org.deeplearning4j.nn.conf.preprocessor

Class CnnToFeedForwardPreProcessor

java.lang.Object

org.deeplearning4j.nn.conf.preprocessor.CnnToFeedForwardPreProcessor

All Implemented Interfaces:

Serializable, Cloneable, InputPreProcessor

public class CnnToFeedForwardPreProcessor
extends Object
implements InputPreProcessor

A preprocessor to allow CNN and standard feed-forward network layers to be used together.
For example, CNN -> Denselayer
This does two things:
(b) Reshapes 4d activations out of CNN layer, with shape [numExamples, numChannels, inputHeight, inputWidth]) (for CNN2DFormat.NCHW format activations) or shape [numExamples, inputHeight, inputWidth, numChannels] (for CNN2DFormat.NHWC) format activations) into 2d activations (with shape [numExamples, inputHeight*inputWidth*numChannels]) for use in feed forward layer. (a) Reshapes epsilons (weights*deltas) out of FeedFoward layer (which is 2D or 3D with shape [numExamples, inputHeight*inputWidth*numChannels]) into 4d epsilons (with shape [numExamples, numChannels, inputHeight, inputWidth] or [numExamples, inputHeight, inputWidth, numChannels]) suitable to feed into CNN layers.
Note: numChannels is equivalent to channels or featureMaps referenced in different literature

Author:

Adam Gibson

See Also:

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

Field Summary

Fields

Modifier and Type	Field and Description
protected CNN2DFormat	format
protected long	inputHeight
protected long	inputWidth
protected long	numChannels

Constructor Summary

Constructors

Constructor and Description
CnnToFeedForwardPreProcessor()
CnnToFeedForwardPreProcessor(long inputHeight, long inputWidth)
CnnToFeedForwardPreProcessor(long inputHeight, long inputWidth, long numChannels)
CnnToFeedForwardPreProcessor(long inputHeight, long inputWidth, long numChannels, CNN2DFormat format)

Method Summary

Modifier and Type	Method and Description
INDArray	backprop(INDArray epsilons, int miniBatchSize, LayerWorkspaceMgr workspaceMgr) Reverse the preProcess during backprop.
CnnToFeedForwardPreProcessor	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

Field Detail

inputHeight

protected long inputHeight

inputWidth

protected long inputWidth

numChannels

protected long numChannels

format

protected CNN2DFormat format

Constructor Detail

CnnToFeedForwardPreProcessor

public CnnToFeedForwardPreProcessor(long inputHeight,
                                    long inputWidth,
                                    long numChannels,
                                    CNN2DFormat format)

Parameters:

inputHeight - the columns

inputWidth - the rows

numChannels - the channels

CnnToFeedForwardPreProcessor

public CnnToFeedForwardPreProcessor(long inputHeight,
                                    long inputWidth)

CnnToFeedForwardPreProcessor

public CnnToFeedForwardPreProcessor(long inputHeight,
                                    long inputWidth,
                                    long numChannels)

CnnToFeedForwardPreProcessor

public CnnToFeedForwardPreProcessor()

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 epsilons,
                         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:

epsilons - 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 CnnToFeedForwardPreProcessor 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