Java Code Examples for org.deeplearning4j.nn.graph.ComputationGraph#score()

@Test
public void testLambdaConf() {
    double[] lambdas = new double[] {0.1, 0.01};
    double[] results = new double[2];
    int numClasses = 2;

    INDArray input = Nd4j.rand(150, 4);
    INDArray labels = Nd4j.zeros(150, numClasses);
    Random r = new Random(12345);
    for (int i = 0; i < 150; i++) {
        labels.putScalar(i, r.nextInt(numClasses), 1.0);
    }
    ComputationGraph graph;

    for (int i = 0; i < lambdas.length; i++) {
        graph = getGraph(numClasses, lambdas[i]);
        graph.setInput(0, input);
        graph.setLabel(0, labels);
        graph.computeGradientAndScore();
        results[i] = graph.score();
    }

    assertNotEquals(results[0], results[1]);
}
 

/**
 * Score based on the loss function
 * @param model the model to score with
 * @param testData the test data to score
 * @param average whether to average the score
 *                for the whole batch or not
 * @return the score for the given test set
 */
public static double score(ComputationGraph model, MultiDataSetIterator testData, boolean average) {
    //TODO: do this properly taking into account division by N, L1/L2 etc
    double sumScore = 0.0;
    int totalExamples = 0;
    while (testData.hasNext()) {
        MultiDataSet ds = testData.next();
        long numExamples = ds.getFeatures(0).size(0);
        sumScore += numExamples * model.score(ds);
        totalExamples += numExamples;
    }

    if (!average)
        return sumScore;
    return sumScore / totalExamples;
}
 

/**
 * Score based on the loss function
 * @param model the model to score with
 * @param testData the test data to score
 * @param average whether to average the score
 *                for the whole batch or not
 * @return the score for the given test set
 */
public static double score(ComputationGraph model, DataSetIterator testData, boolean average) {
    //TODO: do this properly taking into account division by N, L1/L2 etc
    double sumScore = 0.0;
    int totalExamples = 0;
    while (testData.hasNext()) {
        DataSet ds = testData.next();
        int numExamples = ds.numExamples();

        sumScore += numExamples * model.score(ds);
        totalExamples += numExamples;
    }

    if (!average)
        return sumScore;
    return sumScore / totalExamples;
}
 

@Override
public ParameterAveragingTrainingResult getFinalResult(ComputationGraph network) {
    INDArray updaterState = null;
    if (saveUpdater) {
        ComputationGraphUpdater u = network.getUpdater();
        if (u != null)
            updaterState = u.getStateViewArray();
    }

    Nd4j.getExecutioner().commit();

    Collection<StorageMetaData> storageMetaData = null;
    Collection<Persistable> listenerStaticInfo = null;
    Collection<Persistable> listenerUpdates = null;
    if (listenerRouterProvider != null) {
        StatsStorageRouter r = listenerRouterProvider.getRouter();
        if (r instanceof VanillaStatsStorageRouter) { //TODO this is ugly... need to find a better solution
            VanillaStatsStorageRouter ssr = (VanillaStatsStorageRouter) r;
            storageMetaData = ssr.getStorageMetaData();
            listenerStaticInfo = ssr.getStaticInfo();
            listenerUpdates = ssr.getUpdates();
        }
    }

    return new ParameterAveragingTrainingResult(network.params(), updaterState, network.score(), storageMetaData,
                    listenerStaticInfo, listenerUpdates);
}
 

@Override
public Iterator<Tuple2<Long, Double>> call(Iterator<MultiDataSet> dataSetIterator) throws Exception {
    if (!dataSetIterator.hasNext()) {
        return Collections.singletonList(new Tuple2<>(0L, 0.0)).iterator();
    }

    MultiDataSetIterator iter = new IteratorMultiDataSetIterator(dataSetIterator, minibatchSize); //Does batching where appropriate


    ComputationGraph network = new ComputationGraph(ComputationGraphConfiguration.fromJson(json));
    network.init();
    INDArray val = params.value().unsafeDuplication(); //.value() is shared by all executors on single machine -> OK, as params are not changed in score function
    if (val.length() != network.numParams(false))
        throw new IllegalStateException(
                        "Network did not have same number of parameters as the broadcast set parameters");
    network.setParams(val);

