Examples with ComputeGraph edu.iu.dsc.tws.api.compute.graph.ComputeGraph used on opensource projects

Example 46 with ComputeGraph

use of edu.iu.dsc.tws.api.compute.graph.ComputeGraph in project twister2 by DSC-SPIDAL.

the class SvmSgdAdvancedRunner method executeTrainingDataLoadingTaskGraph.

/**
 * This method loads the training data in a distributed mode
 * dataStreamerParallelism is the amount of parallelism used
 * in loaded the data in parallel.
 *
 * @return twister2 DataObject containing the training data
 */
public DataObject<Object> executeTrainingDataLoadingTaskGraph() {
    DataObject<Object> data = null;
    DataObjectSource sourceTask = new DataObjectSource(Context.TWISTER2_DIRECT_EDGE, this.svmJobParameters.getTrainingDataDir());
    DataObjectSink sinkTask = new DataObjectSink();
    trainingBuilder.addSource(Constants.SimpleGraphConfig.DATA_OBJECT_SOURCE, sourceTask, dataStreamerParallelism);
    ComputeConnection firstGraphComputeConnection = trainingBuilder.addCompute(Constants.SimpleGraphConfig.DATA_OBJECT_SINK, sinkTask, dataStreamerParallelism);
    firstGraphComputeConnection.direct(Constants.SimpleGraphConfig.DATA_OBJECT_SOURCE).viaEdge(Context.TWISTER2_DIRECT_EDGE).withDataType(MessageTypes.OBJECT);
    trainingBuilder.setMode(OperationMode.BATCH);
    ComputeGraph datapointsTaskGraph = trainingBuilder.build();
    datapointsTaskGraph.setGraphName("training-data-loading-graph");
    ExecutionPlan firstGraphExecutionPlan = taskExecutor.plan(datapointsTaskGraph);
    taskExecutor.execute(datapointsTaskGraph, firstGraphExecutionPlan);
    data = taskExecutor.getOutput(datapointsTaskGraph, firstGraphExecutionPlan, Constants.SimpleGraphConfig.DATA_OBJECT_SINK);
    if (data == null) {
        throw new NullPointerException("Something Went Wrong in Loading Training Data");
    } else {
        LOG.info("Training Data Total Partitions : " + data.getPartitions().length);
    }
    return data;
}

Example 47 with ComputeGraph

use of edu.iu.dsc.tws.api.compute.graph.ComputeGraph in project twister2 by DSC-SPIDAL.

the class SvmSgdAdvancedRunner method executeIterativeTrainingGraph.

