Java Frame.add方法代码示例

import water.fvec.Frame; //导入方法依赖的package包/类
static Vec unifyFrame(DRFModel.DRFParameters drf, Frame fr, PrepData prep, boolean classification) {
  int idx = prep.prep(fr);
  if( idx < 0 ) { idx = ~idx; }
  String rname = fr._names[idx];
  drf._response_column = fr.names()[idx];

  Vec resp = fr.vecs()[idx];
  Vec ret = null;
  if (classification) {
    ret = fr.remove(idx);
    fr.add(rname,resp.toCategoricalVec());
  } else {
    fr.remove(idx);
    fr.add(rname,resp);
  }
  return ret;
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
static Vec unifyFrame(DeepLearningParameters drf, Frame fr, PrepData prep, boolean classification) {
  int idx = prep.prep(fr);
  if( idx < 0 ) { idx = ~idx; }
  String rname = fr._names[idx];
  drf._response_column = fr.names()[idx];

  Vec resp = fr.vecs()[idx];
  Vec ret = null;
  if (classification) {
    ret = fr.remove(idx);
    fr.add(rname,resp.toCategoricalVec());
  } else {
    fr.remove(idx);
    fr.add(rname,resp);
  }
  return ret;
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
@Override protected Frame transformImpl(Frame f) {
  ((ASTExec)_ast._asts[1])._asts[1] = AST.newASTFrame(f);
  Frame fr = Exec.execute(_ast).getFrame();
  _newCol = _newNames==null?new String[fr.numCols()]:_newNames;
  _newColTypes = toJavaPrimitive(fr.anyVec().get_type_str());
  if( (_multiColReturn=fr.numCols() > 1) ) {
    for(int i=0;i<_newCol.length;i++) {
      if(_newNames==null) _newCol[i] = f.uniquify(i > 0 ? _newCol[i - 1] : _oldCol);
      f.add(_newCol[i], fr.vec(i));
    }
    if( _inplace ) f.remove(f.find(_oldCol)).remove();
  } else {
    _newCol = new String[]{_inplace ? _oldCol : f.uniquify(_oldCol)};
    if( _inplace ) f.replace(f.find(_oldCol), fr.anyVec()).remove();
    else          f.add(_newNames==null?_newCol[0]:_newNames[0], fr.anyVec());
  }
  DKV.put(f);
  return f;
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
@Test public void testModelAdaptMissing() {
  AModel.AParms p = new AModel.AParms();
  AModel.AOutput o = new AModel.AOutput();

  Vec cat = vec(new String[]{"A","B"},0,1,0,1);
  Frame trn = new Frame();
  trn.add("cat",cat);
  o._names = trn.names();
  o._domains = trn.domains();
  trn.remove();
  AModel am = new AModel(Key.make(),p,o);
  
  Frame tst = new Frame();
  tst.add("cat", cat.makeCon(Double.NaN)); // All NAN/missing column
  Frame adapt = new Frame(tst);
  String[] warns = am.adaptTestForTrain(adapt,true, true);
  Assert.assertTrue(warns.length == 0); // No errors during adaption

  Model.cleanup_adapt( adapt, tst );
  tst.remove();
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
@Test public void testModelAdaptConvert() {
  AModel.AParms p = new AModel.AParms();
  AModel.AOutput o = new AModel.AOutput();

  Frame trn = new Frame();
  trn.add("dog",vec(new String[]{"A","B"},0,1,0,1));
  o._names = trn.names();
  o._domains = trn.domains();
  trn.remove();
  AModel am = new AModel(Key.make(),p,o);
  
  Frame tst = new Frame();
  tst.add("dog",vec(2, 3, 2, 3));
  Frame adapt = new Frame(tst);
  boolean saw_iae = false;
  try { am.adaptTestForTrain(adapt, true, true); }
  catch( IllegalArgumentException iae ) { saw_iae = true; }
  Assert.assertTrue(saw_iae);

  Model.cleanup_adapt( adapt, tst );
  tst.remove();
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
public void buildModelOutput() {
  final int vecSize = _parms._vecSize;
  Futures fs = new Futures();
  String[] colNames = new String[vecSize];
  Vec[] vecs = new Vec[vecSize];
  Key keys[] = Vec.VectorGroup.VG_LEN1.addVecs(vecs.length);

