public static void main(String[] args) {
try {
String corpusFile = args[0];
String goldSet = args[1];
File outputDir = new File(args[2]);
SystemConfig systemConfig = DriverUtils.configure(args);
systemConfig.setAnnotationSetName(Constants.GS_NP, goldSet);
Trainer trainer = new Trainer(systemConfig);
FeatureGenerator featureGenerator = new FeatureGenerator(systemConfig);
// get corpus
Corpus c = DriverUtils.loadFiles(corpusFile);
Preprocessor preprocessor = new Preprocessor(systemConfig);
preprocessor.preprocess(c, false);
// generate features
String featureSetName = featureGenerator.generateFeatures(c, true);
// train classifier
Classifier classifier = trainer.runLearner(c, outputDir, featureSetName);
System.out.println("classifier trained: " + classifier.getName());
} catch (IOException e) {
e.printStackTrace();
} catch (ConfigurationException e) {
e.printStackTrace();
}
}
static boolean computeLogMi(
FeatureGenerator featureGen,
double lambda[],
DoubleMatrix2D Mi_YY,
DoubleMatrix1D Ri_Y,
boolean takeExp,
boolean reuseM,
boolean initMDone) {
if (reuseM && initMDone) {
Mi_YY = null;
} else initMDone = false;
if (Mi_YY != null) Mi_YY.assign(0);
Ri_Y.assign(0);
while (featureGen.hasNext()) {
Feature feature = featureGen.next();
int f = feature.index();
int yp = feature.y();
int yprev = feature.yprev();
float val = feature.value();
// System.out.println(feature.toString());
if (yprev < 0) {
// this is a single state feature.
double oldVal = Ri_Y.getQuick(yp);
Ri_Y.setQuick(yp, oldVal + lambda[f] * val);
} else if (Mi_YY != null) {
Mi_YY.setQuick(yprev, yp, Mi_YY.getQuick(yprev, yp) + lambda[f] * val);
initMDone = true;
}
}
if (takeExp) {
for (int r = Ri_Y.size() - 1; r >= 0; r--) {
Ri_Y.setQuick(r, expE(Ri_Y.getQuick(r)));
if (Mi_YY != null)
for (int c = Mi_YY.columns() - 1; c >= 0; c--) {
Mi_YY.setQuick(r, c, expE(Mi_YY.getQuick(r, c)));
}
}
}
return initMDone;
}
static boolean computeLogMi(
FeatureGenerator featureGen,
double lambda[],
DataSequence dataSeq,
int i,
DoubleMatrix2D Mi_YY,
DoubleMatrix1D Ri_Y,
boolean takeExp,
boolean reuseM,
boolean initMDone) {
featureGen.startScanFeaturesAt(dataSeq, i);
return computeLogMi(featureGen, lambda, Mi_YY, Ri_Y, takeExp, reuseM, initMDone);
}
protected void init(CRF model, DataIter data, double[] l) {
edgeGen = model.edgeGen;
lambda = l;
numY = model.numY;
diter = data;
featureGenerator = model.featureGenerator;
numF = featureGenerator.numFeatures();
gradLogli = new double[numF];
diag = new double[numF]; // needed by the optimizer
ExpF = new double[lambda.length];
initMatrices();
reuseM = params.reuseM;
if (params.miscOptions.getProperty("cache", "false").equals("true")) {
featureGenCache = new FeatureGenCache(featureGenerator);
featureGenerator = featureGenCache;
