier and Type	Method	Description
static void	clearCustomLayers()	Clear all custom layers
static void	clearLambdaLayers()	Clear all lambda layers
void	copyWeightsToLayer(Layer layer)	Copy Keras layer weights to DL4J Layer.
String	getClassName()	Get Keras layer class name.
KerasLayer.DimOrder	getDimOrder()	Get Keras layer backend dimension order.
List<String>	getInboundLayerNames()	Get list of inbound layers.
InputPreProcessor	getInputPreprocessor(InputType... inputType)	Gets appropriate DL4J InputPreProcessor for given InputTypes.
int[]	getInputShape()	Get layer input shape.
Integer	getKerasMajorVersion()	Get Keras major version of this layer.
Layer	getLayer()	Gets corresponding DL4J Layer, if any.
String	getLayerName()	Get Keras layer name.
protected long	getNInFromConfig(Map<String,? extends KerasLayer> previousLayers)	Some DL4J layers need explicit specification of number of inputs, which Keras does infer.
int	getNumParams()	Returns number of trainable parameters in layer.
InputType	getOutputType(InputType... inputType)	Get layer output type.
GraphVertex	getVertex()	Gets corresponding DL4J Vertex, if any.
Map<String,INDArray>	getWeights()
boolean	isInputPreProcessor()	Whether this Keras layer maps to a DL4J InputPreProcessor.
boolean	isLayer()	Whether this Keras layer maps to a DL4J Layer.
boolean	isValidInboundLayer()	Indicates whether this layer a valid inbound layer.
boolean	isVertex()	Whether this Keras layer maps to a DL4J Vertex.
static void	registerCustomLayer(String layerName, Class<? extends KerasLayer> configClass)	Register a custom layer
static void	registerLambdaLayer(String lambdaLayerName, SameDiffLambdaLayer sameDiffLambdaLayer)	Register a lambda layer
void	setInboundLayerNames(List<String> inboundLayerNames)	Set list of inbound layers.
void	setLayer(Layer layer)
void	setWeights(Map<String,INDArray> weights)	Set weights for Keras layer.
boolean	usesRegularization()	Indicates whether layer uses regularization.

Methods inherited from class java.lang.Object

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

Field Detail

className

protected String className

layerName

protected String layerName

inputShape

protected int[] inputShape

dimOrder

protected KerasLayer.DimOrder dimOrder

inboundLayerNames

protected List<String> inboundLayerNames

outboundLayerNames

protected List<String> outboundLayerNames

layer

protected Layer layer

vertex

protected GraphVertex vertex

weights

protected Map<String,INDArray> weights

weightL1Regularization

protected double weightL1Regularization

weightL2Regularization

protected double weightL2Regularization

dropout

protected double dropout

kerasMajorVersion

protected Integer kerasMajorVersion

conf

protected KerasLayerConfiguration conf

originalLayerConfig

protected Map<String,Object> originalLayerConfig

Constructor Detail

KerasLayer

protected KerasLayer(Integer kerasVersion)
              throws UnsupportedKerasConfigurationException

Constructor with Keras version only.

Parameters:

kerasVersion - major Keras version (1 or 2)

Throws:

UnsupportedKerasConfigurationException - Unsupported Keras configuration

KerasLayer

protected KerasLayer()
              throws UnsupportedKerasConfigurationException

Default constructor.

Throws:

UnsupportedKerasConfigurationException - Unsupported Keras configuration

KerasLayer

protected KerasLayer(Map<String,Object> layerConfig)
              throws InvalidKerasConfigurationException,
                     UnsupportedKerasConfigurationException

Constructor.

Parameters:

layerConfig - dictionary containing Keras layer configuration

Throws:

InvalidKerasConfigurationException

UnsupportedKerasConfigurationException

KerasLayer

protected KerasLayer(Map<String,Object> layerConfig,
                     boolean enforceTrainingConfig)
              throws InvalidKerasConfigurationException,
                     UnsupportedKerasConfigurationException

Constructor. "enforceTrainingConfig" parameter controls whether layer is built for training. This controls behavior of certain exceptions. In training mode, passing an unsupported regularizer will generate an error. In non-training mode, it generates only a warning.

