Search in sources :

Example 1 with GuaguaParquetMapReduceClient

use of ml.shifu.shifu.guagua.GuaguaParquetMapReduceClient in project shifu by ShifuML.

the class TrainModelProcessor method runDistributedTrain.

protected int runDistributedTrain() throws IOException, InterruptedException, ClassNotFoundException {
    LOG.info("Started {}distributed training.", isDryTrain ? "dry " : "");
    int status = 0;
    Configuration conf = new Configuration();
    SourceType sourceType = super.getModelConfig().getDataSet().getSource();
    final List<String> args = new ArrayList<String>();
    GridSearch gs = new GridSearch(modelConfig.getTrain().getParams(), modelConfig.getTrain().getGridConfigFileContent());
    prepareCommonParams(gs.hasHyperParam(), args, sourceType);
    String alg = super.getModelConfig().getTrain().getAlgorithm();
    // add tmp models folder to config
    FileSystem fileSystem = ShifuFileUtils.getFileSystemBySourceType(sourceType);
    Path tmpModelsPath = fileSystem.makeQualified(new Path(super.getPathFinder().getPathBySourceType(new Path(Constants.TMP, Constants.DEFAULT_MODELS_TMP_FOLDER), sourceType)));
    args.add(String.format(CommonConstants.MAPREDUCE_PARAM_FORMAT, CommonConstants.SHIFU_TMP_MODELS_FOLDER, tmpModelsPath.toString()));
    int baggingNum = isForVarSelect ? 1 : super.getModelConfig().getBaggingNum();
    if (modelConfig.isClassification()) {
        int classes = modelConfig.getTags().size();
        if (classes == 2) {
            // binary classification, only need one job
            baggingNum = 1;
        } else {
            if (modelConfig.getTrain().isOneVsAll()) {
                // one vs all multiple classification, we need multiple bagging jobs to do ONEVSALL
                baggingNum = modelConfig.getTags().size();
            } else {
            // native classification, using bagging from setting job, no need set here
            }
        }
        if (baggingNum != super.getModelConfig().getBaggingNum()) {
            LOG.warn("'train:baggingNum' is set to {} because of ONEVSALL multiple classification.", baggingNum);
        }
    }
    boolean isKFoldCV = false;
    Integer kCrossValidation = this.modelConfig.getTrain().getNumKFold();
    if (kCrossValidation != null && kCrossValidation > 0) {
        isKFoldCV = true;
        baggingNum = modelConfig.getTrain().getNumKFold();
        if (baggingNum != super.getModelConfig().getBaggingNum() && gs.hasHyperParam()) {
            // if it is grid search mode, then kfold mode is disabled
            LOG.warn("'train:baggingNum' is set to {} because of k-fold cross validation is enabled by 'numKFold' not -1.", baggingNum);
        }
    }
    long start = System.currentTimeMillis();
    boolean isParallel = Boolean.valueOf(Environment.getProperty(Constants.SHIFU_DTRAIN_PARALLEL, SHIFU_DEFAULT_DTRAIN_PARALLEL)).booleanValue();
    GuaguaMapReduceClient guaguaClient;
    int[] inputOutputIndex = DTrainUtils.getInputOutputCandidateCounts(modelConfig.getNormalizeType(), this.columnConfigList);
    int inputNodeCount = inputOutputIndex[0] == 0 ? inputOutputIndex[2] : inputOutputIndex[0];
    int candidateCount = inputOutputIndex[2];
    boolean isAfterVarSelect = (inputOutputIndex[0] != 0);
    // cache all feature list for sampling features
    List<Integer> allFeatures = NormalUtils.getAllFeatureList(this.columnConfigList, isAfterVarSelect);
    if (modelConfig.getNormalize().getIsParquet()) {
        guaguaClient = new GuaguaParquetMapReduceClient();
        // set required field list to make sure we only load selected columns.