  //allocate
  NewChunk cs[] = new NewChunk[vecs.length];
  AppendableVec avs[] = new AppendableVec[vecs.length];
  for (int i = 0; i < vecs.length; i++) {
    avs[i] = new AppendableVec(keys[i], Vec.T_NUM);
    cs[i] = new NewChunk(avs[i], 0);
  }
  //fill in vector values
  for( int i = 0; i < _modelInfo._vocabSize; i++ ) {
    for (int j=0;  j < vecSize; j++) {
      cs[j].addNum(_modelInfo._syn0[i * vecSize + j]);
    }
  }

  //finalize vectors
  final int rowLayout = avs[0].compute_rowLayout();
  for (int i = 0; i < vecs.length; i++) {
    colNames[i] = new String("V"+i);
    cs[i].close(0, fs);
    vecs[i] = avs[i].close(rowLayout,fs);
  }

  fs.blockForPending();
  Frame fr = new Frame(_w2vKey = Key.make("w2v"));
  //FIXME this ties the word count frame to this one which means
  //FIXME one can't be deleted without destroying the other
  fr.add("Word", (_parms._vocabKey.get()).vec(0));
  fr.add(colNames, vecs);
  DKV.put(_w2vKey, fr);
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
@Test
public void testCollisionOfDRFParamsChecksum() {
  Frame fr = null;
  try {
    fr = parse_test_file("smalldata/junit/cars.csv");
    fr.remove("name").remove();
    Vec old = fr.remove("economy (mpg)");
    fr.add("economy (mpg)", old); // response to last column
    DKV.put(fr);

    // {"_model_id":null,"_train":{"name":"_83da9e0754c5eb9f6b812fe17e7945e5","type":"Key"},"_valid":null,"_nfolds":0,"_keep_cross_validation_predictions":false,"_fold_assignment":"AUTO","_distribution":"AUTO","_tweedie_power":1.5,"_ignored_columns":null,"_ignore_const_cols":true,"_weights_column":null,"_offset_column":null,"_fold_column":null,"_score_each_iteration":false,"_response_column":"economy (mpg)","_balance_classes":false,"_max_after_balance_size":5.0,"_class_sampling_factors":null,"_max_hit_ratio_k":10,"_max_confusion_matrix_size":20,"_checkpoint":null,"_ntrees":9,"_max_depth":15,"_min_rows":1.0,"_nbins":20,"_nbins_cats":1024,"_r2_stopping":0.999999,"_seed":-4522296119273841674,"_nbins_top_level":1024,"_build_tree_one_node":false,"_initial_score_interval":4000,"_score_interval":4000,"_mtries":3,"_sample_rate":0.6499997,"_binomial_double_trees":false}
    DRFModel.DRFParameters params1 = new DRFModel.DRFParameters();
    params1._train = fr._key;
    params1._response_column = "economy (mpg)";
    params1._seed = -4522296119273841674L;
    params1._mtries = 3;
    params1._max_depth = 15;
    params1._ntrees = 9;
    params1._sample_rate = 0.6499997f;

