Search in sources :

Example 31 with Well19937c

use of org.apache.commons.math3.random.Well19937c in project GDSC-SMLM by aherbert.

the class SCMOSLikelihoodWrapperTest method functionComputesTargetGradientPerDatum.

private void functionComputesTargetGradientPerDatum(Gaussian2DFunction f1, int targetParameter) {
    int[] indices = f1.gradientIndices();
    int gradientIndex = findGradientIndex(f1, targetParameter);
    double[] dyda = new double[indices.length];
    double[] a;
    SCMOSLikelihoodWrapper ff1;
    int n = maxx * maxx;
    int count = 0, total = 0;
    RandomDataGenerator rdg = new RandomDataGenerator(new Well19937c(30051977));
    for (double background : testbackground) for (double signal1 : testsignal1) for (double angle1 : testangle1) for (double cx1 : testcx1) for (double cy1 : testcy1) for (double[] w1 : testw1) {
        a = createParameters(background, signal1, angle1, cx1, cy1, w1[0], w1[1]);
        // Create y as a function we would want to move towards
        double[] a2 = a.clone();
        a2[targetParameter] *= 1.1;
        f1.initialise(a2);
        double[] data = new double[n];
        for (int i = 0; i < n; i++) {
            // Simulate sCMOS camera
            double u = f1.eval(i);
            data[i] = rdg.nextPoisson(u) * g[i] + rdg.nextGaussian(o[i], sd[i]);
        }
        ff1 = new SCMOSLikelihoodWrapper(f1, a, data, n, var, g, o);
        // Numerically solve gradient. 
        // Calculate the step size h to be an exact numerical representation
        final double xx = a[targetParameter];
        // Get h to minimise roundoff error
        double h = Precision.representableDelta(xx, h_);
        for (int x : testx) for (int y : testy) {
            int i = y * maxx + x;
            a[targetParameter] = xx;
            ff1.likelihood(getVariables(indices, a), dyda, i);
            // Evaluate at (x+h) and (x-h)
            a[targetParameter] = xx + h;
            double value2 = ff1.likelihood(getVariables(indices, a), i);
            a[targetParameter] = xx - h;
            double value3 = ff1.likelihood(getVariables(indices, a), i);
            double gradient = (value2 - value3) / (2 * h);
            boolean ok = Math.signum(gradient) == Math.signum(dyda[gradientIndex]) || Math.abs(gradient - dyda[gradientIndex]) < 0.1;
            //		dyda[gradientIndex]);
            if (!ok)
                Assert.assertTrue(NAME[targetParameter] + ": " + gradient + " != " + dyda[gradientIndex], ok);
            ok = eqPerDatum.almostEqualRelativeOrAbsolute(gradient, dyda[gradientIndex]);
            if (ok)
                count++;
            total++;
        }
    }
    double p = (100.0 * count) / total;
    logf("Per Datum %s : %s = %d / %d (%.2f)\n", f1.getClass().getSimpleName(), NAME[targetParameter], count, total, p);
    Assert.assertTrue(NAME[targetParameter] + " fraction too low per datum: " + p, p > 90);
}
Also used : RandomDataGenerator(org.apache.commons.math3.random.RandomDataGenerator) Well19937c(org.apache.commons.math3.random.Well19937c)

Example 32 with Well19937c

use of org.apache.commons.math3.random.Well19937c in project GDSC-SMLM by aherbert.

the class SCMOSLikelihoodWrapperTest method canComputePValue.

