SAETrainer (boone)

java.lang.Object
- boone.Trainer
- - boone.training.SAETrainer

All Implemented Interfaces:

Storable, java.lang.Cloneable
```
public class SAETrainer
extends Trainer
```
The SAE trainer can transform a regular feed-forward network into a stacked auto-encoder. The SAE trainer provides the infrastructure to enable layer-wise training internally uses existing training algorithms. Each layer is trained to be auto-encoding. The training of an individual layer is performed in three phases. The extraction phase, training phase and the reset phase.

Author:

Team stacked auto-encoder

Field Summary

Fields
Modifier and Type	Field and Description
`protected NeuronList`	`hiddenLayer` Hidden layer
`protected NeuralNet`	`partNet` Layer that is trained - extracted net
`protected Trainer`	`partTrainer` The trainer internally used to train each layer
`protected boolean`	`print` Flag indicating how much training information is printed to console
`protected int`	`steps` The number of steps used for training
`protected boolean`	`wasInputLayer` Flag indicating if the current input layer is the input layer of whole net
`protected boolean`	`wasOutputLayer` Flag indicating if the current hidden layer is the output layer of whole net

Fields inherited from class boone.Trainer
currentEpoch, epochs, learnRate, net, shuffle, stepMode, testData, trainingData, trainingSignalGenerator

Constructor Summary

Constructors
Constructor and Description
`SAETrainer()` Create a new SAE trainer using a `SquareError`.
`SAETrainer(Trainer trainer, int steps)` Create a new SAE trainer using a `SquareError`, a given trainer and a given number of steps

Method Summary

All Methods Instance Methods Concrete Methods
Modifier and Type	Method and Description
`protected double`	`calcAdaptation(BrainPart part)` Not used.
`protected void`	`endBatch()` Updates weights and biases using the batch gradient information.
`protected void`	`endTrain()` Do some house keeping after all epochs have been trained.
`protected void`	`extractHiddenLayer(NeuralNet partNet, int layer)` Extract hidden layer (neurons) to setup part net hidden layer Hidden layer is inserted by REFERENCE into part net
`protected void`	`extractInputOutputLayer(NeuralNet partNet, int layer)` Extract input layer (neurons and links) to setup part net input and output layer Input layer is inserted by REFERENCE into part net Copy of input layer is inserted as COPY into part net
`protected java.util.List<java.lang.Double>`	`getHiddenLayerOutput(java.util.List<java.lang.Double> outputs)`
`protected NeuronList`	`getNeuronsOfLayer(int layer)` Extract all neurons that are part of same layer
`protected PatternSet`	`getNextSet(NeuralNet partNet, PatternSet patterns)` Generate the next pattern set from current pattern set by extracting hidden layer activation.
`protected int`	`numberOfLayers()` Return number of layers
`protected void`	`reset()` Resets the trainer to initial values in order to start a new training procedure.
`protected void`	`resetLayer(int ioLayer)` Reset STRUCTURE of net (activation function, bias, input or output...) by resetting of neurons in part net AFTER training
`protected void`	`resetTrainer()` Trainer gets new part net to train
`void`	`setPrint(boolean printing)`
`void`	`train()` Training feed-forward network layer-wise to become auto-encoding.
`protected void`	`train(java.util.List<java.lang.Double> inputPattern, java.util.List<java.lang.Double> targetPattern)` Trains a single pattern once.
`protected void`	`updatePartTrainer()` Pass configuration of SAE trainer down to part trainer for AE net
`protected void`	`updatePartTrainerData()` Pass training data of SAE trainer down to part trainer for AE net

Methods inherited from class boone.Trainer
clone, fromXML, getBatchSize, getEpochs, getLearnRate, getNetwork, getRank, getTestData, getTrainingData, getTrainingSignalGenerator, getWinningNeuron, setBatchSize, setEpochs, setLearnRate, setNetwork, setShuffle, setStepMode, setTestData, setTrainingData, setTrainingSignalGenerator, test, test, toString, toXML, trainPattern, updateLinks, updateNeurons

Methods inherited from class java.lang.Object
equals, finalize, getClass, hashCode, notify, notifyAll, wait, wait, wait