    List<Tuple2<Long, Double>> out = new ArrayList<>();
    while (iter.hasNext()) {
        MultiDataSet ds = iter.next();
        double score = network.score(ds, false);

        long numExamples = ds.getFeatures(0).size(0);
        out.add(new Tuple2<>(numExamples, score * numExamples));
    }

    Nd4j.getExecutioner().commit();

    return out.iterator();
}
 

@Override
public Iterator<Tuple2<Long, Double>> call(Iterator<DataSet> dataSetIterator) throws Exception {
    if (!dataSetIterator.hasNext()) {
        return Collections.singletonList(new Tuple2<>(0L, 0.0)).iterator();
    }

    DataSetIterator iter = new IteratorDataSetIterator(dataSetIterator, minibatchSize); //Does batching where appropriate

    ComputationGraph network = new ComputationGraph(ComputationGraphConfiguration.fromJson(json));
    network.init();
    INDArray val = params.value().unsafeDuplication(); //.value() is shared by all executors on single machine -> OK, as params are not changed in score function
    if (val.length() != network.numParams(false))
        throw new IllegalStateException(
                        "Network did not have same number of parameters as the broadcast set parameters");
    network.setParams(val);

    List<Tuple2<Long, Double>> out = new ArrayList<>();
    while (iter.hasNext()) {
        DataSet ds = iter.next();
        double score = network.score(ds, false);

        long numExamples = ds.getFeatures().size(0);
        out.add(new Tuple2<>(numExamples, score * numExamples));
    }

    Nd4j.getExecutioner().commit();

    return out.iterator();
}
 

@Test
    public void elementWiseMultiplicationLayerTest(){

        for(Activation a : new Activation[]{Activation.IDENTITY, Activation.TANH}) {

            ComputationGraphConfiguration conf = new NeuralNetConfiguration.Builder()
                    .dataType(DataType.DOUBLE)
                    .optimizationAlgo(OptimizationAlgorithm.CONJUGATE_GRADIENT).updater(new NoOp())
                    .seed(12345L)
                    .weightInit(new UniformDistribution(0, 1))
                    .graphBuilder()
                    .addInputs("features")
                    .addLayer("dense", new DenseLayer.Builder().nIn(4).nOut(4)
                            .activation(Activation.TANH)
                            .build(), "features")
                    .addLayer("elementWiseMul", new ElementWiseMultiplicationLayer.Builder().nIn(4).nOut(4)
                            .activation(a)
                            .build(), "dense")
                    .addLayer("loss", new LossLayer.Builder(LossFunctions.LossFunction.COSINE_PROXIMITY)
                            .activation(Activation.IDENTITY).build(), "elementWiseMul")
                    .setOutputs("loss")
                    .build();

            ComputationGraph netGraph = new ComputationGraph(conf);
            netGraph.init();

            log.info("params before learning: " + netGraph.getLayer(1).paramTable());

            //Run a number of iterations of learning manually make some pseudo data
            //the ides is simple: since we do a element wise multiplication layer (just a scaling), we want the cos sim
            // is mainly decided by the fourth value, if everything runs well, we will get a large weight for the fourth value

            INDArray features = Nd4j.create(new double[][]{{1, 2, 3, 4}, {1, 2, 3, 1}, {1, 2, 3, 0}});
            INDArray labels = Nd4j.create(new double[][]{{1, 1, 1, 8}, {1, 1, 1, 2}, {1, 1, 1, 1}});

            netGraph.setInputs(features);
            netGraph.setLabels(labels);
            netGraph.computeGradientAndScore();
            double scoreBefore = netGraph.score();