/**
 * This method executes the iterative training graph
 * Training is done in parallel depending on the parallelism factor given
 * In this implementation the data loading parallelism and data computing or
 * training parallelism is same. It is the general model to keep them equal. But
 * you can increase the parallelism the way you want. But it is adviced to keep these
 * values equal. Dynamic parallelism in training is not yet tested fully in Twister2 Framework.
 *
 * @return Twister2 DataObject{@literal <double[]>} containing the reduced weight vector
 */
public DataObject<double[]> executeIterativeTrainingGraph() {
    DataObject<double[]> trainedWeight = null;
    dataStreamer = new InputDataStreamer(this.operationMode, svmJobParameters.isDummy(), this.binaryBatchModel);
    iterativeSVMCompute = new IterativeSVMCompute(this.binaryBatchModel, this.operationMode);
    svmReduce = new SVMReduce(this.operationMode);
    trainingBuilder.addSource(Constants.SimpleGraphConfig.DATASTREAMER_SOURCE, dataStreamer, dataStreamerParallelism);
    ComputeConnection svmComputeConnection = trainingBuilder.addCompute(Constants.SimpleGraphConfig.SVM_COMPUTE, iterativeSVMCompute, svmComputeParallelism);
    ComputeConnection svmReduceConnection = trainingBuilder.addCompute(Constants.SimpleGraphConfig.SVM_REDUCE, svmReduce, reduceParallelism);
    svmComputeConnection.direct(Constants.SimpleGraphConfig.DATASTREAMER_SOURCE).viaEdge(Constants.SimpleGraphConfig.DATA_EDGE).withDataType(MessageTypes.OBJECT);
    // svmReduceConnection
    // .reduce(Constants.SimpleGraphConfig.SVM_COMPUTE, Constants.SimpleGraphConfig.REDUCE_EDGE,
    // new ReduceAggregator(), DataType.OBJECT);
    svmReduceConnection.allreduce(Constants.SimpleGraphConfig.SVM_COMPUTE).viaEdge(Constants.SimpleGraphConfig.REDUCE_EDGE).withReductionFunction(new ReduceAggregator()).withDataType(MessageTypes.OBJECT);
    trainingBuilder.setMode(operationMode);
    ComputeGraph graph = trainingBuilder.build();
    graph.setGraphName("training-graph");
    ExecutionPlan plan = taskExecutor.plan(graph);
    IExecutor ex = taskExecutor.createExecution(graph, plan);
    // iteration is being decoupled from the computation task
    for (int i = 0; i < this.binaryBatchModel.getIterations(); i++) {
        taskExecutor.addInput(graph, plan, Constants.SimpleGraphConfig.DATASTREAMER_SOURCE, Constants.SimpleGraphConfig.INPUT_DATA, trainingData);
        taskExecutor.addInput(graph, plan, Constants.SimpleGraphConfig.DATASTREAMER_SOURCE, Constants.SimpleGraphConfig.INPUT_WEIGHT_VECTOR, inputWeightVector);
        inputWeightVector = taskExecutor.getOutput(graph, plan, Constants.SimpleGraphConfig.SVM_REDUCE);
        ex.execute();
    }
    ex.closeExecution();
    LOG.info("Task Graph Executed !!! ");
    if (workerId == 0) {
        trainedWeight = retrieveWeightVectorFromTaskGraph(graph, plan);
        this.trainedWeightVector = trainedWeight;
    }
    return trainedWeight;
}

Example 48 with ComputeGraph

use of edu.iu.dsc.tws.api.compute.graph.ComputeGraph in project twister2 by DSC-SPIDAL.

the class SvmSgdAdvancedRunner method executeWeightVectorLoadingTaskGraph.

/**
 * This method loads the training data in a distributed mode
 * dataStreamerParallelism is the amount of parallelism used
 * in loaded the data in parallel.
 *
 * @return twister2 DataObject containing the training data
 */
public DataObject<Object> executeWeightVectorLoadingTaskGraph() {
    DataObject<Object> data = null;
    DataObjectSource sourceTask = new DataObjectSource(Context.TWISTER2_DIRECT_EDGE, this.svmJobParameters.getWeightVectorDataDir());
    DataObjectSink sinkTask = new DataObjectSink();
    trainingBuilder.addSource(Constants.SimpleGraphConfig.DATA_OBJECT_SOURCE, sourceTask, dataStreamerParallelism);
    ComputeConnection firstGraphComputeConnection = trainingBuilder.addCompute(Constants.SimpleGraphConfig.DATA_OBJECT_SINK, sinkTask, dataStreamerParallelism);
    firstGraphComputeConnection.direct(Constants.SimpleGraphConfig.DATA_OBJECT_SOURCE).viaEdge(Context.TWISTER2_DIRECT_EDGE).withDataType(MessageTypes.OBJECT);
    trainingBuilder.setMode(OperationMode.BATCH);
    ComputeGraph datapointsTaskGraph = trainingBuilder.build();
    datapointsTaskGraph.setGraphName("weight-vector-loading-graph");
    ExecutionPlan firstGraphExecutionPlan = taskExecutor.plan(datapointsTaskGraph);
    taskExecutor.execute(datapointsTaskGraph, firstGraphExecutionPlan);
    data = taskExecutor.getOutput(datapointsTaskGraph, firstGraphExecutionPlan, Constants.SimpleGraphConfig.DATA_OBJECT_SINK);
    if (data == null) {
        throw new NullPointerException("Something Went Wrong in Loading Weight Vector");
    } else {
        LOG.info("Training Data Total Partitions : " + data.getPartitions().length);
    }
    return data;
}

