marytts.tools.dbselection.CoverageDefinition Maven / Gradle / Ivy
The newest version!
/**
* Copyright 2007 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 possiblePhoneTypes;
/*
* keep track of the coverage development over time by adding a the current coverage value each time the cover is updated
*/
private List phoneCoverageInTime;
private List diphoneCoverageInTime;
private List overallCoverageInTime;
/* set of covered phones/simple diphones/clustered diphones */
private Set phonesInCover;
private Set simpleDiphonesInCover;
/* number of simple prosodic variations in cover */
private int numSimpleFeatVectsInCover;
/* the featureDefinition for the feature vectors */
private final FeatureDefinition featDef;
/* the number of sentences in the corpus */
private int numSentences;
/* the phones that are not in the corpus and have to be ignored */
private Set phonesToIgnore;
/* the possible phone values */
private String[] possiblePhoneArray;
/* the possible next phone values */
private String[] possibleNextPhoneArray;
/* the possible next phone values */
private String[] possibleProsodyArray;
private String[][] possibleDiphones;
private String[][][] possibleDiphonesProsody;
/* For printing statistics, count the number of occurrences of each diphone in the text corpus */
private int[][] diphoneFrequencies;
private int[] phoneFrequencies;
private CoverageFeatureProvider cfProvider;
/**
* Build a new coverage definition and read in the config file
*
*
*
* @param featDef
* the feature definition for the vectors
* @param cfProvider
* coverage feature provider
* @param configFile
* optionally, the coverage config file name. if this is null, default settings will be used.
* @throws Exception
* Exception
*/
public CoverageDefinition(FeatureDefinition featDef, CoverageFeatureProvider cfProvider, String configFile) throws Exception {
this.featDef = featDef;
this.cfProvider = cfProvider;
readConfigFile(featDef, configFile);
setupFeatureIndexes();
initializeVariables();
}
/**
* @param featDef
* featDef
* @param configFile
* configFile
* @throws Exception
* Exception
*/
private void readConfigFile(FeatureDefinition featDef, String configFile) throws Exception {
try {
InputStream inStream;
if (configFile != null) {
inStream = new FileInputStream(new File(configFile));
} else {
inStream = getClass().getResourceAsStream("covDef.config");
}
BufferedReader configIn = new BufferedReader(new InputStreamReader(inStream, "UTF-8"));
String line;
int numparams = 0;
// loop over the lines of the config file
while ((line = configIn.readLine()) != null) {
if (!line.startsWith("#") && !line.equals("")) {
StringTokenizer tok = new StringTokenizer(line);
String key = tok.nextToken();
String value = tok.nextToken();
if (key.equals("simpleDiphones")) {
if (value.equals("true")) {
simpleDiphones = true;
} else {
simpleDiphones = false;
}
numparams++;
continue;
}
if (key.equals("frequency")) {
if (value.equals("none")) {
considerFrequency = false;
} else {
considerFrequency = true;
frequencySetting = value;
}
numparams++;
continue;
}
if (key.equals("sentenceLength")) {
if (value.equals("none")) {
considerSentenceLength = false;
} else {
considerSentenceLength = true;
maxSentLengthAllowed = Integer.parseInt(value);
minSentLengthAllowed = Integer.parseInt(tok.nextToken());
}
numparams++;
continue;
}
if (key.equals("wantedWeight")) {
phoneLevelWeight = Double.parseDouble(value);
diphoneLevelWeight = Double.parseDouble(tok.nextToken());
prosodyLevelWeight = Double.parseDouble(tok.nextToken());
numparams++;
continue;
}
if (key.equals("wantedWeightDecrease")) {
wantedWeightDecrease = Double.parseDouble(value);
numparams++;
continue;
}
if (key.equals("missingPhones")) {
phonesToIgnore = new HashSet();
phonesToIgnore.add(0); // The non-existing phone "0"
// phoneFeatIndex = featDef.getFeatureIndex("phone");
// phonesToIgnore.add(
// new Integer(featDef.getFeatureValueAsByte(phoneFeatIndex,value)));
while (tok.hasMoreTokens()) {
phonesToIgnore.add(new Integer(featDef.getFeatureValueAsByte(phoneFeatIndex, tok.nextToken())));
}
numparams++;
}
}
}
if (numparams < 6) {
throw new Exception("Error reading coverage Definition Config File: " + configFile + " there are only "
+ numparams + " instead of 6 settings");
}
} catch (Exception e) {
e.printStackTrace();
throw new Exception("Could not read coverage Definition Config File: " + configFile);
}
}
/**
*
*/
private void setupFeatureIndexes() {
System.out.println("TARGETFEATURES used:");
numTargetFeaturesUsed = 0;
for (int i = 0; i < featDef.getNumberOfFeatures(); i++) {
if (featDef.getFeatureName(i).contentEquals("phone")) {
phoneFeatIndex = featDef.getFeatureIndex("phone");
numTargetFeaturesUsed++;
System.out.println(" feature(" + i + ")=" + featDef.getFeatureName(i));
} else if (featDef.getFeatureName(i).contentEquals("next_phone")) {
diphoneFeatIndex = featDef.getFeatureIndex("next_phone");
numTargetFeaturesUsed++;
System.out.println(" feature(" + i + ")=" + featDef.getFeatureName(i));
} else if (featDef.getFeatureName(i).contentEquals("selection_prosody")) {
prosodyIndex = featDef.getFeatureIndex("selection_prosody");
numTargetFeaturesUsed++;
