@Override
public void performExperiment() throws IOException {
BilinearLearnerParameters params = new BilinearLearnerParameters();
params.put(BilinearLearnerParameters.ETA0_U, 0.02);
params.put(BilinearLearnerParameters.ETA0_W, 0.02);
params.put(BilinearLearnerParameters.LAMBDA, 0.001);
params.put(BilinearLearnerParameters.BICONVEX_TOL, 0.01);
params.put(BilinearLearnerParameters.BICONVEX_MAXITER, 10);
params.put(BilinearLearnerParameters.BIAS, true);
params.put(BilinearLearnerParameters.ETA0_BIAS, 0.5);
params.put(BilinearLearnerParameters.WINITSTRAT, new SingleValueInitStrat(0.1));
params.put(BilinearLearnerParameters.UINITSTRAT, new SparseZerosInitStrategy());
BillMatlabFileDataGenerator bmfdg =
new BillMatlabFileDataGenerator(new File(BILL_DATA()), 98, true);
prepareExperimentLog(params);
for (int i = 0; i < bmfdg.nFolds(); i++) {
logger.debug("Fold: " + i);
BilinearSparseOnlineLearner learner = new BilinearSparseOnlineLearner(params);
learner.reinitParams();
bmfdg.setFold(i, Mode.TEST);
List<Pair<Matrix>> testpairs = new ArrayList<Pair<Matrix>>();
while (true) {
Pair<Matrix> next = bmfdg.generate();
if (next == null) break;
testpairs.add(next);
}
logger.debug("...training");
bmfdg.setFold(i, Mode.TRAINING);
int j = 0;
while (true) {
Pair<Matrix> next = bmfdg.generate();
if (next == null) break;
logger.debug("...trying item " + j++);
learner.process(next.firstObject(), next.secondObject());
Matrix u = learner.getU();
Matrix w = learner.getW();
Matrix bias = MatrixFactory.getDenseDefault().copyMatrix(learner.getBias());
BilinearEvaluator eval = new RootMeanSumLossEvaluator();
eval.setLearner(learner);
double loss = eval.evaluate(testpairs);
logger.debug(String.format("Saving learner, Fold %d, Item %d", i, j));
File learnerOut = new File(FOLD_ROOT(i), String.format("learner_%d", j));
IOUtils.writeBinary(learnerOut, learner);
logger.debug("W row sparcity: " + SandiaMatrixUtils.rowSparcity(w));
logger.debug("U row sparcity: " + SandiaMatrixUtils.rowSparcity(u));
Boolean biasMode = learner.getParams().getTyped(BilinearLearnerParameters.BIAS);
if (biasMode) {
logger.debug("Bias: " + SandiaMatrixUtils.diag(bias));
}
logger.debug(String.format("... loss: %f", loss));
}
}
}
public double sumLoss(
List<Pair<Matrix>> pairs, Matrix u, Matrix w, Matrix bias, BilinearLearnerParameters params) {
LossFunction loss = params.getTyped(BilinearLearnerParameters.LOSS);
loss = new MatLossFunction(loss);
double total = 0;
int i = 0;
int ntasks = 0;
for (Pair<Matrix> pair : pairs) {
Matrix X = pair.firstObject();
Matrix Y = pair.secondObject();
SparseMatrix Yexp = BilinearSparseOnlineLearner.expandY(Y);
Matrix expectedAll = u.transpose().times(X.transpose()).times(w);
loss.setY(Yexp);
loss.setX(expectedAll);
if (bias != null) loss.setBias(bias);
logger.debug("Testing pair: " + i);
total += loss.eval(null); // Assums an identity w.
i++;
ntasks += Y.getNumColumns();
}
total /= ntasks;
return Math.sqrt(total);
}
