Examples with GBM hex.tree.gbm.GBM used on opensource projects

Example 6 with GBM

use of hex.tree.gbm.GBM in project h2o-3 by h2oai.

the class WorkFlowTest method testWorkFlow.

// End-to-end workflow test:
// 1- load set of files, train, test, holdout
// 2- light data munging
// 3- build model on train; using test as validation
// 4- score on holdout set
//
// If files are missing, silently fail - as the files are big and this is not
// yet a junit test
private void testWorkFlow(String[] files) {
    try {
        Scope.enter();
        // 1- Load datasets
        Frame data = load_files("data.hex", files);
        if (data == null)
            return;
        // -------------------------------------------------
        // 2- light data munging
        // Convert start time to: Day since the Epoch
        Vec startime = data.vec("starttime");
        data.add(new TimeSplit().doIt(startime));
        // Now do a monster Group-By.  Count bike starts per-station per-day
        Vec days = data.vec("Days");
        long start = System.currentTimeMillis();
        Frame bph = new CountBikes(days).doAll(days, data.vec("start station name")).makeFrame(Key.make("bph.hex"));
        System.out.println("Groupby took " + (System.currentTimeMillis() - start));
        System.out.println(bph);
        System.out.println(bph.toString(10000, 20));
        data.remove();
        QuantileModel.QuantileParameters quantile_parms = new QuantileModel.QuantileParameters();
        quantile_parms._train = bph._key;
        Job<QuantileModel> job2 = new Quantile(quantile_parms).trainModel();
        QuantileModel quantile = job2.get();
        job2.remove();
        System.out.println(Arrays.deepToString(quantile._output._quantiles));
        quantile.remove();
        // Split into train, test and holdout sets
        Key[] keys = new Key[] { Key.make("train.hex"), Key.make("test.hex"), Key.make("hold.hex") };
        double[] ratios = new double[] { 0.6, 0.3, 0.1 };
        Frame[] frs = ShuffleSplitFrame.shuffleSplitFrame(bph, keys, ratios, 1234567689L);
        Frame train = frs[0];
        Frame test = frs[1];
        Frame hold = frs[2];
        bph.remove();
        System.out.println(train);
        System.out.println(test);
        // -------------------------------------------------
        // 3- build model on train; using test as validation
        // ---
        // Gradient Boosting Machine
        GBMModel.GBMParameters gbm_parms = new GBMModel.GBMParameters();
        // base Model.Parameters
        gbm_parms._train = train._key;
        gbm_parms._valid = test._key;
        // default is false
        gbm_parms._score_each_iteration = false;
        // SupervisedModel.Parameters
        gbm_parms._response_column = "bikes";
        // SharedTreeModel.Parameters
        // default is 50, 1000 is 0.90, 10000 is 0.91
        gbm_parms._ntrees = 500;
        // default is 5
        gbm_parms._max_depth = 6;
        // default
        gbm_parms._min_rows = 10;
        // default
        gbm_parms._nbins = 20;
        // GBMModel.Parameters
        // default
        gbm_parms._distribution = DistributionFamily.gaussian;
        // default
        gbm_parms._learn_rate = 0.1f;
        // Train model; block for results
        Job<GBMModel> job = new GBM(gbm_parms).trainModel();
        GBMModel gbm = job.get();
        job.remove();
        // ---
        // Build a GLM model also
        GLMModel.GLMParameters glm_parms = new GLMModel.GLMParameters(GLMModel.GLMParameters.Family.gaussian);
        // base Model.Parameters
        glm_parms._train = train._key;
        glm_parms._valid = test._key;
        // default is false
        glm_parms._score_each_iteration = false;
        // SupervisedModel.Parameters
        glm_parms._response_column = "bikes";
        // GLMModel.Parameters
        glm_parms._use_all_factor_levels = true;
        // Train model; block for results
        Job<GLMModel> glm_job = new GLM(glm_parms).trainModel();
        GLMModel glm = glm_job.get();
        glm_job.remove();
        // -------------------------------------------------
        // 4- Score on holdout set & report
        gbm.score(train).remove();
        glm.score(train).remove();
        // Cleanup
        train.remove();
        test.remove();
        hold.remove();
    } finally {
        Scope.exit();
    }
}

Example 7 with GBM

use of hex.tree.gbm.GBM in project h2o-3 by h2oai.

the class XValPredictionsCheck method testXValPredictions.