} else featureGenCache = null;
}
public void compute() {
alpha_Y.assign(1);
initMDone = false;
boolean doScaling = params.doScaling;
if ((beta_Y == null) || (beta_Y.length < dataSeq.length())) {
beta_Y = new DenseDoubleMatrix1D[2 * dataSeq.length()];
for (int i = 0; i < beta_Y.length; i++) beta_Y[i] = new DenseDoubleMatrix1D(numY);
scale = new double[2 * dataSeq.length()];
scale[dataSeq.length() - 1] = (doScaling) ? numY : 1;
beta_Y[dataSeq.length() - 1].assign(1.0 / scale[dataSeq.length() - 1]);
}
beta.add(beta_Y[dataSeq.length() - 1]);
// System.out.println("Beta "+beta_Y[3].toString());
for (int i = dataSeq.length() - 1; i > 0; i--) {
if (params.debugLvl > 2) {
Util.printDbg("Features fired");
// featureGenerator.startScanFeaturesAt(dataSeq, i);
// while (featureGenerator.hasNext()) {
// Feature feature = featureGenerator.next();
// Util.printDbg(feature.toString());
// }
}
// compute the Mi matrix
// System.out.println("MI previous" +Mi_YY.toString());
initMDone =
Trainer.computeLogMi(
featureGenerator, lambda, dataSeq, i, Mi_YY, Ri_Y, true, reuseM, initMDone);
// System.out.println("MI "+Mi_YY.toString());
tmp_Y.assign(beta_Y[i]);
tmp_Y.assign(Ri_Y, multFunc);
RobustMath.Mult(Mi_YY, tmp_Y, beta_Y[i - 1], 1, 0, false, edgeGen);
// need to scale the beta-s to avoid overflow
scale[i - 1] = doScaling ? beta_Y[i - 1].zSum() : 1;
if ((scale[i - 1] < 1) && (scale[i - 1] > -1)) scale[i - 1] = 1;
constMultiplier.multiplicator = 1.0 / scale[i - 1];
beta_Y[i - 1].assign(constMultiplier);
// System.out.println("Beta "+beta_Y[i - 1].toString() + " ");
beta.add(beta_Y[i - 1]);
}
double thisSeqLogli = 0;
System.out.println("\n");
// Mi_YY.assign(0);
alpha_temp = new DenseDoubleMatrix1D[2 * dataSeq.length()];
for (int i = 0; i < dataSeq.length(); i++) alpha_temp[i] = new DenseDoubleMatrix1D(numY);
for (int i = 0; i < dataSeq.length(); i++) {
// compute the Mi matrix
//
initMDone =
Trainer.computeLogMi(
featureGenerator, lambda, dataSeq, i, Mi_YY, Ri_Y, true, reuseM, initMDone);
// System.out.println("MI: " + Mi_YY.toString());
// find features that fire at this position..
featureGenerator.startScanFeaturesAt(dataSeq, i);
if (i > 0) {
tmp_Y.assign(alpha_Y);
RobustMath.Mult(Mi_YY, tmp_Y, newAlpha_Y, 1, 0, true, edgeGen);
// Mi_YY.zMult(tmp_Y, newAlpha_Y,1,0,true);
newAlpha_Y.assign(Ri_Y, multFunc);
} else {
newAlpha_Y.assign(Ri_Y);
}
while (featureGenerator.hasNext()) {
Feature feature = featureGenerator.next();
int f = feature.index();
int yp = feature.y();
int yprev = feature.yprev();
float val = feature.value();
if ((dataSeq.y(i) == yp)
&& (((i - 1 >= 0) && (yprev == dataSeq.y(i - 1))) || (yprev < 0))) {
thisSeqLogli += val * lambda[f];
}
}
alpha_Y.assign(newAlpha_Y);
// now scale the alpha-s to avoid overflow problems.