Example 49 with ComputeGraph

use of edu.iu.dsc.tws.api.compute.graph.ComputeGraph in project twister2 by DSC-SPIDAL.

the class MultiComputeTasksGraphExample method execute.

@Override
public void execute() {
    LOG.log(Level.INFO, "Task worker starting: " + workerId);
    ComputeGraphBuilder builder = ComputeGraphBuilder.newBuilder(config);
    int parallel = Integer.parseInt((String) config.get(DataObjectConstants.PARALLELISM_VALUE));
    SourceTask sourceTask = new SourceTask();
    FirstComputeTask firstComputeTask = new FirstComputeTask();
    SecondComputeTask secondComputeTask = new SecondComputeTask();
    ReduceTask reduceTask = new ReduceTask();
    String dataDirectory = (String) config.get(DataObjectConstants.DINPUT_DIRECTORY) + workerId;
    String centroidDirectory = (String) config.get(DataObjectConstants.CINPUT_DIRECTORY) + workerId;
    int dimension = Integer.parseInt((String) config.get(DataObjectConstants.DIMENSIONS));
    int numFiles = Integer.parseInt((String) config.get(DataObjectConstants.NUMBER_OF_FILES));
    int dsize = Integer.parseInt((String) config.get(DataObjectConstants.DSIZE));
    int csize = Integer.parseInt((String) config.get(DataObjectConstants.CSIZE));
    String type = config.getStringValue(DataObjectConstants.FILE_TYPE);
    LOG.info("Input Values:" + dataDirectory + centroidDirectory + dimension + numFiles);
    KMeansUtils.generateDataPoints(config, dimension, numFiles, dsize, csize, dataDirectory, centroidDirectory, type);
    // Adding the user-defined constraints to the graph
    Map<String, String> sourceTaskConstraintsMap = new HashMap<>();
    // sourceTaskConstraintsMap.put(Context.TWISTER2_MAX_TASK_INSTANCES_PER_WORKER, "2");
    Map<String, String> computeTaskConstraintsMap = new HashMap<>();
    // computeTaskConstraintsMap.put(Context.TWISTER2_MAX_TASK_INSTANCES_PER_WORKER, "2");
    Map<String, String> sinkTaskConstraintsMap = new HashMap<>();
    // sinkTaskConstraintsMap.put(Context.TWISTER2_MAX_TASK_INSTANCES_PER_WORKER, "2");
    builder.addSource("source", sourceTask, parallel);
    ComputeConnection firstComputeConnection = builder.addCompute("firstcompute", firstComputeTask, parallel);
    ComputeConnection secondComputeConnection = builder.addCompute("secondcompute", secondComputeTask, parallel);
    ComputeConnection reduceConnection = builder.addCompute("compute", reduceTask, parallel);
    firstComputeConnection.direct("source").viaEdge("fdirect").withDataType(MessageTypes.OBJECT);
    secondComputeConnection.direct("source").viaEdge("sdirect").withDataType(MessageTypes.OBJECT);
    reduceConnection.allreduce("firstcompute").viaEdge("freduce").withReductionFunction(new Aggregator()).withDataType(MessageTypes.OBJECT).connect().allreduce("secondcompute").viaEdge("sreduce").withReductionFunction(new Aggregator()).withDataType(MessageTypes.OBJECT);
    builder.setMode(OperationMode.BATCH);
    // Adding graph and node level constraints
    // builder.addNodeConstraints("source", sourceTaskConstraintsMap);
    // builder.addNodeConstraints("firstcompute", computeTaskConstraintsMap);
    // builder.addNodeConstraints("secondcompute", computeTaskConstraintsMap);
    // builder.addNodeConstraints("sink", sinkTaskConstraintsMap);
    builder.addGraphConstraints(Context.TWISTER2_MAX_TASK_INSTANCES_PER_WORKER, "4");
    ComputeGraph graph = builder.build();
    LOG.info("%%% Graph Constraints:%%%" + graph.getGraphConstraints());
    ExecutionPlan plan = taskExecutor.plan(graph);
    taskExecutor.execute(graph, plan);
}

