org.deeplearning4j.nn.conf.preprocessor

Class FeedForwardToRnnPreProcessor

java.lang.Object

org.deeplearning4j.nn.conf.preprocessor.FeedForwardToRnnPreProcessor

All Implemented Interfaces:

Serializable, Cloneable, InputPreProcessor

public class FeedForwardToRnnPreProcessor
extends Object
implements InputPreProcessor

A preprocessor to allow RNN and feed-forward network layers to be used together.
For example, DenseLayer -> GravesLSTM
This does two things:
(a) Reshapes activations out of FeedFoward layer (which is 2D with shape [miniBatchSize*timeSeriesLength,layerSize]) into 3d activations (with shape [miniBatchSize,layerSize,timeSeriesLength]) suitable to feed into RNN layers.
(b) Reshapes 3d epsilons (weights*deltas from RNN layer, with shape [miniBatchSize,layerSize,timeSeriesLength]) into 2d epsilons (with shape [miniBatchSize*timeSeriesLength,layerSize]) for use in feed forward layer

Author:

Alex Black

See Also:

for opposite case (i.e., GravesLSTM -> DenseLayer etc), Serialized Form

Constructor Summary

Constructors

Constructor and Description
FeedForwardToRnnPreProcessor(RNNFormat rnnDataFormat)

Method Summary

Modifier and Type	Method and Description
INDArray	backprop(INDArray output, int miniBatchSize, LayerWorkspaceMgr workspaceMgr) Reverse the preProcess during backprop.
FeedForwardToRnnPreProcessor	clone()
Pair<INDArray,MaskState>	feedForwardMaskArray(INDArray maskArray, MaskState currentMaskState, int minibatchSize)
InputType	getOutputType(InputType inputType) For a given type of input to this preprocessor, what is the type of the output?
INDArray	preProcess(INDArray input, int miniBatchSize, LayerWorkspaceMgr workspaceMgr) Pre preProcess input/activations for a multi layer network

Methods inherited from class java.lang.Object

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

Constructor Detail

FeedForwardToRnnPreProcessor

public FeedForwardToRnnPreProcessor(RNNFormat rnnDataFormat)

Method Detail

preProcess

public INDArray preProcess(INDArray input,
                           int miniBatchSize,
                           LayerWorkspaceMgr workspaceMgr)

Description copied from interface: InputPreProcessor

Pre preProcess input/activations for a multi layer network

Specified by:

preProcess in interface InputPreProcessor

Parameters:

input - the input to pre preProcess

miniBatchSize - Minibatch size

workspaceMgr - Workspace manager

Returns:

the processed input. Note that the returned array should be placed in the ArrayType.ACTIVATIONS workspace via the workspace manager

backprop

public INDArray backprop(INDArray output,
                         int miniBatchSize,
                         LayerWorkspaceMgr workspaceMgr)

Description copied from interface: InputPreProcessor

Reverse the preProcess during backprop. Process Gradient/epsilons before passing them to the layer below.

Specified by:

backprop in interface InputPreProcessor

Parameters:

output - which is a pair of the gradient and epsilon

miniBatchSize - Minibatch size

workspaceMgr - Workspace manager

Returns:

the reverse of the pre preProcess step (if any). Note that the returned array should be placed in ArrayType.ACTIVATION_GRAD workspace via the workspace manager

clone

public FeedForwardToRnnPreProcessor clone()

Specified by:

clone in interface InputPreProcessor

Overrides:

clone in class Object

getOutputType

public InputType getOutputType(InputType inputType)

Description copied from interface: InputPreProcessor

For a given type of input to this preprocessor, what is the type of the output?

Specified by:

getOutputType in interface InputPreProcessor

Parameters:

inputType - Type of input for the preprocessor

Returns:

Type of input after applying the preprocessor

feedForwardMaskArray

public Pair<INDArray,MaskState> feedForwardMaskArray(INDArray maskArray,
                                                     MaskState currentMaskState,
                                                     int minibatchSize)

Specified by:

feedForwardMaskArray in interface InputPreProcessor