Examples with SparkDl4jMultiLayer org.deeplearning4j.spark.impl.multilayer.SparkDl4jMultiLayer used on opensource projects

Example 6 with SparkDl4jMultiLayer

use of org.deeplearning4j.spark.impl.multilayer.SparkDl4jMultiLayer in project deeplearning4j by deeplearning4j.

the class TestSparkMultiLayerParameterAveraging method testDistributedScoring.

@Test
public void testDistributedScoring() {
    MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().regularization(true).l1(0.1).l2(0.1).seed(123).updater(Updater.NESTEROVS).learningRate(0.1).momentum(0.9).list().layer(0, new org.deeplearning4j.nn.conf.layers.DenseLayer.Builder().nIn(nIn).nOut(3).activation(Activation.TANH).build()).layer(1, new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(LossFunctions.LossFunction.MCXENT).nIn(3).nOut(nOut).activation(Activation.SOFTMAX).build()).backprop(true).pretrain(false).build();
    SparkDl4jMultiLayer sparkNet = new SparkDl4jMultiLayer(sc, conf, new ParameterAveragingTrainingMaster(true, numExecutors(), 1, 10, 1, 0));
    MultiLayerNetwork netCopy = sparkNet.getNetwork().clone();
    int nRows = 100;
    INDArray features = Nd4j.rand(nRows, nIn);
    INDArray labels = Nd4j.zeros(nRows, nOut);
    Random r = new Random(12345);
    for (int i = 0; i < nRows; i++) {
        labels.putScalar(new int[] { i, r.nextInt(nOut) }, 1.0);
    }
    INDArray localScoresWithReg = netCopy.scoreExamples(new DataSet(features, labels), true);
    INDArray localScoresNoReg = netCopy.scoreExamples(new DataSet(features, labels), false);
    List<Tuple2<String, DataSet>> dataWithKeys = new ArrayList<>();
    for (int i = 0; i < nRows; i++) {
        DataSet ds = new DataSet(features.getRow(i).dup(), labels.getRow(i).dup());
        dataWithKeys.add(new Tuple2<>(String.valueOf(i), ds));
    }
    JavaPairRDD<String, DataSet> dataWithKeysRdd = sc.parallelizePairs(dataWithKeys);
    JavaPairRDD<String, Double> sparkScoresWithReg = sparkNet.scoreExamples(dataWithKeysRdd, true, 4);
    JavaPairRDD<String, Double> sparkScoresNoReg = sparkNet.scoreExamples(dataWithKeysRdd, false, 4);
    Map<String, Double> sparkScoresWithRegMap = sparkScoresWithReg.collectAsMap();
    Map<String, Double> sparkScoresNoRegMap = sparkScoresNoReg.collectAsMap();
    for (int i = 0; i < nRows; i++) {
        double scoreRegExp = localScoresWithReg.getDouble(i);
        double scoreRegAct = sparkScoresWithRegMap.get(String.valueOf(i));
        assertEquals(scoreRegExp, scoreRegAct, 1e-5);
        double scoreNoRegExp = localScoresNoReg.getDouble(i);
        double scoreNoRegAct = sparkScoresNoRegMap.get(String.valueOf(i));
        assertEquals(scoreNoRegExp, scoreNoRegAct, 1e-5);
    //            System.out.println(scoreRegExp + "\t" + scoreRegAct + "\t" + scoreNoRegExp + "\t" + scoreNoRegAct);
    }
    List<DataSet> dataNoKeys = new ArrayList<>();
    for (int i = 0; i < nRows; i++) {
        dataNoKeys.add(new DataSet(features.getRow(i).dup(), labels.getRow(i).dup()));
    }
    JavaRDD<DataSet> dataNoKeysRdd = sc.parallelize(dataNoKeys);
    List<Double> scoresWithReg = new ArrayList<>(sparkNet.scoreExamples(dataNoKeysRdd, true, 4).collect());
    List<Double> scoresNoReg = new ArrayList<>(sparkNet.scoreExamples(dataNoKeysRdd, false, 4).collect());
    Collections.sort(scoresWithReg);
    Collections.sort(scoresNoReg);
    double[] localScoresWithRegDouble = localScoresWithReg.data().asDouble();
    double[] localScoresNoRegDouble = localScoresNoReg.data().asDouble();
    Arrays.sort(localScoresWithRegDouble);
    Arrays.sort(localScoresNoRegDouble);
    for (int i = 0; i < localScoresWithRegDouble.length; i++) {
        assertEquals(localScoresWithRegDouble[i], scoresWithReg.get(i), 1e-5);
        assertEquals(localScoresNoRegDouble[i], scoresNoReg.get(i), 1e-5);
    //System.out.println(localScoresWithRegDouble[i] + "\t" + scoresWithReg.get(i) + "\t" + localScoresNoRegDouble[i] + "\t" + scoresNoReg.get(i));
    }
}

Example 7 with SparkDl4jMultiLayer

use of org.deeplearning4j.spark.impl.multilayer.SparkDl4jMultiLayer in project deeplearning4j by deeplearning4j.

the class TestSparkMultiLayerParameterAveraging method testSeedRepeatability.