private void canComputePValue(BaseNonLinearFunction nlf) {
    System.out.println(nlf.name);
    int n = maxx * maxx;
    double[] a = new double[] { 1 };
    // Simulate sCMOS camera
    nlf.initialise(a);
    RandomDataGenerator rdg = new RandomDataGenerator(new Well19937c(30051977));
    double[] k = Utils.newArray(n, 0, 1.0);
    for (int i = 0; i < n; i++) {
        double u = nlf.eval(i);
        if (u > 0)
            u = rdg.nextPoisson(u);
        k[i] = u * g[i] + rdg.nextGaussian(o[i], sd[i]);
    }
    SCMOSLikelihoodWrapper f = new SCMOSLikelihoodWrapper(nlf, a, k, n, var, g, o);
    double oll = f.computeObservedLikelihood();
    double oll2 = 0;
    double[] op = new double[n];
    for (int j = 0; j < n; j++) {
        op[j] = SCMOSLikelihoodWrapper.likelihood((k[j] - o[j]) / g[j], var[j], g[j], o[j], k[j]);
        oll2 -= Math.log(op[j]);
    }
    System.out.printf("oll=%f, oll2=%f\n", oll, oll2);
    Assert.assertEquals("Observed Log-likelihood", oll2, oll, oll2 * 1e-10);
    double min = Double.POSITIVE_INFINITY;
    double mina = 0;
    for (int i = 5; i <= 15; i++) {
        a[0] = (double) i / 10;
        double ll = f.likelihood(a);
        double llr = f.computeLogLikelihoodRatio(ll);
        BigDecimal product = new BigDecimal(1);
        double ll2 = 0;
        for (int j = 0; j < n; j++) {
            double p1 = SCMOSLikelihoodWrapper.likelihood(nlf.eval(j), var[j], g[j], o[j], k[j]);
            ll2 -= Math.log(p1);
            double ratio = p1 / op[j];
            product = product.multiply(new BigDecimal(ratio));
        }
        double llr2 = -2 * Math.log(product.doubleValue());
        double q = f.computeQValue(ll);
        System.out.printf("a=%f, ll=%f, ll2=%f, llr=%f, llr2=%f, product=%s, p=%f\n", a[0], ll, ll2, llr, llr2, product.round(new MathContext(4)).toString(), q);
        if (min > ll) {
            min = ll;
            mina = a[0];
        }
        // too small to store in a double.
        if (product.doubleValue() > 0)
            Assert.assertEquals("Log-likelihood", llr, llr2, llr * 1e-10);
    }
    Assert.assertEquals("min", 1, mina, 0);
}
Also used : RandomDataGenerator(org.apache.commons.math3.random.RandomDataGenerator) Well19937c(org.apache.commons.math3.random.Well19937c) BigDecimal(java.math.BigDecimal) MathContext(java.math.MathContext)

Example 33 with Well19937c

use of org.apache.commons.math3.random.Well19937c in project GDSC-SMLM by aherbert.

the class PCPALMMolecules method runSimulation.