    // {"_model_id":null,"_train":{"name":"_83da9e0754c5eb9f6b812fe17e7945e5","type":"Key"},"_valid":null,"_nfolds":0,"_keep_cross_validation_predictions":false,"_fold_assignment":"AUTO","_distribution":"AUTO","_tweedie_power":1.5,"_ignored_columns":null,"_ignore_const_cols":true,"_weights_column":null,"_offset_column":null,"_fold_column":null,"_score_each_iteration":false,"_response_column":"economy (mpg)","_balance_classes":false,"_max_after_balance_size":5.0,"_class_sampling_factors":null,"_max_hit_ratio_k":10,"_max_confusion_matrix_size":20,"_checkpoint":null,"_ntrees":13,"_max_depth":1,"_min_rows":1.0,"_nbins":20,"_nbins_cats":1024,"_r2_stopping":0.999999,"_seed":-4522296119273841674,"_nbins_top_level":1024,"_build_tree_one_node":false,"_initial_score_interval":4000,"_score_interval":4000,"_mtries":1,"_sample_rate":0.6499997,"_binomial_double_trees":false}
    DRFModel.DRFParameters params2 = new DRFModel.DRFParameters();
    params2._train = fr._key;
    params2._response_column = "economy (mpg)";
    params2._seed = -4522296119273841674L;
    params2._mtries = 1;
    params2._max_depth = 1;
    params2._ntrees = 13;
    params2._sample_rate = 0.6499997f;
    long csum1 = params1.checksum();
    long csum2 = params2.checksum();
    Assert.assertNotEquals("Checksums shoudl be different", csum1, csum2);
  } finally {
    if (fr != null) {
      fr.remove();
    }
  }
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
Val exec_check( Env env, Env.StackHelp stk, Frame tst, AST ast, Frame xfr ) {
  Val val = ast.exec(env);
  if( val.isFrame() ) {
    Frame fr = stk.track(val).getFrame();
    if( tst.numCols() != fr.numCols() || tst.numRows() != fr.numRows() )
      throw new IllegalArgumentException("ifelse test frame and other frames must match dimensions, found "+tst+" and "+fr);
    xfr.add(fr);
  }
  return val;
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
private Frame randSubIterInPlace(DataInfo dinfo, SVDModel model) {
  DataInfo yinfo = null;
  Frame yqfrm = null;

  try {
    // 1) Initialize Y = AG where G ~ N(0,1) and compute Y = QR factorization
    update(1, "Initializing random subspace of training data Y");
    double[][] gt = ArrayUtils.gaussianArray(_parms._nv, _ncolExp, _parms._seed);
    RandSubInit rtsk = new RandSubInit(self(), dinfo, gt);
    rtsk.doAll_numericResult(_parms._nv, dinfo._adaptedFrame);
    yqfrm = rtsk.outputFrame(Key.make(), null, null);   // Alternates between Y and Q from Y = QR

    // Make input frame [A,Q] where A = read-only training data, Y = A \tilde{Q}, Q from Y = QR factorization
    // Note: If A is n by p (p = num cols with categoricals expanded), then \tilde{Q} is p by k and Q is n by k
    //       Q frame is used to save both intermediate Y calculation and final orthonormal Q matrix
    Frame aqfrm = new Frame(dinfo._adaptedFrame);
    aqfrm.add(yqfrm);

    // Calculate Cholesky of Y Gram to get R' = L matrix
    update(1, "Computing QR factorization of Y");
    yinfo = new DataInfo(Key.make(), yqfrm, null, true, DataInfo.TransformType.NONE, true, false, false);
    DKV.put(yinfo._key, yinfo);
    LinearAlgebraUtils.computeQInPlace(self(), yinfo);

    model._output._iterations = 0;
    while (model._output._iterations < _parms._max_iterations) {
      if(!isRunning()) break;
      update(1, "Iteration " + String.valueOf(model._output._iterations+1) + " of randomized subspace iteration");

      // 2) Form \tilde{Y}_j = A'Q_{j-1} and compute \tilde{Y}_j = \tilde{Q}_j \tilde{R}_j factorization
      SMulTask stsk = new SMulTask(dinfo, _parms._nv);
      stsk.doAll(aqfrm);

      Matrix ysmall = new Matrix(stsk._atq);
      QRDecomposition ysmall_qr = new QRDecomposition(ysmall);
      double[][] qtilde = ysmall_qr.getQ().getArray();

      // 3) [A,Q_{j-1}] -> [A,Y_j]: Form Y_j = A\tilde{Q}_j and compute Y_j = Q_jR_j factorization
      BMulInPlaceTask tsk = new BMulInPlaceTask(dinfo, ArrayUtils.transpose(qtilde));
      tsk.doAll(aqfrm);
      LinearAlgebraUtils.computeQInPlace(self(), yinfo);
      model._output._iterations++;
      model.update(self());
    }
  } catch( Throwable t ) {
    Job thisJob = DKV.getGet(_key);
    if (thisJob._state == JobState.CANCELLED) {
      Log.info("Job cancelled by user.");
    } else {
      t.printStackTrace();
      failed(t);
      throw t;
    }
  } finally {
    if( yinfo != null ) yinfo.remove();
  }
  return yqfrm;
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
@Test
public void testDuplicatesCarsGrid() {
  Grid grid = null;
  Frame fr = null;
  Vec old = null;
  try {
    fr = parse_test_file("smalldata/junit/cars_20mpg.csv");
    fr.remove("name").remove(); // Remove unique id
    old = fr.remove("economy");
    fr.add("economy", old); // response to last column
    DKV.put(fr);

