databionics.esom.train
Class SOM

java.lang.Object
  databionics.esom.train.SOM

Direct Known Subclasses:

Batch2SOM, BatchSOM, HybridBatchSOM, KBatchSOM, OnlineSOM, SlowBatchSOM

public abstract class SOM
extends java.lang.Object

Abstract base class for a self organizing map training algorithm. It is a container for

• the training data, possibly with a permutation view per epoch
• the distance function in the data space
• the grid of neurons, representing the topology and distance function on the grid
• the neighborhood kernel function
• values and cooling strategies for radius and learning rate
• a list of indices of the best matching neurons for the data vectors

The basic training algorithm is implemented in train() with many callback functions to implement different variants of the algorithm in subclasses, some of which already have a default implementation and should be called via super unless you know what you are doing.

• beforeEpoch()
• beforeSearch(data_index)
• bm_index = findBestMatch(data_index)
• beforeUpdate(bm_index)
• update(data_pattern, bm_index)
• afterUpdate(bm_index, data_index)
• afterEpoch()

Field Summary

protected cern.colt.list.IntArrayList	bestmatches indices of best matching neurons for data points
protected BmSearch	bmSearch the search system
protected BmSearchStat	bmStat the bm search statistic
protected boolean	bmStatNeeded
protected boolean	center flag for centering the map
protected ClsFile	classes
protected cern.colt.list.IntArrayList	count
cern.colt.matrix.DoubleMatrix2D	data matrix with training data
protected Descriptives	descriptives statistics about data
protected cern.colt.function.ThresholdVectorVectorFunction	distanceFunction distance function in data space
protected cern.colt.matrix.DoubleMatrix1D	distances current distances between bestmatches and datapoints
protected int	epoch current epoch number 0..epochs-1, only used internally
protected int	epochs number of training epochs
protected Grid	grid grid of weight vectors
protected java.lang.String	initMethod kind of initialising
protected cern.colt.list.IntArrayList	keys list with unique keys
protected static org.apache.log4j.Logger	log interface to log4j system
protected LrnFile	lrn
protected Neighborhood	neighborhood neighborhood kernel function
protected cern.colt.matrix.DoubleMatrix1D	neuron current neuron, only used internally
protected java.text.NumberFormat	nf number format for messages
protected boolean	offline
protected cern.colt.list.IntArrayList	oldBestmatches indices of best matching neurons for data points of last epoch
protected boolean	online
protected cern.colt.matrix.DoubleMatrix1D	pattern current data pattern, only used internally
protected cern.colt.list.IntArrayList	permutation current permutation of the data vectors, only used internally
protected int	permutations number of possible permutations of the data vectors, only used internally
protected boolean	permute flag whether to permute data vectors before each epoch
protected double	qerror current quantization error
protected int	radius current neighborhood radius, only used internally
protected Cooling	radiusCooling cooling strategy of neighborhood radius
protected cern.jet.random.engine.RandomEngine	random random number generator, only used internally
protected double	rate current learning rate, only used internally
protected Cooling	rateCooling cooling strategy of learning rate
protected int	saveEpoch number of epochs after which a bm and wts file will be saved
protected boolean	saveEpoches Boolean for saving during epochs
protected java.lang.String	savePrefix String with name of files during epochs
protected boolean	saveUMatrix Boolean for saving U-Matrix during epochs
protected cern.colt.matrix.DoubleMatrix2D	view possibly permuted view on training data, only used internally

Constructor Summary

SOM()
Standard constructor.