private void runSimulation(boolean resultsAvailable) {
    if (resultsAvailable && !showSimulationDialog())
        return;
    startLog();
    log("Simulation parameters");
    if (blinkingDistribution == 3) {
        log("  - Clusters = %d", nMolecules);
        log("  - Simulation size = %s um", Utils.rounded(simulationSize, 4));
        log("  - Molecules/cluster = %s", Utils.rounded(blinkingRate, 4));
        log("  - Blinking distribution = %s", BLINKING_DISTRIBUTION[blinkingDistribution]);
        log("  - p-Value = %s", Utils.rounded(p, 4));
    } else {
        log("  - Molecules = %d", nMolecules);
        log("  - Simulation size = %s um", Utils.rounded(simulationSize, 4));
        log("  - Blinking rate = %s", Utils.rounded(blinkingRate, 4));
        log("  - Blinking distribution = %s", BLINKING_DISTRIBUTION[blinkingDistribution]);
    }
    log("  - Average precision = %s nm", Utils.rounded(sigmaS, 4));
    log("  - Clusters simulation = " + CLUSTER_SIMULATION[clusterSimulation]);
    if (clusterSimulation > 0) {
        log("  - Cluster number = %s +/- %s", Utils.rounded(clusterNumber, 4), Utils.rounded(clusterNumberSD, 4));
        log("  - Cluster radius = %s nm", Utils.rounded(clusterRadius, 4));
    }
    final double nmPerPixel = 100;
    double width = simulationSize * 1000.0;
    // Allow a border of 3 x sigma for +/- precision
    //if (blinkingRate > 1)
    width -= 3 * sigmaS;
    RandomGenerator randomGenerator = new Well19937c(System.currentTimeMillis() + System.identityHashCode(this));
    RandomDataGenerator dataGenerator = new RandomDataGenerator(randomGenerator);
    UniformDistribution dist = new UniformDistribution(null, new double[] { width, width, 0 }, randomGenerator.nextInt());
    molecules = new ArrayList<Molecule>(nMolecules);
    // Create some dummy results since the calibration is required for later analysis
    results = new MemoryPeakResults();
    results.setCalibration(new gdsc.smlm.results.Calibration(nmPerPixel, 1, 100));
    results.setSource(new NullSource("Molecule Simulation"));
    results.begin();
    int count = 0;
    // Generate a sequence of coordinates
    ArrayList<double[]> xyz = new ArrayList<double[]>((int) (nMolecules * 1.1));
    Statistics statsRadius = new Statistics();
    Statistics statsSize = new Statistics();
    String maskTitle = TITLE + " Cluster Mask";
    ByteProcessor bp = null;
    double maskScale = 0;
    if (clusterSimulation > 0) {
        // Simulate clusters.
        // Note: In the Veatch et al. paper (Plos 1, e31457) correlation functions are built using circles
        // with small radii of 4-8 Arbitrary Units (AU) or large radii of 10-30 AU. A fluctuations model is
        // created at T = 1.075 Tc. It is not clear exactly how the particles are distributed.
        // It may be that a mask is created first using the model. The particles are placed on the mask using
        // a specified density. This simulation produces a figure to show either a damped cosine function
        // (circles) or an exponential (fluctuations). The number of particles in each circle may be randomly
        // determined just by density. The figure does not discuss the derivation of the cluster size 
        // statistic.
        // 
        // If this plugin simulation is run with a uniform distribution and blinking rate of 1 then the damped
        // cosine function is reproduced. The curve crosses g(r)=1 at a value equivalent to the average
        // distance to the centre-of-mass of each drawn cluster, not the input cluster radius parameter (which 
        // is a hard upper limit on the distance to centre).
        final int maskSize = lowResolutionImageSize;
        int[] mask = null;
        // scale is in nm/pixel
        maskScale = width / maskSize;
        ArrayList<double[]> clusterCentres = new ArrayList<double[]>();
        int totalSteps = 1 + (int) Math.ceil(nMolecules / clusterNumber);
        if (clusterSimulation == 2 || clusterSimulation == 3) {
            // Clusters are non-overlapping circles
            // Ensure the circles do not overlap by using an exclusion mask that accumulates 
            // out-of-bounds pixels by drawing the last cluster (plus some border) on an image. When no
            // more pixels are available then stop generating molecules.
            // This is done by cumulatively filling a mask and using the MaskDistribution to select 
            // a new point. This may be slow but it works.
            // TODO - Allow clusters of different sizes...
            mask = new int[maskSize * maskSize];
            Arrays.fill(mask, 255);
            MaskDistribution maskDistribution = new MaskDistribution(mask, maskSize, maskSize, 0, maskScale, maskScale, randomGenerator);
            double[] centre;
            IJ.showStatus("Computing clusters mask");
            int roiRadius = (int) Math.round((clusterRadius * 2) / maskScale);
            if (clusterSimulation == 3) {
                // Generate a mask of circles then sample from that.
                // If we want to fill the mask completely then adjust the total steps to be the number of 