        RequiredFieldList requiredFieldList = new RequiredFieldList();
        boolean hasCandidates = CommonUtils.hasCandidateColumns(columnConfigList);
        for (ColumnConfig columnConfig : super.columnConfigList) {
            if (columnConfig.isTarget()) {
                requiredFieldList.add(new RequiredField(columnConfig.getColumnName(), columnConfig.getColumnNum(), null, DataType.FLOAT));
            } else {
                if (inputNodeCount == candidateCount) {
                    // no any variables are selected
                    if (!columnConfig.isMeta() && !columnConfig.isTarget() && CommonUtils.isGoodCandidate(columnConfig, hasCandidates)) {
                        requiredFieldList.add(new RequiredField(columnConfig.getColumnName(), columnConfig.getColumnNum(), null, DataType.FLOAT));
                    }
                } else {
                    if (!columnConfig.isMeta() && !columnConfig.isTarget() && columnConfig.isFinalSelect()) {
                        requiredFieldList.add(new RequiredField(columnConfig.getColumnName(), columnConfig.getColumnNum(), null, DataType.FLOAT));
                    }
                }
            }
        }
        // weight is added manually
        requiredFieldList.add(new RequiredField("weight", columnConfigList.size(), null, DataType.DOUBLE));
        args.add(String.format(CommonConstants.MAPREDUCE_PARAM_FORMAT, "parquet.private.pig.required.fields", serializeRequiredFieldList(requiredFieldList)));
        args.add(String.format(CommonConstants.MAPREDUCE_PARAM_FORMAT, "parquet.private.pig.column.index.access", "true"));
    } else {
        guaguaClient = new GuaguaMapReduceClient();
    }
    int parallelNum = Integer.parseInt(Environment.getProperty(CommonConstants.SHIFU_TRAIN_BAGGING_INPARALLEL, "5"));
    int parallelGroups = 1;
    if (gs.hasHyperParam()) {
        parallelGroups = (gs.getFlattenParams().size() % parallelNum == 0 ? gs.getFlattenParams().size() / parallelNum : gs.getFlattenParams().size() / parallelNum + 1);
        baggingNum = gs.getFlattenParams().size();
        LOG.warn("'train:baggingNum' is set to {} because of grid search enabled by settings in 'train#params'.", gs.getFlattenParams().size());
    } else {
        parallelGroups = baggingNum % parallelNum == 0 ? baggingNum / parallelNum : baggingNum / parallelNum + 1;
    }
    LOG.info("Distributed trainning with baggingNum: {}", baggingNum);
    List<String> progressLogList = new ArrayList<String>(baggingNum);
    boolean isOneJobNotContinuous = false;
    for (int j = 0; j < parallelGroups; j++) {
        int currBags = baggingNum;
        if (gs.hasHyperParam()) {
            if (j == parallelGroups - 1) {
                currBags = gs.getFlattenParams().size() % parallelNum == 0 ? parallelNum : gs.getFlattenParams().size() % parallelNum;
            } else {
                currBags = parallelNum;
            }
        } else {
            if (j == parallelGroups - 1) {
                currBags = baggingNum % parallelNum == 0 ? parallelNum : baggingNum % parallelNum;
            } else {
                currBags = parallelNum;
            }
        }
        for (int k = 0; k < currBags; k++) {
            int i = j * parallelNum + k;
            if (gs.hasHyperParam()) {
                LOG.info("Start the {}th grid search job with params: {}", i, gs.getParams(i));
            } else if (isKFoldCV) {
                LOG.info("Start the {}th k-fold cross validation job with params.", i);
            }
            List<String> localArgs = new ArrayList<String>(args);
            // set name for each bagging job.