constMultiplier.multiplicator = 1.0 / scale[i];
alpha_Y.assign(constMultiplier);
alpha_temp[i].assign(newAlpha_Y);
alpha_temp[i].assign(constMultiplier);
// System.out.println("ALpha "+alpha_Y.toString());
alpha.add(alpha_temp[i]);
if (params.debugLvl > 2) {
System.out.println("Alpha-i " + alpha_Y.toString());
System.out.println("Ri " + Ri_Y.toString());
System.out.println("Mi " + Mi_YY.toString());
System.out.println("Beta-i " + beta_Y[i].toString());
}
}
Zx = alpha_Y.zSum();
System.out.println("Zx: " + Zx);
} /* end of computeBeta */
protected double computeFunctionGradientLL(double lambda[], double grad[]) {
double logli = 0;
try {
for (int f = 0; f < lambda.length; f++) {
grad[f] = -1 * lambda[f] * params.invSigmaSquare;
logli -= ((lambda[f] * lambda[f]) * params.invSigmaSquare) / 2;
}
diter.startScan();
if (featureGenCache != null) featureGenCache.startDataScan();
for (int numRecord = 0; diter.hasNext(); numRecord++) {
DataSequence dataSeq = (DataSequence) diter.next();
if (featureGenCache != null) featureGenCache.nextDataIndex();
if (params.debugLvl > 1) {
Util.printDbg("Read next seq: " + numRecord + " logli " + logli);
}
alpha_Y.assign(0);
for (int f = 0; f < lambda.length; f++) ExpF[f] = RobustMath.LOG0;
if ((beta_Y == null) || (beta_Y.length < dataSeq.length())) {
beta_Y = new DenseDoubleMatrix1D[2 * dataSeq.length()];
for (int i = 0; i < beta_Y.length; i++) beta_Y[i] = new DenseDoubleMatrix1D(numY);
}
// compute beta values in a backward scan.
// also scale beta-values to 1 to avoid numerical problems.
beta_Y[dataSeq.length() - 1].assign(0);
for (int i = dataSeq.length() - 1; i > 0; i--) {
if (params.debugLvl > 2) {
/* Util.printDbg("Features fired");
featureGenerator.startScanFeaturesAt(dataSeq, i);
while (featureGenerator.hasNext()) {
Feature feature = featureGenerator.next();
Util.printDbg(feature.toString());
}
*/
}
// compute the Mi matrix
initMDone =
computeLogMi(
featureGenerator, lambda, dataSeq, i, Mi_YY, Ri_Y, false, reuseM, initMDone);
tmp_Y.assign(beta_Y[i]);
tmp_Y.assign(Ri_Y, sumFunc);
RobustMath.logMult(Mi_YY, tmp_Y, beta_Y[i - 1], 1, 0, false, edgeGen);
}
double thisSeqLogli = 0;
for (int i = 0; i < dataSeq.length(); i++) {
// compute the Mi matrix
initMDone =
computeLogMi(
featureGenerator, lambda, dataSeq, i, Mi_YY, Ri_Y, false, reuseM, initMDone);
// find features that fire at this position..
featureGenerator.startScanFeaturesAt(dataSeq, i);
if (i > 0) {
tmp_Y.assign(alpha_Y);
RobustMath.logMult(Mi_YY, tmp_Y, newAlpha_Y, 1, 0, true, edgeGen);
newAlpha_Y.assign(Ri_Y, sumFunc);
} else {
newAlpha_Y.assign(Ri_Y);
}
while (featureGenerator.hasNext()) {
Feature feature = featureGenerator.next();
int f = feature.index();
int yp = feature.y();
int yprev = feature.yprev();
float val = feature.value();
if ((dataSeq.y(i) == yp)
&& (((i - 1 >= 0) && (yprev == dataSeq.y(i - 1))) || (yprev < 0))) {
grad[f] += val;
thisSeqLogli += val * lambda[f];
if (params.debugLvl > 2) {
System.out.println("Feature fired " + f + " " + feature);
}
}
if (yprev < 0) {
ExpF[f] =
RobustMath.logSumExp(
ExpF[f], newAlpha_Y.get(yp) + RobustMath.log(val) + beta_Y[i].get(yp));
} else {
ExpF[f] =
RobustMath.logSumExp(
ExpF[f],
alpha_Y.get(yprev)
+ Ri_Y.get(yp)
+ Mi_YY.get(yprev, yp)
+ RobustMath.log(val)
+ beta_Y[i].get(yp));
}
}
alpha_Y.assign(newAlpha_Y);
if (params.debugLvl > 2) {
System.out.println("Alpha-i " + alpha_Y.toString());
System.out.println("Ri " + Ri_Y.toString());
System.out.println("Mi " + Mi_YY.toString());
System.out.println("Beta-i " + beta_Y[i].toString());
}
}
double lZx = RobustMath.logSumExp(alpha_Y);
thisSeqLogli -= lZx;
logli += thisSeqLogli;
// update grad.