                // circles that can fit inside the mask.
                totalSteps = (int) (maskSize * maskSize / (Math.PI * Math.pow(clusterRadius / maskScale, 2)));
            }
            while ((centre = maskDistribution.next()) != null && clusterCentres.size() < totalSteps) {
                IJ.showProgress(clusterCentres.size(), totalSteps);
                // The mask returns the coordinates with the centre of the image at 0,0
                centre[0] += width / 2;
                centre[1] += width / 2;
                clusterCentres.add(centre);
                // Fill in the mask around the centre to exclude any more circles that could overlap
                double cx = centre[0] / maskScale;
                double cy = centre[1] / maskScale;
                fillMask(mask, maskSize, (int) cx, (int) cy, roiRadius, 0);
                //Utils.display("Mask", new ColorProcessor(maskSize, maskSize, mask));
                try {
                    maskDistribution = new MaskDistribution(mask, maskSize, maskSize, 0, maskScale, maskScale, randomGenerator);
                } catch (IllegalArgumentException e) {
                    // This can happen when there are no more non-zero pixels
                    log("WARNING: No more room for clusters on the mask area (created %d of estimated %d)", clusterCentres.size(), totalSteps);
                    break;
                }
            }
            IJ.showProgress(1);
            IJ.showStatus("");
        } else {
            // Pick centres randomly from the distribution 
            while (clusterCentres.size() < totalSteps) clusterCentres.add(dist.next());
        }
        if (showClusterMask || clusterSimulation == 3) {
            // Show the mask for the clusters
            if (mask == null)
                mask = new int[maskSize * maskSize];
            else
                Arrays.fill(mask, 0);
            int roiRadius = (int) Math.round((clusterRadius) / maskScale);
            for (double[] c : clusterCentres) {
                double cx = c[0] / maskScale;
                double cy = c[1] / maskScale;
                fillMask(mask, maskSize, (int) cx, (int) cy, roiRadius, 1);
            }
            if (clusterSimulation == 3) {
                // We have the mask. Now pick points at random from the mask.
                MaskDistribution maskDistribution = new MaskDistribution(mask, maskSize, maskSize, 0, maskScale, maskScale, randomGenerator);
                // Allocate each molecule position to a parent circle so defining clusters.
                int[][] clusters = new int[clusterCentres.size()][];
                int[] clusterSize = new int[clusters.length];
                for (int i = 0; i < nMolecules; i++) {
                    double[] centre = maskDistribution.next();
                    // The mask returns the coordinates with the centre of the image at 0,0
                    centre[0] += width / 2;
                    centre[1] += width / 2;
                    xyz.add(centre);
                    // Output statistics on cluster size and number.
                    // TODO - Finding the closest cluster could be done better than an all-vs-all comparison
                    double max = distance2(centre, clusterCentres.get(0));
                    int cluster = 0;
                    for (int j = 1; j < clusterCentres.size(); j++) {
                        double d2 = distance2(centre, clusterCentres.get(j));
                        if (d2 < max) {
                            max = d2;
                            cluster = j;
                        }
                    }
                    // Assign point i to cluster
                    centre[2] = cluster;
                    if (clusterSize[cluster] == 0) {
                        clusters[cluster] = new int[10];
                    }
                    if (clusters[cluster].length <= clusterSize[cluster]) {
                        clusters[cluster] = Arrays.copyOf(clusters[cluster], (int) (clusters[cluster].length * 1.5));
                    }
                    clusters[cluster][clusterSize[cluster]++] = i;
                }
                // Generate real cluster size statistics
                for (int j = 0; j < clusterSize.length; j++) {
                    final int size = clusterSize[j];
                    if (size == 0)
                        continue;
                    statsSize.add(size);
                    if (size == 1) {
                        statsRadius.add(0);
                        continue;
                    }
                    // Find centre of cluster and add the distance to each point
                    double[] com = new double[2];
                    for (int n = 0; n < size; n++) {
                        double[] xy = xyz.get(clusters[j][n]);
                        for (int k = 0; k < 2; k++) com[k] += xy[k];
                    }
                    for (int k = 0; k < 2; k++) com[k] /= size;
                    for (int n = 0; n < size; n++) {
                        double dx = xyz.get(clusters[j][n])[0] - com[0];
