Examples with WorkerConfiguration org.deeplearning4j.spark.api.WorkerConfiguration used on opensource projects

Example 1 with WorkerConfiguration

use of org.deeplearning4j.spark.api.WorkerConfiguration in project deeplearning4j by deeplearning4j.

the class ExecuteWorkerFlatMapAdapter method call.

@Override
public Iterable<R> call(Iterator<DataSet> dataSetIterator) throws Exception {
    WorkerConfiguration dataConfig = worker.getDataConfiguration();
    final boolean isGraph = dataConfig.isGraphNetwork();
    boolean stats = dataConfig.isCollectTrainingStats();
    StatsCalculationHelper s = (stats ? new StatsCalculationHelper() : null);
    if (stats)
        s.logMethodStartTime();
    if (!dataSetIterator.hasNext()) {
        if (stats) {
            s.logReturnTime();
            Pair<R, SparkTrainingStats> pair = worker.getFinalResultNoDataWithStats();
            pair.getFirst().setStats(s.build(pair.getSecond()));
            return Collections.singletonList(pair.getFirst());
        } else {
            return Collections.singletonList(worker.getFinalResultNoData());
        }
    }
    int batchSize = dataConfig.getBatchSizePerWorker();
    final int prefetchCount = dataConfig.getPrefetchNumBatches();
    DataSetIterator batchedIterator = new IteratorDataSetIterator(dataSetIterator, batchSize);
    if (prefetchCount > 0) {
        batchedIterator = new AsyncDataSetIterator(batchedIterator, prefetchCount);
    }
    try {
        MultiLayerNetwork net = null;
        ComputationGraph graph = null;
        if (stats)
            s.logInitialModelBefore();
        if (isGraph)
            graph = worker.getInitialModelGraph();
        else
            net = worker.getInitialModel();
        if (stats)
            s.logInitialModelAfter();
        int miniBatchCount = 0;
        int maxMinibatches = (dataConfig.getMaxBatchesPerWorker() > 0 ? dataConfig.getMaxBatchesPerWorker() : Integer.MAX_VALUE);
        while (batchedIterator.hasNext() && miniBatchCount++ < maxMinibatches) {
            if (stats)
                s.logNextDataSetBefore();
            DataSet next = batchedIterator.next();
            if (stats)
                s.logNextDataSetAfter(next.numExamples());
            if (stats) {
                s.logProcessMinibatchBefore();
                Pair<R, SparkTrainingStats> result;
                if (isGraph)
                    result = worker.processMinibatchWithStats(next, graph, !batchedIterator.hasNext());
                else
                    result = worker.processMinibatchWithStats(next, net, !batchedIterator.hasNext());
                s.logProcessMinibatchAfter();
                if (result != null) {
                    //Terminate training immediately
                    s.logReturnTime();
                    SparkTrainingStats workerStats = result.getSecond();
                    SparkTrainingStats returnStats = s.build(workerStats);
                    result.getFirst().setStats(returnStats);
                    return Collections.singletonList(result.getFirst());
                }
            } else {
                R result;
                if (isGraph)
                    result = worker.processMinibatch(next, graph, !batchedIterator.hasNext());
                else
                    result = worker.processMinibatch(next, net, !batchedIterator.hasNext());
                if (result != null) {
                    //Terminate training immediately
                    return Collections.singletonList(result);
                }
            }
        }
        //For some reason, we didn't return already. Normally this shouldn't happen
        if (stats) {
            s.logReturnTime();
            Pair<R, SparkTrainingStats> pair;
            if (isGraph)
                pair = worker.getFinalResultWithStats(graph);
            else
                pair = worker.getFinalResultWithStats(net);
            pair.getFirst().setStats(s.build(pair.getSecond()));
            return Collections.singletonList(pair.getFirst());
        } else {
            if (isGraph)
                return Collections.singletonList(worker.getFinalResult(graph));
            else
                return Collections.singletonList(worker.getFinalResult(net));
        }
    } finally {
        //Make sure we shut down the async thread properly...
        if (batchedIterator instanceof AsyncDataSetIterator) {
            ((AsyncDataSetIterator) batchedIterator).shutdown();
        }
        if (Nd4j.getExecutioner() instanceof GridExecutioner)
            ((GridExecutioner) Nd4j.getExecutioner()).flushQueueBlocking();
    }
}