Method Summary

void	afterEpoch() Things to be done after each epoch.
void	afterUpdate(int index, int row) Things to be done after updating a weight vector.
void	beforeEpoch() Things to be done before each epoch.
void	beforeSearch(int row) Things to be done before searching for a best match.
void	beforeUpdate(int index) Things to be done before updating a weight vector Empty by default.
double	calcQerror()
protected void	centerMap() center map at neuron with highest density
void	cool() Cool down the parameters
cern.colt.list.IntArrayList	getBestMatches() Get the indices of best matching neurons for data points Available after calling train().
BmSearch	getBmSearch()
cern.colt.matrix.DoubleMatrix2D	getData() Get the matrix with the training data.
Descriptives	getDescriptives() Get the statistics about the training data.
double	getDistance(cern.colt.matrix.DoubleMatrix1D data, int index) Get the distance between a dataVector and a neuron.
double	getDistance(int row, int index) Get the distance between a data row and a neuron.
cern.colt.function.ThresholdVectorVectorFunction	getDistanceFunction() Get the distance function in data space
cern.colt.matrix.DoubleMatrix1D	getDistances()
int	getEpochs() Get the number of training epochs.
Grid	getGrid() Get the grid of weight vectors.
cern.colt.list.IntArrayList	getKeys() Get the keys for the data patterns.
Neighborhood	getNeighborhood() Get the neighborhood kernel function.
java.text.NumberFormat	getNumberFormat() Get the number format
cern.colt.list.IntArrayList	getOldBestmatches()
cern.colt.matrix.DoubleMatrix1D	getPattern(int row) Access to a data pattern by row number
cern.colt.list.IntArrayList	getPermutation() Get the current permuation
boolean	getPermute() Get the flag whether to permute data vectors before each epoch.
int	getRadius()
Cooling	getRadiusCooling() Get the cooling strategy of the neighborhood radius.
Cooling	getRateCooling() Get the cooling strategy of learning rate.
void	init() Initialize the grid of weight vectors with the current initialization method.
boolean	isBmStatNeeded()
boolean	isCenter()
boolean	loadCls(java.lang.String filename)
boolean	loadData(java.lang.String filename) Load the training data and keys from a *.lrn file
void	saveBestMatches(java.lang.String filename) Save indices of best matching neurons to a *.bm file Available after calling train().
void	setBestMatches(BMFile bms) Set the indices of best matching neurons for data points
void	setBestMatches(cern.colt.list.IntArrayList bms) Set the indices of best matching neurons for data points
void	setBmSearch(BmSearch bmSearch)
void	setBmStatNeeded(boolean bmStatNeeded)
void	setCenter(boolean center)
void	setData(cern.colt.matrix.DoubleMatrix2D v) Set the matrix with the training data, patterns in rows.
void	setDistanceFunction(cern.colt.function.ThresholdVectorVectorFunction v) Set the distance function in data space
void	setDistances(cern.colt.matrix.DoubleMatrix1D distances)
void	setEpochs(int v) Set the number of training epochs.
void	setGrid(Grid v) Set the grid of weight vectors.
void	setInit(java.lang.String initMethod) Set the method of initialization for the weights
void	setKeys(cern.colt.list.IntArrayList v) Set the list of unique keys for the data patterns.
void	setNeighborhood(Neighborhood v) Set the neighborhood kernel function.
void	setNumberFormat(java.text.NumberFormat v) Set the number format
void	setOldBestmatches(cern.colt.list.IntArrayList oldBestmatches)
void	setOnline(boolean online)
void	setPermute(boolean v) Set the flag whether to permute data vectors before each epoch.
void	setRadiusCooling(Cooling v) Set the cooling strategy of the neighborhood radius.
void	setRateCooling(Cooling v) Set the cooling strategy of learning rate.
void	setSaveEpoch(int e) Set number of epochs after which a *.bm and a *.wts are saved
void	setSaveEpochBoolean() sets the boolean flag true, because we are saving during epochs
void	setSavePrefix(java.lang.String name) Set the prefix for the filename used to regularily save *.bm, *.wts and *.umx files
void	setSaveUMatrix(boolean flag) set the flag for saving a U-Matrix every nth epoch
boolean	stop() Stopping criterion
void	train() General training algorithm for the SOM.
abstract void	update(cern.colt.matrix.DoubleMatrix1D vector, int bm, int pos) Things to be done, after finding a bestmatch, e.g.
void	updateNeighborhood(cern.colt.matrix.DoubleMatrix1D vector, int bm) Update a neuron and its neighborhood.
void	updateNeuron(int index, cern.colt.matrix.DoubleMatrix1D vector, double weight) Update a single weight vector.