            String msg;
            for (int epoch = 0; epoch < 5; epoch++)
                netGraph.fit(new INDArray[]{features}, new INDArray[]{labels});
            netGraph.computeGradientAndScore();
            double scoreAfter = netGraph.score();
            //Can't test in 'characteristic mode of operation' if not learning
            msg = "elementWiseMultiplicationLayerTest() - score did not (sufficiently) decrease during learning - activationFn="
                    + "Id" + ", lossFn=" + "Cos-sim" + ", outputActivation=" + "Id"
                    + ", doLearningFirst=" + "true" + " (before=" + scoreBefore
                    + ", scoreAfter=" + scoreAfter + ")";
            assertTrue(msg, scoreAfter < 0.8 * scoreBefore);

//        expectation in case linear regression(with only element wise multiplication layer): large weight for the fourth weight
            log.info("params after learning: " + netGraph.getLayer(1).paramTable());

            boolean gradOK = GradientCheckUtil.checkGradients(new GradientCheckUtil.GraphConfig().net(netGraph).inputs(new INDArray[]{features})
                    .labels(new INDArray[]{labels}));

            msg = "elementWiseMultiplicationLayerTest() - activationFn=" + "ID" + ", lossFn=" + "Cos-sim"
                    + ", outputActivation=" + "Id" + ", doLearningFirst=" + "true";
            assertTrue(msg, gradOK);

            TestUtils.testModelSerialization(netGraph);
        }
    }
 

@Test
public void testOutputLayerMaskingCG(){
    Nd4j.getRandom().setSeed(12345);
    //Idea: RNN input, global pooling, OutputLayer - with "per example" mask arrays

    int mb = 10;
    int tsLength = 5;
    ComputationGraphConfiguration conf = new NeuralNetConfiguration.Builder()
            .dataType(DataType.DOUBLE)
            .weightInit(new NormalDistribution(0,2))
            .updater(new NoOp())
            .graphBuilder()
            .addInputs("in")
            .layer("0", new LSTM.Builder().nIn(3).nOut(3).build(), "in")
            .layer("1", new GlobalPoolingLayer.Builder().poolingType(PoolingType.AVG).build(), "0")
            .layer("out", new OutputLayer.Builder().nIn(3).nOut(3).activation(Activation.SOFTMAX).build(), "1")
            .setOutputs("out")
            .setInputTypes(InputType.recurrent(3))
            .build();

    ComputationGraph net = new ComputationGraph(conf);
    net.init();

    INDArray f = Nd4j.rand(new int[]{mb, 3, tsLength});
    INDArray l = TestUtils.randomOneHot(mb, 3);
    INDArray lm = TestUtils.randomBernoulli(mb, 1);

    int attempts = 0;
    while(attempts++ < 1000 && lm.sumNumber().intValue() == 0){
        lm = TestUtils.randomBernoulli(mb, 1);
    }
    assertTrue("Could not generate non-zero mask after " + attempts + " attempts", lm.sumNumber().intValue() > 0);

    boolean gradOK = GradientCheckUtil.checkGradients(new GradientCheckUtil.GraphConfig().net(net).inputs(new INDArray[]{f})
            .labels(new INDArray[]{l}).labelMask(new INDArray[]{lm}));
    assertTrue(gradOK);

    //Also ensure score doesn't depend on masked feature or label values
    double score = net.score(new DataSet(f,l,null,lm));

    for( int i=0; i<mb; i++ ){
        if(lm.getDouble(i) != 0.0){
            continue;
        }

        INDArray fView = f.get(interval(i,i,true), all(),all());
        fView.assign(Nd4j.rand(fView.shape()));

        INDArray lView = l.get(interval(i,i,true), all());
        lView.assign(TestUtils.randomOneHot(1, lView.size(1)));

        double score2 = net.score(new DataSet(f,l,null,lm));

        assertEquals(String.valueOf(i), score, score2, 1e-8);
    }
}
 

@Test
    public void testGradientBNWithCNNandSubsamplingCompGraph() {
        //Parameterized test, testing combinations of:
        // (a) activation function
        // (b) Whether to test at random initialization, or after some learning (i.e., 'characteristic mode of operation')
        // (c) Loss function (with specified output activations)
        // (d) l1 and l2 values
        Activation[] activFns = {Activation.TANH, Activation.IDENTITY};
        boolean doLearningFirst = true;

