[Breaking] Don't drop trees during DART prediction by default (#5115)

* Simplify DropTrees calling logic

* Add `training` parameter for prediction method.

* [Breaking]: Add `training` to C API.

* Change for R and Python custom objective.

* Correct comment.

Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu>
Co-authored-by: Jiaming Yuan <jm.yuan@outlook.com>

tests/python/test_basic_models.py

@@ -44,7 +44,8 @@ class TestModels(unittest.TestCase):
     def test_dart(self):
         dtrain = xgb.DMatrix(dpath + 'agaricus.txt.train')
         dtest = xgb.DMatrix(dpath + 'agaricus.txt.test')
-        param = {'max_depth': 5, 'objective': 'binary:logistic', 'booster': 'dart', 'verbosity': 1}
+        param = {'max_depth': 5, 'objective': 'binary:logistic',
+                 'eval_metric': 'logloss', 'booster': 'dart', 'verbosity': 1}
         # specify validations set to watch performance
         watchlist = [(dtest, 'eval'), (dtrain, 'train')]
         num_round = 2
@@ -52,7 +53,8 @@ class TestModels(unittest.TestCase):
         # this is prediction
         preds = bst.predict(dtest, ntree_limit=num_round)
         labels = dtest.get_label()
-        err = sum(1 for i in range(len(preds)) if int(preds[i] > 0.5) != labels[i]) / float(len(preds))
+        err = sum(1 for i in range(len(preds))
+                  if int(preds[i] > 0.5) != labels[i]) / float(len(preds))
         # error must be smaller than 10%
         assert err < 0.1
 
@@ -68,18 +70,31 @@ class TestModels(unittest.TestCase):
         # assert they are the same
         assert np.sum(np.abs(preds2 - preds)) == 0
 
+        def my_logloss(preds, dtrain):
+            labels = dtrain.get_label()
+            return 'logloss', np.sum(
+                np.log(np.where(labels, preds, 1 - preds)))
+
+        # check whether custom evaluation metrics work
+        bst = xgb.train(param, dtrain, num_round, watchlist,
+                        feval=my_logloss)
+        preds3 = bst.predict(dtest, ntree_limit=num_round)
+        assert all(preds3 == preds)
+
         # check whether sample_type and normalize_type work
         num_round = 50
         param['verbosity'] = 0
         param['learning_rate'] = 0.1
         param['rate_drop'] = 0.1
         preds_list = []
-        for p in [[p0, p1] for p0 in ['uniform', 'weighted'] for p1 in ['tree', 'forest']]:
+        for p in [[p0, p1] for p0 in ['uniform', 'weighted']
+                  for p1 in ['tree', 'forest']]:
             param['sample_type'] = p[0]
             param['normalize_type'] = p[1]
             bst = xgb.train(param, dtrain, num_round, watchlist)
             preds = bst.predict(dtest, ntree_limit=num_round)
-            err = sum(1 for i in range(len(preds)) if int(preds[i] > 0.5) != labels[i]) / float(len(preds))
+            err = sum(1 for i in range(len(preds))
+                      if int(preds[i] > 0.5) != labels[i]) / float(len(preds))
             assert err < 0.1
             preds_list.append(preds)
 

tests/python/test_ranking.py

@@ -135,7 +135,7 @@ class TestRanking(unittest.TestCase):
         # specify validations set to watch performance
         watchlist = [(self.dtest, 'eval'), (self.dtrain, 'train')]
         bst = xgboost.train(self.params, self.dtrain, num_boost_round=2500,
-                        early_stopping_rounds=10, evals=watchlist)
+                            early_stopping_rounds=10, evals=watchlist)
         assert bst.best_score > 0.98
 
     def test_cv(self):