System.out.println(" feature(" + i + ")=" + featDef.getFeatureName(i));
} else
System.out.println(" NO implementation in CoverageDefinition for the feature =" + featDef.getFeatureName(i));
}
numPhoneValues = featDef.getNumberOfValues(phoneFeatIndex);
numPhoneValuesMinusIgnored = numPhoneValues - phonesToIgnore.size() - 1;
numPossibleSimpleDiphones = numPhoneValuesMinusIgnored * (numPhoneValuesMinusIgnored + 1);
numProsodyValues = featDef.getNumberOfValues(prosodyIndex);
possiblePhoneArray = featDef.getPossibleValues(phoneFeatIndex);
possibleNextPhoneArray = featDef.getPossibleValues(diphoneFeatIndex);
possibleProsodyArray = featDef.getPossibleValues(prosodyIndex);
// For efficiency, we build all strings once -- much better than doing it all the time:
possibleDiphones = new String[possiblePhoneArray.length][possibleNextPhoneArray.length];
possibleDiphonesProsody = new String[possiblePhoneArray.length][possibleNextPhoneArray.length][possibleProsodyArray.length];
for (int i = 0; i < possiblePhoneArray.length; i++) {
for (int j = 0; j < possibleNextPhoneArray.length; j++) {
String diphone = possiblePhoneArray[i] + "_" + possibleNextPhoneArray[j];
possibleDiphones[i][j] = diphone;
for (int k = 0; k < possibleProsodyArray.length; k++) {
possibleDiphonesProsody[i][j][k] = diphone + "_" + possibleProsodyArray[k];
}
}
}
}
/**
*
*/
private void initializeVariables() {
// initialise several variables
numSelectedFeatVects = 0;
numSentencesInCover = 0;
maxSentLengthInCover = 0;
minSentLengthInCover = 20;
phoneCoverageInTime = new ArrayList();
diphoneCoverageInTime = new ArrayList();
overallCoverageInTime = new ArrayList();
phonesInCover = new HashSet();
simpleDiphonesInCover = new HashSet();
numSimpleFeatVectsInCover = 0;
}
/**
* Compute the coverage of the corpus, build and fill the cover sets. This will iterate once through the entire corpus, to
* compute the maximally achievable coverage with this corpus.
*
* @throws IOException
* IOException
*/
public void initialiseCoverage() throws IOException {
// stuff used for counting the phones and diphones
possiblePhoneTypes = new HashSet();
Set simpleFeatVectTypes = new HashSet();
int numPhoneTypes = 0;
numSimpleDiphoneTypes = 0;
numTokens = 0;
averageSentLength = 0.0;
maxSentLength = 0;
minSentLength = 20;
phoneFrequencies = new int[possiblePhoneArray.length];
diphoneFrequencies = new int[possiblePhoneArray.length][possibleNextPhoneArray.length];
// build Cover
buildCover();
numSentences = cfProvider.getNumSentences();
int tenPercent = Math.max(numSentences / 10, 1);
// loop over the feature vectors
System.out.println("\nAnalysing feature vectors of " + numSentences + " sentences:");
for (int index = 0; index < numSentences; index++) {
if ((index % tenPercent) == 0 && index != 0) {
int percentage = index / tenPercent;
System.out.print(" " + percentage + "0% ");
}
// for each vector, get the values for the relevant features
// add them to the list of possible values
byte[] features = cfProvider.getCoverageFeatures(index); // the feature vectors of one sentence
if (features == null) {
System.err.println("WARNING: null features for sentence id " + cfProvider.getID(index) + " -- skipping");
continue;
}
int numFeatVects = features.length / numTargetFeaturesUsed;
// compute statistics of sentence length
averageSentLength += numFeatVects;
if (numFeatVects > maxSentLength)
maxSentLength = numFeatVects;
if (numFeatVects < minSentLength)
minSentLength = numFeatVects;
// System.out.println("Analysing feature vectors of sentence id " + cfProvider.getID(index) + ": numFeatVects=" +
// numFeatVects);
// Loop over all feature vectors in current sentence:
for (int i = 0; i < numFeatVects; i++) {
numTokens++;
// first deal with current phone
// byte nextPhonebyte = getVectorValue(vectorBuf,i,phoneFeatIndex);
byte nextPhonebyte = features[i * numTargetFeaturesUsed + phoneFeatIndex];
// System.out.println("i=" + i + " phone="
// + featDef.getFeatureValueAsString(myphoneFeatIndex,nextPhonebyte));
// add 1 to the frequency value of the phone
phoneFrequencies[nextPhonebyte]++;
// deal with current diphone
// byte nextnextPhonebyte = getVectorValue(vectorBuf,i,diphoneFeatIndex);
byte nextnextPhonebyte = features[i * numTargetFeaturesUsed + diphoneFeatIndex];
String simpleDiphone = possibleDiphones[nextPhonebyte][nextnextPhonebyte];
// add 1 to the frequency value of the diphone
diphoneFrequencies[nextPhonebyte][nextnextPhonebyte]++;
// deal with current diphone
// byte prosodyValue = getVectorValue(vectorBuf,i,prosodyIndex);
byte prosodyValue = features[i * numTargetFeaturesUsed + prosodyIndex];
simpleFeatVectTypes.add(possibleDiphonesProsody[nextPhonebyte][nextnextPhonebyte][prosodyValue]);
// save feature vector in simple diphone tree
// CoverLeaf leaf = goDownTree(vectorBuf,i,false);
// leaf.addPossibleInstance();
CoverLeaf leaf = (CoverLeaf) simpleCover.children[nextPhonebyte].children[nextnextPhonebyte].children[prosodyValue];
leaf.maxNumFeatVects++;
}
}
System.out.println(" 100% ");
// count phones
numPhoneTypes = getPhonesInCorpus().size();
// count diphones
numSimpleDiphoneTypes = 0;
for (int i = 0; i < possibleDiphones.length; i++) {