@Test
public void testSeedRepeatability() throws Exception {
    MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().seed(12345).updater(Updater.RMSPROP).optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT).iterations(1).weightInit(WeightInit.XAVIER).list().layer(0, new org.deeplearning4j.nn.conf.layers.DenseLayer.Builder().nIn(4).nOut(4).activation(Activation.TANH).build()).layer(1, new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(LossFunctions.LossFunction.MCXENT).nIn(4).nOut(3).activation(Activation.SOFTMAX).build()).pretrain(false).backprop(true).build();
    Nd4j.getRandom().setSeed(12345);
    MultiLayerNetwork n1 = new MultiLayerNetwork(conf);
    n1.init();
    Nd4j.getRandom().setSeed(12345);
    MultiLayerNetwork n2 = new MultiLayerNetwork(conf);
    n2.init();
    Nd4j.getRandom().setSeed(12345);
    MultiLayerNetwork n3 = new MultiLayerNetwork(conf);
    n3.init();
    SparkDl4jMultiLayer sparkNet1 = new SparkDl4jMultiLayer(sc, n1, new ParameterAveragingTrainingMaster.Builder(1).workerPrefetchNumBatches(5).batchSizePerWorker(5).averagingFrequency(1).repartionData(Repartition.Always).rngSeed(12345).build());
    //Training master IDs are only unique if they are created at least 1 ms apart...
    Thread.sleep(100);
    SparkDl4jMultiLayer sparkNet2 = new SparkDl4jMultiLayer(sc, n2, new ParameterAveragingTrainingMaster.Builder(1).workerPrefetchNumBatches(5).batchSizePerWorker(5).averagingFrequency(1).repartionData(Repartition.Always).rngSeed(12345).build());
    Thread.sleep(100);
    SparkDl4jMultiLayer sparkNet3 = new SparkDl4jMultiLayer(sc, n3, new ParameterAveragingTrainingMaster.Builder(1).workerPrefetchNumBatches(5).batchSizePerWorker(5).averagingFrequency(1).repartionData(Repartition.Always).rngSeed(98765).build());
    List<DataSet> data = new ArrayList<>();
    DataSetIterator iter = new IrisDataSetIterator(1, 150);
    while (iter.hasNext()) data.add(iter.next());
    JavaRDD<DataSet> rdd = sc.parallelize(data);
    sparkNet1.fit(rdd);
    sparkNet2.fit(rdd);
    sparkNet3.fit(rdd);
    INDArray p1 = sparkNet1.getNetwork().params();
    INDArray p2 = sparkNet2.getNetwork().params();
    INDArray p3 = sparkNet3.getNetwork().params();
    sparkNet1.getTrainingMaster().deleteTempFiles(sc);
    sparkNet2.getTrainingMaster().deleteTempFiles(sc);
    sparkNet3.getTrainingMaster().deleteTempFiles(sc);
    assertEquals(p1, p2);
    assertNotEquals(p1, p3);
}

Example 8 with SparkDl4jMultiLayer

use of org.deeplearning4j.spark.impl.multilayer.SparkDl4jMultiLayer in project deeplearning4j by deeplearning4j.

the class TestSparkMultiLayerParameterAveraging method testRunIteration.

@Test
public void testRunIteration() {
    DataSet dataSet = new IrisDataSetIterator(5, 5).next();
    List<DataSet> list = dataSet.asList();
    JavaRDD<DataSet> data = sc.parallelize(list);
    SparkDl4jMultiLayer sparkNetCopy = new SparkDl4jMultiLayer(sc, getBasicConf(), new ParameterAveragingTrainingMaster(true, numExecutors(), 1, 5, 1, 0));
    MultiLayerNetwork networkCopy = sparkNetCopy.fit(data);
    INDArray expectedParams = networkCopy.params();
    SparkDl4jMultiLayer sparkNet = getBasicNetwork();
    MultiLayerNetwork network = sparkNet.fit(data);
    INDArray actualParams = network.params();
    assertEquals(expectedParams.size(1), actualParams.size(1));
}

Example 9 with SparkDl4jMultiLayer

use of org.deeplearning4j.spark.impl.multilayer.SparkDl4jMultiLayer in project deeplearning4j by deeplearning4j.

the class TestSparkMultiLayerParameterAveraging method testUpdaters.

