hex

Class Layer

java.lang.Object

water.Iced

hex.Layer

All Implemented Interfaces:

java.lang.Cloneable, Freezable

Direct Known Subclasses:

Layer.Input, Layer.Maxout, Layer.Output, Layer.Rectifier, Layer.Tanh

public abstract class Layer
extends Iced

Neural network layer.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
class	Layer.Dropout Helper class for dropout, only to be used from within a Layer
static class	Layer.Input
static class	Layer.Linear Linear output layer is used for regression Rows with missing values in the response column will be ignored
static class	Layer.Maxout
static class	Layer.MaxoutDropout
static class	Layer.Output
static class	Layer.Rectifier
static class	Layer.RectifierDropout
static class	Layer.RectifierPrime
static class	Layer.Softmax Softmax output layer is used for classification Rows with missing values in the response column will be ignored
static class	Layer.Tanh
static class	Layer.TanhDropout
static class	Layer.TanhPrime Apply tanh to the weights' transpose.
static class	Layer.VecLinear
static class	Layer.VecsInput
static class	Layer.VecSoftmax

Field Summary

Fields

Modifier and Type	Field and Description
protected float[]	_a
protected float[]	_b
protected float[]	_bm
protected float[]	_e
protected Layer	_previous
protected float[]	_w
protected float[]	_wm
static DocGen.FieldDoc[]	DOC_FIELDS
static float	missing_float_value
static int	missing_int_value We need a way to encode a missing value in the neural net forward/back-propagation scheme.
NeuralNet	params
int	units

Constructor Summary

Constructors

Constructor and Description
Layer()

Method Summary

Methods

Modifier and Type	Method and Description
protected abstract void	bprop()
Layer	clone()
void	close()
protected abstract void	fprop(long seed, boolean training)
void	init(Layer[] ls, int index, boolean weights)
void	init(Layer[] ls, int index, NeuralNet p)
float	momentum(long n)
float	rate(long n)
static void	shareWeights(Layer[] src, Layer[] dst)
static void	shareWeights(Layer src, Layer dst)
AutoBuffer	writeJSON(AutoBuffer bb)

Methods inherited from class water.Iced

frozenType, init, newInstance, read, toDocField, write, writeJSONFields

Methods inherited from class java.lang.Object

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

Field Detail

DOC_FIELDS

public static DocGen.FieldDoc[] DOC_FIELDS

units

@Request.API(help="Number of neurons")
public int units

params

public NeuralNet params

_a

protected transient float[] _a

_e

protected transient float[] _e

_w

protected transient float[] _w

_wm

protected transient float[] _wm

_b

protected transient float[] _b

_bm

protected transient float[] _bm

_previous

protected transient Layer _previous

missing_int_value

public static final int missing_int_value

We need a way to encode a missing value in the neural net forward/back-propagation scheme. For simplicity and performance, we simply use the largest values to encode a missing value. If we run into exactly one of those values with regular neural net updates, then we're very likely also running into overflow problems, which will trigger a NaN somewhere, which will be caught and lead to automatic job cancellation.

See Also:

Constant Field Values

missing_float_value

public static final float missing_float_value

See Also:

Constant Field Values

Constructor Detail

Layer

public Layer()

Method Detail

init

public final void init(Layer[] ls,
        int index,
        NeuralNet p)

init

public void init(Layer[] ls,
        int index,
        boolean weights)

close

public void close()

fprop

protected abstract void fprop(long seed,
         boolean training)

bprop

protected abstract void bprop()

rate

public float rate(long n)

momentum

public float momentum(long n)

clone

public Layer clone()

Overrides:

clone in class Iced

shareWeights

public static void shareWeights(Layer src,
                Layer dst)

shareWeights

public static void shareWeights(Layer[] src,
                Layer[] dst)

writeJSON

public AutoBuffer writeJSON(AutoBuffer bb)

Overrides:

writeJSON in class Iced