                        double dy = xyz.get(clusters[j][n])[1] - com[1];
                        statsRadius.add(Math.sqrt(dx * dx + dy * dy));
                    }
                }
            }
            if (showClusterMask) {
                bp = new ByteProcessor(maskSize, maskSize);
                for (int i = 0; i < mask.length; i++) if (mask[i] != 0)
                    bp.set(i, 128);
                Utils.display(maskTitle, bp);
            }
        }
        // Use the simulated cluster centres to create clusters of the desired size
        if (clusterSimulation == 1 || clusterSimulation == 2) {
            for (double[] clusterCentre : clusterCentres) {
                int clusterN = (int) Math.round((clusterNumberSD > 0) ? dataGenerator.nextGaussian(clusterNumber, clusterNumberSD) : clusterNumber);
                if (clusterN < 1)
                    continue;
                //double[] clusterCentre = dist.next();
                if (clusterN == 1) {
                    // No need for a cluster around a point
                    xyz.add(clusterCentre);
                    statsRadius.add(0);
                    statsSize.add(1);
                } else {
                    // Generate N random points within a circle of the chosen cluster radius.
                    // Locate the centre-of-mass and the average distance to the centre.
                    double[] com = new double[3];
                    int j = 0;
                    while (j < clusterN) {
                        // Generate a random point within a circle uniformly
                        // http://stackoverflow.com/questions/5837572/generate-a-random-point-within-a-circle-uniformly
                        double t = 2.0 * Math.PI * randomGenerator.nextDouble();
                        double u = randomGenerator.nextDouble() + randomGenerator.nextDouble();
                        double r = clusterRadius * ((u > 1) ? 2 - u : u);
                        double x = r * Math.cos(t);
                        double y = r * Math.sin(t);
                        double[] xy = new double[] { clusterCentre[0] + x, clusterCentre[1] + y };
                        xyz.add(xy);
                        for (int k = 0; k < 2; k++) com[k] += xy[k];
                        j++;
                    }
                    // Add the distance of the points from the centre of the cluster.
                    // Note this does not account for the movement due to precision.
                    statsSize.add(j);
                    if (j == 1) {
                        statsRadius.add(0);
                    } else {
                        for (int k = 0; k < 2; k++) com[k] /= j;
                        while (j > 0) {
                            double dx = xyz.get(xyz.size() - j)[0] - com[0];
                            double dy = xyz.get(xyz.size() - j)[1] - com[1];
                            statsRadius.add(Math.sqrt(dx * dx + dy * dy));
                            j--;
                        }
                    }
                }
            }
        }
    } else {
        // Random distribution
        for (int i = 0; i < nMolecules; i++) xyz.add(dist.next());
    }
    // The Gaussian sigma should be applied so the overall distance from the centre
    // ( sqrt(x^2+y^2) ) has a standard deviation of sigmaS?
    final double sigma1D = sigmaS / Math.sqrt(2);
    // Show optional histograms
    StoredDataStatistics intraDistances = null;
    StoredData blinks = null;
    if (showHistograms) {
        int capacity = (int) (xyz.size() * blinkingRate);
        intraDistances = new StoredDataStatistics(capacity);
        blinks = new StoredData(capacity);
    }
    Statistics statsSigma = new Statistics();
    for (int i = 0; i < xyz.size(); i++) {
        int nOccurrences = getBlinks(dataGenerator, blinkingRate);
        if (showHistograms)
            blinks.add(nOccurrences);
        final int size = molecules.size();
        // Get coordinates in nm
        final double[] moleculeXyz = xyz.get(i);
        if (bp != null && nOccurrences > 0) {
            bp.putPixel((int) Math.round(moleculeXyz[0] / maskScale), (int) Math.round(moleculeXyz[1] / maskScale), 255);
        }
        while (nOccurrences-- > 0) {
            final double[] localisationXy = Arrays.copyOf(moleculeXyz, 2);
            // Add random precision
            if (sigma1D > 0) {
                final double dx = dataGenerator.nextGaussian(0, sigma1D);
                final double dy = dataGenerator.nextGaussian(0, sigma1D);
                localisationXy[0] += dx;
                localisationXy[1] += dy;
                if (!dist.isWithinXY(localisationXy))
                    continue;
                // Calculate mean-squared displacement
                statsSigma.add(dx * dx + dy * dy);
            }
            final double x = localisationXy[0];
            final double y = localisationXy[1];
            molecules.add(new Molecule(x, y, i, 1));
            // Store in pixels
            float[] params = new float[7];
            params[Gaussian2DFunction.X_POSITION] = (float) (x / nmPerPixel);