Parameters:

layerConfig - dictionary containing Keras layer configuration

enforceTrainingConfig - whether layer should be built for training (controls certain exceptions)

Throws:

InvalidKerasConfigurationException

UnsupportedKerasConfigurationException

Method Detail

registerLambdaLayer

public static void registerLambdaLayer(String lambdaLayerName,
                                       SameDiffLambdaLayer sameDiffLambdaLayer)

Register a lambda layer

Parameters:

lambdaLayerName - name of the lambda layer in the serialized Keras model

sameDiffLambdaLayer - SameDiffLambdaLayer instance to map to Keras Lambda layer

clearLambdaLayers

public static void clearLambdaLayers()

Clear all lambda layers

registerCustomLayer

public static void registerCustomLayer(String layerName,
                                       Class<? extends KerasLayer> configClass)

Register a custom layer

Parameters:

layerName - name of custom layer class

configClass - class of custom layer

clearCustomLayers

public static void clearCustomLayers()

Clear all custom layers

getKerasMajorVersion

public Integer getKerasMajorVersion()

Get Keras major version of this layer.

Returns:

Keras version as integer

getClassName

public String getClassName()

Get Keras layer class name.

Returns:

Keras layer class name

getLayerName

public String getLayerName()

Get Keras layer name.

Returns:

layer name

getInputShape

public int[] getInputShape()

Get layer input shape.

Returns:

input shape

getDimOrder

public KerasLayer.DimOrder getDimOrder()

Get Keras layer backend dimension order.

Returns:

Keras layer (backend) dimension order

getInboundLayerNames

public List<String> getInboundLayerNames()

Get list of inbound layers.

Returns:

list of inbound layer names

setInboundLayerNames

public void setInboundLayerNames(List<String> inboundLayerNames)

Set list of inbound layers.

Parameters:

inboundLayerNames - list of inbound layer names

getNumParams

public int getNumParams()

Returns number of trainable parameters in layer.

Returns:

number of trainable parameters

usesRegularization

public boolean usesRegularization()

Indicates whether layer uses regularization.

Returns:

boolean

setWeights

public void setWeights(Map<String,INDArray> weights)
                throws InvalidKerasConfigurationException

Set weights for Keras layer.

Parameters:

weights - Map of named NDArrays

Throws:

InvalidKerasConfigurationException

getWeights

public Map<String,INDArray> getWeights()

copyWeightsToLayer

public void copyWeightsToLayer(Layer layer)
                        throws InvalidKerasConfigurationException

Copy Keras layer weights to DL4J Layer.

Parameters:

layer - DL4J layer

Throws:

InvalidKerasConfigurationException - Invalid Keras configuration

isLayer

public boolean isLayer()

Whether this Keras layer maps to a DL4J Layer.

Returns:

true or false

getLayer

public Layer getLayer()

Gets corresponding DL4J Layer, if any.

Returns:

DL4J Layer

See Also:

Layer

setLayer

public void setLayer(Layer layer)

isVertex

public boolean isVertex()

Whether this Keras layer maps to a DL4J Vertex.

Returns:

true or false

getVertex

public GraphVertex getVertex()

Gets corresponding DL4J Vertex, if any.

Returns:

DL4J Vertex

See Also:

GraphVertex

isInputPreProcessor

public boolean isInputPreProcessor()

Whether this Keras layer maps to a DL4J InputPreProcessor.

Returns:

true or false

getNInFromConfig

protected long getNInFromConfig(Map<String,? extends KerasLayer> previousLayers)
                         throws UnsupportedKerasConfigurationException

Some DL4J layers need explicit specification of number of inputs, which Keras does infer. This method searches through previous layers until a FeedForwardLayer is found. These layers have nOut values that subsequently correspond to the nIn value of this layer.

Parameters:

previousLayers -

Returns:

Throws:

UnsupportedKerasConfigurationException

getInputPreprocessor

public InputPreProcessor getInputPreprocessor(InputType... inputType)
                                       throws InvalidKerasConfigurationException

Gets appropriate DL4J InputPreProcessor for given InputTypes.

Parameters:

inputType - Array of InputTypes

Returns:

DL4J InputPreProcessor

Throws:

InvalidKerasConfigurationException - Invalid Keras configuration

See Also:

InputPreProcessor

getOutputType

public InputType getOutputType(InputType... inputType)
                        throws InvalidKerasConfigurationException,
                               UnsupportedKerasConfigurationException

Get layer output type.

Parameters:

inputType - Array of InputTypes

Returns:

output type as InputType

Throws:

InvalidKerasConfigurationException - Invalid Keras configuration

UnsupportedKerasConfigurationException

isValidInboundLayer

public boolean isValidInboundLayer()
                            throws InvalidKerasConfigurationException

Indicates whether this layer a valid inbound layer. Currently, only (known) DL4J Layers and inputs are valid inbound layers. "Preprocessor" layers (reshaping, merging, etc.) are replaced by their own inbound layers.

Returns:

boolean indicating whether layer is valid inbound layer

Throws:

InvalidKerasConfigurationException

See Also:

Layer