PUBDEV-8266 XGBoost alias rewriting (#5641)

PUBDEV-8266 fix alias rewriting

h2o-extensions/xgboost/src/main/java/hex/tree/xgboost/XGBoost.java

@@ -231,6 +231,13 @@ public class XGBoost extends ModelBuilder<XGBoostModel,XGBoostModel.XGBoostParam
     checkPositiveRate("colsample_bylevel", _parms._colsample_bylevel);
     checkPositiveRate("colsample_bynode", _parms._colsample_bynode);
     checkPositiveRate("colsample_bytree", _parms._colsample_bytree);
+    checkColumnAlias("col_sample_rate", _parms._col_sample_rate, "colsample_bylevel", _parms._colsample_bylevel, 1);
+    checkColumnAlias("col_sample_rate_per_tree", _parms._col_sample_rate_per_tree, "colsample_bytree", _parms._colsample_bytree, 1);
+    checkColumnAlias("sample_rate", _parms._sample_rate, "subsample", _parms._subsample, 1);
+    checkColumnAlias("learn_rate", _parms._learn_rate, "subsample", _parms._eta, 0.3);
+    checkColumnAlias("max_abs_leafnode_pred", _parms._max_abs_leafnode_pred, "max_delta_step", _parms._max_delta_step,0);
+    checkColumnAlias("ntrees", _parms._ntrees, "n_estimators", _parms._n_estimators, 0);
+
 
     if (_parms._scale_pos_weight != 1) {
       if (_nclass != 2)
@@ -268,6 +275,15 @@ private void checkPositiveRate(String paramName, double rateValue) {
       error("_" + paramName, paramName + " must be between 0 (exclusive) and 1 (inclusive)");
   }
 
+  private void checkColumnAlias(String paramName, double paramValue, String aliasName, double aliasValue, double defaultValue) {
+    if (paramValue != defaultValue && aliasValue != defaultValue && paramValue != aliasValue) {
+      error("_" + paramName, paramName + " and its alias " + aliasName + " are both set to different value than default value. Set " + aliasName + " to default value (" + defaultValue + "), to use " + paramName + " actual value.");
+    }
+    if (aliasValue != defaultValue){
+      warn("_"+paramName, "Using user-provided parameter "+aliasName+" instead of "+paramName+".\"");
+    }
+  }
+
   @Override
   protected void checkEarlyStoppingReproducibility() {
     if (_parms._score_tree_interval == 0 && !_parms._score_each_iteration) {

h2o-extensions/xgboost/src/main/java/hex/tree/xgboost/XGBoostModel.java

@@ -307,7 +307,6 @@ public static Map<String, Object> createParamsMap(XGBoostParameters p, int nClas
       p._n_estimators = p._ntrees;
     }
     if (p._eta != 0.3) {
-      LOG.info("Using user-provided parameter eta instead of learn_rate.");
       params.put("eta", p._eta);
       p._learn_rate = p._eta;
     } else {
@@ -321,23 +320,20 @@ public static Map<String, Object> createParamsMap(XGBoostParameters p, int nClas
       params.put("silent", p._quiet_mode);
     }
     if (p._subsample != 1.0) {
-      LOG.info("Using user-provided parameter subsample instead of sample_rate.");
       params.put("subsample", p._subsample);
       p._sample_rate = p._subsample;
     } else {
       params.put("subsample", p._sample_rate);
       p._subsample = p._sample_rate;
     }
     if (p._colsample_bytree != 1.0) {
-      LOG.info("Using user-provided parameter colsample_bytree instead of col_sample_rate_per_tree.");
       params.put("colsample_bytree", p._colsample_bytree);
       p._col_sample_rate_per_tree = p._colsample_bytree;
     } else {
       params.put("colsample_bytree", p._col_sample_rate_per_tree);
       p._colsample_bytree = p._col_sample_rate_per_tree;
     }
     if (p._colsample_bylevel != 1.0) {
-      LOG.info("Using user-provided parameter colsample_bylevel instead of col_sample_rate.");
       params.put("colsample_bylevel", p._colsample_bylevel);
       p._col_sample_rate = p._colsample_bylevel;
     } else {
@@ -348,7 +344,6 @@ public static Map<String, Object> createParamsMap(XGBoostParameters p, int nClas
       params.put("colsample_bynode", p._colsample_bynode);
     }    
     if (p._max_delta_step != 0) {
-      LOG.info("Using user-provided parameter max_delta_step instead of max_abs_leafnode_pred.");
       params.put("max_delta_step", p._max_delta_step);
       p._max_abs_leafnode_pred = p._max_delta_step;
     } else {