for (int j = 0; j < possibleDiphones[i].length; j++) {
if (diphoneFrequencies[i][j] > 0) {
numSimpleDiphoneTypes++;
}
}
}
// compute average sentence length
averageSentLength = averageSentLength / (double) numSentences;
// calculate cover size
numPossibleSimpleDiphones = numPhoneValuesMinusIgnored * (numPhoneValuesMinusIgnored + 1);
numSimpleLeaves = numPossibleSimpleDiphones * numProsodyValues;
// number of feature vector types
numSimpleFeatVectTypes = simpleFeatVectTypes.size();
// compute coverage of corpus
possiblePhoneCoverage = (double) numPhoneTypes / (double) (numPhoneValuesMinusIgnored);
possibleSimpleDiphoneCoverage = numSimpleDiphoneTypes / (double) numPossibleSimpleDiphones;
possibleOverallSimpleCoverage = (double) numSimpleFeatVectTypes / (double) numSimpleLeaves;
// calculate relative frequency for each node
// rel. freq. = freq / all tokens
if (simpleDiphones) {
computeRelativeFrequency(simpleCover, numTokens);
}
}
/**
* Build the trees that represent the cover sets
*
*/
private void buildCover() {
simpleCover = new CoverNode((byte) numPhoneValues, wantedWeightDecrease);
// compute all possible combinations
for (int k = 0; k < possiblePhoneArray.length; k++) {
if (phonesToIgnore.contains(new Integer(k)))
continue;
// find out the index of the current phone
byte nextIndex = (byte) k;
// add a node for the phonetic identity of the next phone
CoverNode nextSimpleChild = new CoverNode((byte) numPhoneValues, wantedWeightDecrease);
// set the weight that determines how many instances
// are wanted of this phone
nextSimpleChild.setWantedWeight(phoneLevelWeight);
simpleCover.addChild(nextSimpleChild, nextIndex);
byte numGrandChildren = nextSimpleChild.getNumChildren();
// go through the grandchildren of simpleCover
for (byte i = 0; i < numGrandChildren; i++) {
// each grandchild is a prosody node
CoverNode prosodyNode = new CoverNode((byte) numProsodyValues, wantedWeightDecrease);
// set the weight that determines how many instances
// are wanted of this diphone
prosodyNode.setWantedWeight(diphoneLevelWeight);
nextSimpleChild.addChild(prosodyNode, i);
// go through the children of the prosody node
for (byte j = 0; j < numProsodyValues; j++) {
// each child is a leaf
CoverLeaf prosodyChild = new CoverLeaf(wantedWeightDecrease);
// set the weight that determines how many instances
// are wanted of this prosody variation
prosodyChild.setWantedWeight(prosodyLevelWeight);
prosodyNode.addChild(prosodyChild, j);
}
}
}
}
/**
* Get descriptive statistics for the full corpus. These values are independent of the selection of any sentences.
*
* @return cs
*/
public CoverageStatistics getCorpusStatistics() {
CoverageStatistics cs = new CoverageStatistics();
cs.coveredPhones = getPhonesInCorpus();
cs.allPhones = getAllPhones();
cs.coveredDiphones = getDiphonesInCorpus();
cs.numPossibleDiphones = numPossibleSimpleDiphones;
cs.numCoveredDiphonesWithProsody = numSimpleFeatVectTypes;
cs.numPossibleDiphonesWithProsody = numSimpleLeaves;
cs.numTokens = numTokens;
return cs;
}
public Set getPhonesInCorpus() {
Set phones = new TreeSet();
int phoneFeatureIndex = featDef.getFeatureIndex("phone");
for (int i = 0; i < phoneFrequencies.length; i++) {
if (phoneFrequencies[i] > 0) {
phones.add(featDef.getFeatureValueAsString(phoneFeatureIndex, i));
}
}
return phones;
}
public Set getAllPhones() {
Set phones = new TreeSet();
int phoneFeatureIndex = featDef.getFeatureIndex("phone");
int i = 0;
for (String ph : featDef.getPossibleValues(phoneFeatureIndex)) {
if (!phonesToIgnore.contains(i)) {
phones.add(ph);
}
i++;
}
return phones;
}
public Set getDiphonesInCorpus() {
Set diphones = new HashSet();
for (int i = 0; i < phoneFrequencies.length; i++) {
for (int j = 0; j < phoneFrequencies.length; j++) {
if (diphoneFrequencies[i][j] > 0) {
diphones.add(possibleDiphones[i][j]);
}
}
}
return diphones;
}
/**
* Go down the cover tree according to the values in the feature vector
*
* @param simpleDiphones
* if true, go down simple cover tree, else go down clustered cover tree
* @param vectors
* the feature vectors
* @param index
* the index of the current feature vector
* @param addNewFeatureVector
* if true, decrease wantedWeights of the nodes you pass
* @return the leaf that you arrived at
*/
private CoverLeaf goDownTree(byte[] vectors, int index, boolean addNewFeatureVector) {
// go down to phone level
// byte nextIndex = getVectorValue(vectors,index,phoneFeatIndex);
byte nextIndex = vectors[index * numTargetFeaturesUsed + phoneFeatIndex];
// CoverNode nextNode = simpleCover.getChild(nextIndex);
CoverNode nextNode = simpleCover.children[nextIndex];
if (addNewFeatureVector)
nextNode.decreaseWantedWeight();
// go down to diphone level
// nextIndex = getVectorValue(vectors,index,diphoneFeatIndex);
nextIndex = vectors[index * numTargetFeaturesUsed + diphoneFeatIndex];
// nextNode = nextNode.getChild(nextIndex);
nextNode = nextNode.children[nextIndex];
if (addNewFeatureVector)
nextNode.decreaseWantedWeight();
// go down to prosody level
// nextIndex = getVectorValue(vectors,index,prosodyIndex);
nextIndex = vectors[index * numTargetFeaturesUsed + prosodyIndex];
// nextNode = nextNode.getChild(nextIndex);
nextNode = nextNode.children[nextIndex];