        LossFunctions.LossFunction[] lossFunctions = {LossFunctions.LossFunction.NEGATIVELOGLIKELIHOOD};
        Activation[] outputActivations = {Activation.SOFTMAX}; //i.e., lossFunctions[i] used with outputActivations[i] here

        double[] l2vals = {0.0, 0.1};
        double[] l1vals = {0.0, 0.2}; //i.e., use l2vals[j] with l1vals[j]

        Nd4j.getRandom().setSeed(12345);
        int minibatch = 10;
        int depth = 2;
        int hw = 5;
        int nOut = 3;
        INDArray input = Nd4j.rand(new int[]{minibatch, depth, hw, hw});
        INDArray labels = Nd4j.zeros(minibatch, nOut);
        Random r = new Random(12345);
        for (int i = 0; i < minibatch; i++) {
            labels.putScalar(i, r.nextInt(nOut), 1.0);
        }

        DataSet ds = new DataSet(input, labels);

        for (boolean useLogStd : new boolean[]{true, false}) {
            for (Activation afn : activFns) {
                for (int i = 0; i < lossFunctions.length; i++) {
                    for (int j = 0; j < l2vals.length; j++) {
                        LossFunctions.LossFunction lf = lossFunctions[i];
                        Activation outputActivation = outputActivations[i];

                        ComputationGraphConfiguration conf = new NeuralNetConfiguration.Builder().seed(12345)
                                .dataType(DataType.DOUBLE)
                                .optimizationAlgo(OptimizationAlgorithm.LINE_GRADIENT_DESCENT)
                                .updater(new NoOp())
                                .dist(new UniformDistribution(-2, 2)).seed(12345L).graphBuilder()
                                .addInputs("in")
                                .addLayer("0", new ConvolutionLayer.Builder(2, 2).stride(1, 1).nOut(3)
                                        .activation(afn).build(), "in")
                                .addLayer("1", new BatchNormalization.Builder().useLogStd(useLogStd).build(), "0")
                                .addLayer("2", new SubsamplingLayer.Builder(SubsamplingLayer.PoolingType.MAX)
                                        .kernelSize(2, 2).stride(1, 1).build(), "1")
                                .addLayer("3", new BatchNormalization.Builder().useLogStd(useLogStd).build(), "2")
                                .addLayer("4", new ActivationLayer.Builder().activation(afn).build(), "3")
                                .addLayer("5", new OutputLayer.Builder(lf).activation(outputActivation)
                                        .nOut(nOut).build(), "4")
                                .setOutputs("5").setInputTypes(InputType.convolutional(hw, hw, depth))
                                .build();

                        ComputationGraph net = new ComputationGraph(conf);
                        net.init();
                        String name = new Object() {
                        }.getClass().getEnclosingMethod().getName();

                        if (doLearningFirst) {
                            //Run a number of iterations of learning
                            net.setInput(0, ds.getFeatures());
                            net.setLabels(ds.getLabels());
                            net.computeGradientAndScore();
                            double scoreBefore = net.score();
                            for (int k = 0; k < 20; k++)
                                net.fit(ds);
                            net.computeGradientAndScore();
                            double scoreAfter = net.score();
                            //Can't test in 'characteristic mode of operation' if not learning
                            String msg = name
                                    + " - score did not (sufficiently) decrease during learning - activationFn="
                                    + afn + ", lossFn=" + lf + ", outputActivation=" + outputActivation
                                    + ", doLearningFirst= " + doLearningFirst + " (before=" + scoreBefore
                                    + ", scoreAfter=" + scoreAfter + ")";
                            assertTrue(msg, scoreAfter < 0.9 * scoreBefore);
                        }

                        System.out.println(name + " - activationFn=" + afn + ", lossFn=" + lf
                                + ", outputActivation=" + outputActivation + ", doLearningFirst="
                                + doLearningFirst + ", l1=" + l1vals[j] + ", l2=" + l2vals[j]);
//                        for (int k = 0; k < net.getNumLayers(); k++)
//                            System.out.println("Layer " + k + " # params: " + net.getLayer(k).numParams());