            params[Gaussian2DFunction.Y_POSITION] = (float) (y / nmPerPixel);
            results.addf(i + 1, (int) x, (int) y, 0, 0, 0, params, null);
        }
        if (molecules.size() > size) {
            count++;
            if (showHistograms) {
                int newCount = molecules.size() - size;
                if (newCount == 1) {
                    //intraDistances.add(0);
                    continue;
                }
                // Get the distance matrix between these molecules
                double[][] matrix = new double[newCount][newCount];
                for (int ii = size, x = 0; ii < molecules.size(); ii++, x++) {
                    for (int jj = size + 1, y = 1; jj < molecules.size(); jj++, y++) {
                        final double d2 = molecules.get(ii).distance2(molecules.get(jj));
                        matrix[x][y] = matrix[y][x] = d2;
                    }
                }
                // Get the maximum distance for particle linkage clustering of this molecule
                double max = 0;
                for (int x = 0; x < newCount; x++) {
                    // Compare to all-other molecules and get the minimum distance 
                    // needed to join at least one
                    double linkDistance = Double.POSITIVE_INFINITY;
                    for (int y = 0; y < newCount; y++) {
                        if (x == y)
                            continue;
                        if (matrix[x][y] < linkDistance)
                            linkDistance = matrix[x][y];
                    }
                    // Check if this is larger 
                    if (max < linkDistance)
                        max = linkDistance;
                }
                intraDistances.add(Math.sqrt(max));
            }
        }
    }
    results.end();
    if (bp != null)
        Utils.display(maskTitle, bp);
    // Used for debugging
    //System.out.printf("  * Molecules = %d (%d activated)\n", xyz.size(), count);
    //if (clusterSimulation > 0)
    //	System.out.printf("  * Cluster number = %s +/- %s. Radius = %s +/- %s\n",
    //			Utils.rounded(statsSize.getMean(), 4), Utils.rounded(statsSize.getStandardDeviation(), 4),
    //			Utils.rounded(statsRadius.getMean(), 4), Utils.rounded(statsRadius.getStandardDeviation(), 4));
    log("Simulation results");
    log("  * Molecules = %d (%d activated)", xyz.size(), count);
    log("  * Blinking rate = %s", Utils.rounded((double) molecules.size() / xyz.size(), 4));
    log("  * Precision (Mean-displacement) = %s nm", (statsSigma.getN() > 0) ? Utils.rounded(Math.sqrt(statsSigma.getMean()), 4) : "0");
    if (showHistograms) {
        if (intraDistances.getN() == 0) {
            log("  * Mean Intra-Molecule particle linkage distance = 0 nm");
            log("  * Fraction of inter-molecule particle linkage @ 0 nm = 0 %%");
        } else {
            plot(blinks, "Blinks/Molecule", true);
            double[][] intraHist = plot(intraDistances, "Intra-molecule particle linkage distance", false);
            // Determine 95th and 99th percentile
            int p99 = intraHist[0].length - 1;
            double limit1 = 0.99 * intraHist[1][p99];
            double limit2 = 0.95 * intraHist[1][p99];
            while (intraHist[1][p99] > limit1 && p99 > 0) p99--;
            int p95 = p99;
            while (intraHist[1][p95] > limit2 && p95 > 0) p95--;
            log("  * Mean Intra-Molecule particle linkage distance = %s nm (95%% = %s, 99%% = %s, 100%% = %s)", Utils.rounded(intraDistances.getMean(), 4), Utils.rounded(intraHist[0][p95], 4), Utils.rounded(intraHist[0][p99], 4), Utils.rounded(intraHist[0][intraHist[0].length - 1], 4));
            if (distanceAnalysis) {
                performDistanceAnalysis(intraHist, p99);
            }
        }
    }
    if (clusterSimulation > 0) {
        log("  * Cluster number = %s +/- %s", Utils.rounded(statsSize.getMean(), 4), Utils.rounded(statsSize.getStandardDeviation(), 4));
        log("  * Cluster radius = %s +/- %s nm (mean distance to centre-of-mass)", Utils.rounded(statsRadius.getMean(), 4), Utils.rounded(statsRadius.getStandardDeviation(), 4));
    }
}
Also used : ByteProcessor(ij.process.ByteProcessor) TDoubleArrayList(gnu.trove.list.array.TDoubleArrayList) ArrayList(java.util.ArrayList) MaskDistribution(gdsc.smlm.model.MaskDistribution) Well19937c(org.apache.commons.math3.random.Well19937c) RandomGenerator(org.apache.commons.math3.random.RandomGenerator) MemoryPeakResults(gdsc.smlm.results.MemoryPeakResults) NullSource(gdsc.smlm.results.NullSource) RandomDataGenerator(org.apache.commons.math3.random.RandomDataGenerator) UniformDistribution(gdsc.smlm.model.UniformDistribution) StoredDataStatistics(gdsc.core.utils.StoredDataStatistics) Statistics(gdsc.core.utils.Statistics) StoredDataStatistics(gdsc.core.utils.StoredDataStatistics) DescriptiveStatistics(org.apache.commons.math3.stat.descriptive.DescriptiveStatistics) WeightedObservedPoint(org.apache.commons.math3.fitting.WeightedObservedPoint) ClusterPoint(gdsc.core.clustering.ClusterPoint) StoredData(gdsc.core.utils.StoredData)