- Field Detail
  - partTrainer
```
protected Trainer partTrainer
```
    The trainer internally used to train each layer
  - partNet
```
protected NeuralNet partNet
```
    Layer that is trained - extracted net
  - hiddenLayer
```
protected NeuronList hiddenLayer
```
    Hidden layer
  - wasInputLayer
```
protected boolean wasInputLayer
```
    Flag indicating if the current input layer is the input layer of whole net
  - wasOutputLayer
```
protected boolean wasOutputLayer
```
    Flag indicating if the current hidden layer is the output layer of whole net
  - steps
```
protected int steps
```
    The number of steps used for training
  - print
```
protected boolean print
```
    Flag indicating how much training information is printed to console
- Constructor Detail
  - SAETrainer
```
public SAETrainer()
```
    Create a new SAE trainer using a SquareError.
  - SAETrainer
```
public SAETrainer(Trainer trainer,
                  int steps)
```
    Create a new SAE trainer using a SquareError, a given trainer and a given number of steps
- Method Detail
  - train
```
protected void train(java.util.List<java.lang.Double> inputPattern,
                     java.util.List<java.lang.Double> targetPattern)
```
    Description copied from class: Trainer
    
    Trains a single pattern once.
    
    Specified by:
    
    train in class Trainer
    
    Parameters:
    
    inputPattern - an input pattern
    
    targetPattern - a target pattern (may be null)
  - calcAdaptation
```
protected double calcAdaptation(BrainPart part)
```
    Not used.
    
    Specified by:
    
    calcAdaptation in class Trainer
    
    Parameters:
    
    part - the brain part to be adapted
    
    Returns:
    
    the adaptation value
  - reset
```
protected void reset()
```
    Description copied from class: Trainer
    
    Resets the trainer to initial values in order to start a new training procedure.
    
    Specified by:
    
    reset in class Trainer
  - endBatch
```
protected void endBatch()
```
    Description copied from class: Trainer
    
    Updates weights and biases using the batch gradient information.
    
    Overrides:
    
    endBatch in class Trainer
  - endTrain
```
protected void endTrain()
```
    Description copied from class: Trainer
    
    Do some house keeping after all epochs have been trained.
    
    Specified by:
    
    endTrain in class Trainer
  - train
```
public void train()
```
    Training feed-forward network layer-wise to become auto-encoding. For each layer: - extract - train - reset and setup
    
    Overrides:
    
    train in class Trainer
  - resetLayer
```
protected void resetLayer(int ioLayer)
```
    Reset STRUCTURE of net (activation function, bias, input or output...) by resetting of neurons in part net AFTER training
    
    Parameters:
    
    ioLayer - actual layer number of input layer of part net
  - resetTrainer
```
protected void resetTrainer()
```
    Trainer gets new part net to train
  - updatePartTrainer
```
protected void updatePartTrainer()
```
    Pass configuration of SAE trainer down to part trainer for AE net
  - updatePartTrainerData
```
protected void updatePartTrainerData()
```
    Pass training data of SAE trainer down to part trainer for AE net
  - getNeuronsOfLayer
```
protected NeuronList getNeuronsOfLayer(int layer)
```
    Extract all neurons that are part of same layer
    
    Parameters:
    
    layer - to get neurons from
  - extractInputOutputLayer
```
protected void extractInputOutputLayer(NeuralNet partNet,
                                       int layer)
```
    Extract input layer (neurons and links) to setup part net input and output layer Input layer is inserted by REFERENCE into part net Copy of input layer is inserted as COPY into part net
    
    Parameters:
    
    layer - input layer to extract
  - extractHiddenLayer
```
protected void extractHiddenLayer(NeuralNet partNet,
                                  int layer)
```
    Extract hidden layer (neurons) to setup part net hidden layer Hidden layer is inserted by REFERENCE into part net
    
    Parameters:
    
    layer - hidden layer to extract
  - numberOfLayers
```
protected int numberOfLayers()
```
    Return number of layers
  - getNextSet
```
protected PatternSet getNextSet(NeuralNet partNet,
                                PatternSet patterns)
```
    Generate the next pattern set from current pattern set by extracting hidden layer activation. Test auto-encoding property
    
    Parameters:
    
    partNet - current part net
    
    patterns - current patterns
  - getHiddenLayerOutput
```
protected java.util.List<java.lang.Double> getHiddenLayerOutput(java.util.List<java.lang.Double> outputs)
```
  - setPrint
```
public void setPrint(boolean printing)
```

Class SAETrainer

Field Summary

Fields inherited from class boone.Trainer

Constructor Summary

Method Summary

Methods inherited from class boone.Trainer

Methods inherited from class java.lang.Object

Field Detail

partTrainer

partNet

hiddenLayer

wasInputLayer

wasOutputLayer

steps

print

Constructor Detail

SAETrainer

SAETrainer

Method Detail

train

calcAdaptation

reset

endBatch

endTrain

train

resetLayer

resetTrainer

updatePartTrainer

updatePartTrainerData

getNeuronsOfLayer

extractInputOutputLayer

extractHiddenLayer

numberOfLayers

getNextSet

getHiddenLayerOutput

setPrint