                        //Mean and variance vars are not gradient checkable; mean/variance "gradient" is used to implement running mean/variance calc
                        //i.e., runningMean = decay * runningMean + (1-decay) * batchMean
                        //However, numerical gradient will be 0 as forward pass doesn't depend on this "parameter"
                        Set<String> excludeParams = new HashSet<>(Arrays.asList("1_mean", "1_var", "3_mean", "3_var", "1_log10stdev", "3_log10stdev"));
                        boolean gradOK = GradientCheckUtil.checkGradients(new GradientCheckUtil.GraphConfig().net(net).inputs(new INDArray[]{input})
                                .labels(new INDArray[]{labels}).excludeParams(excludeParams));

                        assertTrue(gradOK);
                        TestUtils.testModelSerialization(net);
                    }
                }
            }
        }
    }
 

@Test
public void testSameDiffLamdaLayerBasic(){
    for(WorkspaceMode wsm : new WorkspaceMode[]{WorkspaceMode.ENABLED, WorkspaceMode.NONE}) {
        log.info("--- Workspace Mode: {} ---", wsm);


        Nd4j.getRandom().setSeed(12345);
        ComputationGraphConfiguration conf = new NeuralNetConfiguration.Builder()
                .trainingWorkspaceMode(wsm)
                .inferenceWorkspaceMode(wsm)
                .seed(12345)
                .updater(new Adam(0.01))
                .graphBuilder()
                .addInputs("in")
                .addLayer("0", new DenseLayer.Builder().nIn(5).nOut(5).activation(Activation.TANH).build(), "in")
                .addLayer("1", new SameDiffSimpleLambdaLayer(), "0")
                .addLayer("2", new OutputLayer.Builder().nIn(5).nOut(5).activation(Activation.SOFTMAX)
                        .lossFunction(LossFunctions.LossFunction.MCXENT).build(), "1")
                .setOutputs("2")
                .build();

        //Equavalent, not using SameDiff Lambda:
        ComputationGraphConfiguration confStd = new NeuralNetConfiguration.Builder()
                .trainingWorkspaceMode(wsm)
                .inferenceWorkspaceMode(wsm)
                .seed(12345)
                .updater(new Adam(0.01))
                .graphBuilder()
                .addInputs("in")
                .addLayer("0", new DenseLayer.Builder().nIn(5).nOut(5).activation(Activation.TANH).build(), "in")
                .addVertex("1", new ShiftVertex(1.0), "0")
                .addVertex("2", new ScaleVertex(2.0), "1")
                .addLayer("3", new OutputLayer.Builder().nIn(5).nOut(5).activation(Activation.SOFTMAX)
                        .lossFunction(LossFunctions.LossFunction.MCXENT).build(), "2")
                .setOutputs("3")
                .build();

        ComputationGraph lambda = new ComputationGraph(conf);
        lambda.init();

        ComputationGraph std = new ComputationGraph(confStd);
        std.init();

        lambda.setParams(std.params());

        INDArray in = Nd4j.rand(3, 5);
        INDArray labels = TestUtils.randomOneHot(3, 5);
        DataSet ds = new DataSet(in, labels);

        INDArray outLambda = lambda.outputSingle(in);
        INDArray outStd = std.outputSingle(in);

        assertEquals(outLambda, outStd);

        double scoreLambda = lambda.score(ds);
        double scoreStd = std.score(ds);

        assertEquals(scoreStd, scoreLambda, 1e-6);

        for (int i = 0; i < 3; i++) {
            lambda.fit(ds);
            std.fit(ds);

            String s = String.valueOf(i);
            assertEquals(s, std.params(), lambda.params());
            assertEquals(s, std.getFlattenedGradients(), lambda.getFlattenedGradients());
        }

        ComputationGraph loaded = TestUtils.testModelSerialization(lambda);
        outLambda = loaded.outputSingle(in);
        outStd = std.outputSingle(in);

        assertEquals(outStd, outLambda);

        //Sanity check on different minibatch sizes:
        INDArray newIn = Nd4j.vstack(in, in);
        INDArray outMbsd = lambda.output(newIn)[0];
        INDArray outMb = std.output(newIn)[0];
        assertEquals(outMb, outMbsd);
    }
}
 