    // Setup random hyperparameter search space
    HashMap<String, Object[]> hyperParms = new HashMap<>();
    hyperParms.put("_distribution", new Distribution.Family[]{Distribution.Family.gaussian});
    hyperParms.put("_ntrees", new Integer[]{5, 5});
    hyperParms.put("_max_depth", new Integer[]{2, 2});
    hyperParms.put("_learn_rate", new Float[]{.1f, .1f});

    // Fire off a grid search
    GBMModel.GBMParameters params = new GBMModel.GBMParameters();
    params._train = fr._key;
    params._response_column = "economy";

    GridSearch gs = GridSearch.startGridSearch(params, hyperParms, GBM_MODEL_FACTORY);
    grid = (Grid) gs.get();

    // Check that duplicate model have not been constructed
    Model[] models = grid.getModels();
    assertTrue("Number of returned models has to be > 0", models.length > 0);
    // But all off them should be same
    Key<Model> modelKey = models[0]._key;
    for (Model m : models) {
      assertTrue("Number of constructed models has to be equal to 1", modelKey == m._key);
    }
  } finally {
    if (old != null) {
      old.remove();
    }
    if (fr != null) {
      fr.remove();
    }
    if (grid != null) {
      grid.remove();
    }
  }
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
@Ignore("PUBDEV-1829")
public void testCheckpointReconstruction4BinomialPUBDEV1829() {
  Frame tr = parse_test_file("smalldata/jira/gbm_checkpoint_train.csv");
  Frame val = parse_test_file("smalldata/jira/gbm_checkpoint_valid.csv");

  Vec old = null;

  tr.remove("name").remove();
  tr.remove("economy").remove();
  val.remove("name").remove();
  val.remove("economy").remove();

  old = tr.remove("economy_20mpg");
  tr.add("economy_20mpg", old);
  DKV.put(tr);

  old = val.remove("economy_20mpg");
  val.add("economy_20mpg", old);
  DKV.put(val);

  GBMModel model = null;
  GBMModel modelFromCheckpoint = null;
  GBMModel modelFinal = null;

  try {
    GBMModel.GBMParameters gbmParams = new GBMModel.GBMParameters();
    gbmParams._model_id = Key.make("Initial model");
    gbmParams._train = tr._key;
    gbmParams._valid = val._key;
    gbmParams._response_column = "economy_20mpg";
    gbmParams._ntrees = 5;
    gbmParams._max_depth = 5;
    gbmParams._min_rows = 10;
    gbmParams._score_each_iteration = true;
    gbmParams._seed = 42;
    model = new GBM(gbmParams).trainModel().get();

    GBMModel.GBMParameters gbmFromCheckpointParams = new GBMModel.GBMParameters();
    gbmFromCheckpointParams._model_id = Key.make("Model from checkpoint");
    gbmFromCheckpointParams._train = tr._key;
    gbmFromCheckpointParams._valid = val._key;
    gbmFromCheckpointParams._response_column = "economy_20mpg";
    gbmFromCheckpointParams._ntrees = 10;
    gbmFromCheckpointParams._checkpoint = model._key;
    gbmFromCheckpointParams._score_each_iteration = true;
    gbmFromCheckpointParams._max_depth = 5;
    gbmFromCheckpointParams._min_rows = 10;
    gbmFromCheckpointParams._seed = 42;
    modelFromCheckpoint = new GBM(gbmFromCheckpointParams).trainModel().get();

    // Compute a separated model containing the same number of trees as a model built from checkpoint
    GBMModel.GBMParameters gbmFinalParams = new GBMModel.GBMParameters();
    gbmFinalParams._model_id = Key.make("Validation model");
    gbmFinalParams._train = tr._key;
    gbmFinalParams._valid = val._key;
    gbmFinalParams._response_column = "economy_20mpg";
    gbmFinalParams._ntrees = 10;
    gbmFinalParams._score_each_iteration = true;
    gbmFinalParams._max_depth = 5;
    gbmFinalParams._min_rows = 10;
    gbmFinalParams._seed = 42;
    modelFinal = new GBM(gbmFinalParams).trainModel().get();