Example 34 with Well19937c

use of org.apache.commons.math3.random.Well19937c in project GDSC-SMLM by aherbert.

the class PCPALMFitting method runBoundedOptimiser.

private PointValuePair runBoundedOptimiser(double[][] gr, double[] initialSolution, double[] lB, double[] uB, SumOfSquaresModelFunction function) {
    // Create the functions to optimise
    ObjectiveFunction objective = new ObjectiveFunction(new SumOfSquaresMultivariateFunction(function));
    ObjectiveFunctionGradient gradient = new ObjectiveFunctionGradient(new SumOfSquaresMultivariateVectorFunction(function));
    final boolean debug = false;
    // Try a BFGS optimiser since this will produce a deterministic solution and can respect bounds.
    PointValuePair optimum = null;
    boundedEvaluations = 0;
    final MaxEval maxEvaluations = new MaxEval(2000);
    MultivariateOptimizer opt = null;
    for (int iteration = 0; iteration <= fitRestarts; iteration++) {
        try {
            opt = new BFGSOptimizer();
            final double relativeThreshold = 1e-6;
            // Configure maximum step length for each dimension using the bounds
            double[] stepLength = new double[lB.length];
            for (int i = 0; i < stepLength.length; i++) stepLength[i] = (uB[i] - lB[i]) * 0.3333333;
            // The GoalType is always minimise so no need to pass this in
            optimum = opt.optimize(maxEvaluations, gradient, objective, new InitialGuess((optimum == null) ? initialSolution : optimum.getPointRef()), new SimpleBounds(lB, uB), new BFGSOptimizer.GradientTolerance(relativeThreshold), new BFGSOptimizer.StepLength(stepLength));
            if (debug)
                System.out.printf("BFGS Iter %d = %g (%d)\n", iteration, optimum.getValue(), opt.getEvaluations());
        } catch (TooManyEvaluationsException e) {
            // No need to restart
            break;
        } catch (RuntimeException e) {
            // No need to restart
            break;
        } finally {
            boundedEvaluations += opt.getEvaluations();
        }
    }
    // Try a CMAES optimiser which is non-deterministic. To overcome this we perform restarts.
    // CMAESOptimiser based on Matlab code:
    // https://www.lri.fr/~hansen/cmaes.m
    // Take the defaults from the Matlab documentation
    //Double.NEGATIVE_INFINITY;
    double stopFitness = 0;
    boolean isActiveCMA = true;
    int diagonalOnly = 0;
    int checkFeasableCount = 1;
    //Well19937c();
    RandomGenerator random = new Well44497b();
    boolean generateStatistics = false;
    ConvergenceChecker<PointValuePair> checker = new SimpleValueChecker(1e-6, 1e-10);
    // The sigma determines the search range for the variables. It should be 1/3 of the initial search region.
    double[] range = new double[lB.length];
    for (int i = 0; i < lB.length; i++) range[i] = (uB[i] - lB[i]) / 3;
    OptimizationData sigma = new CMAESOptimizer.Sigma(range);
    OptimizationData popSize = new CMAESOptimizer.PopulationSize((int) (4 + Math.floor(3 * Math.log(initialSolution.length))));
    SimpleBounds bounds = new SimpleBounds(lB, uB);
    opt = new CMAESOptimizer(maxEvaluations.getMaxEval(), stopFitness, isActiveCMA, diagonalOnly, checkFeasableCount, random, generateStatistics, checker);
    // Restart the optimiser several times and take the best answer.
    for (int iteration = 0; iteration <= fitRestarts; iteration++) {
        try {
            // Start from the initial solution
            PointValuePair constrainedSolution = opt.optimize(new InitialGuess(initialSolution), objective, GoalType.MINIMIZE, bounds, sigma, popSize, maxEvaluations);
            if (debug)
                System.out.printf("CMAES Iter %d initial = %g (%d)\n", iteration, constrainedSolution.getValue(), opt.getEvaluations());
            boundedEvaluations += opt.getEvaluations();
            if (optimum == null || constrainedSolution.getValue() < optimum.getValue()) {
                optimum = constrainedSolution;
            }
        } catch (TooManyEvaluationsException e) {
        } catch (TooManyIterationsException e) {
        } finally {
            boundedEvaluations += maxEvaluations.getMaxEval();
        }
        if (optimum == null)
            continue;
        try {
            // Also restart from the current optimum
            PointValuePair constrainedSolution = opt.optimize(new InitialGuess(optimum.getPointRef()), objective, GoalType.MINIMIZE, bounds, sigma, popSize, maxEvaluations);
            if (debug)
                System.out.printf("CMAES Iter %d restart = %g (%d)\n", iteration, constrainedSolution.getValue(), opt.getEvaluations());
            if (constrainedSolution.getValue() < optimum.getValue()) {
                optimum = constrainedSolution;
            }
        } catch (TooManyEvaluationsException e) {
        } catch (TooManyIterationsException e) {
        } finally {
            boundedEvaluations += maxEvaluations.getMaxEval();
        }
    }
    return optimum;
}
Also used : MultivariateOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.MultivariateOptimizer) MaxEval(org.apache.commons.math3.optim.MaxEval) InitialGuess(org.apache.commons.math3.optim.InitialGuess) SimpleBounds(org.apache.commons.math3.optim.SimpleBounds) ObjectiveFunction(org.apache.commons.math3.optim.nonlinear.scalar.ObjectiveFunction) SimpleValueChecker(org.apache.commons.math3.optim.SimpleValueChecker) BFGSOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.gradient.BFGSOptimizer) RandomGenerator(org.apache.commons.math3.random.RandomGenerator) PointValuePair(org.apache.commons.math3.optim.PointValuePair) TooManyEvaluationsException(org.apache.commons.math3.exception.TooManyEvaluationsException) TooManyIterationsException(org.apache.commons.math3.exception.TooManyIterationsException) CMAESOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.noderiv.CMAESOptimizer) ObjectiveFunctionGradient(org.apache.commons.math3.optim.nonlinear.scalar.ObjectiveFunctionGradient) Well44497b(org.apache.commons.math3.random.Well44497b) OptimizationData(org.apache.commons.math3.optim.OptimizationData)

Example 35 with Well19937c

use of org.apache.commons.math3.random.Well19937c in project GDSC-SMLM by aherbert.