@Test
public void testSameDiffLamdaVertexBasic(){
    for(WorkspaceMode wsm : new WorkspaceMode[]{WorkspaceMode.ENABLED, WorkspaceMode.NONE}) {
        log.info("--- Workspace Mode: {} ---", wsm);

        Nd4j.getRandom().setSeed(12345);
        ComputationGraphConfiguration conf = new NeuralNetConfiguration.Builder()
                .trainingWorkspaceMode(wsm)
                .inferenceWorkspaceMode(wsm)
                .dataType(DataType.DOUBLE)
                .seed(12345)
                .updater(new Adam(0.01))
                .graphBuilder()
                .addInputs("in1", "in2")
                .addLayer("0", new DenseLayer.Builder().nIn(5).nOut(5).activation(Activation.TANH).build(), "in1")
                .addLayer("1", new DenseLayer.Builder().nIn(5).nOut(5).activation(Activation.TANH).build(), "in2")
                .addVertex("lambda", new SameDiffSimpleLambdaVertex(), "0", "1")
                .addLayer("2", new OutputLayer.Builder().nIn(5).nOut(5).activation(Activation.SOFTMAX)
                        .lossFunction(LossFunctions.LossFunction.MCXENT).build(), "lambda")
                .setOutputs("2")
                .build();

        //Equavalent, not using SameDiff Lambda:
        ComputationGraphConfiguration confStd = new NeuralNetConfiguration.Builder()
                .trainingWorkspaceMode(wsm)
                .inferenceWorkspaceMode(wsm)
                .dataType(DataType.DOUBLE)
                .seed(12345)
                .updater(new Adam(0.01))
                .graphBuilder()
                .addInputs("in1", "in2")
                .addLayer("0", new DenseLayer.Builder().nIn(5).nOut(5).activation(Activation.TANH).build(), "in1")
                .addLayer("1", new DenseLayer.Builder().nIn(5).nOut(5).activation(Activation.TANH).build(), "in2")
                .addVertex("elementwise", new ElementWiseVertex(ElementWiseVertex.Op.Product), "0", "1")
                .addLayer("3", new OutputLayer.Builder().nIn(5).nOut(5).activation(Activation.SOFTMAX)
                        .lossFunction(LossFunctions.LossFunction.MCXENT).build(), "elementwise")
                .setOutputs("3")
                .build();

        ComputationGraph lambda = new ComputationGraph(conf);
        lambda.init();

        ComputationGraph std = new ComputationGraph(confStd);
        std.init();

        lambda.setParams(std.params());

        INDArray in1 = Nd4j.rand(3, 5);
        INDArray in2 = Nd4j.rand(3, 5);
        INDArray labels = TestUtils.randomOneHot(3, 5);
        MultiDataSet mds = new org.nd4j.linalg.dataset.MultiDataSet(new INDArray[]{in1, in2}, new INDArray[]{labels});

        INDArray outLambda = lambda.output(in1, in2)[0];
        INDArray outStd = std.output(in1, in2)[0];

        assertEquals(outLambda, outStd);

        double scoreLambda = lambda.score(mds);
        double scoreStd = std.score(mds);

        assertEquals(scoreStd, scoreLambda, 1e-6);

        for (int i = 0; i < 3; i++) {
            lambda.fit(mds);
            std.fit(mds);

            String s = String.valueOf(i);
            assertEquals(s, std.params(), lambda.params());
            assertEquals(s, std.getFlattenedGradients(), lambda.getFlattenedGradients());
        }

        ComputationGraph loaded = TestUtils.testModelSerialization(lambda);
        outLambda = loaded.output(in1, in2)[0];
        outStd = std.output(in1, in2)[0];

        assertEquals(outStd, outLambda);

        //Sanity check on different minibatch sizes:
        INDArray newIn1 = Nd4j.vstack(in1, in1);
        INDArray newIn2 = Nd4j.vstack(in2, in2);
        INDArray outMbsd = lambda.output(newIn1, newIn2)[0];
        INDArray outMb = std.output(newIn1, newIn2)[0];
        assertEquals(outMb, outMbsd);
    }
}
 

