|
||||||||||
PREV CLASS NEXT CLASS | FRAMES NO FRAMES | |||||||||
SUMMARY: NESTED | FIELD | CONSTR | METHOD | DETAIL: FIELD | CONSTR | METHOD |
java.lang.Objectorg.joone.engine.Layer
org.joone.engine.SimpleLayer
org.joone.engine.ContextLayer
public class ContextLayer
The context layer is similar to the linear layer except that it has an auto-recurrent connection between its output and input. MG: Added ability to save and restore the initial state. Also changed so that initial state is used when first pattern is requested. TODO: Context neurons are usually used in the input layer of Jordan networks together with non-recursive neurons. So there are two questions: - how can this layer be combined with an non-recursive layer? - is there already a mechanism how to feed the network output back in the neurons of this layer?
Field Summary | |
---|---|
protected Matrix |
initialState
The initial state |
Fields inherited from class org.joone.engine.Layer |
---|
bias, gradientInps, gradientOuts, inps, inputPatternListeners, learnable, learning, m_batch, monitor, myLearner, outputPatternListeners, outs, running, step, STOP_FLAG |
Constructor Summary | |
---|---|
ContextLayer()
|
|
ContextLayer(java.lang.String name)
|
Method Summary | |
---|---|
void |
backward(double[] pattern)
Reverse transfer function of the component. |
void |
forward(double[] pattern)
Transfer function to recall a result on a trained net |
double |
getBeta()
Getter for property beta. |
double |
getDefaultState()
Return the default state of a node in this layer, such as 0 for a tanh or 0.5 for a sigmoid layer |
double |
getDerivative(int i)
Similar to the backward message and used by RTRL |
Matrix |
getInitialState()
Get the initial state, also use this to change, in place, the initial state |
double |
getMaximumState()
Return maximum value of a node in this layer |
double |
getMinimumState()
Return minimum value of a node in this layer |
double |
getTimeConstant()
Getter for property timeConstant. |
java.util.Collection |
Inspections()
It doesn't make sense to return biases for this layer |
void |
resetInitialState()
Reset the initial state of this context layer. |
void |
setBeta(double beta)
Setter for property beta. |
protected void |
setDimensions()
Use this to also allocate space for the initial state vector |
void |
setTimeConstant(double timeConstant)
Setter for property timeConstant. |
Methods inherited from class org.joone.engine.SimpleLayer |
---|
getLearningRate, getLrate, getMomentum, setLrate, setMomentum, setMonitor |
Methods inherited from class java.lang.Object |
---|
clone, equals, getClass, hashCode, notify, notifyAll, wait, wait, wait |
Field Detail |
---|
protected Matrix initialState
Constructor Detail |
---|
public ContextLayer()
public ContextLayer(java.lang.String name)
Method Detail |
---|
public void backward(double[] pattern)
Layer
backward
in class SimpleLayer
pattern
- input pattern on which to apply the transfer functionpublic double getDerivative(int i)
getDerivative
in class Layer
public void forward(double[] pattern)
Layer
forward
in class Layer
pattern
- input pattern to which to apply the rtransfer functionpublic double getBeta()
public void setBeta(double beta)
beta
- New value of property beta.public double getTimeConstant()
public void setTimeConstant(double timeConstant)
timeConstant
- New value of property timeConstant.public java.util.Collection Inspections()
Inspections
in interface Inspectable
Inspections
in class Layer
org.joone.Inspection
protected void setDimensions()
setDimensions
in class SimpleLayer
public Matrix getInitialState()
public void resetInitialState()
public double getDefaultState()
Layer
getDefaultState
in class Layer
public double getMinimumState()
Layer
getMinimumState
in class Layer
public double getMaximumState()
Layer
getMaximumState
in class Layer
|
||||||||||
PREV CLASS NEXT CLASS | FRAMES NO FRAMES | |||||||||
SUMMARY: NESTED | FIELD | CONSTR | METHOD | DETAIL: FIELD | CONSTR | METHOD |