Examples with RnnOutputLayer org.deeplearning4j.nn.conf.layers.RnnOutputLayer used on opensource projects

Example 1 with RnnOutputLayer

use of org.deeplearning4j.nn.conf.layers.RnnOutputLayer in project deeplearning4j by deeplearning4j.

the class TestVariableLengthTSCG method testVariableLengthSimple.

@Test
public void testVariableLengthSimple() {
    //Test: Simple RNN layer + RNNOutputLayer
    //Length of 4 for standard
    //Length of 5 with last time step output mask set to 0
    //Expect the same gradients etc in both cases...
    int[] miniBatchSizes = { 1, 2, 5 };
    int nOut = 1;
    Random r = new Random(12345);
    for (int nExamples : miniBatchSizes) {
        Nd4j.getRandom().setSeed(12345);
        ComputationGraphConfiguration conf = new NeuralNetConfiguration.Builder().optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT).iterations(1).updater(Updater.SGD).learningRate(0.1).seed(12345).graphBuilder().addInputs("in").addLayer("0", new GravesLSTM.Builder().activation(Activation.TANH).nIn(2).nOut(2).build(), "in").addLayer("1", new RnnOutputLayer.Builder().lossFunction(LossFunctions.LossFunction.MSE).nIn(2).nOut(1).build(), "0").setOutputs("1").build();
        ComputationGraph net = new ComputationGraph(conf);
        net.init();
        INDArray in1 = Nd4j.rand(new int[] { nExamples, 2, 4 });
        INDArray in2 = Nd4j.rand(new int[] { nExamples, 2, 5 });
        in2.put(new INDArrayIndex[] { NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.interval(0, 3, true) }, in1);
        assertEquals(in1, in2.get(NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.interval(0, 4)));
        INDArray labels1 = Nd4j.rand(new int[] { nExamples, 1, 4 });
        INDArray labels2 = Nd4j.create(nExamples, 1, 5);
        labels2.put(new INDArrayIndex[] { NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.interval(0, 3, true) }, labels1);
        assertEquals(labels1, labels2.get(NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.interval(0, 4)));
        INDArray labelMask = Nd4j.ones(nExamples, 5);
        for (int j = 0; j < nExamples; j++) {
            labelMask.putScalar(new int[] { j, 4 }, 0);
        }
        net.setInput(0, in1);
        net.setLabel(0, labels1);
        net.computeGradientAndScore();
        double score1 = net.score();
        Gradient g1 = net.gradient();
        net.setInput(0, in2);
        net.setLabel(0, labels2);
        net.setLayerMaskArrays(null, new INDArray[] { labelMask });
        net.computeGradientAndScore();
        double score2 = net.score();
        Gradient g2 = net.gradient();
        //Scores and gradients should be identical for two cases (given mask array)
        assertEquals(score1, score2, 0.0);
        Map<String, INDArray> g1map = g1.gradientForVariable();
        Map<String, INDArray> g2map = g2.gradientForVariable();
        for (String s : g1map.keySet()) {
            INDArray g1s = g1map.get(s);
            INDArray g2s = g2map.get(s);
            assertEquals(s, g1s, g2s);
        }
        // (a) score, (b) gradients
        for (int i = 0; i < nExamples; i++) {
            for (int j = 0; j < nOut; j++) {
                double d = r.nextDouble();
                labels2.putScalar(new int[] { i, j, 4 }, d);
            }
            net.setLabel(0, labels2);
            net.computeGradientAndScore();
            double score2a = net.score();
            Gradient g2a = net.gradient();
            assertEquals(score2, score2a, 0.0);
            for (String s : g2map.keySet()) {
                INDArray g2s = g2map.get(s);
                INDArray g2sa = g2a.getGradientFor(s);
                assertEquals(s, g2s, g2sa);
            }
        }
    }
}

Example 2 with RnnOutputLayer

use of org.deeplearning4j.nn.conf.layers.RnnOutputLayer in project deeplearning4j by deeplearning4j.

the class ComputationGraphTestRNN method testRnnTimeStepMultipleInOut.