@Test
public void testComputationGraphTypeConversion() {

    for (DataType dt : new DataType[]{DataType.DOUBLE, DataType.FLOAT, DataType.HALF}) {
        Nd4j.setDefaultDataTypes(dt, dt);

        ComputationGraphConfiguration conf = new NeuralNetConfiguration.Builder()
                .seed(12345)
                .weightInit(WeightInit.XAVIER)
                .updater(new Adam(0.01))
                .dataType(DataType.DOUBLE)
                .graphBuilder()
                .addInputs("in")
                .layer("l0", new DenseLayer.Builder().activation(Activation.TANH).nIn(10).nOut(10).build(), "in")
                .layer("l1", new DenseLayer.Builder().activation(Activation.TANH).nIn(10).nOut(10).build(), "l0")
                .layer("out", new OutputLayer.Builder().nIn(10).nOut(10).activation(Activation.SOFTMAX).lossFunction(LossFunctions.LossFunction.MCXENT).build(), "l1")
                .setOutputs("out")
                .build();

        ComputationGraph net = new ComputationGraph(conf);
        net.init();

        INDArray inD = Nd4j.rand(DataType.DOUBLE, 1, 10);
        INDArray lD = Nd4j.create(DataType.DOUBLE, 1, 10);
        net.fit(new DataSet(inD, lD));

        INDArray outDouble = net.outputSingle(inD);
        net.setInput(0, inD);
        net.setLabels(lD);
        net.computeGradientAndScore();
        double scoreDouble = net.score();
        INDArray grads = net.getFlattenedGradients();
        INDArray u = net.getUpdater().getStateViewArray();
        assertEquals(DataType.DOUBLE, net.params().dataType());
        assertEquals(DataType.DOUBLE, grads.dataType());
        assertEquals(DataType.DOUBLE, u.dataType());


        ComputationGraph netFloat = net.convertDataType(DataType.FLOAT);
        netFloat.initGradientsView();
        assertEquals(DataType.FLOAT, netFloat.params().dataType());
        assertEquals(DataType.FLOAT, netFloat.getFlattenedGradients().dataType());
        assertEquals(DataType.FLOAT, netFloat.getUpdater(true).getStateViewArray().dataType());
        INDArray inF = inD.castTo(DataType.FLOAT);
        INDArray lF = lD.castTo(DataType.FLOAT);
        INDArray outFloat = netFloat.outputSingle(inF);
        netFloat.setInput(0, inF);
        netFloat.setLabels(lF);
        netFloat.computeGradientAndScore();
        double scoreFloat = netFloat.score();
        INDArray gradsFloat = netFloat.getFlattenedGradients();
        INDArray uFloat = netFloat.getUpdater().getStateViewArray();

        assertEquals(scoreDouble, scoreFloat, 1e-6);
        assertEquals(outDouble.castTo(DataType.FLOAT), outFloat);
        assertEquals(grads.castTo(DataType.FLOAT), gradsFloat);
        INDArray uCast = u.castTo(DataType.FLOAT);
        assertTrue(uCast.equalsWithEps(uFloat, 1e-4));

        ComputationGraph netFP16 = net.convertDataType(DataType.HALF);
        netFP16.initGradientsView();
        assertEquals(DataType.HALF, netFP16.params().dataType());
        assertEquals(DataType.HALF, netFP16.getFlattenedGradients().dataType());
        assertEquals(DataType.HALF, netFP16.getUpdater(true).getStateViewArray().dataType());