Example 2 with WorkerConfiguration

use of org.deeplearning4j.spark.api.WorkerConfiguration in project deeplearning4j by deeplearning4j.

the class ExecuteWorkerMultiDataSetFlatMapAdapter method call.

@Override
public Iterable<R> call(Iterator<MultiDataSet> dataSetIterator) throws Exception {
    WorkerConfiguration dataConfig = worker.getDataConfiguration();
    boolean stats = dataConfig.isCollectTrainingStats();
    StatsCalculationHelper s = (stats ? new StatsCalculationHelper() : null);
    if (stats)
        s.logMethodStartTime();
    if (!dataSetIterator.hasNext()) {
        if (stats)
            s.logReturnTime();
        //Sometimes: no data
        return Collections.emptyList();
    }
    int batchSize = dataConfig.getBatchSizePerWorker();
    final int prefetchCount = dataConfig.getPrefetchNumBatches();
    MultiDataSetIterator batchedIterator = new IteratorMultiDataSetIterator(dataSetIterator, batchSize);
    if (prefetchCount > 0) {
        batchedIterator = new AsyncMultiDataSetIterator(batchedIterator, prefetchCount);
    }
    try {
        if (stats)
            s.logInitialModelBefore();
        ComputationGraph net = worker.getInitialModelGraph();
        if (stats)
            s.logInitialModelAfter();
        int miniBatchCount = 0;
        int maxMinibatches = (dataConfig.getMaxBatchesPerWorker() > 0 ? dataConfig.getMaxBatchesPerWorker() : Integer.MAX_VALUE);
        while (batchedIterator.hasNext() && miniBatchCount++ < maxMinibatches) {
            if (stats)
                s.logNextDataSetBefore();
            MultiDataSet next = batchedIterator.next();
            if (stats)
                s.logNextDataSetAfter(next.getFeatures(0).size(0));
            if (stats) {
                s.logProcessMinibatchBefore();
                Pair<R, SparkTrainingStats> result = worker.processMinibatchWithStats(next, net, !batchedIterator.hasNext());
                s.logProcessMinibatchAfter();
                if (result != null) {
                    //Terminate training immediately
                    s.logReturnTime();
                    SparkTrainingStats workerStats = result.getSecond();
                    SparkTrainingStats returnStats = s.build(workerStats);
                    result.getFirst().setStats(returnStats);
                    return Collections.singletonList(result.getFirst());
                }
            } else {
                R result = worker.processMinibatch(next, net, !batchedIterator.hasNext());
                if (result != null) {
                    //Terminate training immediately
                    return Collections.singletonList(result);
                }
            }
        }
        //For some reason, we didn't return already. Normally this shouldn't happen
        if (stats) {
            s.logReturnTime();
            Pair<R, SparkTrainingStats> pair = worker.getFinalResultWithStats(net);
            pair.getFirst().setStats(s.build(pair.getSecond()));
            return Collections.singletonList(pair.getFirst());
        } else {
            return Collections.singletonList(worker.getFinalResult(net));
        }
    } finally {
        //Make sure we shut down the async thread properly...
        if (batchedIterator instanceof AsyncMultiDataSetIterator) {
            ((AsyncMultiDataSetIterator) batchedIterator).shutdown();
        }
        if (Nd4j.getExecutioner() instanceof GridExecutioner)
            ((GridExecutioner) Nd4j.getExecutioner()).flushQueueBlocking();
    }
}