@Test
public void testRnnTimeStepMultipleInOut() {
    //Test rnnTimeStep functionality with multiple inputs and outputs...
    Nd4j.getRandom().setSeed(12345);
    int timeSeriesLength = 12;
    //4 layer network: 2 GravesLSTM + DenseLayer + RnnOutputLayer. Hence also tests preprocessors.
    //Network architecture: lstm0 -> Dense -> RnnOutputLayer0
    // and lstm1 -> Dense -> RnnOutputLayer1
    ComputationGraphConfiguration conf = new NeuralNetConfiguration.Builder().seed(12345).graphBuilder().addInputs("in0", "in1").addLayer("lstm0", new org.deeplearning4j.nn.conf.layers.GravesLSTM.Builder().nIn(5).nOut(6).activation(Activation.TANH).weightInit(WeightInit.DISTRIBUTION).dist(new NormalDistribution(0, 0.5)).build(), "in0").addLayer("lstm1", new org.deeplearning4j.nn.conf.layers.GravesLSTM.Builder().nIn(4).nOut(5).activation(Activation.TANH).weightInit(WeightInit.DISTRIBUTION).dist(new NormalDistribution(0, 0.5)).build(), "in1").addLayer("dense", new DenseLayer.Builder().nIn(6 + 5).nOut(9).activation(Activation.TANH).weightInit(WeightInit.DISTRIBUTION).dist(new NormalDistribution(0, 0.5)).build(), "lstm0", "lstm1").addLayer("out0", new RnnOutputLayer.Builder(LossFunctions.LossFunction.MCXENT).weightInit(WeightInit.DISTRIBUTION).nIn(9).nOut(3).activation(Activation.SOFTMAX).weightInit(WeightInit.DISTRIBUTION).dist(new NormalDistribution(0, 0.5)).build(), "dense").addLayer("out1", new RnnOutputLayer.Builder(LossFunctions.LossFunction.MCXENT).weightInit(WeightInit.DISTRIBUTION).nIn(9).nOut(4).activation(Activation.SOFTMAX).weightInit(WeightInit.DISTRIBUTION).dist(new NormalDistribution(0, 0.5)).build(), "dense").setOutputs("out0", "out1").inputPreProcessor("dense", new RnnToFeedForwardPreProcessor()).inputPreProcessor("out0", new FeedForwardToRnnPreProcessor()).inputPreProcessor("out1", new FeedForwardToRnnPreProcessor()).pretrain(false).backprop(true).build();
    ComputationGraph graph = new ComputationGraph(conf);
    graph.init();
    INDArray input0 = Nd4j.rand(new int[] { 3, 5, timeSeriesLength });
    INDArray input1 = Nd4j.rand(new int[] { 3, 4, timeSeriesLength });
    Map<String, INDArray> allOutputActivations = graph.feedForward(new INDArray[] { input0, input1 }, true);
    INDArray fullActLSTM0 = allOutputActivations.get("lstm0");
    INDArray fullActLSTM1 = allOutputActivations.get("lstm1");
    INDArray fullActOut0 = allOutputActivations.get("out0");
    INDArray fullActOut1 = allOutputActivations.get("out1");
    assertArrayEquals(new int[] { 3, 6, timeSeriesLength }, fullActLSTM0.shape());
    assertArrayEquals(new int[] { 3, 5, timeSeriesLength }, fullActLSTM1.shape());
    assertArrayEquals(new int[] { 3, 3, timeSeriesLength }, fullActOut0.shape());
    assertArrayEquals(new int[] { 3, 4, timeSeriesLength }, fullActOut1.shape());
    int[] inputLengths = { 1, 2, 3, 4, 6, 12 };
    //Should get the same result regardless of step size; should be identical to standard forward pass
    for (int i = 0; i < inputLengths.length; i++) {
        int inLength = inputLengths[i];
        //each of length inLength
        int nSteps = timeSeriesLength / inLength;
        graph.rnnClearPreviousState();
        for (int j = 0; j < nSteps; j++) {
            int startTimeRange = j * inLength;
            int endTimeRange = startTimeRange + inLength;
            INDArray inputSubset0 = input0.get(NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.interval(startTimeRange, endTimeRange));
            if (inLength > 1)
                assertTrue(inputSubset0.size(2) == inLength);
            INDArray inputSubset1 = input1.get(NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.interval(startTimeRange, endTimeRange));
            if (inLength > 1)
                assertTrue(inputSubset1.size(2) == inLength);
            INDArray[] outArr = graph.rnnTimeStep(inputSubset0, inputSubset1);
            assertEquals(2, outArr.length);
            INDArray out0 = outArr[0];
            INDArray out1 = outArr[1];
            INDArray expOutSubset0;
            if (inLength == 1) {
                int[] sizes = new int[] { fullActOut0.size(0), fullActOut0.size(1), 1 };
                expOutSubset0 = Nd4j.create(sizes);
                expOutSubset0.tensorAlongDimension(0, 1, 0).assign(fullActOut0.get(NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.point(startTimeRange)));
            } else {
                expOutSubset0 = fullActOut0.get(NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.interval(startTimeRange, endTimeRange));
            }
            INDArray expOutSubset1;
            if (inLength == 1) {
                int[] sizes = new int[] { fullActOut1.size(0), fullActOut1.size(1), 1 };
                expOutSubset1 = Nd4j.create(sizes);
                expOutSubset1.tensorAlongDimension(0, 1, 0).assign(fullActOut1.get(NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.point(startTimeRange)));
            } else {
                expOutSubset1 = fullActOut1.get(NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.interval(startTimeRange, endTimeRange));
            }
            assertEquals(expOutSubset0, out0);
            assertEquals(expOutSubset1, out1);
            Map<String, INDArray> currLSTM0State = graph.rnnGetPreviousState("lstm0");
            Map<String, INDArray> currLSTM1State = graph.rnnGetPreviousState("lstm1");
            INDArray lastActL0 = currLSTM0State.get(GravesLSTM.STATE_KEY_PREV_ACTIVATION);
            INDArray lastActL1 = currLSTM1State.get(GravesLSTM.STATE_KEY_PREV_ACTIVATION);
            INDArray expLastActL0 = fullActLSTM0.tensorAlongDimension(endTimeRange - 1, 1, 0);
            INDArray expLastActL1 = fullActLSTM1.tensorAlongDimension(endTimeRange - 1, 1, 0);
            assertEquals(expLastActL0, lastActL0);
            assertEquals(expLastActL1, lastActL1);
        }
    }
}

Example 3 with RnnOutputLayer

use of org.deeplearning4j.nn.conf.layers.RnnOutputLayer in project deeplearning4j by deeplearning4j.

the class MultiLayerTestRNN method testRnnTimeStepGravesLSTM.