            localArgs.add("-n");
            localArgs.add(String.format("Shifu Master-Workers %s Training Iteration: %s id:%s", alg, super.getModelConfig().getModelSetName(), i));
            LOG.info("Start trainer with id: {}", i);
            String modelName = getModelName(i);
            Path modelPath = fileSystem.makeQualified(new Path(super.getPathFinder().getModelsPath(sourceType), modelName));
            Path bModelPath = fileSystem.makeQualified(new Path(super.getPathFinder().getNNBinaryModelsPath(sourceType), modelName));
            // check if job is continuous training, this can be set multiple times and we only get last one
            boolean isContinuous = false;
            if (gs.hasHyperParam()) {
                isContinuous = false;
            } else {
                int intContinuous = checkContinuousTraining(fileSystem, localArgs, modelPath, modelConfig.getTrain().getParams());
                if (intContinuous == -1) {
                    LOG.warn("Model with index {} with size of trees is over treeNum, such training will not be started.", i);
                    continue;
                } else {
                    isContinuous = (intContinuous == 1);
                }
            }
            // training
            if (gs.hasHyperParam() || isKFoldCV) {
                isContinuous = false;
            }
            if (!isContinuous && !isOneJobNotContinuous) {
                isOneJobNotContinuous = true;
                // delete all old models if not continuous
                String srcModelPath = super.getPathFinder().getModelsPath(sourceType);
                String mvModelPath = srcModelPath + "_" + System.currentTimeMillis();
                LOG.info("Old model path has been moved to {}", mvModelPath);
                fileSystem.rename(new Path(srcModelPath), new Path(mvModelPath));
                fileSystem.mkdirs(new Path(srcModelPath));
                FileSystem.getLocal(conf).delete(new Path(super.getPathFinder().getModelsPath(SourceType.LOCAL)), true);
            }
            if (NNConstants.NN_ALG_NAME.equalsIgnoreCase(alg)) {
                // tree related parameters initialization
                Map<String, Object> params = gs.hasHyperParam() ? gs.getParams(i) : this.modelConfig.getTrain().getParams();
                Object fssObj = params.get("FeatureSubsetStrategy");
                FeatureSubsetStrategy featureSubsetStrategy = null;
                double featureSubsetRate = 0d;
                if (fssObj != null) {
                    try {
                        featureSubsetRate = Double.parseDouble(fssObj.toString());
                        // no need validate featureSubsetRate is in (0,1], as already validated in ModelInspector
                        featureSubsetStrategy = null;
                    } catch (NumberFormatException ee) {
                        featureSubsetStrategy = FeatureSubsetStrategy.of(fssObj.toString());
                    }
                } else {
                    LOG.warn("FeatureSubsetStrategy is not set, set to ALL by default.");
                    featureSubsetStrategy = FeatureSubsetStrategy.ALL;
                    featureSubsetRate = 0;
                }
                Set<Integer> subFeatures = null;
                if (isContinuous) {
                    BasicFloatNetwork existingModel = (BasicFloatNetwork) ModelSpecLoaderUtils.getBasicNetwork(ModelSpecLoaderUtils.loadModel(modelConfig, modelPath, ShifuFileUtils.getFileSystemBySourceType(this.modelConfig.getDataSet().getSource())));
                    if (existingModel == null) {
                        subFeatures = new HashSet<Integer>(getSubsamplingFeatures(allFeatures, featureSubsetStrategy, featureSubsetRate, inputNodeCount));
                    } else {
                        subFeatures = existingModel.getFeatureSet();
                    }
                } else {
                    subFeatures = new HashSet<Integer>(getSubsamplingFeatures(allFeatures, featureSubsetStrategy, featureSubsetRate, inputNodeCount));
                }
                if (subFeatures == null || subFeatures.size() == 0) {
                    localArgs.add(String.format(CommonConstants.MAPREDUCE_PARAM_FORMAT, CommonConstants.SHIFU_NN_FEATURE_SUBSET, ""));
                } else {
                    localArgs.add(String.format(CommonConstants.MAPREDUCE_PARAM_FORMAT, CommonConstants.SHIFU_NN_FEATURE_SUBSET, StringUtils.join(subFeatures, ',')));
                    LOG.debug("Size: {}, list: {}.", subFeatures.size(), StringUtils.join(subFeatures, ','));
                }
