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;
}
/** Creates a new instance of testmatrix */
public GLSsolver(double[][] p_MatrixgleichNull) throws IllegalArgumentException {
// --------------------------------------
// Kontrolle, ob Eingabematrix rechteckig
// --------------------------------------
int nplus1 = p_MatrixgleichNull[0].length;
for (int i = 1; i < p_MatrixgleichNull.length; i++) { // Zeilen i
if (p_MatrixgleichNull[i].length != nplus1) {
System.err.println(
"Programmfehler: Matrix des GLS ist nicht rechteckig! (im solver entdeckt)");
throw new IllegalArgumentException();
}
}
if (nplus1 <= 1) throw new IllegalArgumentException("keine Unbekannte"); // keine Unbekannte!!!
// Umgeht einen Fehler in der colt-Bibliothek // TODO wenn behoben, Workaround entfernen
// ------
int anzGl = p_MatrixgleichNull.length;
if (anzGl < nplus1 - 1) { // anzGleichungen < anz Unbekannte
if (debug) System.out.println("WorkAround fuer Fehler in colt: 0 = 0 Gleichungen anhaengen");
anzGl = nplus1 - 1; // = Anzahl Unbek, 0 0 0 ... 0 = 0 Zeile angehängt
}
// -------------------------
// Daten in A und b einlesen
// -------------------------
// so dass A*x = b
A = new DenseDoubleMatrix2D(anzGl, (nplus1 - 1));
DenseDoubleMatrix2D b = new DenseDoubleMatrix2D(anzGl, 1);
for (int i = 0; i < p_MatrixgleichNull.length; i++) { // Zeilen i
for (int j = 0; j < nplus1 - 1; j++) { // Spalten
A.set(i, j, p_MatrixgleichNull[i][j]);
}
b.set(i, 0, -p_MatrixgleichNull[i][nplus1 - 1]);
}
if (debug) {
System.out.println(" A = " + A.toString());
System.out.println(" b = " + b.toString());
System.out.println("");
}
// --------------
// LR - Zerlegung
// --------------
LUDecomposition ALU = new LUDecomposition(A);
if (debug) System.out.println(ALU.toString());
DoubleMatrix2D L = ALU.getL();
R = ALU.getU();
int[] piv = ALU.getPivot();
Algebra alg = new Algebra();
// if (debug) System.out.println("L = " + L.toString());
// if (debug) System.out.println("Kontrolle L*R = " + alg.mult(L,R).toString());
// if (debug) System.out.println("Kontrolle P*b = " + alg.permute(b, piv, null) );
//
// if (debug) System.out.println("Rx = c: R = " + R.toString());
// if (debug) System.out.println("alg.permute(b, piv, null) = " + alg.permute(b, piv,
// null).toString());
c = alg.solve(L, alg.permute(b, piv, null)); // TODO: kann zu Problemen führen,
// wenn weniger Gleichungen als Unbek --> s.Workaround oben
if (debug) System.out.println("Lc = Pb: c = " + c.toString());
if (debug) {
System.out.println("Rang A: " + alg.rank(A));
System.out.println("Rang R: " + alg.rank(R));
}
assert (alg.rank(A) == alg.rank(R)) : "Rang von A ungleich Rang von R --> Programmfehler";
anzUnbestParam = A.columns() - alg.rank(A);
if (debug) System.out.println("Anz unbest Parameter: " + anzUnbestParam);
}
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;
}