@Test
public void testUpdaters() {
    SparkDl4jMultiLayer sparkNet = getBasicNetwork();
    MultiLayerNetwork netCopy = sparkNet.getNetwork().clone();
    netCopy.fit(data);
    Updater expectedUpdater = netCopy.conf().getLayer().getUpdater();
    double expectedLR = netCopy.conf().getLayer().getLearningRate();
    double expectedMomentum = netCopy.conf().getLayer().getMomentum();
    Updater actualUpdater = sparkNet.getNetwork().conf().getLayer().getUpdater();
    sparkNet.fit(sparkData);
    double actualLR = sparkNet.getNetwork().conf().getLayer().getLearningRate();
    double actualMomentum = sparkNet.getNetwork().conf().getLayer().getMomentum();
    assertEquals(expectedUpdater, actualUpdater);
    assertEquals(expectedLR, actualLR, 0.01);
    assertEquals(expectedMomentum, actualMomentum, 0.01);
}

Example 10 with SparkDl4jMultiLayer

use of org.deeplearning4j.spark.impl.multilayer.SparkDl4jMultiLayer in project deeplearning4j by deeplearning4j.

the class TestTrainingStatsCollection method testStatsCollection.

@Test
public void testStatsCollection() throws Exception {
    int nWorkers = 4;
    SparkConf sparkConf = new SparkConf();
    sparkConf.setMaster("local[" + nWorkers + "]");
    sparkConf.setAppName("Test");
    JavaSparkContext sc = new JavaSparkContext(sparkConf);
    try {
        MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT).iterations(1).list().layer(0, new DenseLayer.Builder().nIn(10).nOut(10).build()).layer(1, new OutputLayer.Builder().nIn(10).nOut(10).build()).pretrain(false).backprop(true).build();
        int miniBatchSizePerWorker = 10;
        int averagingFrequency = 5;
        int numberOfAveragings = 3;
        int totalExamples = nWorkers * miniBatchSizePerWorker * averagingFrequency * numberOfAveragings;
        Nd4j.getRandom().setSeed(12345);
        List<DataSet> list = new ArrayList<>();
        for (int i = 0; i < totalExamples; i++) {
            INDArray f = Nd4j.rand(1, 10);
            INDArray l = Nd4j.rand(1, 10);
            DataSet ds = new DataSet(f, l);
            list.add(ds);
        }
        JavaRDD<DataSet> rdd = sc.parallelize(list);
        rdd.repartition(4);
        ParameterAveragingTrainingMaster tm = new ParameterAveragingTrainingMaster.Builder(nWorkers, 1).averagingFrequency(averagingFrequency).batchSizePerWorker(miniBatchSizePerWorker).saveUpdater(true).workerPrefetchNumBatches(0).repartionData(Repartition.Always).build();
        SparkDl4jMultiLayer sparkNet = new SparkDl4jMultiLayer(sc, conf, tm);
        sparkNet.setCollectTrainingStats(true);
        sparkNet.fit(rdd);
        //Collect the expected keys:
        List<String> expectedStatNames = new ArrayList<>();
        Class<?>[] classes = new Class[] { CommonSparkTrainingStats.class, ParameterAveragingTrainingMasterStats.class, ParameterAveragingTrainingWorkerStats.class };
        String[] fieldNames = new String[] { "columnNames", "columnNames", "columnNames" };
        for (int i = 0; i < classes.length; i++) {
            Field field = classes[i].getDeclaredField(fieldNames[i]);
            field.setAccessible(true);
            Object f = field.get(null);
            Collection<String> c = (Collection<String>) f;
            expectedStatNames.addAll(c);
        }
        System.out.println(expectedStatNames);
        SparkTrainingStats stats = sparkNet.getSparkTrainingStats();
        Set<String> actualKeySet = stats.getKeySet();
        assertEquals(expectedStatNames.size(), actualKeySet.size());
        for (String s : stats.getKeySet()) {
            assertTrue(expectedStatNames.contains(s));
            assertNotNull(stats.getValue(s));
        }
        String statsAsString = stats.statsAsString();
        System.out.println(statsAsString);
        //One line per stat
        assertEquals(actualKeySet.size(), statsAsString.split("\n").length);
        //Go through nested stats
        //First: master stats
        assertTrue(stats instanceof ParameterAveragingTrainingMasterStats);