for (int f = 0; f < grad.length; f++) {
grad[f] -= RobustMath.exp(ExpF[f] - lZx);
}
if (params.debugLvl > 1) {
System.out.println(
"Sequence "
+ thisSeqLogli
+ " logli "
+ logli
+ " log(Zx) "
+ lZx
+ " Zx "
+ Math.exp(lZx));
}
}
if (params.debugLvl > 2) {
for (int f = 0; f < lambda.length; f++) System.out.print(lambda[f] + " ");
System.out.println(" :x");
for (int f = 0; f < lambda.length; f++) System.out.print(grad[f] + " ");
System.out.println(" :g");
}
if (params.debugLvl > 0)
Util.printDbg(
"Iteration "
+ icall
+ " log-likelihood "
+ logli
+ " norm(grad logli) "
+ norm(grad)
+ " norm(x) "
+ norm(lambda));
} catch (Exception e) {
System.out.println("Alpha-i " + alpha_Y.toString());
System.out.println("Ri " + Ri_Y.toString());
System.out.println("Mi " + Mi_YY.toString());
e.printStackTrace();
System.exit(0);
}
return logli;
}
protected double computeFunctionGradient(double lambda[], double grad[]) {
initMDone = false;
if (params.trainerType.equals("ll")) return computeFunctionGradientLL(lambda, grad);
double logli = 0;
try {
for (int f = 0; f < lambda.length; f++) {
grad[f] = -1 * lambda[f] * params.invSigmaSquare;
logli -= ((lambda[f] * lambda[f]) * params.invSigmaSquare) / 2;
}
boolean doScaling = params.doScaling;
diter.startScan();
if (featureGenCache != null) featureGenCache.startDataScan();
int numRecord = 0;
for (numRecord = 0; diter.hasNext(); numRecord++) {
DataSequence dataSeq = (DataSequence) diter.next();
if (featureGenCache != null) featureGenCache.nextDataIndex();
if (params.debugLvl > 1) {
Util.printDbg("Read next seq: " + numRecord + " logli " + logli);
}
alpha_Y.assign(1);
for (int f = 0; f < lambda.length; f++) ExpF[f] = 0;
if ((beta_Y == null) || (beta_Y.length < dataSeq.length())) {
beta_Y = new DenseDoubleMatrix1D[2 * dataSeq.length()];
for (int i = 0; i < beta_Y.length; i++) beta_Y[i] = new DenseDoubleMatrix1D(numY);
scale = new double[2 * dataSeq.length()];
}
// compute beta values in a backward scan.
// also scale beta-values to 1 to avoid numerical problems.
scale[dataSeq.length() - 1] = (doScaling) ? numY : 1;
beta_Y[dataSeq.length() - 1].assign(1.0 / scale[dataSeq.length() - 1]);
for (int i = dataSeq.length() - 1; i > 0; i--) {
if (params.debugLvl > 2) {
Util.printDbg("Features fired");
// featureGenerator.startScanFeaturesAt(dataSeq, i);
// while (featureGenerator.hasNext()) {
// Feature feature = featureGenerator.next();
// Util.printDbg(feature.toString());
// }
}
// compute the Mi matrix
initMDone =
computeLogMi(
featureGenerator, lambda, dataSeq, i, Mi_YY, Ri_Y, true, reuseM, initMDone);
tmp_Y.assign(beta_Y[i]);
tmp_Y.assign(Ri_Y, multFunc);
RobustMath.Mult(Mi_YY, tmp_Y, beta_Y[i - 1], 1, 0, false, edgeGen);
// Mi_YY.zMult(tmp_Y, beta_Y[i-1]);
// need to scale the beta-s to avoid overflow
scale[i - 1] = doScaling ? beta_Y[i - 1].zSum() : 1;
if ((scale[i - 1] < 1) && (scale[i - 1] > -1)) scale[i - 1] = 1;
constMultiplier.multiplicator = 1.0 / scale[i - 1];
beta_Y[i - 1].assign(constMultiplier);
}
double thisSeqLogli = 0;
for (int i = 0; i < dataSeq.length(); i++) {
// compute the Mi matrix
initMDone =
computeLogMi(
featureGenerator, lambda, dataSeq, i, Mi_YY, Ri_Y, true, reuseM, initMDone);
// find features that fire at this position..