@Test
public void testRnnTimeStepGravesLSTM() {
    Nd4j.getRandom().setSeed(12345);
    int timeSeriesLength = 12;
    //4 layer network: 2 GravesLSTM + DenseLayer + RnnOutputLayer. Hence also tests preprocessors.
    MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().seed(12345).list().layer(0, new org.deeplearning4j.nn.conf.layers.GravesLSTM.Builder().nIn(5).nOut(7).activation(Activation.TANH).weightInit(WeightInit.DISTRIBUTION).dist(new NormalDistribution(0, 0.5)).build()).layer(1, new org.deeplearning4j.nn.conf.layers.GravesLSTM.Builder().nIn(7).nOut(8).activation(Activation.TANH).weightInit(WeightInit.DISTRIBUTION).dist(new NormalDistribution(0, 0.5)).build()).layer(2, new DenseLayer.Builder().nIn(8).nOut(9).activation(Activation.TANH).weightInit(WeightInit.DISTRIBUTION).dist(new NormalDistribution(0, 0.5)).build()).layer(3, new RnnOutputLayer.Builder(LossFunction.MCXENT).weightInit(WeightInit.DISTRIBUTION).nIn(9).nOut(4).activation(Activation.SOFTMAX).weightInit(WeightInit.DISTRIBUTION).dist(new NormalDistribution(0, 0.5)).build()).inputPreProcessor(2, new RnnToFeedForwardPreProcessor()).inputPreProcessor(3, new FeedForwardToRnnPreProcessor()).build();
    MultiLayerNetwork mln = new MultiLayerNetwork(conf);
    INDArray input = Nd4j.rand(new int[] { 3, 5, timeSeriesLength });
    List<INDArray> allOutputActivations = mln.feedForward(input, true);
    INDArray fullOutL0 = allOutputActivations.get(1);
    INDArray fullOutL1 = allOutputActivations.get(2);
    INDArray fullOutL3 = allOutputActivations.get(4);
    int[] inputLengths = { 1, 2, 3, 4, 6, 12 };
    //Should get the same result regardless of step size; should be identical to standard forward pass
    for (int i = 0; i < inputLengths.length; i++) {
        int inLength = inputLengths[i];
        //each of length inLength
        int nSteps = timeSeriesLength / inLength;
        mln.rnnClearPreviousState();
        //Reset; should be set by rnnTimeStep method
        mln.setInputMiniBatchSize(1);
        for (int j = 0; j < nSteps; j++) {
            int startTimeRange = j * inLength;
            int endTimeRange = startTimeRange + inLength;
            INDArray inputSubset;
            if (inLength == 1) {
                //Workaround to nd4j bug
                int[] sizes = new int[] { input.size(0), input.size(1), 1 };
                inputSubset = Nd4j.create(sizes);
                inputSubset.tensorAlongDimension(0, 1, 0).assign(input.get(NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.point(startTimeRange)));
            } else {
                inputSubset = input.get(NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.interval(startTimeRange, endTimeRange));
            }
            if (inLength > 1)
                assertTrue(inputSubset.size(2) == inLength);
            INDArray out = mln.rnnTimeStep(inputSubset);
            INDArray expOutSubset;
            if (inLength == 1) {
                int[] sizes = new int[] { fullOutL3.size(0), fullOutL3.size(1), 1 };
                expOutSubset = Nd4j.create(sizes);
                expOutSubset.tensorAlongDimension(0, 1, 0).assign(fullOutL3.get(NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.point(startTimeRange)));
            } else {
                expOutSubset = fullOutL3.get(NDArrayIndex.all(), NDArrayIndex.all(), NDArrayIndex.interval(startTimeRange, endTimeRange));
            }
            assertEquals(expOutSubset, out);
            Map<String, INDArray> currL0State = mln.rnnGetPreviousState(0);
            Map<String, INDArray> currL1State = mln.rnnGetPreviousState(1);
            INDArray lastActL0 = currL0State.get(GravesLSTM.STATE_KEY_PREV_ACTIVATION);
            INDArray lastActL1 = currL1State.get(GravesLSTM.STATE_KEY_PREV_ACTIVATION);
            INDArray expLastActL0 = fullOutL0.tensorAlongDimension(endTimeRange - 1, 1, 0);
            INDArray expLastActL1 = fullOutL1.tensorAlongDimension(endTimeRange - 1, 1, 0);
            assertEquals(expLastActL0, lastActL0);
            assertEquals(expLastActL1, lastActL1);
        }
    }
}