        ParameterAveragingTrainingMasterStats masterStats = (ParameterAveragingTrainingMasterStats) stats;
        List<EventStats> exportTimeStats = masterStats.getParameterAveragingMasterExportTimesMs();
        assertEquals(1, exportTimeStats.size());
        assertDurationGreaterZero(exportTimeStats);
        assertNonNullFields(exportTimeStats);
        assertExpectedNumberMachineIdsJvmIdsThreadIds(exportTimeStats, 1, 1, 1);
        List<EventStats> countRddTime = masterStats.getParameterAveragingMasterCountRddSizeTimesMs();
        //occurs once per fit
        assertEquals(1, countRddTime.size());
        assertDurationGreaterEqZero(countRddTime);
        assertNonNullFields(countRddTime);
        //should occur only in master once
        assertExpectedNumberMachineIdsJvmIdsThreadIds(countRddTime, 1, 1, 1);
        List<EventStats> broadcastCreateTime = masterStats.getParameterAveragingMasterBroadcastCreateTimesMs();
        assertEquals(numberOfAveragings, broadcastCreateTime.size());
        assertDurationGreaterEqZero(broadcastCreateTime);
        assertNonNullFields(broadcastCreateTime);
        //only 1 thread for master
        assertExpectedNumberMachineIdsJvmIdsThreadIds(broadcastCreateTime, 1, 1, 1);
        List<EventStats> fitTimes = masterStats.getParameterAveragingMasterFitTimesMs();
        //i.e., number of times fit(JavaRDD<DataSet>) was called
        assertEquals(1, fitTimes.size());
        assertDurationGreaterZero(fitTimes);
        assertNonNullFields(fitTimes);
        //only 1 thread for master
        assertExpectedNumberMachineIdsJvmIdsThreadIds(fitTimes, 1, 1, 1);
        List<EventStats> splitTimes = masterStats.getParameterAveragingMasterSplitTimesMs();
        //Splitting of the data set is executed once only (i.e., one fit(JavaRDD<DataSet>) call)
        assertEquals(1, splitTimes.size());
        assertDurationGreaterEqZero(splitTimes);
        assertNonNullFields(splitTimes);
        //only 1 thread for master
        assertExpectedNumberMachineIdsJvmIdsThreadIds(splitTimes, 1, 1, 1);
        List<EventStats> aggregateTimesMs = masterStats.getParamaterAveragingMasterAggregateTimesMs();
        assertEquals(numberOfAveragings, aggregateTimesMs.size());
        assertDurationGreaterEqZero(aggregateTimesMs);
        assertNonNullFields(aggregateTimesMs);
        //only 1 thread for master
        assertExpectedNumberMachineIdsJvmIdsThreadIds(aggregateTimesMs, 1, 1, 1);
        List<EventStats> processParamsTimesMs = masterStats.getParameterAveragingMasterProcessParamsUpdaterTimesMs();
        assertEquals(numberOfAveragings, processParamsTimesMs.size());
        assertDurationGreaterEqZero(processParamsTimesMs);
        assertNonNullFields(processParamsTimesMs);
        //only 1 thread for master
        assertExpectedNumberMachineIdsJvmIdsThreadIds(processParamsTimesMs, 1, 1, 1);
        List<EventStats> repartitionTimesMs = masterStats.getParameterAveragingMasterRepartitionTimesMs();
        assertEquals(numberOfAveragings, repartitionTimesMs.size());
        assertDurationGreaterEqZero(repartitionTimesMs);
        assertNonNullFields(repartitionTimesMs);
        //only 1 thread for master
        assertExpectedNumberMachineIdsJvmIdsThreadIds(repartitionTimesMs, 1, 1, 1);
        //Second: Common spark training stats
        SparkTrainingStats commonStats = masterStats.getNestedTrainingStats();
        assertNotNull(commonStats);
        assertTrue(commonStats instanceof CommonSparkTrainingStats);
        CommonSparkTrainingStats cStats = (CommonSparkTrainingStats) commonStats;
        List<EventStats> workerFlatMapTotalTimeMs = cStats.getWorkerFlatMapTotalTimeMs();
        assertEquals(numberOfAveragings * nWorkers, workerFlatMapTotalTimeMs.size());
        assertDurationGreaterZero(workerFlatMapTotalTimeMs);
        assertNonNullFields(workerFlatMapTotalTimeMs);
        assertExpectedNumberMachineIdsJvmIdsThreadIds(workerFlatMapTotalTimeMs, 1, 1, nWorkers);