featureGenerator.startScanFeaturesAt(dataSeq, i);
if (i > 0) {
tmp_Y.assign(alpha_Y);
RobustMath.Mult(Mi_YY, tmp_Y, newAlpha_Y, 1, 0, true, edgeGen);
// Mi_YY.zMult(tmp_Y, newAlpha_Y,1,0,true);
newAlpha_Y.assign(Ri_Y, multFunc);
} else {
newAlpha_Y.assign(Ri_Y);
}
while (featureGenerator.hasNext()) {
Feature feature = featureGenerator.next();
int f = feature.index();
int yp = feature.y();
int yprev = feature.yprev();
float val = feature.value();
if ((dataSeq.y(i) == yp)
&& (((i - 1 >= 0) && (yprev == dataSeq.y(i - 1))) || (yprev < 0))) {
grad[f] += val;
thisSeqLogli += val * lambda[f];
}
if (yprev < 0) {
ExpF[f] += newAlpha_Y.get(yp) * val * beta_Y[i].get(yp);
} else {
ExpF[f] +=
alpha_Y.get(yprev)
* Ri_Y.get(yp)
* Mi_YY.get(yprev, yp)
* val
* beta_Y[i].get(yp);
}
}
alpha_Y.assign(newAlpha_Y);
// now scale the alpha-s to avoid overflow problems.
constMultiplier.multiplicator = 1.0 / scale[i];
alpha_Y.assign(constMultiplier);
if (params.debugLvl > 2) {
System.out.println("Alpha-i " + alpha_Y.toString());
System.out.println("Ri " + Ri_Y.toString());
System.out.println("Mi " + Mi_YY.toString());
System.out.println("Beta-i " + beta_Y[i].toString());
}
}
double Zx = alpha_Y.zSum();
thisSeqLogli -= log(Zx);
// correct for the fact that alpha-s were scaled.
for (int i = 0; i < dataSeq.length(); i++) {
thisSeqLogli -= log(scale[i]);
}
logli += thisSeqLogli;
// update grad.
for (int f = 0; f < grad.length; f++) grad[f] -= ExpF[f] / Zx;
if (params.debugLvl > 1) {
System.out.println(
"Sequence "
+ thisSeqLogli
+ " logli "
+ logli
+ " log(Zx) "
+ Math.log(Zx)
+ " Zx "
+ Zx);
}
}
if (params.debugLvl > 2) {
for (int f = 0; f < lambda.length; f++) System.out.print(lambda[f] + " ");
System.out.println(" :x");
for (int f = 0; f < lambda.length; f++)
System.out.println(featureGenerator.featureName(f) + " " + grad[f] + " ");
System.out.println(" :g");
}
if (params.debugLvl > 0)
Util.printDbg(
"Iter "
+ icall
+ " log likelihood "
+ logli
+ " norm(grad logli) "
+ norm(grad)
+ " norm(x) "
+ norm(lambda));
if (icall == 0) {
System.out.println("Number of training records" + numRecord);
}
} catch (Exception e) {
System.out.println("Alpha-i " + alpha_Y.toString());
System.out.println("Ri " + Ri_Y.toString());
System.out.println("Mi " + Mi_YY.toString());
e.printStackTrace();
System.exit(0);
}
return logli;
}