if (addNewFeatureVector)
nextNode.decreaseWantedWeight();
if (!(nextNode instanceof CoverLeaf)) {
// something went wrong
throw new Error("Went down cover tree for feature vector" + " and did not end up on leaf!");
}
return (CoverLeaf) nextNode;
}
/**
* Compute the relative frequency of each node in the corpus
*
* @param node
* the node to compute the frequency for
* @param allTokens
* total number of tokens in the corpus
* @return the frequency for the given node
*/
private double computeRelativeFrequency(CoverNode node, double allTokens) {
double freq = 0;
if (node instanceof CoverLeaf) {
// compute the relative frequency for this leaf
int numPossInstances = ((CoverLeaf) node).maxNumFeatVects();
freq = (double) numPossInstances / allTokens;
if (considerFrequency) {
if (frequencySetting.equals("1minus")) {
node.setFrequencyWeight(1 - freq);
} else {
if (frequencySetting.equals("inverse")) {
node.setFrequencyWeight(1 / freq);
} else {
node.setFrequencyWeight(freq);
}
}
}
} else {
// frequency is the sum of the frequency of the children
byte numChildren = node.getNumChildren();
// go through children
for (byte i = 0; i < numChildren; i++) {
CoverNode child = node.getChild(i);
if (child == null)
continue;
freq += computeRelativeFrequency(child, allTokens);
}
if (considerFrequency) {
if (frequencySetting.equals("1minus")) {
node.setFrequencyWeight(1 - freq);
} else {
if (frequencySetting.equals("inverse")) {
node.setFrequencyWeight(1 / freq);
} else {
node.setFrequencyWeight(freq);
}
}
}
}
return freq;
}
/**
* Print a statistic of the unit distribution in the corpus
*
* @param out
* the print writer to print to
* @throws Exception
* Exception
*/
public void printTextCorpusStatistics(PrintWriter out) throws Exception {
DecimalFormat df = new DecimalFormat("0.00000");
out.println("*********************" + "\n* Unit distribution *" + "\n*********************\n\n");
/* print out the sentence length statistics */
out.println("Number of sentences : " + numSentences);
out.println("Average sentence length : " + averageSentLength);
out.println("Maximum sentence length : " + maxSentLength);
out.println("Minimum sentence length : " + minSentLength);
/* print out coverage statistics */
CoverageStatistics stats = getCorpusStatistics();
out.println(stats);
/*
* out.println("\nClustered Coverage:"); out.println("phones: "+df.format(possiblePhoneCoverage));
* out.println("diphones: "+df.format(possibleClusteredDiphoneCoverage));
* out.println("overall: "+df.format(possibleOverallClusteredCoverage)+"\n\n");
*/
if (possiblePhoneCoverage < 1) {
out.println("The following phones are missing: ");
for (int k = 1; k < possiblePhoneArray.length; k++) {
String nextPhone = possiblePhoneArray[k];
if (phonesToIgnore.contains(new Integer(k)))
continue;
if (!possiblePhoneTypes.contains(nextPhone)) {
out.print(nextPhone + " ");
}
}
out.print("\n");
}
out.println("\n\nDiphones and their frequencies :\n");
out.println("Simple diphones:\n");
printDiphones(out);
out.flush();
out.close();
}
/**
* Print the settings of the config file
*
* @param out
* the PrintWriter to print to
*/
public void printSettings(PrintWriter out) {
/* print out setings */
out.println("\nSettings of Coverage Definition:");
out.println("simpleDiphones " + Boolean.toString(simpleDiphones));
if (considerFrequency) {
out.println("frequency " + frequencySetting);
} else {
out.println("frequency none");
}
out.println("considerSentenceLength " + Boolean.toString(considerSentenceLength));
out.println("phoneLevelWeight " + phoneLevelWeight);
out.println("diphoneLevelWeight " + diphoneLevelWeight);
out.println("prosodyLevelWeight " + prosodyLevelWeight);
out.println("divideWantedWeightBy " + wantedWeightDecrease);
if (considerSentenceLength) {
out.println("maxSentenceLength " + maxSentLengthAllowed);
out.println("minSentenceLength " + minSentLengthAllowed);
}
}
/**
* Print the diphone distribution of the corpus
*
* @param out
* the PrintWriter to print to
* @param ph2Frequency
* maps from diphones to their frequency
*/
private void printDiphones(PrintWriter out) {
DecimalFormat df = new DecimalFormat("0.00000");
// Sort phones according to their frequencies
TreeMap> freq2Diphones = new TreeMap>(Collections.reverseOrder());
Map freq2Prob = new HashMap();
for (int i = 0; i < possibleDiphones.length; i++) {
for (int j = 0; j < possibleDiphones[i].length; j++) {
String diphone = possibleDiphones[i][j];
int freq = diphoneFrequencies[i][j];
if (!freq2Diphones.containsKey(freq)) {
List phoneList = new ArrayList();
phoneList.add(diphone);
freq2Diphones.put(freq, phoneList);
double prob = (double) freq * 100.0 / (double) numTokens;
freq2Prob.put(freq, prob);
} else {
List phoneList = freq2Diphones.get(freq);
phoneList.add(diphone);
}
}
}
// output phones and their frequencies
Set frequencies = freq2Diphones.keySet();
for (Integer nextFreq : frequencies) {
Double nextProb = freq2Prob.get(nextFreq);
List nextPhoneList = freq2Diphones.get(nextFreq);
for (int i = 0; i < nextPhoneList.size(); i++) {
out.print(nextPhoneList.get(i));
out.print(" : ");
out.print(nextFreq);
out.print(", ");
out.println(df.format(nextProb));
}
}
}
/**
