/**
* Query a regression algorithm using equilateral encoding.
*
* @param alg The algorithm being used.
* @param theTrainingData The training data.
* @param items The category items classified.
* @param high The high value.
* @param low The low value.
*/
public static void queryEquilateral(
final RegressionAlgorithm alg,
final List<BasicData> theTrainingData,
final Map<String, Integer> items,
final double high,
final double low) {
// first, we need to invert the items. Right now it maps from category to index. We need index
// to category.
final Map<Integer, String> invMap = new HashMap<>();
for (final Map.Entry<String, Integer> entry : items.entrySet()) {
invMap.put(entry.getValue(), entry.getKey());
}
// now we can query
final Equilateral eq = new Equilateral(items.size(), high, low);
for (final BasicData data : theTrainingData) {
final double[] output = alg.computeRegression(data.getInput());
final int idealIndex = eq.decode(data.getIdeal());
final int actualIndex = eq.decode(output);
System.out.println(
Arrays.toString(data.getInput())
+ " -> "
+ invMap.get(actualIndex)
+ ", Ideal: "
+ invMap.get(idealIndex));
}
}
/** {@inheritDoc} */
@Override
public void iteration() {
this.network.setNetworkTraining(true);
// alert the layers that a new batch is starting.
for (Layer layer : this.network.getLayers()) {
layer.trainingBatch(this.stochastic);
}
// begin the iteration
this.gradients.reset();
this.errorCalc.clear();
int iterationSize =
this.batchSize == 0 ? this.training.size() : Math.min(this.batchSize, this.training.size());
for (int i = 0; i < iterationSize; i++) {
BasicData element;
if (isOnlineTraining()) {
if (this.stochastic != null) {
int stochasticIndex = this.stochastic.nextInt(0, this.training.size());
element = this.training.get(stochasticIndex);
} else {
element = this.training.get(this.currentIndex++);
}
} else {
element = this.training.get(i);
}
this.gradients.process(this.errorCalc, element.getInput(), element.getIdeal());
}
if (this.currentIndex > this.training.size() || this.batchSize == 0) {
this.currentIndex = 0;
}
this.currentError = this.errorCalc.calculate();
for (int i = 0; i < this.network.getWeights().length; i++) {
double delta;
if (this.nesterovUpdate) {
double prevNesterov = this.lastDelta[i];
this.lastDelta[i] =
(this.momentum * prevNesterov) + (this.gradients.getGradients()[i] * this.learningRate);
delta = (this.momentum * prevNesterov) - ((1 + this.momentum) * this.lastDelta[i]);
} else {
delta =
(this.gradients.getGradients()[i] * -this.learningRate)
+ (this.lastDelta[i] * this.momentum);
this.lastDelta[i] = delta;
}
this.network.getWeights()[i] += delta;
}
this.network.setNetworkTraining(false);
}
/**
* Query a regression algorithm and see how close it matches the training data.
*
* @param alg The algorithm to evaluate.
* @param theTrainingData The training data.
*/
public static void query(final RegressionAlgorithm alg, final List<BasicData> theTrainingData) {
for (final BasicData data : theTrainingData) {
final double[] output = alg.computeRegression(data.getInput());
System.out.println(
Arrays.toString(data.getInput())
+ " -> "
+ Arrays.toString(output)
+ ", Ideal: "
+ Arrays.toString(data.getIdeal()));
}
}