package com.etsy.conjecture.model;
import com.etsy.conjecture.data.LazyVector;
import com.etsy.conjecture.data.StringKeyedVector;
import static com.google.common.base.Preconditions.checkArgument;
import com.etsy.conjecture.Utilities;
import com.etsy.conjecture.data.LabeledInstance;
import java.util.Map;
import java.util.Iterator;
/**
* Current search ads control. Remove after current exp.
*/
public class ControlOptimizer extends SGDOptimizer {
private StringKeyedVector summedGradients = new StringKeyedVector();
@Override
public StringKeyedVector getUpdate(LabeledInstance instance) {
StringKeyedVector gradients = model.getGradients(instance);
StringKeyedVector updateVec = new StringKeyedVector();
Iterator<Map.Entry<String, Double>> it = gradients.iterator();
while (it.hasNext()) {
Map.Entry<String,Double> pairs = (Map.Entry)it.next();
String feature = pairs.getKey();
double gradient = pairs.getValue();
double featureLearningRate = updateAndGetFeatureLearningRate(feature, gradient);
updateVec.setCoordinate(feature, gradient * -featureLearningRate);
}
return updateVec;
}
/**
* Update adaptive feature specific learning rates
*/
public double updateAndGetFeatureLearningRate(String feature, double gradient) {
double gradUpdate = 0.0;
if (summedGradients.containsKey(feature)) {
gradUpdate = gradient * gradient;
} else {
/**
* Unmentioned in the literature, but initializing
* the squared gradient at 1.0 rather than 0.0
* helps avoid oscillation.
*/
gradUpdate = 1d+(gradient * gradient);
}
summedGradients.addToCoordinate(feature, gradUpdate);
return getFeatureLearningRate(feature);
}
public double getFeatureLearningRate(String feature) {
return initialLearningRate/Math.sqrt(summedGradients.getCoordinate(feature));
}
@Override
public void teardown() {
summedGradients = new StringKeyedVector();
}
}