hex.deeplearning

Class Neurons.Softmax

java.lang.Object

hex.deeplearning.Neurons

hex.deeplearning.Neurons.Output

hex.deeplearning.Neurons.Softmax

Enclosing class:

Neurons

public static class Neurons.Softmax
extends Neurons.Output

Output neurons for classification - Softmax

Nested Class Summary

Nested classes/interfaces inherited from class hex.deeplearning.Neurons

Neurons.ExpRectifier, Neurons.ExpRectifierDropout, Neurons.Input, Neurons.Linear, Neurons.Maxout, Neurons.MaxoutDropout, Neurons.Output, Neurons.Rectifier, Neurons.RectifierDropout, Neurons.Softmax, Neurons.Tanh, Neurons.TanhDropout

Field Summary

Fields inherited from class hex.deeplearning.Neurons

_a, _avg_a, _b, _bEA, _dropout, _e, _index, _input, _origa, _previous, _w, _wEA, params, units

Constructor Summary

Constructors

Constructor and Description
Neurons.Softmax(int units)

Method Summary

Methods

Modifier and Type	Method and Description
protected void	fprop(long seed, boolean training, int n) Forward propagation
protected void	setOutputLayerGradient(double target, int mb, int n) Part of backpropagation for classification Update every weight as follows: w += -rate * dE/dw Compute dE/dw via chain rule: dE/dw = dE/dy * dy/dnet * dnet/dw, where net = sum(xi*wi)+b and y = activation function

Methods inherited from class hex.deeplearning.Neurons.Output

bprop

Methods inherited from class hex.deeplearning.Neurons

autoEncoderGradient, bpropOutputLayer, init, momentum, momentum, rate, toString

Methods inherited from class java.lang.Object

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

Constructor Detail

Neurons.Softmax

public Neurons.Softmax(int units)

Method Detail

fprop

protected void fprop(long seed,
         boolean training,
         int n)

Description copied from class: Neurons

Forward propagation

Specified by:

fprop in class Neurons

Parameters:

seed - For seeding the RNG inside (for dropout)

training - Whether training is done or just testing (no need for dropout)

n - number of actually trained samples in this mini-batch

setOutputLayerGradient

protected void setOutputLayerGradient(double target,
                          int mb,
                          int n)

Part of backpropagation for classification Update every weight as follows: w += -rate * dE/dw Compute dE/dw via chain rule: dE/dw = dE/dy * dy/dnet * dnet/dw, where net = sum(xi*wi)+b and y = activation function

Overrides:

setOutputLayerGradient in class Neurons

Parameters:

target - actual class label (integer)