    CompressedTree[][] treesFromCheckpoint = getTrees(modelFromCheckpoint);
    CompressedTree[][] treesFromFinalModel = getTrees(modelFinal);
    assertTreeEquals("The model created from checkpoint and corresponding model created from scratch should have the same trees!",
            treesFromCheckpoint, treesFromFinalModel, true);
  } finally {
    if (tr!=null) tr.delete();
    if (val!=null) val.delete();
    if (old != null) old.remove();
    if (model!=null) model.delete();
    if (modelFromCheckpoint!=null) modelFromCheckpoint.delete();
    if (modelFinal!=null) modelFinal.delete();
  }
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
@Test
public void testNfoldsConsecutiveModelsSame() {
  Frame tfr = null;
  Vec old = null;
  DRFModel drf1 = null;
  DRFModel drf2 = null;

  Scope.enter();
  try {
    tfr = parse_test_file("smalldata/junit/cars_20mpg.csv");
    tfr.remove("name").remove(); // Remove unique id
    tfr.remove("economy").remove();
    old = tfr.remove("economy_20mpg");
    tfr.add("economy_20mpg", old.toCategoricalVec()); // response to last column
    DKV.put(tfr);

    DRFModel.DRFParameters parms = new DRFModel.DRFParameters();
    parms._train = tfr._key;
    parms._response_column = "economy_20mpg";
    parms._min_rows = 2;
    parms._max_depth = 2;
    parms._nfolds = 3;
    parms._ntrees = 3;
    parms._seed = 77777;

    DRF job1 = new DRF(parms);
    drf1 = job1.trainModel().get();

    DRF job2 = new DRF(parms);
    drf2 = job2.trainModel().get();

    ModelMetricsBinomial mm1 = (ModelMetricsBinomial)drf1._output._cross_validation_metrics;
    ModelMetricsBinomial mm2 = (ModelMetricsBinomial)drf2._output._cross_validation_metrics;
    assertEquals(mm1.auc()._auc, mm2.auc()._auc, 1e-12);
    assertEquals(mm1.mse(), mm2.mse(), 1e-12);
    assertEquals(mm1.r2(), mm2.r2(), 1e-12);
    assertEquals(mm1.logloss(), mm2.logloss(), 1e-12);

    job1.remove();
    job2.remove();
  } finally {
    if (tfr != null) tfr.remove();
    if (old != null) old.remove();
    if (drf1 != null) {
      drf1.deleteCrossValidationModels();
      drf1.delete();
    }
    if (drf2 != null) {
      drf2.deleteCrossValidationModels();
      drf2.delete();
    }
    Scope.exit();
  }
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
@Test
public void testMTrys() {
  Frame tfr = null;
  Vec old = null;
  DRFModel drf1 = null;

  for (int i=1; i<=6; ++i) {
    Scope.enter();
    try {
      tfr = parse_test_file("smalldata/junit/cars_20mpg.csv");
      tfr.remove("name").remove(); // Remove unique id
      tfr.remove("economy").remove();
      old = tfr.remove("economy_20mpg");
      tfr.add("economy_20mpg", old.toCategoricalVec()); // response to last column
      DKV.put(tfr);

      DRFModel.DRFParameters parms = new DRFModel.DRFParameters();
      parms._train = tfr._key;
      parms._response_column = "economy_20mpg";
      parms._min_rows = 2;
      parms._ntrees = 5;
      parms._max_depth = 5;
      parms._nfolds = 3;
      parms._mtries = i;

      DRF job1 = new DRF(parms);
      drf1 = job1.trainModel().get();

