private double multiLL(DoubleMatrix2D coeffs, Node dep, List<Node> indep) {
DoubleMatrix2D indepData =
factory2D.make(internalData.subsetColumns(indep).getDoubleData().toArray());
List<Node> depList = new ArrayList<>();
depList.add(dep);
DoubleMatrix2D depData =
factory2D.make(internalData.subsetColumns(depList).getDoubleData().toArray());
int N = indepData.rows();
DoubleMatrix2D probs =
Algebra.DEFAULT.mult(factory2D.appendColumns(factory2D.make(N, 1, 1.0), indepData), coeffs);
probs =
factory2D
.appendColumns(factory2D.make(indepData.rows(), 1, 1.0), probs)
.assign(Functions.exp);
double ll = 0;
for (int i = 0; i < N; i++) {
DoubleMatrix1D curRow = probs.viewRow(i);
curRow.assign(Functions.div(curRow.zSum()));
ll += Math.log(curRow.get((int) depData.get(i, 0)));
}
return ll;
}
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 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;
}