* Print statistics of the selected sentences and a table of coverage development over time
*
* @param distributionFile
* the file to print the statistics to
* @param developmentFile
* the file to print the coverage development to
* @param logDevelopment
* if true, print development file
* @throws Exception
* Exception
*/
public void printSelectionDistribution(String distributionFile, String developmentFile, boolean logDevelopment)
throws Exception {
PrintWriter disOut = new PrintWriter(new FileWriter(new File(distributionFile)));
/* print settings */
DecimalFormat df = new DecimalFormat("0.00000");
disOut.println("\nSettings of Coverage Definition:");
disOut.println("simpleDiphones " + Boolean.toString(simpleDiphones));
if (considerFrequency) {
disOut.println("frequency " + frequencySetting);
} else {
disOut.println("frequency none");
}
disOut.println("considerSentenceLength " + Boolean.toString(considerSentenceLength));
disOut.println("phoneLevelWeight " + phoneLevelWeight);
disOut.println("diphoneLevelWeight " + diphoneLevelWeight);
disOut.println("prosodyLevelWeight " + prosodyLevelWeight);
disOut.println("maxSentenceLength " + maxSentLengthAllowed);
disOut.println("minSentenceLength " + minSentLengthAllowed);
disOut.println("divideWantedWeightBy " + wantedWeightDecrease);
/* print results */
disOut.println("\nResults:");
disOut.println("Num sent in cover : " + numSentencesInCover);
double avSentLength = (double) numSelectedFeatVects / (double) numSentencesInCover;
disOut.println("Avg sent length : " + df.format(avSentLength));
disOut.println("Max sent length : " + maxSentLengthInCover);
disOut.println("Min sent length : " + minSentLengthInCover);
/* print distribution info */
double phoneCov = (double) phonesInCover.size() / (double) numPhoneValuesMinusIgnored;
double simpleDiphoneCov = (double) simpleDiphonesInCover.size() / (double) numPossibleSimpleDiphones;
double overallSimpleCov = (double) numSimpleFeatVectsInCover / (double) numSimpleLeaves;
// double clusteredDiphoneCov = (double)clusteredDiphonesInCover.size()/(double)numPossibleClusteredDiphones;
// double overallClusteredCov = (double)numClusteredFeatVectsInCover/(double)numClusteredLeaves;
disOut.println("phones: " + df.format(phoneCov) + " (" + df.format(possiblePhoneCoverage) + ")");
disOut.println("Simple Coverage:");
disOut.println("phones: " + df.format(phoneCov) + " (" + df.format(possiblePhoneCoverage) + ")");
disOut.println("diphones: " + df.format(simpleDiphoneCov) + " (" + df.format(possibleSimpleDiphoneCoverage) + ")");
disOut.println("overall: " + df.format(overallSimpleCov) + " (" + df.format(possibleOverallSimpleCoverage) + ")");
/*
* disOut.println("Clustered Coverage:");
* //disOut.println("phones: "+df.format(phoneCov)+" ("+df.format(possiblePhoneCoverage)+")");
* disOut.println("diphones: "+df.format(clusteredDiphoneCov) +" ("+df.format(possibleClusteredDiphoneCoverage)+")");
* disOut.println("overall: "+df.format(overallClusteredCov) +" ("+df.format(possibleOverallClusteredCoverage)+")");
*/
disOut.flush();
disOut.close();
/* print coverage development over time */
if (logDevelopment) {
PrintWriter devOut = new PrintWriter(new FileWriter(new File(developmentFile)));
devOut.println("\toverall coverage\tdiphone coverage\tphone coverage");
for (int i = 0; i < overallCoverageInTime.size(); i++) {
devOut.print(i + "\t" + df.format(overallCoverageInTime.get(i)) + "\t" + df.format(diphoneCoverageInTime.get(i))
+ "\t" + df.format(phoneCoverageInTime.get(i)) + "\n");
}
devOut.flush();
devOut.close();
}
}
public void printResultToLog(PrintWriter logOut) {
/* print settings */
DecimalFormat df = new DecimalFormat("0.00000");
logOut.println("simpleDiphones " + Boolean.toString(simpleDiphones));
if (considerFrequency) {
logOut.println("frequency " + frequencySetting);
} else {
logOut.println("frequency none");
}
logOut.println("considerSentenceLength " + Boolean.toString(considerSentenceLength));
logOut.println("phoneLevelWeight " + phoneLevelWeight);
logOut.println("diphoneLevelWeight " + diphoneLevelWeight);
logOut.println("prosodyLevelWeight " + prosodyLevelWeight);
logOut.println("maxSentenceLength " + maxSentLengthAllowed);
logOut.println("minSentenceLength " + minSentLengthAllowed);
logOut.println("divideWantedWeightBy " + wantedWeightDecrease);
logOut.println("\nNum sent in cover : " + numSentencesInCover);
double avSentLength = (double) numSelectedFeatVects / (double) numSentencesInCover;
logOut.println("Avg sent length : " + df.format(avSentLength));
logOut.println("Max sent length : " + maxSentLengthInCover);
logOut.println("Min sent length : " + minSentLengthInCover);
/* print distribution info */
double phoneCov = (double) phonesInCover.size() / (double) numPhoneValuesMinusIgnored;
double simpleDiphoneCov = (double) simpleDiphonesInCover.size() / (double) numPossibleSimpleDiphones;
double overallSimpleCov = (double) numSimpleFeatVectsInCover / (double) numSimpleLeaves;
logOut.println("phones: " + df.format(phoneCov) + " (" + df.format(possiblePhoneCoverage) + ")");
logOut.println("Simple Coverage:");
// logOut.println("phones: "+df.format(phoneCov)+" ("+df.format(possiblePhoneCoverage)+")");