        List<EventStats> workerFlatMapGetInitialModelTimeMs = cStats.getWorkerFlatMapGetInitialModelTimeMs();
        assertEquals(numberOfAveragings * nWorkers, workerFlatMapGetInitialModelTimeMs.size());
        assertDurationGreaterEqZero(workerFlatMapGetInitialModelTimeMs);
        assertNonNullFields(workerFlatMapGetInitialModelTimeMs);
        assertExpectedNumberMachineIdsJvmIdsThreadIds(workerFlatMapGetInitialModelTimeMs, 1, 1, nWorkers);
        List<EventStats> workerFlatMapDataSetGetTimesMs = cStats.getWorkerFlatMapDataSetGetTimesMs();
        int numMinibatchesProcessed = workerFlatMapDataSetGetTimesMs.size();
        //1 for every time we get a data set
        int expectedNumMinibatchesProcessed = numberOfAveragings * nWorkers * averagingFrequency;
        //Sometimes random split is just bad - some executors might miss out on getting the expected amount of data
        assertTrue(numMinibatchesProcessed >= expectedNumMinibatchesProcessed - 5);
        List<EventStats> workerFlatMapProcessMiniBatchTimesMs = cStats.getWorkerFlatMapProcessMiniBatchTimesMs();
        assertTrue(workerFlatMapProcessMiniBatchTimesMs.size() >= numberOfAveragings * nWorkers * averagingFrequency - 5);
        assertDurationGreaterEqZero(workerFlatMapProcessMiniBatchTimesMs);
        assertNonNullFields(workerFlatMapDataSetGetTimesMs);
        assertExpectedNumberMachineIdsJvmIdsThreadIds(workerFlatMapDataSetGetTimesMs, 1, 1, nWorkers);
        //Third: ParameterAveragingTrainingWorker stats
        SparkTrainingStats paramAvgStats = cStats.getNestedTrainingStats();
        assertNotNull(paramAvgStats);
        assertTrue(paramAvgStats instanceof ParameterAveragingTrainingWorkerStats);
        ParameterAveragingTrainingWorkerStats pStats = (ParameterAveragingTrainingWorkerStats) paramAvgStats;
        List<EventStats> parameterAveragingWorkerBroadcastGetValueTimeMs = pStats.getParameterAveragingWorkerBroadcastGetValueTimeMs();
        assertEquals(numberOfAveragings * nWorkers, parameterAveragingWorkerBroadcastGetValueTimeMs.size());
        assertDurationGreaterEqZero(parameterAveragingWorkerBroadcastGetValueTimeMs);
        assertNonNullFields(parameterAveragingWorkerBroadcastGetValueTimeMs);
        assertExpectedNumberMachineIdsJvmIdsThreadIds(parameterAveragingWorkerBroadcastGetValueTimeMs, 1, 1, nWorkers);
        List<EventStats> parameterAveragingWorkerInitTimeMs = pStats.getParameterAveragingWorkerInitTimeMs();
        assertEquals(numberOfAveragings * nWorkers, parameterAveragingWorkerInitTimeMs.size());
        assertDurationGreaterEqZero(parameterAveragingWorkerInitTimeMs);
        assertNonNullFields(parameterAveragingWorkerInitTimeMs);
        assertExpectedNumberMachineIdsJvmIdsThreadIds(parameterAveragingWorkerInitTimeMs, 1, 1, nWorkers);
        List<EventStats> parameterAveragingWorkerFitTimesMs = pStats.getParameterAveragingWorkerFitTimesMs();
        assertTrue(parameterAveragingWorkerFitTimesMs.size() >= numberOfAveragings * nWorkers * averagingFrequency - 5);
        assertDurationGreaterEqZero(parameterAveragingWorkerFitTimesMs);
        assertNonNullFields(parameterAveragingWorkerFitTimesMs);
        assertExpectedNumberMachineIdsJvmIdsThreadIds(parameterAveragingWorkerFitTimesMs, 1, 1, nWorkers);
        assertNull(pStats.getNestedTrainingStats());
        //Finally: try exporting stats
        String tempDir = System.getProperty("java.io.tmpdir");
        String outDir = FilenameUtils.concat(tempDir, "dl4j_testTrainingStatsCollection");
        stats.exportStatFiles(outDir, sc.sc());
        String htmlPlotsPath = FilenameUtils.concat(outDir, "AnalysisPlots.html");
        StatsUtils.exportStatsAsHtml(stats, htmlPlotsPath, sc);
        ByteArrayOutputStream baos = new ByteArrayOutputStream();
        StatsUtils.exportStatsAsHTML(stats, baos);
        baos.close();
        byte[] bytes = baos.toByteArray();
        String str = new String(bytes, "UTF-8");
    //            System.out.println(str);
    } finally {
        sc.stop();
    }
}