/**
* Copyright 2006 DFKI GmbH.
* All Rights Reserved. Use is subject to license terms.
*
* This file is part of MARY TTS.
*
* MARY TTS is free software: you can redistribute it and/or modify
* it under the terms of the GNU Lesser General Public License as published by
* the Free Software Foundation, version 3 of the License.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public License
* along with this program. If not, see <http://www.gnu.org/licenses/>.
*
*/
package marytts.unitselection.select;
import java.io.BufferedReader;
import java.io.IOException;
import java.io.InputStream;
import java.io.InputStreamReader;
import marytts.exceptions.MaryConfigurationException;
import marytts.features.FeatureDefinition;
import marytts.features.FeatureProcessorManager;
import marytts.features.FeatureVector;
import marytts.features.TargetFeatureComputer;
import marytts.server.MaryProperties;
import marytts.signalproc.display.Histogram;
import marytts.unitselection.data.FeatureFileReader;
import marytts.unitselection.data.Unit;
import marytts.unitselection.weightingfunctions.WeightFunc;
import marytts.unitselection.weightingfunctions.WeightFunctionManager;
import marytts.util.MaryUtils;
public class FFRTargetCostFunction implements TargetCostFunction {
protected WeightFunc[] weightFunction;
protected TargetFeatureComputer targetFeatureComputer;
protected FeatureVector[] featureVectors;
protected FeatureDefinition featureDefinition;
protected boolean[] weightsNonZero;
protected boolean debugShowCostGraph = false;
protected double[] cumulWeightedCosts = null;
protected int nCostComputations = 0;
public FFRTargetCostFunction() {
}
/**
* Compute the goodness-of-fit of a given unit for a given target.
*
* @param target
* target
* @param unit
* unit
* @return a non-negative number; smaller values mean better fit, i.e. smaller cost.
*/
public double cost(Target target, Unit unit) {
return cost(target, unit, featureDefinition, weightFunction);
}
protected double cost(Target target, Unit unit, FeatureDefinition weights, WeightFunc[] weightFunctions) {
nCostComputations++; // for debug
FeatureVector targetFeatures = target.getFeatureVector();
assert targetFeatures != null : "Target " + target + " does not have pre-computed feature vector";
FeatureVector unitFeatures = featureVectors[unit.index];
int nBytes = targetFeatures.byteValuedDiscreteFeatures.length;
int nShorts = targetFeatures.shortValuedDiscreteFeatures.length;
int nFloats = targetFeatures.continuousFeatures.length;
assert nBytes == unitFeatures.byteValuedDiscreteFeatures.length;
assert nShorts == unitFeatures.shortValuedDiscreteFeatures.length;
assert nFloats == unitFeatures.continuousFeatures.length;
float[] weightVector = weights.getFeatureWeights();
// Now the actual computation
double cost = 0;
// byte-valued features:
if (nBytes > 0) {
for (int i = 0; i < nBytes; i++) {
if (weightsNonZero[i]) {
float weight = weightVector[i];
if (featureDefinition.hasSimilarityMatrix(i)) {
byte targetFeatValueIndex = targetFeatures.byteValuedDiscreteFeatures[i];
byte unitFeatValueIndex = unitFeatures.byteValuedDiscreteFeatures[i];
float similarity = featureDefinition.getSimilarity(i, unitFeatValueIndex, targetFeatValueIndex);
cost += similarity * weight;
if (debugShowCostGraph)
cumulWeightedCosts[i] += similarity * weight;
} else if (targetFeatures.byteValuedDiscreteFeatures[i] != unitFeatures.byteValuedDiscreteFeatures[i]) {
cost += weight;
if (debugShowCostGraph)
cumulWeightedCosts[i] += weight;
}
}
}
}
// short-valued features:
if (nShorts > 0) {
for (int i = nBytes, n = nBytes + nShorts; i < n; i++) {
if (weightsNonZero[i]) {
float weight = weightVector[i];
// if (targetFeatures.getShortFeature(i) != unitFeatures.getShortFeature(i)) {
if (targetFeatures.shortValuedDiscreteFeatures[i - nBytes] != unitFeatures.shortValuedDiscreteFeatures[i
- nBytes]) {
cost += weight;
if (debugShowCostGraph)
cumulWeightedCosts[i] += weight;
}
}
}
}
// continuous features:
if (nFloats > 0) {
int nDiscrete = nBytes + nShorts;
for (int i = nDiscrete, n = nDiscrete + nFloats; i < n; i++) {
if (weightsNonZero[i]) {
float weight = weightVector[i];
// float a = targetFeatures.getContinuousFeature(i);
float a = targetFeatures.continuousFeatures[i - nDiscrete];
// float b = unitFeatures.getContinuousFeature(i);
float b = unitFeatures.continuousFeatures[i - nDiscrete];
// if (!Float.isNaN(a) && !Float.isNaN(b)) {
// Implementation of isNaN() is: (v != v).
if (!(a != a) && !(b != b)) {
double myCost = weightFunctions[i - nDiscrete].cost(a, b);
cost += weight * myCost;
if (debugShowCostGraph) {
cumulWeightedCosts[i] += weight * myCost;
}
} // and if it is NaN, simply compute no cost
}
}
}
return cost;
}
/**
* Compute the goodness-of-fit between given unit and given target for a given feature
*
* @param target
* target unit
* @param unit
* candidate unit
* @param featureName
* feature name
* @return a non-negative number; smaller values mean better fit, i.e. smaller cost.