logOut.println("diphones: " + df.format(simpleDiphoneCov) + " (" + df.format(possibleSimpleDiphoneCoverage) + ")");
logOut.println("overall: " + df.format(overallSimpleCov) + " (" + df.format(possibleOverallSimpleCoverage) + ")");
/*
* logOut.println("Clustered Coverage:");
* //logOut.println("phones: "+df.format(phoneCov)+" ("+df.format(possiblePhoneCoverage)+")");
* logOut.println("diphones: "+df.format(clusteredDiphoneCov) +" ("+df.format(possibleClusteredDiphoneCoverage)+")");
* logOut.println("overall: "+df.format(overallClusteredCov) +" ("+df.format(possibleOverallClusteredCoverage)+")\n\n");
*/
}
/**
* Add the feature vectors for one sentence to the cover
*
* @param features
* the feature vectors to add
*/
public void updateCover(byte[] features) {
int numFeatVects = features.length / numTargetFeaturesUsed;
// loop through the feature vectors
for (int i = 0; i < numFeatVects; i++) {
/* update simpleCover */
CoverLeaf leaf = goDownTree(features, i, true);
// if this is the first feature vector in this leaf
// decrease cover size
if (leaf.getNumFeatureVectors() == 0) {
numSimpleFeatVectsInCover++;
}
leaf.addFeatureVector();
String phone = possiblePhoneArray[getVectorValue(features, i, phoneFeatIndex)];
// update coverage statistics
String diphone = phone + "_" + possibleNextPhoneArray[getVectorValue(features, i, diphoneFeatIndex)];
phonesInCover.add(phone);
simpleDiphonesInCover.add(diphone);
}
// update phone coverage statistics
double phoneCoverage = (double) phonesInCover.size() / (double) numPhoneValuesMinusIgnored;
phoneCoverageInTime.add(new Double(phoneCoverage));
// update diphone and overall coverage statistics
if (simpleDiphones) {
double diphoneCoverage = (double) simpleDiphonesInCover.size() / (double) numPossibleSimpleDiphones;
double overallCoverage = (double) numSimpleFeatVectsInCover / (double) numSimpleLeaves;
diphoneCoverageInTime.add(new Double(diphoneCoverage));
overallCoverageInTime.add(new Double(overallCoverage));
}
// compute statistics of sentence length
numSentencesInCover++;
if (numFeatVects > maxSentLengthInCover)
maxSentLengthInCover = numFeatVects;
if (numFeatVects < minSentLengthInCover)
minSentLengthInCover = numFeatVects;
numSelectedFeatVects += numFeatVects;
}
/**
* Check if cover has maximum simple diphone coverage
*
* @return true if cover has maximum simple diphone coverage
*/
public boolean reachedMaxSimpleDiphones() {
return simpleDiphonesInCover.size() >= numSimpleDiphoneTypes;
}
/**
* Check if cover has maximum simple prosody coverage
*
* @return true if cover has maximum simple prosody coverage
*/
public boolean reachedMaxSimpleProsody() {
return numSimpleFeatVectsInCover == numSimpleFeatVectTypes;
}
/**
* Get the usefulness of the given feature vectors Usefulness of a feature vector is defined as the sum of the score for the
* feature vectors on all levels of the tree. On each level, the score is the product of the two weights of the node. The
* first weight reflects the frequency/ inverted frequency of the value associated with the node in the corpus (→
* frequencyWeight). The second weight reflects how much an instance of a feature vector containing the associated value is
* wanted in the cover (→ wantedWeight).
*
* @param featureVectors
* the feature vectors
* @return the usefulness
*/
public double usefulnessOfFVs(byte[] featureVectors) {
double usefulness = 0.0;
// int numFeatureVectors = featureVectors.length/4;
int numFeatureVectors = featureVectors.length / numTargetFeaturesUsed;
if (considerSentenceLength) {
// too long sentences are useless
if (numFeatureVectors > maxSentLengthAllowed)
return -1.0;
// too short sentences are useless as well
if (numFeatureVectors < minSentLengthAllowed)
return -1.0;
}
// loop over the feature vectors
// System.out.print("Usefulness = ");
// we cannot trust that all bytes in the feature vector are meaningful -- therefore,
// it is not guaranteed that numFeatureVectors * numTargetFeaturesUsed == featureVectors.length!!
for (int pos = 0, max = numFeatureVectors * numTargetFeaturesUsed; pos < max; pos += numTargetFeaturesUsed) {
double u = 0;
// get the associated leaf
// go down to phone level
// byte nextIndex = getVectorValue(featureVectors,i,phoneFeatIndex);
byte nextIndex = featureVectors[pos + phoneFeatIndex];
CoverNode nextNode = simpleCover.children[nextIndex];
// double relFreq = nextNode.getFrequencyWeight();
// double wantedWeight = nextNode.getWantedWeight();
// System.out.print(" +"+relFreq+"*"+wantedWeight);
// u += nextNode.frequencyWeight * nextNode.wantedWeight;
u += nextNode.usefulness;
// go down to diphone level
// nextIndex = getVectorValue(featureVectors,i,diphoneFeatIndex);
nextIndex = featureVectors[pos + diphoneFeatIndex];
nextNode = nextNode.children[nextIndex];
// relFreq = nextNode.getFrequencyWeight();
// wantedWeight = nextNode.getWantedWeight();
// System.out.print(" +"+relFreq+"*"+wantedWeight);
// u += nextNode.frequencyWeight * nextNode.wantedWeight;
u += nextNode.usefulness;
// go down to prosody level
// nextIndex = getVectorValue(featureVectors,i,prosodyIndex);
nextIndex = featureVectors[pos + prosodyIndex];
nextNode = nextNode.children[nextIndex];
