@Override
public double[] computeErrors() {
double[] errors = backPropagationNetwork.getErrors();
for (int i = 0; i < backPropagationNetwork.getNumberOfOutputCells(); i++)
errors[i] =
Math.abs(
backPropagationNetwork.getTargets()[i] - backPropagationNetwork.getOutputCells()[i]);
backPropagationNetwork.setErrors(errors);
return errors;
}
public BackPropagationController(
BackPropagationNetwork backPropagationNetwork, CsvWriter csvWriter) {
this.backPropagationNetwork = backPropagationNetwork;
this.csvWriter = csvWriter;
tempHiddens = new double[backPropagationNetwork.getNumberOfHiddenCells()];
th = new Thread();
}
@Override
public void stopBackpropagation() {
if (th.isAlive()) {
String stoppedMessage = "Backpropagation Stopped";
String endMessage = "End result(" + backPropagationNetwork.getEpoch() + "):\n";
for (int i = 0; i < backPropagationNetwork.getOutputCells().length; i++) {
endMessage +=
String.format("Output %d : %.8f -- ", i, backPropagationNetwork.getOutputCells()[i]);
}
endMessage += "\n";
for (int i = 0; i < backPropagationNetwork.getTargets().length; i++) {
endMessage +=
String.format("Target %d : %.8f -- ", i, backPropagationNetwork.getTargets()[i]);
}
th.interrupt();
JavaFxView.backPropagationStatus(stoppedMessage);
logger.trace(stoppedMessage);
logger.trace(endMessage);
csvWriter.saveCurrentNetworkAsCsv();
}
}
@Override
public void startBackpropagation() {
if (!backPropagationNetwork.isInitialized()) {
String errorMessage = "Network has not been initialized yet.";
logger.error(errorMessage);
return;
}
String runningMessage = "Running BackPropagation.";
logger.trace(runningMessage);
Task task =
new Task<Void>() {
@Override
public Void call() throws Exception {
for (int i = 0; i <= backPropagationNetwork.getMAX_EPOCH(); i++) {
final int epoch = i;
Platform.runLater(
() -> {
if (epoch == 0) JavaFxView.backPropagationStatus(runningMessage);
backPropagationNetwork.setEpoch(epoch);
computeOutputs();
updateWeights();
JavaFxView.changeValues();
// Stop if errortreshold is reached
double[] errors = backPropagationNetwork.getErrors();
double maxError = backPropagationNetwork.getErrorTreshold();
boolean allOutputsAreGood = false;
for (double error : errors) {
if (error < maxError) allOutputsAreGood = true;
else {
allOutputsAreGood = false;
break;
}
}
if (allOutputsAreGood) stopBackpropagation();
if (epoch == backPropagationNetwork.getMAX_EPOCH()) stopBackpropagation();
});
Thread.sleep(WAIT_TIME_MILLIS);
}
return null;
}
};
th = new Thread(task);
th.setDaemon(true);
th.start();
}
@Override
public double[] computeOutputs() {
int numberOfInputCells = backPropagationNetwork.getNumberOfInputCells();
int numberOfHiddenCells = backPropagationNetwork.getNumberOfHiddenCells();
int numberOfOutputCells = backPropagationNetwork.getNumberOfOutputCells();
// initialize hidden cells
for (int i = 0; i < numberOfHiddenCells; i++) backPropagationNetwork.setHiddenCell(i, 0);
for (int i = 0; i < numberOfHiddenCells; i++) {
for (int j = 0; j < numberOfInputCells; j++) {
backPropagationNetwork.setHiddenCell(
i,
backPropagationNetwork.getHiddenCells()[i]
+ (backPropagationNetwork.getInputCells()[j]
* backPropagationNetwork.getIhWeights()[j][i]));
}
backPropagationNetwork.setHiddenCell(
i, backPropagationNetwork.getHiddenCells()[i] + backPropagationNetwork.gethBiases()[i]);
}
for (int i = 0; i < numberOfHiddenCells; i++)
tempHiddens[i] = hyperTanFunction(backPropagationNetwork.getHiddenCells()[i]);
double outputCells[] = backPropagationNetwork.getOutputCells();
for (int i = 0; i < numberOfOutputCells; i++) {
for (int j = 0; j < numberOfHiddenCells; j++)
outputCells[i] += (tempHiddens[j] * backPropagationNetwork.getHoWeights()[j][i]);
outputCells[i] += backPropagationNetwork.getoBiases()[i];
}