            }
            localArgs.add(String.format(CommonConstants.MAPREDUCE_PARAM_FORMAT, CommonConstants.GUAGUA_OUTPUT, modelPath.toString()));
            localArgs.add(String.format(CommonConstants.MAPREDUCE_PARAM_FORMAT, Constants.SHIFU_NN_BINARY_MODEL_PATH, bModelPath.toString()));
            if (gs.hasHyperParam() || isKFoldCV) {
                // k-fold cv need val error
                Path valErrPath = fileSystem.makeQualified(new Path(super.getPathFinder().getValErrorPath(sourceType), "val_error_" + i));
                localArgs.add(String.format(CommonConstants.MAPREDUCE_PARAM_FORMAT, CommonConstants.GS_VALIDATION_ERROR, valErrPath.toString()));
            }
            localArgs.add(String.format(CommonConstants.MAPREDUCE_PARAM_FORMAT, CommonConstants.SHIFU_TRAINER_ID, String.valueOf(i)));
            final String progressLogFile = getProgressLogFile(i);
            progressLogList.add(progressLogFile);
            localArgs.add(String.format(CommonConstants.MAPREDUCE_PARAM_FORMAT, CommonConstants.SHIFU_DTRAIN_PROGRESS_FILE, progressLogFile));
            String hdpVersion = HDPUtils.getHdpVersionForHDP224();
            if (StringUtils.isNotBlank(hdpVersion)) {
                localArgs.add(String.format(CommonConstants.MAPREDUCE_PARAM_FORMAT, "hdp.version", hdpVersion));
                HDPUtils.addFileToClassPath(HDPUtils.findContainingFile("hdfs-site.xml"), conf);
                HDPUtils.addFileToClassPath(HDPUtils.findContainingFile("core-site.xml"), conf);
                HDPUtils.addFileToClassPath(HDPUtils.findContainingFile("mapred-site.xml"), conf);
                HDPUtils.addFileToClassPath(HDPUtils.findContainingFile("yarn-site.xml"), conf);
            }
            if (isParallel) {
                guaguaClient.addJob(localArgs.toArray(new String[0]));
            } else {
                TailThread tailThread = startTailThread(new String[] { progressLogFile });
                boolean ret = guaguaClient.createJob(localArgs.toArray(new String[0])).waitForCompletion(true);
                status += (ret ? 0 : 1);
                stopTailThread(tailThread);
            }
        }
        if (isParallel) {
            TailThread tailThread = startTailThread(progressLogList.toArray(new String[0]));
            status += guaguaClient.run();
            stopTailThread(tailThread);
        }
    }
    if (isKFoldCV) {
        // k-fold we also copy model files at last, such models can be used for evaluation
        for (int i = 0; i < baggingNum; i++) {
            String modelName = getModelName(i);
            Path modelPath = fileSystem.makeQualified(new Path(super.getPathFinder().getModelsPath(sourceType), modelName));
            if (ShifuFileUtils.getFileSystemBySourceType(sourceType).exists(modelPath)) {
                copyModelToLocal(modelName, modelPath, sourceType);
            } else {
                LOG.warn("Model {} isn't there, maybe job is failed, for bagging it can be ignored.", modelPath.toString());
                status += 1;
            }
        }
        List<Double> valErrs = readAllValidationErrors(sourceType, fileSystem, kCrossValidation);
        double sum = 0d;
        for (Double err : valErrs) {
            sum += err;
        }
        LOG.info("Average validation error for current k-fold cross validation is {}.", sum / valErrs.size());
        LOG.info("K-fold cross validation on distributed training finished in {}ms.", System.currentTimeMillis() - start);
    } else if (gs.hasHyperParam()) {
        // select the best parameter composite in grid search
        LOG.info("Original grid search params: {}", modelConfig.getParams());
        Map<String, Object> params = findBestParams(sourceType, fileSystem, gs);
        // temp copy all models for evaluation
        for (int i = 0; i < baggingNum; i++) {
            String modelName = getModelName(i);
            Path modelPath = fileSystem.makeQualified(new Path(super.getPathFinder().getModelsPath(sourceType), modelName));
            if (ShifuFileUtils.getFileSystemBySourceType(sourceType).exists(modelPath) && (status == 0)) {
                copyModelToLocal(modelName, modelPath, sourceType);
            } else {
                LOG.warn("Model {} isn't there, maybe job is failed, for bagging it can be ignored.", modelPath.toString());
            }
        }
        LOG.info("The best parameters in grid search is {}", params);
        LOG.info("Grid search on distributed training finished in {}ms.", System.currentTimeMillis() - start);
    } else {
        // copy model files at last.