// relFreq = nextNode.getFrequencyWeight();
// wantedWeight = nextNode.getWantedWeight();
// System.out.print(" +"+relFreq+"*"+wantedWeight+"\n");
// u += nextNode.frequencyWeight * nextNode.wantedWeight;
u += nextNode.usefulness;
usefulness += u;
}
// System.out.print(" = "+usefulness+"\n");
return usefulness / (double) numFeatureVectors;
}
public CoverageFeatureProvider getCoverageFeatureProvider() {
return cfProvider;
}
public byte getVectorValue(byte[] vectors, int vectorIndex, int valueIndex) {
// byte result = vectors[vectorIndex*4+valueIndex];
return vectors[vectorIndex * numTargetFeaturesUsed + valueIndex];
}
/**
* Print the cover sets to the given file
*
* @param filename
* the file to print to
* @throws Exception
* Exception
*/
public void writeCoverageBin(String filename) throws Exception {
DataOutputStream out = new DataOutputStream(new FileOutputStream(new File(filename)));
/* print all the relevant information */
out.writeInt(numTokens);
out.writeInt(numSimpleDiphoneTypes);
out.writeInt(numSimpleFeatVectTypes);
out.writeDouble(averageSentLength);
out.writeInt(maxSentLength);
out.writeInt(minSentLength);
out.writeInt(numSimpleLeaves);
out.writeDouble(possiblePhoneCoverage);
out.writeDouble(possibleSimpleDiphoneCoverage);
out.writeDouble(possibleOverallSimpleCoverage);
out.writeInt(numSentences);
/* print the coverage tree */
writeTreeBin(out, simpleCover);
out.flush();
out.close();
}
/**
* Print the cover tree
*
* @param out
* the output stream to write to
* @param cover
* the tree to print
* @throws IOException
* IOException
*/
private void writeTreeBin(DataOutputStream out, CoverNode cover) throws IOException {
// go down to phone level
byte numChildren = cover.getNumChildren();
double frequencyWeight = cover.getFrequencyWeight();
double wantedWeight = cover.getWantedWeight();
double wantedWeightDecrease = cover.getWantedWeightDecrease();
out.writeByte(numChildren);
for (byte i = 0; i < numChildren; i++) {
if (phonesToIgnore.contains(new Integer(i)))
continue;
CoverNode phoneNode = cover.getChild(i);
frequencyWeight = phoneNode.getFrequencyWeight();
wantedWeight = phoneNode.getWantedWeight();
wantedWeightDecrease = phoneNode.getWantedWeightDecrease();
byte numNextChildren = phoneNode.getNumChildren();
out.writeByte(numNextChildren);
// go down to diphone level
for (byte j = 0; j < numNextChildren; j++) {
CoverNode diphoneNode = phoneNode.getChild(j);
// go down to prosody level
byte numNextNextChildren = diphoneNode.getNumChildren();
out.writeByte(numNextNextChildren);
for (byte k = 0; k < numNextNextChildren; k++) {
CoverLeaf nextLeaf = (CoverLeaf) diphoneNode.getChild(k);
int numVectors = nextLeaf.maxNumFeatVects();
frequencyWeight = nextLeaf.getFrequencyWeight();
wantedWeight = nextLeaf.getWantedWeight();
wantedWeightDecrease = nextLeaf.getWantedWeightDecrease();
out.writeInt(numVectors);
}
}
}
}
/**
* Read the cover sets from the given file
*
* @param filename
* the file containing the cover sets
* @param idSentenceList
* the id of the sentence list
* @throws Exception
* Exception
*/
// public void readCoverageBin(String filename, FeatureDefinition fDef, String[] basenames)throws Exception{
public void readCoverageBin(String filename, int[] idSentenceList) throws Exception {
DataInputStream in = new DataInputStream(new FileInputStream(new File(filename)));
/* read all the relevant information */
numTokens = in.readInt();
numSimpleDiphoneTypes = in.readInt();
numSimpleFeatVectTypes = in.readInt();
averageSentLength = in.readDouble();
maxSentLength = in.readInt();
minSentLength = in.readInt();
numSimpleLeaves = in.readInt();
possiblePhoneCoverage = in.readDouble();
possibleSimpleDiphoneCoverage = in.readDouble();
possibleOverallSimpleCoverage = in.readDouble();
numSentences = in.readInt();
/* print the coverage tree */
readTreeBin(in);
in.close();
System.out.print("Num Tokens: " + numTokens + "\n");
}
/**
* Read a cover tree
*
* @param in
* the inputstream to read from
* @param isSimpleCover
* if true, build the cover tree fro simpleDiphones
* @throws Exception
* Exception
*/
private void readTreeBin(DataInputStream in) throws Exception {
byte numChildren = in.readByte();
double wantedWeight = 0.0;
CoverNode cover = new CoverNode(numChildren, wantedWeightDecrease, wantedWeight);
for (byte i = 0; i < numChildren; i++) {
if (phonesToIgnore.contains(new Integer(i)))
continue;
byte nextNumChildren = in.readByte();
CoverNode diphoneNode = new CoverNode(nextNumChildren, wantedWeightDecrease, phoneLevelWeight);
cover.addChild(diphoneNode, i);
for (byte j = 0; j < nextNumChildren; j++) {
byte nextNextNumChildren = in.readByte();
CoverNode prosodyNode = new CoverNode(nextNextNumChildren, wantedWeightDecrease, diphoneLevelWeight);
diphoneNode.addChild(prosodyNode, j);
for (byte k = 0; k < nextNextNumChildren; k++) {
int numVectors = in.readInt();
CoverLeaf leafNode = new CoverLeaf(wantedWeightDecrease, prosodyLevelWeight, numVectors);
prosodyNode.addChild(leafNode, k);
}
}
}
computeRelativeFrequency(cover, numTokens);
simpleCover = cover;
}
/**
* A node in the cover tree Represents a feature. Number of children is the number of possible values.
