Examples with EventStats org.deeplearning4j.spark.stats.EventStats used on opensource projects

Example 1 with EventStats

use of org.deeplearning4j.spark.stats.EventStats 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();
    }
}

Example 2 with EventStats

use of org.deeplearning4j.spark.stats.EventStats in project deeplearning4j by deeplearning4j.

the class TestTrainingStatsCollection method assertNonNullFields.

private static void assertNonNullFields(List<EventStats> array) {
    for (EventStats e : array) {
        assertNotNull(e.getMachineID());
        assertNotNull(e.getJvmID());
        assertNotNull(e.getDurationMs());
        assertFalse(e.getMachineID().isEmpty());
        assertFalse(e.getJvmID().isEmpty());
        assertTrue(e.getThreadID() > 0);
    }
}

Example 3 with EventStats

use of org.deeplearning4j.spark.stats.EventStats in project deeplearning4j by deeplearning4j.

the class StatsCalculationHelper method build.

public CommonSparkTrainingStats build(SparkTrainingStats masterSpecificStats) {
    List<EventStats> totalTime = new ArrayList<>();
    totalTime.add(new ExampleCountEventStats(methodStartTime, returnTime - methodStartTime, totalExampleCount));
    List<EventStats> initTime = new ArrayList<>();
    initTime.add(new BaseEventStats(initalModelBefore, initialModelAfter - initalModelBefore));
    return new CommonSparkTrainingStats.Builder().trainingMasterSpecificStats(masterSpecificStats).workerFlatMapTotalTimeMs(totalTime).workerFlatMapGetInitialModelTimeMs(initTime).workerFlatMapDataSetGetTimesMs(dataSetGetTimes).workerFlatMapProcessMiniBatchTimesMs(processMiniBatchTimes).build();
}

Example 4 with EventStats

use of org.deeplearning4j.spark.stats.EventStats in project deeplearning4j by deeplearning4j.

the class TestSparkMultiLayerParameterAveraging method testFitViaStringPathsSize1.

@Test
public void testFitViaStringPathsSize1() throws Exception {
    Path tempDir = Files.createTempDirectory("DL4J-testFitViaStringPathsSize1");
    File tempDirF = tempDir.toFile();
    tempDirF.deleteOnExit();
    int dataSetObjSize = 1;
    int batchSizePerExecutor = 25;
    int numSplits = 10;
    int averagingFrequency = 3;
    int totalExamples = numExecutors() * batchSizePerExecutor * numSplits * averagingFrequency;
    DataSetIterator iter = new MnistDataSetIterator(dataSetObjSize, totalExamples, false);
    int i = 0;
    while (iter.hasNext()) {
        File nextFile = new File(tempDirF, i + ".bin");
        DataSet ds = iter.next();
        ds.save(nextFile);
        i++;
    }
    System.out.println("Saved to: " + tempDirF.getAbsolutePath());
    MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().updater(Updater.RMSPROP).optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT).iterations(1).list().layer(0, new org.deeplearning4j.nn.conf.layers.DenseLayer.Builder().nIn(28 * 28).nOut(50).activation(Activation.TANH).build()).layer(1, new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(LossFunctions.LossFunction.MCXENT).nIn(50).nOut(10).activation(Activation.SOFTMAX).build()).pretrain(false).backprop(true).build();
    SparkDl4jMultiLayer sparkNet = new SparkDl4jMultiLayer(sc, conf, new ParameterAveragingTrainingMaster.Builder(numExecutors(), dataSetObjSize).workerPrefetchNumBatches(5).batchSizePerWorker(batchSizePerExecutor).averagingFrequency(averagingFrequency).repartionData(Repartition.Always).build());
    sparkNet.setCollectTrainingStats(true);
    //List files:
    Configuration config = new Configuration();
    FileSystem hdfs = FileSystem.get(tempDir.toUri(), config);
    RemoteIterator<LocatedFileStatus> fileIter = hdfs.listFiles(new org.apache.hadoop.fs.Path(tempDir.toString()), false);
    List<String> paths = new ArrayList<>();
    while (fileIter.hasNext()) {
        String path = fileIter.next().getPath().toString();
        paths.add(path);
    }
    INDArray paramsBefore = sparkNet.getNetwork().params().dup();
    JavaRDD<String> pathRdd = sc.parallelize(paths);
    sparkNet.fitPaths(pathRdd);
    INDArray paramsAfter = sparkNet.getNetwork().params().dup();
    assertNotEquals(paramsBefore, paramsAfter);
    Thread.sleep(2000);
    SparkTrainingStats stats = sparkNet.getSparkTrainingStats();
    //Expect
    System.out.println(stats.statsAsString());
    assertEquals(numSplits, stats.getValue("ParameterAveragingMasterRepartitionTimesMs").size());
    List<EventStats> list = stats.getValue("ParameterAveragingWorkerFitTimesMs");
    assertEquals(numSplits * numExecutors() * averagingFrequency, list.size());
    for (EventStats es : list) {
        ExampleCountEventStats e = (ExampleCountEventStats) es;
        assertTrue(batchSizePerExecutor * averagingFrequency - 10 >= e.getTotalExampleCount());
    }
    sparkNet.getTrainingMaster().deleteTempFiles(sc);
}

Example 5 with EventStats

use of org.deeplearning4j.spark.stats.EventStats in project deeplearning4j by deeplearning4j.

the class TestSparkMultiLayerParameterAveraging method testParameterAveragingMultipleExamplesPerDataSet.

@Test
public void testParameterAveragingMultipleExamplesPerDataSet() throws Exception {
    int dataSetObjSize = 5;
    int batchSizePerExecutor = 25;
    List<DataSet> list = new ArrayList<>();
    DataSetIterator iter = new MnistDataSetIterator(dataSetObjSize, 1000, false);
    while (iter.hasNext()) {
        list.add(iter.next());
    }
    MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().updater(Updater.RMSPROP).optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT).iterations(1).list().layer(0, new org.deeplearning4j.nn.conf.layers.DenseLayer.Builder().nIn(28 * 28).nOut(50).activation(Activation.TANH).build()).layer(1, new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(LossFunctions.LossFunction.MCXENT).nIn(50).nOut(10).activation(Activation.SOFTMAX).build()).pretrain(false).backprop(true).build();
    SparkDl4jMultiLayer sparkNet = new SparkDl4jMultiLayer(sc, conf, new ParameterAveragingTrainingMaster.Builder(numExecutors(), dataSetObjSize).batchSizePerWorker(batchSizePerExecutor).averagingFrequency(1).repartionData(Repartition.Always).build());
    sparkNet.setCollectTrainingStats(true);
    JavaRDD<DataSet> rdd = sc.parallelize(list);
    sparkNet.fit(rdd);
    SparkTrainingStats stats = sparkNet.getSparkTrainingStats();
    List<EventStats> mapPartitionStats = stats.getValue("ParameterAveragingMasterMapPartitionsTimesMs");
    //For an averaging frequency of 1
    int numSplits = list.size() * dataSetObjSize / (numExecutors() * batchSizePerExecutor);
    assertEquals(numSplits, mapPartitionStats.size());
    List<EventStats> workerFitStats = stats.getValue("ParameterAveragingWorkerFitTimesMs");
    for (EventStats e : workerFitStats) {
        ExampleCountEventStats eces = (ExampleCountEventStats) e;
        System.out.println(eces.getTotalExampleCount());
    }
    for (EventStats e : workerFitStats) {
        ExampleCountEventStats eces = (ExampleCountEventStats) e;
        assertEquals(batchSizePerExecutor, eces.getTotalExampleCount());
    }
}