the class FastMLEGradient2ProcedureTest method gradientProcedureLinearIsFasterThanGradientProcedure.

private void gradientProcedureLinearIsFasterThanGradientProcedure(final int nparams) {
    org.junit.Assume.assumeTrue(speedTests || TestSettings.RUN_SPEED_TESTS);
    final int iter = 100;
    rdg = new RandomDataGenerator(new Well19937c(30051977));
    final ArrayList<double[]> paramsList = new ArrayList<double[]>(iter);
    final ArrayList<double[]> yList = new ArrayList<double[]>(iter);
    createData(1, iter, paramsList, yList);
    // Remove the timing of the function call by creating a dummy function
    final Gradient2Function func = new FakeGradientFunction(blockWidth, nparams);
    for (int i = 0; i < paramsList.size(); i++) {
        FastMLEGradient2Procedure p1 = new FastMLEGradient2Procedure(yList.get(i), func);
        p1.computeSecondDerivative(paramsList.get(i));
        p1.computeSecondDerivative(paramsList.get(i));
        FastMLEGradient2Procedure p2 = FastMLEGradient2ProcedureFactory.createUnrolled(yList.get(i), func);
        p2.computeSecondDerivative(paramsList.get(i));
        p2.computeSecondDerivative(paramsList.get(i));
        // Check they are the same
        Assert.assertArrayEquals("D1 " + i, p1.d1, p2.d1, 0);
        Assert.assertArrayEquals("D2 " + i, p1.d2, p2.d2, 0);
    }
    // Realistic loops for an optimisation
    final int loops = 15;
    // Run till stable timing
    Timer t1 = new Timer() {

        @Override
        void run() {
            for (int i = 0, k = 0; i < paramsList.size(); i++) {
                FastMLEGradient2Procedure p1 = new FastMLEGradient2Procedure(yList.get(i), func);
                for (int j = loops; j-- > 0; ) p1.computeSecondDerivative(paramsList.get(k++ % iter));
            }
        }
    };
    long time1 = t1.getTime();
    Timer t2 = new Timer(t1.loops) {

        @Override
        void run() {
            for (int i = 0, k = 0; i < paramsList.size(); i++) {
                FastMLEGradient2Procedure p2 = FastMLEGradient2ProcedureFactory.createUnrolled(yList.get(i), func);
                for (int j = loops; j-- > 0; ) p2.computeSecondDerivative(paramsList.get(k++ % iter));
            }
        }
    };
    long time2 = t2.getTime();
    log("Standard = %d : Unrolled %d = %d : %fx\n", time1, nparams, time2, (1.0 * time1) / time2);
    Assert.assertTrue(time2 < time1 * 1.5);
}
Also used : RandomDataGenerator(org.apache.commons.math3.random.RandomDataGenerator) Gradient2Function(gdsc.smlm.function.Gradient2Function) PrecomputedGradient2Function(gdsc.smlm.function.PrecomputedGradient2Function) ArrayList(java.util.ArrayList) Well19937c(org.apache.commons.math3.random.Well19937c) FakeGradientFunction(gdsc.smlm.function.FakeGradientFunction)

Aggregations

Well19937c (org.apache.commons.math3.random.Well19937c)66 RandomDataGenerator (org.apache.commons.math3.random.RandomDataGenerator)41 ArrayList (java.util.ArrayList)31 RandomGenerator (org.apache.commons.math3.random.RandomGenerator)26 FakeGradientFunction (gdsc.smlm.function.FakeGradientFunction)17 Test (org.junit.Test)17 DoubleEquality (gdsc.core.utils.DoubleEquality)7 Gaussian2DFunction (gdsc.smlm.function.gaussian.Gaussian2DFunction)7 DenseMatrix64F (org.ejml.data.DenseMatrix64F)6 StoredDataStatistics (gdsc.core.utils.StoredDataStatistics)5 PointValuePair (org.apache.commons.math3.optim.PointValuePair)5 TimingService (gdsc.core.test.TimingService)4 Statistics (gdsc.core.utils.Statistics)4 Gradient1Function (gdsc.smlm.function.Gradient1Function)4 ValueProcedure (gdsc.smlm.function.ValueProcedure)4 ErfGaussian2DFunction (gdsc.smlm.function.gaussian.erf.ErfGaussian2DFunction)4 TooManyEvaluationsException (org.apache.commons.math3.exception.TooManyEvaluationsException)4 SimpleValueChecker (org.apache.commons.math3.optim.SimpleValueChecker)4 PseudoRandomGenerator (gdsc.core.utils.PseudoRandomGenerator)3 TurboList (gdsc.core.utils.TurboList)3