* @throws IllegalArgumentException
* if featureName not available in featureDefinition
*/
public double featureCost(Target target, Unit unit, String featureName) {
return featureCost(target, unit, featureName, featureDefinition, weightFunction);
}
protected double featureCost(Target target, Unit unit, String featureName, FeatureDefinition weights,
WeightFunc[] weightFunctions) {
if (!this.featureDefinition.hasFeature(featureName)) {
throw new IllegalArgumentException("this feature does not exists in feature definition");
}
FeatureVector targetFeatures = target.getFeatureVector();
assert targetFeatures != null : "Target " + target + " does not have pre-computed feature vector";
FeatureVector unitFeatures = featureVectors[unit.index];
int nBytes = targetFeatures.byteValuedDiscreteFeatures.length;
int nShorts = targetFeatures.shortValuedDiscreteFeatures.length;
int nFloats = targetFeatures.continuousFeatures.length;
assert nBytes == unitFeatures.byteValuedDiscreteFeatures.length;
assert nShorts == unitFeatures.shortValuedDiscreteFeatures.length;
assert nFloats == unitFeatures.continuousFeatures.length;
int featureIndex = this.featureDefinition.getFeatureIndex(featureName);
float[] weightVector = weights.getFeatureWeights();
double cost = 0;
if (featureIndex < nBytes) {
if (weightsNonZero[featureIndex]) {
float weight = weightVector[featureIndex];
if (featureDefinition.hasSimilarityMatrix(featureIndex)) {
byte targetFeatValueIndex = targetFeatures.byteValuedDiscreteFeatures[featureIndex];
byte unitFeatValueIndex = unitFeatures.byteValuedDiscreteFeatures[featureIndex];
float similarity = featureDefinition.getSimilarity(featureIndex, unitFeatValueIndex, targetFeatValueIndex);
cost = similarity * weight;
if (debugShowCostGraph)
cumulWeightedCosts[featureIndex] += similarity * weight;
} else if (targetFeatures.byteValuedDiscreteFeatures[featureIndex] != unitFeatures.byteValuedDiscreteFeatures[featureIndex]) {
cost = weight;
if (debugShowCostGraph)
cumulWeightedCosts[featureIndex] += weight;
}
}
} else if (featureIndex < nShorts + nBytes) {
if (weightsNonZero[featureIndex]) {
float weight = weightVector[featureIndex];
// if (targetFeatures.getShortFeature(i) != unitFeatures.getShortFeature(i)) {
if (targetFeatures.shortValuedDiscreteFeatures[featureIndex - nBytes] != unitFeatures.shortValuedDiscreteFeatures[featureIndex
- nBytes]) {
cost = weight;
if (debugShowCostGraph)
cumulWeightedCosts[featureIndex] += weight;
}
}
} else {
int nDiscrete = nBytes + nShorts;
if (weightsNonZero[featureIndex]) {
float weight = weightVector[featureIndex];
// float a = targetFeatures.getContinuousFeature(i);
float a = targetFeatures.continuousFeatures[featureIndex - nDiscrete];
// float b = unitFeatures.getContinuousFeature(i);
float b = unitFeatures.continuousFeatures[featureIndex - nDiscrete];
// if (!Float.isNaN(a) && !Float.isNaN(b)) {
// Implementation of isNaN() is: (v != v).
if (!(a != a) && !(b != b)) {
double myCost = weightFunctions[featureIndex - nDiscrete].cost(a, b);
cost = weight * myCost;
if (debugShowCostGraph) {
cumulWeightedCosts[featureIndex] += weight * myCost;
}
} // and if it is NaN, simply compute no cost
}
}
return cost;
}
/**
* Initialise the data needed to do a target cost computation.
*
* @param featureFileName
* name of a file containing the unit features
* @param weightsStream
* an optional weights file -- if non-null, contains feature weights that override the ones present in the feature
* file.