        INDArray inH = inD.castTo(DataType.HALF);
        INDArray lH = lD.castTo(DataType.HALF);
        INDArray outHalf = netFP16.outputSingle(inH);
        netFP16.setInput(0, inH);
        netFP16.setLabels(lH);
        netFP16.computeGradientAndScore();
        double scoreHalf = netFP16.score();
        INDArray gradsHalf = netFP16.getFlattenedGradients();
        INDArray uHalf = netFP16.getUpdater().getStateViewArray();

        assertEquals(scoreDouble, scoreHalf, 1e-4);
        boolean outHalfEq = outDouble.castTo(DataType.HALF).equalsWithEps(outHalf, 1e-3);
        assertTrue(outHalfEq);
        boolean gradsHalfEq = grads.castTo(DataType.HALF).equalsWithEps(gradsHalf, 1e-3);
        assertTrue(gradsHalfEq);
        INDArray uHalfCast = u.castTo(DataType.HALF);
        assertTrue(uHalfCast.equalsWithEps(uHalf, 1e-4));
    }
}
 

@Override
public void onCandidateIteration(CandidateInfo candidateInfo, Object candidate, int iteration) {
    double score;
    long numParams;
    int numLayers;
    String modelConfigJson;
    int totalNumUpdates;
    if(candidate instanceof MultiLayerNetwork){
        MultiLayerNetwork m = (MultiLayerNetwork)candidate;
        score = m.score();
        numParams = m.numParams();
        numLayers = m.getnLayers();
        modelConfigJson = m.getLayerWiseConfigurations().toJson();
        totalNumUpdates = m.getLayerWiseConfigurations().getIterationCount();
    } else if(candidate instanceof ComputationGraph) {
        ComputationGraph cg = (ComputationGraph)candidate;
        score = cg.score();
        numParams = cg.numParams();
        numLayers = cg.getNumLayers();
        modelConfigJson = cg.getConfiguration().toJson();
        totalNumUpdates = cg.getConfiguration().getIterationCount();
    } else {
        score = 0;
        numParams = 0;
        numLayers = 0;
        totalNumUpdates = 0;
        modelConfigJson = "";
    }

    int idx = candidateInfo.getIndex();

    Pair<IntArrayList, FloatArrayList> pair = candidateScoreVsIter.computeIfAbsent(idx, k -> new Pair<>(new IntArrayList(), new FloatArrayList()));

    IntArrayList iter = pair.getFirst();
    FloatArrayList scores = pair.getSecond();

    //Do we need subsampling to avoid having too many data points?
    int subsamplingFreq = candidateScoreVsIterSubsampleFreq.computeIfAbsent(idx, k -> 1);
    if(iteration / subsamplingFreq > MAX_SCORE_VS_ITER_PTS){
        //Double subsampling frequency and re-parse data
        subsamplingFreq *= 2;
        candidateScoreVsIterSubsampleFreq.put(idx, subsamplingFreq);

        IntArrayList newIter = new IntArrayList();
        FloatArrayList newScores = new FloatArrayList();
        for( int i=0; i<iter.size(); i++ ){
            int it = iter.get(i);
            if(it % subsamplingFreq == 0){
                newIter.add(it);
                newScores.add(scores.get(i));
            }
        }

        iter = newIter;
        scores = newScores;
        candidateScoreVsIter.put(idx, new Pair<>(iter, scores));
    }

    if(iteration % subsamplingFreq == 0) {
        iter.add(iteration);
        scores.add((float) score);
    }


    int[] iters = iter.toIntArray();
    float[] fScores = new float[iters.length];
    for( int i=0; i<iters.length; i++ ){
        fScores[i] = scores.get(i);
    }

    ModelInfoPersistable p = new ModelInfoPersistable.Builder()
            .timestamp(candidateInfo.getCreatedTime())
            .sessionId(sessionId)
            .workerId(String.valueOf(candidateInfo.getIndex()))
            .modelIdx(candidateInfo.getIndex())
            .score(candidateInfo.getScore())
            .status(candidateInfo.getCandidateStatus())
            .scoreVsIter(iters, fScores)
            .lastUpdateTime(System.currentTimeMillis())
            .numParameters(numParams)
            .numLayers(numLayers)
            .totalNumUpdates(totalNumUpdates)
            .paramSpaceValues(candidateInfo.getFlatParams())
            .modelConfigJson(modelConfigJson)
            .exceptionStackTrace(candidateInfo.getExceptionStackTrace())
            .build();


    lastModelInfoPersistable.put(candidateInfo.getIndex(), p);
    statsStorage.putUpdate(p);
}