h2o-extensions/xgboost/src/test/java/hex/tree/xgboost/XGBoostTest.java

@@ -2623,4 +2623,140 @@ public void testScalePosWeight() {
     }
   }
 
+  @Test
+  public void testColSampleRate() {
+    Scope.enter();
+    try {
+      XGBoostModel model1, model2;
+      Frame train = parseTestFile("smalldata/gbm_test/ecology_model.csv");
+
+      train.remove("Site").remove();
+      train.remove("Method").remove();
+      train.toCategoricalCol("Angaus");
+      Scope.track(train);
+
+      XGBoostModel.XGBoostParameters parms = new XGBoostModel.XGBoostParameters();
+      parms._train = train._key;
+      parms._valid = train._key;
+      parms._response_column = "Angaus";
+      parms._distribution = multinomial;
+      parms._ntrees = 5;
+      parms._tree_method = XGBoostModel.XGBoostParameters.TreeMethod.hist;
+      parms._seed = 42;
+      parms._col_sample_rate = 0.9;
+      model1 = new hex.tree.xgboost.XGBoost(parms).trainModel().get();
+      Scope.track_generic(model1);
+
+      XGBoostModel.XGBoostParameters parms2 = new XGBoostModel.XGBoostParameters();
+      parms2._train = train._key;
+      parms2._valid = train._key;
+      parms2._response_column = "Angaus";
+      parms2._distribution = multinomial;
+      parms2._ntrees = 5;
+      parms2._tree_method = XGBoostModel.XGBoostParameters.TreeMethod.hist;
+      parms2._seed = 42;
+      parms2._col_sample_rate = 0.1;
+      model2 = new hex.tree.xgboost.XGBoost(parms2).trainModel().get();
+      Scope.track_generic(model2);
+      assertNotEquals(model1._output._training_metrics.rmse(), model2._output._training_metrics.rmse(), 0);
+    } finally {
+      Scope.exit();
+    }
+  }
+
+  @Test
+  public void testColSampleRateSameValue() {
+    Scope.enter();
+    try {
+      Frame train = parseTestFile("smalldata/gbm_test/ecology_model.csv");
+
+      train.remove("Site").remove();
+      train.remove("Method").remove();
+      train.toCategoricalCol("Angaus");
+      Scope.track(train);
+
+      XGBoostModel model1, model2;
+      XGBoostModel.XGBoostParameters parms = new XGBoostModel.XGBoostParameters();
+      parms._train = train._key;
+      parms._valid = train._key;
+      parms._response_column = "Angaus";
+      parms._distribution = multinomial;
+      parms._ntrees = 5;
+      parms._tree_method = XGBoostModel.XGBoostParameters.TreeMethod.hist;
+      parms._seed = 42;
+      parms._col_sample_rate = 0.9;
+      model1 = new hex.tree.xgboost.XGBoost(parms).trainModel().get();
+      Scope.track_generic(model1);