@Test
public void testXValPredictions() {
    final int nfolds = 3;
    Frame tfr = null;
    try {
        // Load data, hack frames
        tfr = parse_test_file("smalldata/iris/iris_wheader.csv");
        Frame foldId = new Frame(new String[] { "foldId" }, new Vec[] { AstKFold.kfoldColumn(tfr.vec("class").makeZero(), nfolds, 543216789) });
        tfr.add(foldId);
        DKV.put(tfr);
        // GBM
        GBMModel.GBMParameters parms = new GBMModel.GBMParameters();
        parms._train = tfr._key;
        parms._response_column = "class";
        parms._ntrees = 1;
        parms._max_depth = 1;
        parms._fold_column = "foldId";
        parms._distribution = DistributionFamily.multinomial;
        parms._keep_cross_validation_predictions = true;
        GBM job = new GBM(parms);
        GBMModel gbm = job.trainModel().get();
        checkModel(gbm, foldId.anyVec(), 3);
        // DRF
        DRFModel.DRFParameters parmsDRF = new DRFModel.DRFParameters();
        parmsDRF._train = tfr._key;
        parmsDRF._response_column = "class";
        parmsDRF._ntrees = 1;
        parmsDRF._max_depth = 1;
        parmsDRF._fold_column = "foldId";
        parmsDRF._distribution = DistributionFamily.multinomial;
        parmsDRF._keep_cross_validation_predictions = true;
        DRF drfJob = new DRF(parmsDRF);
        DRFModel drf = drfJob.trainModel().get();
        checkModel(drf, foldId.anyVec(), 3);
        // GLM
        GLMModel.GLMParameters parmsGLM = new GLMModel.GLMParameters();
        parmsGLM._train = tfr._key;
        parmsGLM._response_column = "sepal_len";
        parmsGLM._fold_column = "foldId";
        parmsGLM._keep_cross_validation_predictions = true;
        GLM glmJob = new GLM(parmsGLM);
        GLMModel glm = glmJob.trainModel().get();
        checkModel(glm, foldId.anyVec(), 1);
        // DL
        DeepLearningModel.DeepLearningParameters parmsDL = new DeepLearningModel.DeepLearningParameters();
        parmsDL._train = tfr._key;
        parmsDL._response_column = "class";
        parmsDL._hidden = new int[] { 1 };
        parmsDL._epochs = 1;
        parmsDL._fold_column = "foldId";
        parmsDL._keep_cross_validation_predictions = true;
        DeepLearning dlJob = new DeepLearning(parmsDL);
        DeepLearningModel dl = dlJob.trainModel().get();
        checkModel(dl, foldId.anyVec(), 3);
    } finally {
        if (tfr != null)
            tfr.remove();
    }
}

Example 8 with GBM

use of hex.tree.gbm.GBM in project h2o-3 by h2oai.

the class PartialDependenceTest method weatherBinary.

@Test
public void weatherBinary() {
    Frame fr = null;
    GBMModel model = null;
    PartialDependence partialDependence = null;
    try {
        // Frame
        fr = parse_test_file("smalldata/junit/weather.csv");
        // Model
        GBMModel.GBMParameters parms = new GBMModel.GBMParameters();
        parms._train = fr._key;
        parms._ignored_columns = new String[] { "Date", "RISK_MM", "EvapMM" };
        parms._response_column = "RainTomorrow";
        model = new GBM(parms).trainModel().get();
        // PartialDependence
        partialDependence = new PartialDependence(Key.<PartialDependence>make());
        partialDependence._nbins = 33;
        partialDependence._cols = new String[] { "Sunshine", "MaxWindPeriod", "WindSpeed9am" };
        partialDependence._model_id = (Key) model._key;
        partialDependence._frame_id = fr._key;
        partialDependence.execImpl().get();
        for (TwoDimTable t : partialDependence._partial_dependence_data) Log.info(t);
    } finally {
        if (fr != null)
            fr.remove();
        if (model != null)
            model.remove();
        if (partialDependence != null)
            partialDependence.remove();
    }
}

Example 9 with GBM

use of hex.tree.gbm.GBM in project h2o-3 by h2oai.

the class PartialDependenceTest method prostateBinaryPickCols.

@Test
public void prostateBinaryPickCols() {
    Frame fr = null;
    GBMModel model = null;
    PartialDependence partialDependence = null;
    try {
        // Frame
        fr = parse_test_file("smalldata/prostate/prostate.csv");
        for (String s : new String[] { "RACE", "GLEASON", "DPROS", "DCAPS", "CAPSULE" }) {
            Vec v = fr.remove(s);
            fr.add(s, v.toCategoricalVec());
            v.remove();
        }
        DKV.put(fr);
        // Model
        GBMModel.GBMParameters parms = new GBMModel.GBMParameters();
        parms._train = fr._key;
        parms._ignored_columns = new String[] { "ID" };
        parms._response_column = "CAPSULE";
        model = new GBM(parms).trainModel().get();
        // PartialDependence
        partialDependence = new PartialDependence(Key.<PartialDependence>make());
        //pick columns manually
        partialDependence._cols = new String[] { "DPROS", "GLEASON" };
        partialDependence._nbins = 10;
        partialDependence._model_id = (Key) model._key;
        partialDependence._frame_id = fr._key;
        partialDependence.execImpl().get();
        for (TwoDimTable t : partialDependence._partial_dependence_data) Log.info(t);
        Assert.assertTrue(partialDependence._partial_dependence_data.length == 2);
    } finally {
        if (fr != null)
            fr.remove();
        if (model != null)
            model.remove();
        if (partialDependence != null)
            partialDependence.remove();
    }
}