* @param featProc
* a feature processor manager which can provide feature processors to compute the features for a target at run
* time
* @throws IOException
* IOException
* @throws MaryConfigurationException
* MaryConfigurationException
*/
@Override
public void load(String featureFileName, InputStream weightsStream, FeatureProcessorManager featProc) throws IOException,
MaryConfigurationException {
FeatureFileReader ffr = FeatureFileReader.getFeatureFileReader(featureFileName);
load(ffr, weightsStream, featProc);
}
@Override
public void load(FeatureFileReader ffr, InputStream weightsStream, FeatureProcessorManager featProc) throws IOException {
this.featureDefinition = ffr.getFeatureDefinition();
this.featureVectors = ffr.getFeatureVectors();
if (weightsStream != null) {
MaryUtils.getLogger("TargetCostFeatures").debug("Overwriting target cost weights from file");
// overwrite weights from file
FeatureDefinition newWeights = new FeatureDefinition(
new BufferedReader(new InputStreamReader(weightsStream, "UTF-8")), true);
if (!newWeights.featureEquals(featureDefinition)) {
throw new IOException("Weights file: feature definition incompatible with feature file");
}
featureDefinition = newWeights;
}
weightFunction = new WeightFunc[featureDefinition.getNumberOfContinuousFeatures()];
WeightFunctionManager wfm = new WeightFunctionManager();
int nDiscreteFeatures = featureDefinition.getNumberOfByteFeatures() + featureDefinition.getNumberOfShortFeatures();
for (int i = 0; i < weightFunction.length; i++) {
String weightFunctionName = featureDefinition.getWeightFunctionName(nDiscreteFeatures + i);
if ("".equals(weightFunctionName))
weightFunction[i] = wfm.getWeightFunction("linear");
else
weightFunction[i] = wfm.getWeightFunction(weightFunctionName);
}
// TODO: If the target feature computer had direct access to the feature definition, it could do some consistency checking
this.targetFeatureComputer = new TargetFeatureComputer(featProc, featureDefinition.getFeatureNames());
rememberWhichWeightsAreNonZero();
if (MaryProperties.getBoolean("debug.show.cost.graph")) {
debugShowCostGraph = true;
cumulWeightedCosts = new double[featureDefinition.getNumberOfFeatures()];
TargetCostReporter tcr2 = new TargetCostReporter(cumulWeightedCosts);
tcr2.showInJFrame("Average weighted target costs", false, false);
tcr2.start();
}
}
protected void rememberWhichWeightsAreNonZero() {
// remember which weights are non-zero
weightsNonZero = new boolean[featureDefinition.getNumberOfFeatures()];
for (int i = 0, n = featureDefinition.getNumberOfFeatures(); i < n; i++) {
weightsNonZero[i] = (featureDefinition.getWeight(i) > 0);
}
}
/**
* Compute the features for a given target, and store them in the target.
*
* @param target
* the target for which to compute the features
* @see Target#getFeatureVector()
*/
public void computeTargetFeatures(Target target) {
FeatureVector fv = targetFeatureComputer.computeFeatureVector(target);
target.setFeatureVector(fv);
}
/**
* Look up the features for a given unit.
*
* @param unit
* a unit in the database
* @return the FeatureVector for target cost computation associated to this unit
*/
public FeatureVector getFeatureVector(Unit unit) {
return featureVectors[unit.index];
}
/**
* Get the string representation of the feature value associated with the given unit
*
* @param unit
* the unit whose feature value is requested
* @param featureName
* name of the feature requested
* @return a string representation of the feature value
* @throws IllegalArgumentException
* if featureName is not a known feature
*/
public String getFeature(Unit unit, String featureName) {
int featureIndex = featureDefinition.getFeatureIndex(featureName);
if (featureDefinition.isByteFeature(featureIndex)) {
byte value = featureVectors[unit.index].getByteFeature(featureIndex);
return featureDefinition.getFeatureValueAsString(featureIndex, value);
} else if (featureDefinition.isShortFeature(featureIndex)) {
short value = featureVectors[unit.index].getShortFeature(featureIndex);
return featureDefinition.getFeatureValueAsString(featureIndex, value);
} else { // continuous -- return float as string
float value = featureVectors[unit.index].getContinuousFeature(featureIndex);
return String.valueOf(value);
}
}
public FeatureDefinition getFeatureDefinition() {
return featureDefinition;
}
public class TargetCostReporter extends Histogram {
private double[] data;
private int lastN = 0;
public TargetCostReporter(double[] data) {
super(0, 1, data);
this.data = data;
}
public void start() {
new Thread() {
public void run() {
while (isVisible()) {
try {
Thread.sleep(500);
} catch (InterruptedException ie) {
}
updateGraph();
}
}
}.start();
}
protected void updateGraph() {
if (nCostComputations == lastN)
return;
lastN = nCostComputations;
double[] newCosts = new double[data.length];
for (int i = 0; i < newCosts.length; i++) {
newCosts[i] = data[i] / nCostComputations;
}
updateData(0, 1, newCosts);
repaint();
}
}
public FeatureVector[] getFeatureVectors() {
return featureVectors;
}
}