Methods inherited from class java.lang.Object

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

Field Detail

log

protected static final org.apache.log4j.Logger log

interface to log4j system

bestmatches

protected cern.colt.list.IntArrayList bestmatches

indices of best matching neurons for data points

oldBestmatches

protected cern.colt.list.IntArrayList oldBestmatches

indices of best matching neurons for data points of last epoch

data

public cern.colt.matrix.DoubleMatrix2D data

matrix with training data

descriptives

protected Descriptives descriptives

statistics about data

view

protected cern.colt.matrix.DoubleMatrix2D view

possibly permuted view on training data, only used internally

keys

protected cern.colt.list.IntArrayList keys

list with unique keys

distanceFunction

protected cern.colt.function.ThresholdVectorVectorFunction distanceFunction

distance function in data space

epoch

protected int epoch

current epoch number 0..epochs-1, only used internally

epochs

protected int epochs

number of training epochs

grid

protected Grid grid

grid of weight vectors

neighborhood

protected Neighborhood neighborhood

neighborhood kernel function

neuron

protected cern.colt.matrix.DoubleMatrix1D neuron

current neuron, only used internally

distances

protected cern.colt.matrix.DoubleMatrix1D distances

current distances between bestmatches and datapoints

nf

protected java.text.NumberFormat nf

number format for messages

pattern

protected cern.colt.matrix.DoubleMatrix1D pattern

current data pattern, only used internally

permutation

protected cern.colt.list.IntArrayList permutation

current permutation of the data vectors, only used internally

permutations

protected int permutations

number of possible permutations of the data vectors, only used internally

permute

protected boolean permute

flag whether to permute data vectors before each epoch

radius

protected int radius

current neighborhood radius, only used internally

radiusCooling

protected Cooling radiusCooling

cooling strategy of neighborhood radius

random

protected cern.jet.random.engine.RandomEngine random

random number generator, only used internally

rate

protected double rate

current learning rate, only used internally

qerror

protected double qerror

current quantization error

rateCooling

protected Cooling rateCooling

cooling strategy of learning rate

saveEpoch

protected int saveEpoch

number of epochs after which a bm and wts file will be saved

initMethod

protected java.lang.String initMethod

kind of initialising

savePrefix

protected java.lang.String savePrefix

String with name of files during epochs

saveEpoches

protected boolean saveEpoches

Boolean for saving during epochs

saveUMatrix

protected boolean saveUMatrix

Boolean for saving U-Matrix during epochs

center

protected boolean center

flag for centering the map

bmSearch

protected BmSearch bmSearch

the search system

bmStat

protected BmSearchStat bmStat

the bm search statistic

bmStatNeeded

protected boolean bmStatNeeded

online

protected boolean online

offline

protected boolean offline

classes

protected ClsFile classes

count

protected cern.colt.list.IntArrayList count

lrn

protected LrnFile lrn

Constructor Detail

SOM

public SOM()

Standard constructor. Sets default values and initializes the random number generator.

Method Detail

init

public void init()

Initialize the grid of weight vectors with the current initialization method.

train

public void train()

General training algorithm for the SOM.

beforeEpoch

public void beforeEpoch()

Things to be done before each epoch. Cools down the parameters, prints some info, and optionally permutes the data patterns.

beforeSearch

public void beforeSearch(int row)

Things to be done before searching for a best match. Empty by default.

Parameters:

row - Row of the data pattern

beforeUpdate

public void beforeUpdate(int index)

Things to be done before updating a weight vector Empty by default.

