org.nd4j.linalg.api.rng.DefaultRandom Java Examples

@Test
public void testLegacyDistribution1() throws Exception {
    NormalDistribution distribution = new NormalDistribution(new DefaultRandom(), 0.0, 1.0);
    INDArray z1 = distribution.sample(new int[] {1, 30000000});

    assertEquals(0.0, z1.meanNumber().doubleValue(), 0.01);
    assertEquals(1.0, z1.stdNumber().doubleValue(), 0.01);
}
 

@Test
public void testLegacyDistribution1() {
    NormalDistribution distribution = new NormalDistribution(new DefaultRandom(), 0.0, 1.0);
    INDArray z1 = distribution.sample(new int[] {1, 1000000});

    assertEquals(0.0, z1.meanNumber().doubleValue(), 0.01);
    assertEquals(1.0, z1.stdNumber().doubleValue(), 0.01);
}
 

private static void testSphereFnMultipleStepsHelper(OptimizationAlgorithm oa, int nOptIter,
                int maxNumLineSearchIter) {
    double[] scores = new double[nOptIter + 1];

    for (int i = 0; i <= nOptIter; i++) {
        Random rng = new DefaultRandom(12345L);
        org.nd4j.linalg.api.rng.distribution.Distribution dist =
                        new org.nd4j.linalg.api.rng.distribution.impl.UniformDistribution(rng, -10, 10);
        NeuralNetConfiguration conf = new NeuralNetConfiguration.Builder()
                        .maxNumLineSearchIterations(maxNumLineSearchIter).updater(new Sgd(0.1))
                        .layer(new DenseLayer.Builder().nIn(1).nOut(1).build()).build();
        conf.addVariable("W"); //Normally done by ParamInitializers, but obviously that isn't done here

        Model m = new SphereFunctionModel(100, dist, conf);
        if (i == 0) {
            m.computeGradientAndScore(LayerWorkspaceMgr.noWorkspaces());
            scores[0] = m.score(); //Before optimization
        } else {
            ConvexOptimizer opt = getOptimizer(oa, conf, m);
            for( int j=0; j<100; j++ ) {
                opt.optimize(LayerWorkspaceMgr.noWorkspaces());
            }
            m.computeGradientAndScore(LayerWorkspaceMgr.noWorkspaces());
            scores[i] = m.score();
            assertTrue(!Double.isNaN(scores[i]) && !Double.isInfinite(scores[i]));
        }
    }

    if (PRINT_OPT_RESULTS) {
        System.out.println("Multiple optimization iterations (" + nOptIter
                        + " opt. iter.) score vs iteration, maxNumLineSearchIter=" + maxNumLineSearchIter + ": "
                        + oa);
        System.out.println(Arrays.toString(scores));
    }

    for (int i = 1; i < scores.length; i++) {
        assertTrue(scores[i] <= scores[i - 1]);
    }
    assertTrue(scores[scores.length - 1] < 1.0); //Very easy function, expect score ~= 0 with any reasonable number of steps/numLineSearchIter
}
 

@Override
protected Distribution getOldFunction(int seed) {
    Random random = new DefaultRandom(seed);
    Distribution distribution = new org.nd4j.linalg.api.rng.distribution.impl.UniformDistribution(random, 0.4D, 4D);
    return distribution;
}
 

@Override
protected Distribution getOldFunction(int seed) {
    Random random = new DefaultRandom(seed);
    Distribution distribution = new org.nd4j.linalg.api.rng.distribution.impl.BinomialDistribution(random, 10, 0.5D);
    return distribution;
}
 

@Override
protected Distribution getOldFunction(int seed) {
    Random random = new DefaultRandom(seed);
    Distribution distribution = new org.nd4j.linalg.api.rng.distribution.impl.NormalDistribution(random, 1D, 5D);
    return distribution;
}
 

public void testSphereFnOptHelper(OptimizationAlgorithm oa, int numLineSearchIter, int nDimensions) {

        if (PRINT_OPT_RESULTS)
            System.out.println("---------\n Alg= " + oa + ", nIter= " + numLineSearchIter + ", nDimensions= "
                            + nDimensions);

        NeuralNetConfiguration conf = new NeuralNetConfiguration.Builder().maxNumLineSearchIterations(numLineSearchIter)
                        .updater(new Sgd(1e-2))
                        .layer(new DenseLayer.Builder().nIn(1).nOut(1).build()).build();
        conf.addVariable("W"); //Normally done by ParamInitializers, but obviously that isn't done here

        Random rng = new DefaultRandom(12345L);
        org.nd4j.linalg.api.rng.distribution.Distribution dist =
                        new org.nd4j.linalg.api.rng.distribution.impl.UniformDistribution(rng, -10, 10);
        Model m = new SphereFunctionModel(nDimensions, dist, conf);
        m.computeGradientAndScore(LayerWorkspaceMgr.noWorkspaces());
        double scoreBefore = m.score();
        assertTrue(!Double.isNaN(scoreBefore) && !Double.isInfinite(scoreBefore));
        if (PRINT_OPT_RESULTS) {
            System.out.println("Before:");
            System.out.println(scoreBefore);
            System.out.println(m.params());
        }

        ConvexOptimizer opt = getOptimizer(oa, conf, m);

        opt.setupSearchState(m.gradientAndScore());
        for( int i=0; i<100; i++ ) {
            opt.optimize(LayerWorkspaceMgr.noWorkspaces());
        }
        m.computeGradientAndScore(LayerWorkspaceMgr.noWorkspaces());
        double scoreAfter = m.score();

        assertTrue(!Double.isNaN(scoreAfter) && !Double.isInfinite(scoreAfter));
        if (PRINT_OPT_RESULTS) {
            System.out.println("After:");
            System.out.println(scoreAfter);
            System.out.println(m.params());
        }

        //Expected behaviour after optimization:
        //(a) score is better (lower) after optimization.
        //(b) Parameters are closer to minimum after optimization (TODO)
        assertTrue("Score did not improve after optimization (b= " + scoreBefore + " ,a= " + scoreAfter + ")",
                        scoreAfter < scoreBefore);
    }
 

private static INDArray initParams(int nDimensions) {
    Random rng = new DefaultRandom(12345L);
    org.nd4j.linalg.api.rng.distribution.Distribution dist =
                    new org.nd4j.linalg.api.rng.distribution.impl.UniformDistribution(rng, -5.12, 5.12);
    return dist.sample(new int[] {1, nDimensions});
}
 

private static INDArray initParams(int nDimensions) {
    Random rng = new DefaultRandom(12345L);
    org.nd4j.linalg.api.rng.distribution.Distribution dist =
                    new org.nd4j.linalg.api.rng.distribution.impl.UniformDistribution(rng, -4.0, 4.0);
    return dist.sample(new int[] {1, nDimensions});
}