/**
* Load a CSV file into a memory dataset.
*
* @param format The CSV format to use.
* @param filename The filename to load.
* @param headers True if there is a header line.
* @param inputSize The input size. Input always comes first in a file.
* @param idealSize The ideal size, 0 for unsupervised.
* @return A NeuralDataSet that holds the contents of the CSV file.
*/
public static MLDataSet loadCSVTOMemory(
CSVFormat format, String filename, boolean headers, int inputSize, int idealSize) {
MLDataSet result = new BasicMLDataSet();
ReadCSV csv = new ReadCSV(filename, headers, format);
while (csv.next()) {
MLData input = null;
MLData ideal = null;
int index = 0;
input = new BasicMLData(inputSize);
for (int i = 0; i < inputSize; i++) {
double d = csv.getDouble(index++);
input.setData(i, d);
}
if (idealSize > 0) {
ideal = new BasicMLData(idealSize);
for (int i = 0; i < idealSize; i++) {
double d = csv.getDouble(index++);
ideal.setData(i, d);
}
}
MLDataPair pair = new BasicMLDataPair(input, ideal);
result.add(pair);
}
return result;
}
/**
* Called when the recognize button is pressed.
*
* @param event The event.
*/
void recognize_actionPerformed(final java.awt.event.ActionEvent event) {
if (this.net == null) {
JOptionPane.showMessageDialog(
this, "I need to be trained first!", "Error", JOptionPane.ERROR_MESSAGE);
return;
}
this.entry.downSample();
final MLData input = new BasicMLData(5 * 7);
int idx = 0;
final SampleData ds = this.sample.getData();
for (int y = 0; y < ds.getHeight(); y++) {
for (int x = 0; x < ds.getWidth(); x++) {
input.setData(idx++, ds.getData(x, y) ? .5 : -.5);
}
}
final int best = this.net.classify(input);
final char map[] = mapNeurons();
JOptionPane.showMessageDialog(
this,
" " + map[best] + " (Neuron #" + best + " fired)",
"That Letter Is",
JOptionPane.PLAIN_MESSAGE);
clear_actionPerformed(null);
}
/** Run method for the background training thread. */
public void trainSOM() {
try {
final int inputNeuron = OCR.DOWNSAMPLE_HEIGHT * OCR.DOWNSAMPLE_WIDTH;
final int outputNeuron = this.letterListModel.size();
final MLDataSet trainingSet = new BasicMLDataSet();
for (int t = 0; t < this.letterListModel.size(); t++) {
final MLData item = new BasicMLData(inputNeuron);
int idx = 0;
final SampleData ds = (SampleData) this.letterListModel.getElementAt(t);
for (int y = 0; y < ds.getHeight(); y++) {
for (int x = 0; x < ds.getWidth(); x++) {
item.setData(idx++, ds.getData(x, y) ? .5 : -.5);
}
}
trainingSet.add(new BasicMLDataPair(item, null));
}
this.net = new SOM(inputNeuron, outputNeuron);
this.net.reset();
SOMClusterCopyTraining train = new SOMClusterCopyTraining(this.net, trainingSet);
train.iteration();
JOptionPane.showMessageDialog(
this, "Training has completed.", "Training", JOptionPane.PLAIN_MESSAGE);
} catch (final Exception e) {
e.printStackTrace();
JOptionPane.showMessageDialog(this, "Error: " + e, "Training", JOptionPane.ERROR_MESSAGE);
}
}
/**
* Used to map neurons to actual letters.
*
* @return The current mapping between neurons and letters as an array.
*/
char[] mapNeurons() {
final char map[] = new char[this.letterListModel.size()];
for (int i = 0; i < map.length; i++) {
map[i] = '?';
}
for (int i = 0; i < this.letterListModel.size(); i++) {
final MLData input = new BasicMLData(5 * 7);
int idx = 0;
final SampleData ds = (SampleData) this.letterListModel.getElementAt(i);
for (int y = 0; y < ds.getHeight(); y++) {
for (int x = 0; x < ds.getWidth(); x++) {
input.setData(idx++, ds.getData(x, y) ? .5 : -.5);
}
}
final int best = this.net.classify(input);
map[best] = ds.getLetter();
}
return map;
}
public MLMethod decode(final NEATPopulation pop, final Substrate substrate, final Genome genome) {
// obtain the CPPN
final NEATCODEC neatCodec = new NEATCODEC();
final NEATNetwork cppn = (NEATNetwork) neatCodec.decode(genome);
final List<NEATLink> linkList = new ArrayList<NEATLink>();
final ActivationFunction[] afs = new ActivationFunction[substrate.getNodeCount()];
final ActivationFunction af = new ActivationSteepenedSigmoid();
// all activation functions are the same
for (int i = 0; i < afs.length; i++) {
afs[i] = af;
}
final double c = this.maxWeight / (1.0 - this.minWeight);
final MLData input = new BasicMLData(cppn.getInputCount());
// First create all of the non-bias links.
for (final SubstrateLink link : substrate.getLinks()) {
final SubstrateNode source = link.getSource();
final SubstrateNode target = link.getTarget();
int index = 0;
for (final double d : source.getLocation()) {
input.setData(index++, d);
}
for (final double d : target.getLocation()) {
input.setData(index++, d);
}
final MLData output = cppn.compute(input);
double weight = output.getData(0);
if (Math.abs(weight) > this.minWeight) {
weight = (Math.abs(weight) - this.minWeight) * c * Math.signum(weight);
linkList.add(new NEATLink(source.getId(), target.getId(), weight));
}
}
// now create biased links
input.clear();
final int d = substrate.getDimensions();
final List<SubstrateNode> biasedNodes = substrate.getBiasedNodes();
for (final SubstrateNode target : biasedNodes) {
for (int i = 0; i < d; i++) {
input.setData(d + i, target.getLocation()[i]);
}
final MLData output = cppn.compute(input);
double biasWeight = output.getData(1);
if (Math.abs(biasWeight) > this.minWeight) {
biasWeight = (Math.abs(biasWeight) - this.minWeight) * c * Math.signum(biasWeight);
linkList.add(new NEATLink(0, target.getId(), biasWeight));
}
}
// check for invalid neural network
if (linkList.size() == 0) {
return null;
}
Collections.sort(linkList);
final NEATNetwork network =
new NEATNetwork(substrate.getInputCount(), substrate.getOutputCount(), linkList, afs);
network.setActivationCycles(substrate.getActivationCycles());
return network;
}
