Examples with NonPositiveDefiniteMatrixException org.apache.commons.math3.linear.NonPositiveDefiniteMatrixException used on opensource projects

Example 1 with NonPositiveDefiniteMatrixException

use of org.apache.commons.math3.linear.NonPositiveDefiniteMatrixException in project GDSC-SMLM by aherbert.

the class TrackPopulationAnalysis method createMixed.

/**
 * Creates the multivariate gaussian mixture as the best of many repeats of the expectation
 * maximisation algorithm.
 *
 * @param data the data
 * @param dimensions the dimensions
 * @param numComponents the number of components
 * @return the multivariate gaussian mixture expectation maximization
 */
private MultivariateGaussianMixtureExpectationMaximization createMixed(final double[][] data, int dimensions, int numComponents) {
    // Fit a mixed multivariate Gaussian with different repeats.
    final UnitSphereSampler sampler = UnitSphereSampler.of(UniformRandomProviders.create(Mixers.stafford13(settings.seed++)), dimensions);
    final LocalList<CompletableFuture<MultivariateGaussianMixtureExpectationMaximization>> results = new LocalList<>(settings.repeats);
    final DoubleDoubleBiPredicate test = createConvergenceTest(settings.relativeError);
    if (settings.debug) {
        ImageJUtils.log("  Fitting %d components", numComponents);
    }
    final Ticker ticker = ImageJUtils.createTicker(settings.repeats, 2, "Fitting...");
    final AtomicInteger failures = new AtomicInteger();
    for (int i = 0; i < settings.repeats; i++) {
        final double[] vector = sampler.sample();
        results.add(CompletableFuture.supplyAsync(() -> {
            final MultivariateGaussianMixtureExpectationMaximization fitter = new MultivariateGaussianMixtureExpectationMaximization(data);
            try {
                // This may also throw the same exceptions due to inversion of the covariance matrix
                final MixtureMultivariateGaussianDistribution initialMixture = MultivariateGaussianMixtureExpectationMaximization.estimate(data, numComponents, point -> {
                    double dot = 0;
                    for (int j = 0; j < dimensions; j++) {
                        dot += vector[j] * point[j];
                    }
                    return dot;
                });
                final boolean result = fitter.fit(initialMixture, settings.maxIterations, test);
                // Log the result. Note: The ImageJ log is synchronized.
                if (settings.debug) {
                    ImageJUtils.log("  Fit: log-likelihood=%s, iter=%d, converged=%b", fitter.getLogLikelihood(), fitter.getIterations(), result);
                }
                return result ? fitter : null;
            } catch (NonPositiveDefiniteMatrixException | SingularMatrixException ex) {
                failures.getAndIncrement();
                if (settings.debug) {
                    ImageJUtils.log("  Fit failed during iteration %d. No variance in a sub-population " + "component (check alpha is not always 1.0).", fitter.getIterations());
                }
            } finally {
                ticker.tick();
            }
            return null;
        }));
    }
    ImageJUtils.finished();
    if (failures.get() != 0 && settings.debug) {
        ImageJUtils.log("  %d component fit failed %d/%d", numComponents, failures.get(), settings.repeats);
    }
    // Collect results and return the best model.
    return results.stream().map(f -> f.join()).filter(f -> f != null).sorted((f1, f2) -> Double.compare(f2.getLogLikelihood(), f1.getLogLikelihood())).findFirst().orElse(null);
}