for (int i = 0; i < numberOfOutputCells; i++) outputCells[i] = sigmoidFunction(outputCells[i]);
backPropagationNetwork.setOutputCells(outputCells);
return outputCells;
}
@Override
public void updateWeights() {
computeErrors();
// region get neccessary cells and values from the network
double[] outputCells = backPropagationNetwork.getOutputCells();
double[] inputCells = backPropagationNetwork.getInputCells();
double[] hiddenCells = backPropagationNetwork.getHiddenCells();
double[] targetValues = backPropagationNetwork.getTargets();
double[][] hoWeights = backPropagationNetwork.getHoWeights();
double[][] ihWeights = backPropagationNetwork.getIhWeights();
double[] hiddenBiases = backPropagationNetwork.gethBiases();
double[] outputBiases = backPropagationNetwork.getoBiases();
double[] outputGradients = backPropagationNetwork.getOutputGradients();
double[] hiddenGradients = backPropagationNetwork.getHiddenGradients();
double derivative;
double sum = 0.0;
double delta;
int numberOfOutputcells = backPropagationNetwork.getNumberOfOutputCells();
int numberOfHiddenCells = backPropagationNetwork.getNumberOfHiddenCells();
int numberOfInputCells = backPropagationNetwork.getNumberOfInputCells();
double learningRate = backPropagationNetwork.getLearningRate();
double momentum = backPropagationNetwork.getMomentum();
double[][] ihPreviousWeightsDelta = backPropagationNetwork.getIhPreviousWeightsDelta();
double[] hPreviousBiasesDelta = backPropagationNetwork.gethPreviousBiasesDelta();
double[][] hoPreviousWeightsDelta = backPropagationNetwork.getHoPreviousWeightsDelta();
double[] oPreviousBiasesDelta = backPropagationNetwork.getoPreviousBiasesDelta();
// endregion
// region Calculate gradients
for (int i = 0; i < numberOfOutputcells; i++) {
derivative = (1 - outputCells[i]) * outputCells[i];
outputGradients[i] = derivative * (targetValues[i] - outputCells[i]);
}
backPropagationNetwork.setOutputGradients(outputGradients);
for (int i = 0; i < numberOfHiddenCells; i++) {
derivative = (1 - tempHiddens[i]) * (1 + tempHiddens[i]);
for (int j = 0; j < backPropagationNetwork.getNumberOfOutputCells(); j++) {
sum += outputGradients[j] * hoWeights[i][j];
}
hiddenGradients[i] = derivative * sum;
}
backPropagationNetwork.setHiddenGradients(hiddenGradients);
// endregion
// region update weights & biases
for (int i = 0; i < numberOfInputCells; i++)
for (int j = 0; j < numberOfHiddenCells; j++) {
delta = learningRate * hiddenGradients[j] * inputCells[i];
ihWeights[i][j] += delta;
ihWeights[i][j] += momentum * ihPreviousWeightsDelta[i][j];
ihPreviousWeightsDelta[i][j] = delta;
}
backPropagationNetwork.setIhWeights(ihWeights);
backPropagationNetwork.setIhPreviousWeightsDelta(ihPreviousWeightsDelta);
for (int i = 0; i < numberOfHiddenCells; i++) {
delta = learningRate * hiddenGradients[i];
hiddenBiases[i] += delta;
hiddenBiases[i] += momentum * hPreviousBiasesDelta[i];
hPreviousBiasesDelta[i] = delta;
}
backPropagationNetwork.sethBiases(hiddenBiases);
backPropagationNetwork.sethPreviousBiasesDelta(hPreviousBiasesDelta);
for (int i = 0; i < numberOfHiddenCells; i++)
for (int j = 0; j < numberOfOutputcells; j++) {
delta = learningRate * outputGradients[j] * hiddenCells[i];
hoWeights[i][j] += delta;
hoWeights[i][j] += momentum * hoPreviousWeightsDelta[i][j];
hoPreviousWeightsDelta[i][j] = delta;
}
backPropagationNetwork.setHoWeights(hoWeights);
backPropagationNetwork.setHoPreviousWeightsDelta(hoPreviousWeightsDelta);
for (int i = 0; i < numberOfOutputcells; i++) {
delta = learningRate * outputGradients[i];
outputBiases[i] += delta;
outputBiases[i] += momentum * oPreviousBiasesDelta[i];
oPreviousBiasesDelta[i] = delta;
}
backPropagationNetwork.setoBiases(outputBiases);
backPropagationNetwork.setoPreviousBiasesDelta(oPreviousBiasesDelta);
// endregion
}