*
* @author Anna Hunecke
*
*/
class CoverNode {
/* children of this node */
protected CoverNode[] children;
/* number of children of this node */
private byte numChildren;
/*
* how much is this node and its children wanted in the cover
*/
protected double wantedWeight;
/*
* frequency/inverted frequency of the node in the corpus
*/
protected double frequencyWeight;
/* number by which the wantedWeight is divided */
protected double wantedWeightDecrease;
/* usefulness is the product of wantedWeight and frequencyWeight. It is here purely for efficiency reasons. */
protected double usefulness;
/**
* Build a new CoverNode Set frequency weight to 1
*/
public CoverNode() {
frequencyWeight = 1;
usefulness = 0;
}
/**
* Build a new CoverNode
*
* @param numChildren
* the number of children
* @param wantedWeightDecrease
* the value by which the wanted weight is divided
* @param wantedWeight
* the wanted weight
*/
public CoverNode(byte numChildren, double wantedWeightDecrease, double wantedWeight) {
this.numChildren = numChildren;
children = new CoverNode[numChildren];
this.wantedWeightDecrease = wantedWeightDecrease;
this.wantedWeight = wantedWeight;
frequencyWeight = 1;
usefulness = frequencyWeight * this.wantedWeight;
}
/**
* Build a new CoverNode
*
* @param values
* the number of values
* @param wantedWeightDecrease
* the wantedWeightDecrease
*/
public CoverNode(byte values, double wantedWeightDecrease) {
children = new CoverNode[values];
numChildren = (byte) children.length;
frequencyWeight = 1;
this.wantedWeightDecrease = wantedWeightDecrease;
usefulness = 0;
}
/**
* Add a new child
*
* @param child
* the child
* @param value
* the position of the child in the children array
*/
public void addChild(CoverNode child, byte value) {
children[value] = child;
}
/**
* Get a child
*
* @param value
* the position of the child in the children array
* @return the child (null, if there is no child at this position)
*/
public CoverNode getChild(byte value) {
return children[value];
}
/**
* Get the number of children
*
* @return the number of children
*/
public byte getNumChildren() {
return numChildren;
}
/**
* Set the wantedWeight
*
* @param wantedWeight
* the new wantedWeight
*/
public void setWantedWeight(double wantedWeight) {
this.wantedWeight = wantedWeight;
usefulness = this.wantedWeight * frequencyWeight;
}
/**
* Get the wantedWeight
*
* @return the wantedWeight
*/
public double getWantedWeight() {
return wantedWeight;
}
/**
* Get the wantedWeightDecrease
*
* @return the wantedWeightDecrease
*/
public double getWantedWeightDecrease() {
return wantedWeightDecrease;
}
/**
* Decrease the wantedWeight by dividing it by wantedWeightDecrease
*
*/
public void decreaseWantedWeight() {
wantedWeight = wantedWeight / wantedWeightDecrease;
usefulness = frequencyWeight * wantedWeight;
}
/**
* Set the frequencyWeight
*
* @param frequencyWeight
* the new frequencyWeight
*/
public void setFrequencyWeight(double frequencyWeight) {
this.frequencyWeight = frequencyWeight;
usefulness = this.frequencyWeight * wantedWeight;
}
/**
* Get the frequencyWeight
*
* @return the frequencyWeight
*/
public double getFrequencyWeight() {
return frequencyWeight;
}
}
/**
* A leaf in the cover tree. Collects the feature vectors that belong to the path that leads to the leaf.
*
* @author Anna Hunecke
*
*/
class CoverLeaf extends CoverNode {
/* the number of feature vectors in this node */
private int numFeatVects;
/*
* the maximimum number of feature vectors that could be in this node (according to the corpus)
*/
private int maxNumFeatVects;
/**
* Build a new cover leaf
*
* @param wantedWeightDecrease
* the wantedWeightDecrease
*/
public CoverLeaf(double wantedWeightDecrease) {
super();
numFeatVects = 0;
maxNumFeatVects = 0;
this.wantedWeightDecrease = wantedWeightDecrease;
frequencyWeight = 1;
}
/**
* Build a new CoverLeaf
*
* @param wantedWeightDecrease
* the wantedWeightDecrease
* @param wantedWeight
* the wanted weight
* @param maxNumFeatVects
* maximum number of feature vectors that can be collected in this leaf
*/
public CoverLeaf(double wantedWeightDecrease, double wantedWeight, int maxNumFeatVects) {
this.wantedWeightDecrease = wantedWeightDecrease;
this.wantedWeight = wantedWeight;
this.maxNumFeatVects = maxNumFeatVects;
frequencyWeight = 1;
}
/**
* Add a new feature vector
*
* @param featureVector
* the new feature vector
*/
public void addFeatureVector() {
numFeatVects++;
}
/**
* Increase the maximum number of feature vectors by one (because we have seen a feature vector for this node in the
* corpus)
*/
public void addPossibleInstance() {
maxNumFeatVects++;
}
/**
* Get the number of feature vectors of this node
*
* @return the number of feature vectors
*/
public int getNumFeatureVectors() {
return numFeatVects;
}
/**
* Get the maximum number of feature vectors of this node
*
* @return the maximum number of feature vectors
*/
public int maxNumFeatVects() {
return maxNumFeatVects;
}
}
public static class CoverageStatistics {
public Set coveredPhones;
public Set allPhones;
public Set coveredDiphones;
public int numPossibleDiphones;
public int numCoveredDiphonesWithProsody;
public int numPossibleDiphonesWithProsody;
public int numTokens;
@Override
public String toString() {
StringWriter sw = new StringWriter();
PrintWriter out = new PrintWriter(sw);
out.println("\nSimple Coverage:");
out.printf(Locale.US, "phones: %.5f\n", coveredPhones.size() / (double) allPhones.size());
out.printf(Locale.US, "diphones: %.5f\n", coveredDiphones.size() / (double) numPossibleDiphones);
out.printf(Locale.US, "overall: %.5f\n", numCoveredDiphonesWithProsody / (double) numPossibleDiphonesWithProsody);
if (coveredPhones.size() < allPhones.size()) {
out.println("The following phones are missing: ");
for (String ph : allPhones) {
if (!coveredPhones.contains(ph)) {
out.print(ph + " ");
}
}
out.println();
out.println("The following phones are present: ");
for (String ph : coveredPhones) {
out.print(ph + " ");
}
out.println();
}
/* print out the diphone statistics */
out.println("\n");
out.println("Number of diphones seen : " + numTokens);
out.println("Number of different diphones : " + coveredDiphones.size() + " out of a theoretical "
+ numPossibleDiphones);
out.close();
return sw.toString();
}
}
}
© 2015 - 2025 Weber Informatics LLC | Privacy Policy