Parameters:

index - Index of neuron to be updated

update

public abstract void update(cern.colt.matrix.DoubleMatrix1D vector,
                            int bm,
                            int pos)

Things to be done, after finding a bestmatch, e.g. in online learning to update the neuron. To be implemented in subclasses.

Parameters:

vector - vector of data

bm - index of bestmatching neuron

updateNeighborhood

public void updateNeighborhood(cern.colt.matrix.DoubleMatrix1D vector,
                               int bm)

Update a neuron and its neighborhood.

Parameters:

vector - vector of data

bm - index of bestmatching neuron

updateNeuron

public void updateNeuron(int index,
                         cern.colt.matrix.DoubleMatrix1D vector,
                         double weight)

Update a single weight vector.

Parameters:

index - Index of neuron to be updated

dataVector - Vector of the data pattern

weight - Weight of neighborhood kernel assigned by distance on grid

afterUpdate

public void afterUpdate(int index,
                        int row)

Things to be done after updating a weight vector. Empty by default.

Parameters:

index - Index of neuron just updated

afterEpoch

public void afterEpoch()

Things to be done after each epoch. Saves *.bm and *.wts every n-th epoch

cool

public void cool()

Cool down the parameters

stop

public boolean stop()

Stopping criterion

centerMap

protected void centerMap()

center map at neuron with highest density

getBestMatches

public cern.colt.list.IntArrayList getBestMatches()

Get the indices of best matching neurons for data points Available after calling train().

Returns:

list of best matches.

setBestMatches

public void setBestMatches(cern.colt.list.IntArrayList bms)

Set the indices of best matching neurons for data points

setBestMatches

public void setBestMatches(BMFile bms)

Set the indices of best matching neurons for data points

saveBestMatches

public void saveBestMatches(java.lang.String filename)

Save indices of best matching neurons to a *.bm file Available after calling train().

getDescriptives

public Descriptives getDescriptives()

Get the statistics about the training data.

Returns:

value of descriptives.

getData

public cern.colt.matrix.DoubleMatrix2D getData()

Get the matrix with the training data.

Returns:

value of data.

setData

public void setData(cern.colt.matrix.DoubleMatrix2D v)

Set the matrix with the training data, patterns in rows. Also updates the number of permutations and initializes array for best matching indices.

Parameters:

v - Value to assign to data.

loadData

public boolean loadData(java.lang.String filename)

Load the training data and keys from a *.lrn file

Parameters:

filename - Name of the file

loadCls

public boolean loadCls(java.lang.String filename)

getDistance

public double getDistance(int row,
                          int index)

Get the distance between a data row and a neuron.

Parameters:

row - Row of data pattern

index - Index of neuron

Returns:

value of distance.

getDistance

public double getDistance(cern.colt.matrix.DoubleMatrix1D data,
                          int index)

Get the distance between a dataVector and a neuron.

Parameters:

data - DataVector

index - Index of neuron

Returns:

value of distance.

getDistanceFunction

public cern.colt.function.ThresholdVectorVectorFunction getDistanceFunction()

Get the distance function in data space

Returns:

value of distanceFunction.

setDistanceFunction

public void setDistanceFunction(cern.colt.function.ThresholdVectorVectorFunction v)

Set the distance function in data space

Parameters:

v - Value to assign to distanceFunction.

calcQerror

public double calcQerror()

getEpochs

public int getEpochs()

Get the number of training epochs.

Returns:

value of epochs.

setEpochs

public void setEpochs(int v)

Set the number of training epochs.

Parameters:

v - Value to assign to epochs.

getGrid

public Grid getGrid()

Get the grid of weight vectors.

Returns:

value of grid.

setGrid

public void setGrid(Grid v)

Set the grid of weight vectors.

Parameters:

v - Value to assign to grid.

getKeys

public cern.colt.list.IntArrayList getKeys()

Get the keys for the data patterns.

Returns:

value of keys.

setKeys

public void setKeys(cern.colt.list.IntArrayList v)

Set the list of unique keys for the data patterns.