        for (int i = 0; i < baggingNum; i++) {
            String modelName = getModelName(i);
            Path modelPath = fileSystem.makeQualified(new Path(super.getPathFinder().getModelsPath(sourceType), modelName));
            if (ShifuFileUtils.getFileSystemBySourceType(sourceType).exists(modelPath) && (status == 0)) {
                copyModelToLocal(modelName, modelPath, sourceType);
            } else {
                LOG.warn("Model {} isn't there, maybe job is failed, for bagging it can be ignored.", modelPath.toString());
            }
        }
        // copy temp model files, for RF/GBT, not to copy tmp models because of larger space needed, for others
        // by default copy tmp models to local
        boolean copyTmpModelsToLocal = Boolean.TRUE.toString().equalsIgnoreCase(Environment.getProperty(Constants.SHIFU_TMPMODEL_COPYTOLOCAL, "true"));
        if (copyTmpModelsToLocal) {
            copyTmpModelsToLocal(tmpModelsPath, sourceType);
        } else {
            LOG.info("Tmp models are not copied into local, please find them in hdfs path: {}", tmpModelsPath);
        }
        LOG.info("Distributed training finished in {}ms.", System.currentTimeMillis() - start);
    }
    if (CommonUtils.isTreeModel(modelConfig.getAlgorithm())) {
        List<BasicML> models = ModelSpecLoaderUtils.loadBasicModels(this.modelConfig, null);
        // compute feature importance and write to local file after models are trained
        Map<Integer, MutablePair<String, Double>> featureImportances = CommonUtils.computeTreeModelFeatureImportance(models);
        String localFsFolder = pathFinder.getLocalFeatureImportanceFolder();
        String localFIPath = pathFinder.getLocalFeatureImportancePath();
        processRollupForFIFiles(localFsFolder, localFIPath);
        CommonUtils.writeFeatureImportance(localFIPath, featureImportances);
    }
    if (status != 0) {
        LOG.error("Error may occurred. There is no model generated. Please check!");
    }
    return status;
}
Also used : Configuration(org.apache.hadoop.conf.Configuration) ColumnConfig(ml.shifu.shifu.container.obj.ColumnConfig) SourceType(ml.shifu.shifu.container.obj.RawSourceData.SourceType) FeatureSubsetStrategy(ml.shifu.shifu.core.dtrain.FeatureSubsetStrategy) BasicML(org.encog.ml.BasicML) GuaguaMapReduceClient(ml.shifu.guagua.mapreduce.GuaguaMapReduceClient) MutablePair(org.apache.commons.lang3.tuple.MutablePair) RequiredField(org.apache.pig.LoadPushDown.RequiredField) FileSystem(org.apache.hadoop.fs.FileSystem) BasicFloatNetwork(ml.shifu.shifu.core.dtrain.dataset.BasicFloatNetwork) Path(org.apache.hadoop.fs.Path) GuaguaParquetMapReduceClient(ml.shifu.shifu.guagua.GuaguaParquetMapReduceClient) GridSearch(ml.shifu.shifu.core.dtrain.gs.GridSearch) RequiredFieldList(org.apache.pig.LoadPushDown.RequiredFieldList)

Aggregations

GuaguaMapReduceClient (ml.shifu.guagua.mapreduce.GuaguaMapReduceClient)1 ColumnConfig (ml.shifu.shifu.container.obj.ColumnConfig)1 SourceType (ml.shifu.shifu.container.obj.RawSourceData.SourceType)1 FeatureSubsetStrategy (ml.shifu.shifu.core.dtrain.FeatureSubsetStrategy)1 BasicFloatNetwork (ml.shifu.shifu.core.dtrain.dataset.BasicFloatNetwork)1 GridSearch (ml.shifu.shifu.core.dtrain.gs.GridSearch)1 GuaguaParquetMapReduceClient (ml.shifu.shifu.guagua.GuaguaParquetMapReduceClient)1 MutablePair (org.apache.commons.lang3.tuple.MutablePair)1 Configuration (org.apache.hadoop.conf.Configuration)1 FileSystem (org.apache.hadoop.fs.FileSystem)1 Path (org.apache.hadoop.fs.Path)1 RequiredField (org.apache.pig.LoadPushDown.RequiredField)1 RequiredFieldList (org.apache.pig.LoadPushDown.RequiredFieldList)1 BasicML (org.encog.ml.BasicML)1