public Vec optimizeColumn(final FuncC1 func, final Vec codingColumn) {
final Vec mu = new ArrayVec(codingColumn.dim());
for (int i = 0; i < mu.dim(); i++) {
final double code = codingColumn.get(i);
if (code == 1.0) mu.set(i, 1.0);
else if (code == -1.0) mu.set(i, 0.0);
else mu.set(i, 0.5);
}
double error = 100500;
while (error > 1e-3) {
final Vec muPrev = VecTools.copy(mu);
final Vec gradient = func.gradient(mu);
VecTools.incscale(mu, gradient, -step);
for (int i = 0; i < mu.dim(); i++) {
final double code = codingColumn.get(i);
final double val = mu.get(i);
if (code == 1.0 || val > 1.0) {
mu.set(i, 1.0);
} else if (code == -1.0 || val < 0) {
mu.set(i, 0);
}
}
System.out.println(mu);
error = VecTools.norm(VecTools.subtract(muPrev, mu));
}
return new ArrayVec(codingColumn.dim());
}
@Override
public Ensemble fit(final VecDataSet learn, final GlobalLoss globalLoss) {
final Vec cursor = new ArrayVec(globalLoss.xdim());
final List<Trans> weakModels = new ArrayList<>(iterationsCount);
final Trans gradient = globalLoss.gradient();
for (int t = 0; t < iterationsCount; t++) {
final Vec gradientValueAtCursor = gradient.trans(cursor);
final L2 localLoss = DataTools.newTarget(factory, gradientValueAtCursor, learn);
final Trans weakModel = weak.fit(learn, localLoss);
weakModels.add(weakModel);
invoke(new Ensemble(weakModels, -step));
VecTools.append(cursor, VecTools.scale(weakModel.transAll(learn.data()), -step));
}
return new Ensemble(weakModels, -step);
}
public CMLMetricOptimization(
final VecDataSet ds,
final BlockwiseMLLLogit target,
final Mx S,
final double c,
final double step) {
this.ds = ds;
this.target = target;
this.step = step;
this.classesIdxs = MCTools.splitClassesIdxs(target.labels());
this.laplacian = VecTools.copy(S);
VecTools.scale(laplacian, -1.0);
for (int i = 0; i < laplacian.rows(); i++) {
final double diagElem = VecTools.sum(S.row(i));
laplacian.adjust(i, i, diagElem);
}
this.c = c;
}
public Mx trainProbs(final Mx codingMatrix, final Func[] binClassifiers) {
final Mx result = new VecBasedMx(codingMatrix.rows(), codingMatrix.columns());
for (int l = 0; l < result.columns(); l++) {
System.out.println("Optimize column " + l);
final FuncC1 columnTargetFunction = new ColumnTargetFunction(binClassifiers[l]);
final Vec muColumn = optimizeColumn(columnTargetFunction, codingMatrix.col(l));
VecTools.assign(result.col(l), muColumn);
}
return result;
}
@Override
public double value(final Vec x) {
final Mx predictMx = (Mx) x;
int count = 0;
for (int i = 0; i < predictMx.rows(); i++) {
if (VecTools.distance(predictMx.row(i), targets.row(i)) < MathTools.EPSILON) {
count++;
}
}
return (double) count / targets.rows();
}
@Override
public double value(final Vec mu) {
double result = 0.0;
for (int i = 0; i < ds.length(); i++) {
final double trans = binClassifier.value(ds.data().row(i));
final double sigmoid = MathTools.sigmoid(trans);
final double underLog =
mu.get(target.label(i)) * sigmoid + (1 - mu.get(target.label(i))) * (1 - sigmoid);
result -= Math.log(underLog);
}
result += c * VecTools.multiply(MxTools.multiply(laplacian, mu), mu);
return result;
}
@Override
public Vec gradient(final Vec mu) {
final Vec grad = new ArrayVec(mu.dim());
for (int k = 0; k < grad.dim(); k++) {
final TIntList idxs = classesIdxs.get(k);
double val = 0.0;
for (final TIntIterator listIter = idxs.iterator(); listIter.hasNext(); ) {
final Vec x = ds.data().row(listIter.next());
final double trans = binClassifier.value(x);
final double sigmoid = MathTools.sigmoid(trans);
val -= (2 * sigmoid - 1) / (mu.get(k) * sigmoid + (1 - mu.get(k)) * (1 - sigmoid));
grad.set(k, val);
}
}
final double norm = VecTools.norm(grad);
VecTools.scale(grad, 1 / norm);
for (int k = 0; k < grad.dim(); k++) {
final double val = VecTools.multiply(laplacian.row(k), mu);
grad.adjust(k, val);
}
return grad;
}
public void backward() {
Mx cnc = null;
if (bias_b != 0) {
cnc = leftContract(output);
} else {
cnc = VecTools.copy(output);
}
difference = MxTools.multiply(MxTools.transpose(cnc), activations);
for (int i = 0; i < difference.dim(); i++) {
difference.set(i, difference.get(i) / activations.rows());
}
input = MxTools.multiply(cnc, weights);
rectifier.grad(activations, activations);
for (int i = 0; i < input.dim(); i++) {
input.set(i, input.get(i) * activations.get(i));
if (dropoutFraction > 0) {
input.set(i, input.get(i) * dropoutMask.get(i));
}
}
}
public void forward() {
if (bias != 0) {
activations = leftExtend(input);
} else {
activations = VecTools.copy(input);
}
output = MxTools.multiply(activations, MxTools.transpose(weights));
rectifier.value(output, output);
if (dropoutFraction > 0) {
if (isTrain) {
dropoutMask = getDropoutMask();
for (int i = 0; i < output.dim(); i++) {
output.set(i, output.get(i) * dropoutMask.get(i));
}
} else {
for (int i = 0; i < output.dim(); i++) {
output.set(i, output.get(i) * (1 - dropoutFraction));
}
}
}
}
public Vec gradientTo(Vec x, Vec to) {
final Vec trans = gradient(x);
VecTools.assign(to, trans);
return to;
}
