deepnetts.net.layers

Class ConvolutionalLayer

java.lang.Object

deepnetts.net.layers.AbstractLayer

deepnetts.net.layers.ConvolutionalLayer

All Implemented Interfaces:

Layer, Serializable

public final class ConvolutionalLayer
extends AbstractLayer

Convolutional layer performs image convolution operation on outputs of a previous layer using filters. This filtering operation is similar like applying image filters in photoshop, but this filters can also be trained to learn image features of interest. Layer include parameters: filter's width, heigh Number of filters / depth Step when applying filters : stride Padding, which is an image border to keep the size of image and avoid information loss padding Stride defaults to 1

See Also:

Serialized Form

Constructor Summary

Constructors

Constructor and Description
ConvolutionalLayer(int channels, Filter filter)
ConvolutionalLayer(int channels, Filter filter, ActivationType activationType)
ConvolutionalLayer(int channels, int filterWidth, int filterHeight) Create a new instance of convolutional layer with specified number of channels filter size, default padding (filter-1)/2, and default stride stride value 1, and specified number of channels.
ConvolutionalLayer(int filterWidth, int filterHeight, int channels, int stride, ActivationType activationType)

Method Summary

Modifier and Type	Method and Description
void	applyWeightChanges() Apply weight changes calculated in backward pass
void	backward() Backward pass for convolutional layer tweaks the weights in filters.
void	forward() Forward pass for convolutional layer.
Tensor	getFilterDeltaWeights()
int	getFilterDepth()
int	getFilterHeight()
Tensor	getFilters()
int	getFilterWidth()
float	getL1WeightSum()
float	getL2WeightSum()
int	getStride()
void	init() Init dimensions, create matrices, filters, weights, biases and all internal structures etc.
void	initTransientFields()
void	setFilters(String filtersStr)
void	setFilters(Tensor filters)
String	toString()

Methods inherited from class deepnetts.net.layers.AbstractLayer

getActivation, getActivationType, getBatchSize, getBiases, getDeltaBiases, getDeltas, getDeltaWeights, getDepth, getGradients, getHeight, getL1Regularization, getL2Regularization, getLearningRate, getMomentum, getNextLayer, getOptimizer, getOptimizerType, getOutputs, getPrevDeltaBiases, getPrevDeltaWeights, getPrevlayer, getWeights, getWidth, isBatchMode, isTrainable, setBatchMode, setBatchSize, setBiases, setDeltas, setL1Regularization, setL2Regularization, setLearningRate, setMomentum, setNextlayer, setOptimizerType, setOutputs, setPrevDeltaWeights, setPrevLayer, setTrainable, setWeights, setWeights

Methods inherited from class java.lang.Object

equals, getClass, hashCode, notify, notifyAll, wait, wait, wait

Constructor Detail

ConvolutionalLayer

public ConvolutionalLayer(int channels,
                          int filterWidth,
                          int filterHeight)

Create a new instance of convolutional layer with specified number of channels filter size, default padding (filter-1)/2, and default stride stride value 1, and specified number of channels. Uses Linear activation function by default.

Parameters:

channels - number of channels which corresponds to the number of image pixel feature that you want to learn and detect

filterWidth -

filterHeight -

ConvolutionalLayer

public ConvolutionalLayer(int channels,
                          Filter filter)

ConvolutionalLayer

public ConvolutionalLayer(int channels,
                          Filter filter,
                          ActivationType activationType)

ConvolutionalLayer

public ConvolutionalLayer(int filterWidth,
                          int filterHeight,
                          int channels,
                          int stride,
                          ActivationType activationType)

Method Detail

init

public void init()

Init dimensions, create matrices, filters, weights, biases and all internal structures etc. Assumes that prevLayer is set in network builder

Specified by:

init in class AbstractLayer

initTransientFields

public void initTransientFields()

Overrides:

initTransientFields in class AbstractLayer

forward

public void forward()

Forward pass for convolutional layer. Performs convolution operation on output from previous layer using filters in this layer, on all channels. Each channel from prev layer has its own filter (3D filter), and every channel in this layer has its 3D filter used to scan all channels in prev layer. Previous layers can be: Input, MaxPooling or Convolutional. For more about convolution see http://www.songho.ca/dsp/convolution/convolution.html

Specified by:

forward in interface Layer

Specified by:

forward in class AbstractLayer

backward

public void backward()

Backward pass for convolutional layer tweaks the weights in filters. Next layer can be: FC, MaxPooling, Conv, (output same as FC), 1D or 3D Prev layer can: Input, pool, conv, all 2D or 3D - all can be as generalized 3D U 2 koraka: 1. povuci delte iz sledeceg lejera, i izracunaj tezinsku sumu delta za sve neurone/outpute u ovom sloju 2. izracunaj promene tezina za sve veze iz prethodnog lejera za svaki neuron/output u ovom sloju

Specified by:

backward in interface Layer

Specified by:

backward in class AbstractLayer

applyWeightChanges

public void applyWeightChanges()

Apply weight changes calculated in backward pass

Specified by:

applyWeightChanges in class AbstractLayer

getFilters

public Tensor getFilters()

setFilters

public void setFilters(Tensor filters)

setFilters

public void setFilters(String filtersStr)

getFilterWidth

public int getFilterWidth()

getFilterHeight

public int getFilterHeight()

getFilterDepth

public int getFilterDepth()

getStride

public int getStride()

getFilterDeltaWeights

public Tensor getFilterDeltaWeights()

getL1WeightSum

public float getL1WeightSum()

Overrides:

getL1WeightSum in class AbstractLayer

getL2WeightSum

public float getL2WeightSum()

Overrides:

getL2WeightSum in class AbstractLayer

toString

public String toString()

Overrides:

toString in class Object