+
+      parms._colsample_bylevel = 0.9;
+      model2 = new hex.tree.xgboost.XGBoost(parms).trainModel().get();
+      Scope.track_generic(model2);
+      assertEquals(model1._output._training_metrics.rmse(), model2._output._training_metrics.rmse(), 0);
+    } finally {
+      Scope.exit();
+    }
+  }
+
+  @Test
+  public void testColSampleRateAndAlias() {
+    Scope.enter();
+    try {
+      XGBoostModel model1;
+      Frame train = parseTestFile("smalldata/gbm_test/ecology_model.csv");
+
+      // Fix training set
+      train.remove("Site").remove();
+      train.remove("Method").remove();
+      train.toCategoricalCol("Angaus");
+      Scope.track(train);
+
+      XGBoostModel.XGBoostParameters parms = new XGBoostModel.XGBoostParameters();
+      parms._train = train._key;
+      parms._valid = train._key;
+      parms._response_column = "Angaus";
+      parms._distribution = multinomial;
+      parms._ntrees = 5;
+      parms._tree_method = XGBoostModel.XGBoostParameters.TreeMethod.hist;
+      parms._seed = 42;
+      parms._col_sample_rate = 0.9;
+      parms._colsample_bylevel = 0.3;
+      model1 = new hex.tree.xgboost.XGBoost(parms).trainModel().get();
+      Scope.track_generic(model1);
+      fail("Model training should fail.");
+    } catch(H2OModelBuilderIllegalArgumentException ex){
+      assertTrue(ex.getMessage().contains("col_sample_rate and its alias colsample_bylevel are both set"));
+    } finally {
+      Scope.exit();
+    }
+  }
+
+  @Test
+  public void testColSampleRateAndAliasSame() {
+    Scope.enter();
+    try {
+      XGBoostModel model1;
+      Frame train = parseTestFile("smalldata/gbm_test/ecology_model.csv");
+
+      train.remove("Site").remove();    
+      train.remove("Method").remove();
+      train.toCategoricalCol("Angaus");
+      Scope.track(train);
+
+      XGBoostModel.XGBoostParameters parms = new XGBoostModel.XGBoostParameters();
+      parms._train = train._key;
+      parms._valid = train._key;
+      parms._response_column = "Angaus";
+      parms._distribution = multinomial;
+      parms._ntrees = 5;
+      parms._tree_method = XGBoostModel.XGBoostParameters.TreeMethod.hist;
+      parms._seed = 42;
+      parms._col_sample_rate = 0.9;
+      parms._colsample_bylevel = 0.9;
+      model1 = new hex.tree.xgboost.XGBoost(parms).trainModel().get();
+      Scope.track_generic(model1);
+      assertEquals(model1._parms._col_sample_rate, model1._parms._colsample_bylevel, 0);
+    } finally {
+      Scope.exit();
+    }
+  }
 }

