private void computeTotalGradient(
Gradients totalGradients, Gradients partialGradients, SetOfIOPairs trainingSet) {
// na zaciatku sa inicializuju gradienty (total)
totalGradients.resetGradients();
// partialGradients.resetGradients();
// Gradients totalGradients = new Gradients(this);
// Gradients partialGradients = new Gradients(this); /***/
for (SetOfIOPairs.IOPair pair : trainingSet.pairs) { // pre kazdy par trenovacej mnoziny
// partialGradients = computeGradient(pair.inputs, pair.outputs);
computeGradient(partialGradients, pair.inputs, pair.outputs);
for (int il = this.numberOfLayers() - 1; il >= 1; il--) { // pre vsetky vrstvy okrem poslednej
NeuralLayer currentLayer = this.getLayer(il);
for (int in = 0;
in < currentLayer.numberOfNeurons();
in++) { // pre vsetky neurony na currentLayer
// upravime gradient prahov :
totalGradients.incrementThreshold(il, in, partialGradients.getThreshold(il, in));
for (int ii = 0;
ii < currentLayer.lowerLayer().numberOfNeurons();
ii++) { // pre vsetky vstupy
totalGradients.incrementWeight(il, in, ii, partialGradients.getWeight(il, in, ii));
}
}
} // end for layer
} // end foreach
// return totalGradients;
} // end method
private void adaptation(
SetOfIOPairs trainingSet, int maxK, double eps, double lambda, double micro) {
// trainingSet : trenovacia mnozina
// maxK : maximalny pocet iteracii
// eps : pozadovana presnost normovanej dlzky gradientu
// lambda : rychlost ucenia (0.1)
// micro : momentovy clen
double delta;
Gradients deltaGradients = new Gradients(this);
Gradients totalGradients = new Gradients(this);
Gradients partialGradients = new Gradients(this);
System.out.println("setting up random weights and thresholds ...");
// prahy a vahy neuronovej siete nastavime na nahodne hodnoty, delta-gradienty vynulujeme (oni
// sa nuluju uz pri init)
for (int il = this.numberOfLayers() - 1;
il >= 1;
il--) { // iteracia cez vsetky vrstvy nadol okrem poslednej
NeuralLayer currentLayer = this.getLayer(il);
for (int in = 0; in < currentLayer.numberOfNeurons(); in++) { // pre kazdy neuron na vrstve
Neuron currentNeuron = currentLayer.getNeuron(in);
currentNeuron.threshold = 2 * this.random() - 1;
// deltaGradients.setThreshold(il,in,0.0);
for (int ii = 0; ii < currentNeuron.numberOfInputs(); ii++) {
currentNeuron.getInput(ii).weight = 2 * this.random() - 1;
// deltaGradients.setWeight(il,in,ii,0.0);
} // end ii
} // end in
} // end il
int currK = 0; // citac iteracii
double currE =
Double.POSITIVE_INFINITY; // pociatocna aktualna presnost bude nekonecna (tendencia
// znizovania)
System.out.println("entering adaptation loop ... (maxK = " + maxK + ")");
while (currK < maxK && currE > eps) {
computeTotalGradient(totalGradients, partialGradients, trainingSet);
for (int il = this.numberOfLayers() - 1;
il >= 1;
il--) { // iteracia cez vsetky vrstvy nadol okrem poslednej
NeuralLayer currentLayer = this.getLayer(il);
for (int in = 0; in < currentLayer.numberOfNeurons(); in++) { // pre kazdy neuron na vrstve
Neuron currentNeuron = currentLayer.getNeuron(in);
delta =
-lambda * totalGradients.getThreshold(il, in)
+ micro * deltaGradients.getThreshold(il, in);
currentNeuron.threshold += delta;
deltaGradients.setThreshold(il, in, delta);
} // end for ii 1
for (int in = 0; in < currentLayer.numberOfNeurons(); in++) { // pre kazdy neuron na vrstve
Neuron currentNeuron = currentLayer.getNeuron(in);
for (int ii = 0; ii < currentNeuron.numberOfInputs(); ii++) { // a pre kazdy vstup neuronu
delta =
-lambda * totalGradients.getWeight(il, in, ii)
+ micro * deltaGradients.getWeight(il, in, ii);
currentNeuron.getInput(ii).weight += delta;
deltaGradients.setWeight(il, in, ii, delta);
} // end for ii
} // end for in 2
} // end for il
currE = totalGradients.getGradientAbs();
currK++;
if (currK % 25 == 0) System.out.println("currK=" + currK + " currE=" + currE);
} // end while
}
private void computeGradient(
Gradients gradients, Vector<Double> inputs, Vector<Double> requiredOutputs) {
// Gradients gradients = new Gradients(this);
activities(inputs);
for (int il = this.numberOfLayers() - 1;
il >= 1;
il--) { // backpropagation cez vsetky vrstvy okrem poslednej
NeuralLayer currentLayer = this.getLayer(il);
if (currentLayer.isLayerTop()) { // ak sa jedna o najvyssiu vrstvu
// pridame gradient prahov pre danu vrstvu do odpovedajuceho vektora a tento gradient
// pocitame cez neurony :
// gradients.thresholds.add(il, new Vector<Double>());
for (int in = 0;
in < currentLayer.numberOfNeurons();
in++) { // pre vsetky neurony na vrstve
Neuron currentNeuron = currentLayer.getNeuron(in);
gradients.setThreshold(
il,
in,
currentNeuron.output
* (1 - currentNeuron.output)
* (currentNeuron.output - requiredOutputs.elementAt(in)));
} // end for each neuron
for (int in = 0; in < currentLayer.numberOfNeurons(); in++) { // for each neuron
Neuron currentNeuron = currentLayer.getNeuron(in);
for (int ii = 0; ii < currentNeuron.numberOfInputs(); ii++) { // for each neuron's input
NeuralInput currentInput = currentNeuron.getInput(ii);
gradients.setWeight(
il,
in,
ii,
gradients.getThreshold(il, in) * currentLayer.lowerLayer().getNeuron(ii).output);
} // end for each input
} // end for each neuron
} else { // ak sa jedna o spodnejsie vrstvy (najnizsiu vrstvu nepocitame, ideme len po 1.)
// pocitame gradient prahov :
// gradients.thresholds.add(il, new Vector<Double>());
for (int in = 0; in < currentLayer.numberOfNeurons(); in++) { // for each neuron
double aux = 0;
// iterujeme cez vsetky axony neuronu (resp. synapsie neuronov na vyssej vrstve)
for (int ia = 0; ia < currentLayer.upperLayer().numberOfNeurons(); ia++) {
aux +=
gradients.getThreshold(il + 1, ia)
* currentLayer.upperLayer().getNeuron(ia).getInput(in).weight;
}
gradients.setThreshold(
il,
in,
currentLayer.getNeuron(in).output * (1 - currentLayer.getNeuron(in).output) * aux);
} // end for each neuron
// pocitame gradienty vah :
for (int in = 0; in < currentLayer.numberOfNeurons(); in++) { // for each neuron
Neuron currentNeuron = currentLayer.getNeuron(in);
for (int ii = 0; ii < currentNeuron.numberOfInputs(); ii++) { // for each neuron's input
NeuralInput currentInput = currentNeuron.getInput(ii);
gradients.setWeight(
il,
in,
ii,
gradients.getThreshold(il, in) * currentLayer.lowerLayer().getNeuron(ii).output);
} // end for each input
} // end for each neuron
} // end layer IF
} // end backgropagation for each layer
// return gradients;
}