Example 50 with ComputeGraph

use of edu.iu.dsc.tws.api.compute.graph.ComputeGraph in project twister2 by DSC-SPIDAL.

the class SvmSgdAdvancedRunner method executeTrainingGraph.

/**
 * This method executes the training graph
 * Training is done in parallel depending on the parallelism factor given
 * In this implementation the data loading parallelism and data computing or
 * training parallelism is same. It is the general model to keep them equal. But
 * you can increase the parallelism the way you want. But it is adviced to keep these
 * values equal. Dynamic parallelism in training is not yet tested fully in Twister2 Framework.
 *
 * @return Twister2 DataObject{@literal <double[]>} containing the reduced weight vector
 */
public DataObject<double[]> executeTrainingGraph() {
    DataObject<double[]> trainedWeight = null;
    dataStreamer = new InputDataStreamer(this.operationMode, svmJobParameters.isDummy(), this.binaryBatchModel);
    svmCompute = new SVMCompute(this.binaryBatchModel, this.operationMode);
    svmReduce = new SVMReduce(this.operationMode);
    trainingBuilder.addSource(Constants.SimpleGraphConfig.DATASTREAMER_SOURCE, dataStreamer, dataStreamerParallelism);
    ComputeConnection svmComputeConnection = trainingBuilder.addCompute(Constants.SimpleGraphConfig.SVM_COMPUTE, svmCompute, svmComputeParallelism);
    ComputeConnection svmReduceConnection = trainingBuilder.addCompute(Constants.SimpleGraphConfig.SVM_REDUCE, svmReduce, reduceParallelism);
    svmComputeConnection.direct(Constants.SimpleGraphConfig.DATASTREAMER_SOURCE).viaEdge(Constants.SimpleGraphConfig.DATA_EDGE).withDataType(MessageTypes.OBJECT);
    // svmReduceConnection
    // .reduce(Constants.SimpleGraphConfig.SVM_COMPUTE, Constants.SimpleGraphConfig.REDUCE_EDGE,
    // new ReduceAggregator(), DataType.OBJECT);
    svmReduceConnection.allreduce(Constants.SimpleGraphConfig.SVM_COMPUTE).viaEdge(Constants.SimpleGraphConfig.REDUCE_EDGE).withReductionFunction(new ReduceAggregator()).withDataType(MessageTypes.OBJECT);
    trainingBuilder.setMode(operationMode);
    ComputeGraph graph = trainingBuilder.build();
    graph.setGraphName("training-graph");
    ExecutionPlan plan = taskExecutor.plan(graph);
    taskExecutor.addInput(graph, plan, Constants.SimpleGraphConfig.DATASTREAMER_SOURCE, Constants.SimpleGraphConfig.INPUT_DATA, trainingData);
    taskExecutor.addInput(graph, plan, Constants.SimpleGraphConfig.DATASTREAMER_SOURCE, Constants.SimpleGraphConfig.INPUT_WEIGHT_VECTOR, inputWeightVector);
    taskExecutor.execute(graph, plan);
    LOG.info("Task Graph Executed !!! ");
    if (workerId == 0) {
        trainedWeight = retrieveWeightVectorFromTaskGraph(graph, plan);
        this.trainedWeightVector = trainedWeight;
    }
    return trainedWeight;
}