h2o-extensions/xgboost/src/test/java/hex/tree/xgboost/remote/SteamExecutorStarterTest.java

@@ -245,8 +245,12 @@ public void testSteamClusterStop() throws Exception {
             expectAndCheckStopResponse(steam, "03_stop_req_02", true);
 
             // building another model
-            params._ntrees = 5;
-            Job<XGBoostModel> model2 = new XGBoost(params).trainModel();
+            XGBoostModel.XGBoostParameters params2 = new XGBoostModel.XGBoostParameters();
+            params2._train = train._key;
+            params2._ntrees = 5;
+            params2._response_column = "AGE";
+            params2._ignored_columns = new String[]{"ID"};
+            Job<XGBoostModel> model2 = new XGBoost(params2).trainModel();
 
             // will request external cluster start again
             Map<String, String> startReq2 = steam.waitToReceiveMessage("start request");

h2o-py/tests/testdir_algos/xgboost/pyunit_synonym_params_xgboost.py

@@ -3,6 +3,7 @@
 sys.path.insert(1,"../../../")
 from tests import pyunit_utils
 from h2o.estimators.xgboost import *
+from h2o.exceptions import H2OResponseError
 
 
 def xgboost_synonym_params():
@@ -24,23 +25,27 @@ def xgboost_synonym_params():
         vals = a[2] 
         print("check parity of %s and %s via %s" % (p1, p2, vals))
         # check default values end up to same value
-        model.train(x=x, y=y, training_frame=df )
+        model.train(x=x, y=y, training_frame=df)
         assert model.parms[p1]['actual_value'] == model.parms[p2]['actual_value']
-        # changing p2 modifies both
+        # changing p2 and p1 is default - should not fail
         setattr(model, p2, vals[0])
-        model.train(x=x, y=y, training_frame=df )
+        model.train(x=x, y=y, training_frame=df)
         assert model.parms[p1]['actual_value'] == vals[0]
         assert model.parms[p1]['actual_value'],  model.parms[p2]['actual_value']
-        # changing p1 modifies both
+        # changing p1 and p2 is not default - should fail
         setattr(model, p1, vals[1])
-        model.train(x=x, y=y, training_frame=df )
-        assert model.parms[p1]['actual_value'] == vals[1]
-        assert model.parms[p1]['actual_value'], model.parms[p2]['actual_value']
-        # changing p2 has no effect since p1 has precedence
+        try:
+            model.train(x=x, y=y, training_frame=df)
+        except H2OResponseError as e:
+            assert "ERRR on field: _"+p2 in str(e), p2+" and its alias "+p1+" are both set"
+            setattr(model, p2, vals[1])
+        # changing p2 since p1 has precedence and is not default - should fail
         setattr(model, p2, vals[2])
-        model.train(x=x, y=y, training_frame=df)
-        assert model.parms[p1]['actual_value'] == vals[1]
-        assert model.parms[p1]['actual_value'], model.parms[p2]['actual_value']
+        try:
+            model.train(x=x, y=y, training_frame=df)
+        except H2OResponseError as e:
+            assert "ERRR on field: _"+p2 in str(e), p2+" and its alias "+p1+" are both set"
+            setattr(model, p2, vals[1])
 
 
 if __name__ == "__main__":

h2o-py/tests/testdir_algos/xgboost/pyunit_xgboost_colsample_bylevel.py

@@ -0,0 +1,63 @@
+from h2o.estimators.xgboost import *
+from tests import pyunit_utils
+from h2o.exceptions import H2OResponseError
+
+
+def test_param_and_alias_are_same(data, x_names, y):
+    assert H2OXGBoostEstimator.available() is True
+
+    num_round = 5
+    params = {
+        'tree_method': 'hist',
+        'ntrees': num_round,
+        'backend': 'cpu',
+        'save_matrix_directory': "/home/mori/Documents/h2o/code/test/xgboost_data/",
+        'seed': 42,
+        'colsample_bylevel': 0.9,
+        'col_sample_rate': 0.9
+    }
+
+    # train h2o XGBoost models
+    h2o_model = H2OXGBoostEstimator(**params)
+    h2o_model.train(x=x_names, y=y, training_frame=data)
+
+    assert h2o_model is not None, "Training should not fail."
+
+
+def test_param_and_alias_are_not_same(data, x_names, y):
+    assert H2OXGBoostEstimator.available() is True
+
+    num_round = 5
+    params = {
+        'tree_method': 'hist',
+        'ntrees': num_round,
+        'backend': 'cpu',
+        'save_matrix_directory': "/home/mori/Documents/h2o/code/test/xgboost_data/",
+        'seed': 42,
+        'colsample_bylevel': 0.9,
+        'col_sample_rate': 0.3
+    }
+
+    # train h2o XGBoost models
+    h2o_model = H2OXGBoostEstimator(**params)
+    try:
+        h2o_model.train(x=x_names, y=y, training_frame=data)
+        assert False, "Training should fail."
+    except H2OResponseError as e:
+        assert "ERRR on field: _col_sample_rate" in str(e), \
+            "col_sample_rate and its alias colsample_bylevel are both set"    
+
+
+def test_alias():
+    data = h2o.import_file(path="../../../../smalldata/gbm_test/ecology_model.csv")
+    y = "Angaus"
+    data[y] = data[y].asfactor()
+    x_names = data.col_names.remove(y)
+    test_param_and_alias_are_same(data, x_names, y)
+    test_param_and_alias_are_not_same(data, x_names, y)
+
+
+if __name__ == "__main__":
+    pyunit_utils.standalone_test(test_alias)
+else:
+    test_alias()