Parameters:

v - Value to assign to keys.

getNeighborhood

public Neighborhood getNeighborhood()

Get the neighborhood kernel function.

Returns:

value of neighborhood.

setNeighborhood

public void setNeighborhood(Neighborhood v)

Set the neighborhood kernel function.

Parameters:

v - Value to assign to neighborhood.

getNumberFormat

public java.text.NumberFormat getNumberFormat()

Get the number format

Returns:

value of number format

setNumberFormat

public void setNumberFormat(java.text.NumberFormat v)

Set the number format

Parameters:

v - value of number format

getPattern

public cern.colt.matrix.DoubleMatrix1D getPattern(int row)

Access to a data pattern by row number

Parameters:

row - Row of data pattern

getPermute

public boolean getPermute()

Get the flag whether to permute data vectors before each epoch.

Returns:

value of permute.

setPermute

public void setPermute(boolean v)

Set the flag whether to permute data vectors before each epoch.

Parameters:

v - Value to assign to permute.

getPermutation

public cern.colt.list.IntArrayList getPermutation()

Get the current permuation

Returns:

value of permutation

getRadiusCooling

public Cooling getRadiusCooling()

Get the cooling strategy of the neighborhood radius.

Returns:

value of radiusCooling.

setRadiusCooling

public void setRadiusCooling(Cooling v)

Set the cooling strategy of the neighborhood radius.

Parameters:

v - Value to assign to radiusCooling.

getRateCooling

public Cooling getRateCooling()

Get the cooling strategy of learning rate.

Returns:

value of rateCooling.

setRateCooling

public void setRateCooling(Cooling v)

Set the cooling strategy of learning rate.

Parameters:

v - Value to assign to rateCooling.

setSaveEpoch

public void setSaveEpoch(int e)

Set number of epochs after which a *.bm and a *.wts are saved

Parameters:

e - number of epochs

setSaveEpochBoolean

public void setSaveEpochBoolean()

sets the boolean flag true, because we are saving during epochs

setInit

public void setInit(java.lang.String initMethod)

Set the method of initialization for the weights

Parameters:

initMethod - name of method allowed names are: norm_mean_2std, uni_mean_2std, uni_min_max, pca

setSavePrefix

public void setSavePrefix(java.lang.String name)

Set the prefix for the filename used to regularily save *.bm, *.wts and *.umx files

Parameters:

name - String with name of file

setSaveUMatrix

public void setSaveUMatrix(boolean flag)

set the flag for saving a U-Matrix every nth epoch

Parameters:

flag - true for save

isCenter

public boolean isCenter()

Returns:

Returns the center.

setCenter

public void setCenter(boolean center)

Parameters:

center - The center to set.

getDistances

public cern.colt.matrix.DoubleMatrix1D getDistances()

Returns:

Returns the distances.

setDistances

public void setDistances(cern.colt.matrix.DoubleMatrix1D distances)

Parameters:

distances - The distances to set.

getBmSearch

public BmSearch getBmSearch()

Returns:

Returns the bmSearch.

setBmSearch

public void setBmSearch(BmSearch bmSearch)

Parameters:

bmSearch - The bmSearch to set.

getOldBestmatches

public cern.colt.list.IntArrayList getOldBestmatches()

Returns:

Returns the oldBestmatches.

setOldBestmatches

public void setOldBestmatches(cern.colt.list.IntArrayList oldBestmatches)

Parameters:

oldBestmatches - The oldBestmatches to set.

getRadius

public int getRadius()

Returns:

Returns the radius.

isBmStatNeeded

public boolean isBmStatNeeded()

Returns:

Returns the bmStatNeeded.

setBmStatNeeded

public void setBmStatNeeded(boolean bmStatNeeded)

Parameters:

bmStatNeeded - The bmStatNeeded to set.

setOnline

public void setOnline(boolean online)

Parameters:

online - The online to set.