      ModelMetricsBinomial mm1 = (ModelMetricsBinomial) drf1._output._cross_validation_metrics;
      Assert.assertTrue(mm1._auc != null);

      job1.remove();
    } finally {
      if (tfr != null) tfr.remove();
      if (old != null) old.remove();
      if (drf1 != null) {
        drf1.deleteCrossValidationModels();
        drf1.delete();
      }
      Scope.exit();
    }
  }
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
@Test
public void testDuplicatesCarsGrid() {
  Grid grid = null;
  Frame fr = null;
  Vec old = null;
  try {
    fr = parse_test_file("smalldata/junit/cars_20mpg.csv");
    fr.remove("name").remove(); // Remove unique id
    old = fr.remove("economy");
    fr.add("economy", old); // response to last column
    DKV.put(fr);

    // Setup random hyperparameter search space
    HashMap<String, Object[]> hyperParms = new HashMap<>();
    hyperParms.put("_ntrees", new Integer[]{5, 5});
    hyperParms.put("_max_depth", new Integer[]{2, 2});
    hyperParms.put("_mtries", new Integer[]{-1, -1});
    hyperParms.put("_sample_rate", new Float[]{.1f, .1f});

    // Fire off a grid search
    DRFModel.DRFParameters params = new DRFModel.DRFParameters();
    params._train = fr._key;
    params._response_column = "economy";

    // Get the Grid for this modeling class and frame
    GridSearch gs = GridSearch.startGridSearch(params, hyperParms, DRF_MODEL_FACTORY);
    grid = (Grid) gs.get();

    // Check that duplicate model have not been constructed
    Model[] models = grid.getModels();
    assertTrue("Number of returned models has to be > 0", models.length > 0);
    // But all off them should be same
    Key<Model> modelKey = models[0]._key;
    for (Model m : models) {
      assertTrue("Number of constructed models has to be equal to 1", modelKey == m._key);
    }
  } finally {
    if (old != null) {
      old.remove();
    }
    if (fr != null) {
      fr.remove();
    }
    if (grid != null) {
      grid.remove();
    }
  }
}
 

import water.fvec.Frame; //导入方法依赖的package包/类
@Test public void elasticAveragingTrivial() {
  DeepLearningParameters dl;
  Frame frTrain;
  int N = 2;
  DeepLearningModel [] models = new DeepLearningModel[N];
  dl = new DeepLearningParameters();
  Scope.enter();
  try {
    for (int i = 0; i < N; ++i) {
      frTrain = parse_test_file("./smalldata/covtype/covtype.20k.data");
      Vec resp = frTrain.lastVec().toCategoricalVec();
      frTrain.remove(frTrain.vecs().length - 1).remove();
      frTrain.add("Response", resp);
      DKV.put(frTrain);
      dl._train = frTrain._key;
      dl._response_column = ((Frame) DKV.getGet(dl._train)).lastVecName();
      dl._export_weights_and_biases = true;
      dl._hidden = new int[]{17, 11};
      dl._quiet_mode = false;

      // make it reproducible
      dl._seed = 1234;
      dl._reproducible = true;

      // only do one M/R iteration, and there's no elastic average yet - so the two paths below should be identical
      dl._epochs = 1;
      dl._train_samples_per_iteration = -1;

      if (i == 0) {
        // no elastic averaging
        dl._elastic_averaging = false;
        dl._elastic_averaging_moving_rate = 0.5; //ignored
        dl._elastic_averaging_regularization = 0.9; //ignored
      } else {
        // no-op elastic averaging
        dl._elastic_averaging = true; //go different path, but don't really do anything because of epochs=1 and train_samples_per_iteration=-1
        dl._elastic_averaging_moving_rate = 0.5; //doesn't matter, it's not used since we only do one M/R iteration and there's no time average
        dl._elastic_averaging_regularization = 0.1; //doesn't matter, since elastic average isn't yet available in first iteration
      }

      // Invoke DL and block till the end
      DeepLearning job = null;
      try {
        job = new DeepLearning(dl);
        // Get the model
        models[i] = job.trainModel().get();
      } finally {
        if (job != null) job.remove();
      }
      frTrain.remove();
    }
    for (int i = 0; i < N; ++i) {
      Log.info(models[i]._output._training_metrics.cm().table().toString());
      Assert.assertEquals(models[i]._output._training_metrics._MSE, models[0]._output._training_metrics._MSE, 1e-6);
    }

  }finally{
    for (int i=0; i<N; ++i)
      if (models[i] != null)
        models[i].delete();
    Scope.exit();
  }
}