Java Code Examples for org.nd4j.linalg.indexing.BooleanIndexing#and()

@Test
public void testBinomialDistribution1() throws Exception {
    Random random1 = Nd4j.getRandomFactory().getNewRandomInstance(119);
    Random random2 = Nd4j.getRandomFactory().getNewRandomInstance(119);

    INDArray z1 = Nd4j.zeros(1000);
    INDArray z2 = Nd4j.zeros(1000);
    INDArray z1Dup = Nd4j.zeros(1000);

    BinomialDistribution op1 = new BinomialDistribution(z1, 5, 0.25);
    BinomialDistribution op2 = new BinomialDistribution(z2, 5, 0.25);

    Nd4j.getExecutioner().exec(op1, random1);
    Nd4j.getExecutioner().exec(op2, random2);

    assertNotEquals(z1Dup, z1);

    assertEquals(z1, z2);

    BooleanIndexing.and(z1, Conditions.lessThanOrEqual(5.0));
    BooleanIndexing.and(z1, Conditions.greaterThanOrEqual(0.0));
}
 

@Test
public void testBinomialDistribution2() throws Exception {
    Random random1 = Nd4j.getRandomFactory().getNewRandomInstance(119);
    Random random2 = Nd4j.getRandomFactory().getNewRandomInstance(119);

    INDArray z1 = Nd4j.zeros(1000);
    INDArray z2 = Nd4j.zeros(1000);
    INDArray z1Dup = Nd4j.zeros(1000);

    INDArray probs = Nd4j.create(new float[] {0.25f, 0.43f, 0.55f, 0.43f, 0.25f});

    BinomialDistribution op1 = new BinomialDistribution(z1, 5, probs);
    BinomialDistribution op2 = new BinomialDistribution(z2, 5, probs);

    Nd4j.getExecutioner().exec(op1, random1);
    Nd4j.getExecutioner().exec(op2, random2);

    assertNotEquals(z1Dup, z1);

    assertEquals(z1, z2);

    BooleanIndexing.and(z1, Conditions.lessThanOrEqual(5.0));
    BooleanIndexing.and(z1, Conditions.greaterThanOrEqual(0.0));
}
 

@Test
public void testBinomialDistribution1() {
    Random random1 = Nd4j.getRandomFactory().getNewRandomInstance(119);
    Random random2 = Nd4j.getRandomFactory().getNewRandomInstance(119);

    INDArray z1 = Nd4j.zeros(1000);
    INDArray z2 = Nd4j.zeros(1000);
    INDArray z1Dup = Nd4j.zeros(1000);

    BinomialDistribution op1 = new BinomialDistribution(z1, 5, 0.25);
    BinomialDistribution op2 = new BinomialDistribution(z2, 5, 0.25);

    Nd4j.getExecutioner().exec(op1, random1);
    Nd4j.getExecutioner().exec(op2, random2);

    assertNotEquals(z1Dup, z1);

    assertEquals(z1, z2);

    BooleanIndexing.and(z1, Conditions.lessThanOrEqual(5.0));
    BooleanIndexing.and(z1, Conditions.greaterThanOrEqual(0.0));
}
 

@Test
public void testBinomialDistribution2() {
    Random random1 = Nd4j.getRandomFactory().getNewRandomInstance(119);
    Random random2 = Nd4j.getRandomFactory().getNewRandomInstance(119);

    INDArray z1 = Nd4j.create(DataType.FLOAT, 1000);
    INDArray z2 = Nd4j.zeros(DataType.FLOAT,1000);
    INDArray z1Dup = Nd4j.zeros(DataType.FLOAT,1000);

    INDArray probs = Nd4j.create(new float[] {0.25f, 0.43f, 0.55f, 0.43f, 0.25f});

    BinomialDistribution op1 = new BinomialDistribution(z1, 5, probs);
    BinomialDistribution op2 = new BinomialDistribution(z2, 5, probs);

    Nd4j.getExecutioner().exec(op1, random1);
    Nd4j.getExecutioner().exec(op2, random2);

    assertNotEquals(z1Dup, z1);

    assertEquals(z1, z2);

    BooleanIndexing.and(z1, Conditions.lessThanOrEqual(5.0));
    BooleanIndexing.and(z1, Conditions.greaterThanOrEqual(0.0));
}
 

/**
 * Optimizes the given loss function with respect to a given PLNet's inputs using gradient descent. Ensures the outcome will be within the range of 0 and 1.
 * Performs gradient descent for a given number of steps starting at a given input, using a linearly decaying learning rate.
 * The inputs that should be optimized can be specified using a 0,1-vector
 * @param plNet					PLNet whose inputs to optimize.
 * @param input					Initial inputs to start the gradient descent procedure from.
 * @param loss					The loss to be minimized.
 * @param initialLearningRate	The initial learning rate.
 * @param finalLearningRate		The value the learning rate should decay to.
 * @param numSteps				The number of steps to perform gradient descent for.
 * @param inputMask				0,1 vector specifying the inputs to optimize, i.e. should have a 1 at the index of any input that should be optimized and a 0 elsewhere.
 * @return						The input optimized with respect to the given loss.
 */
public INDArray optimizeInput(PLNetDyadRanker plNet, INDArray input, InputOptimizerLoss loss, double initialLearningRate, double finalLearningRate, int numSteps,
		INDArray inputMask) {
	INDArray inp = input.dup();
	INDArray alphas = Nd4j.zeros(inp.shape());
	INDArray betas = Nd4j.zeros(inp.shape());
	INDArray ones = Nd4j.ones(inp.shape());
	double output = plNet.getPlNet().output(inp).getDouble(0);
	double incumbentOutput = output;
	INDArray incumbent = inp.dup();
	for (int i = 0; i < numSteps; i++) {
		double lrDecayTerm = (double) i / (double) numSteps;
		double learningRate = (1 - lrDecayTerm) * initialLearningRate + lrDecayTerm * finalLearningRate;
		// Gradient of PLNet
		INDArray grad = computeInputDerivative(plNet, inp, loss);
		// Gradient of KKT term
		grad.subi(alphas);
		grad.addi(betas);
		// Apply gradient to alphas and betas
		alphas.subi(inp);
		betas.addi(inp.sub(ones));
		BooleanIndexing.replaceWhere(alphas, 0.0d, Conditions.lessThan(0.0d));
		BooleanIndexing.replaceWhere(betas, 0.0d, Conditions.lessThan(0.0d));
		grad.muli(inputMask);
		grad.muli(learningRate);
		inp.subi(grad);

		output = plNet.getPlNet().output(inp).getDouble(0);
		if (listener != null) {
			listener.reportOptimizationStep(inp, output);
		}

		INDArray incCheck = inp.dup().muli(inputMask);
		if (output > incumbentOutput && BooleanIndexing.and(incCheck, Conditions.greaterThanOrEqual(0.0d)) && BooleanIndexing.and(incCheck, Conditions.lessThanOrEqual(1.0d))) {
			incumbent = inp.dup();
			incumbentOutput = output;
		}
	}

	return incumbent;
}