fix bug when using dart with init_model (#2251)

* add test

* fix a index bug

src/boosting/dart.hpp

@@ -171,8 +171,8 @@ class DART: public GBDT {
           train_score_updater_->AddScore(models_[curr_tree].get(), cur_tree_id);
         }
         if (!config_->uniform_drop) {
-          sum_weight_ -= tree_weight_[i] * (1.0f / (k + 1.0f));
-          tree_weight_[i] *= (k / (k + 1.0f));
+          sum_weight_ -= tree_weight_[i - num_init_iteration_] * (1.0f / (k + 1.0f));
+          tree_weight_[i - num_init_iteration_] *= (k / (k + 1.0f));
         }
       }
     } else {
@@ -189,8 +189,8 @@ class DART: public GBDT {
           train_score_updater_->AddScore(models_[curr_tree].get(), cur_tree_id);
         }
         if (!config_->uniform_drop) {
-          sum_weight_ -= tree_weight_[i] * (1.0f / (k + config_->learning_rate));;
-          tree_weight_[i] *= (k / (k + config_->learning_rate));
+          sum_weight_ -= tree_weight_[i - num_init_iteration_] * (1.0f / (k + config_->learning_rate));;
+          tree_weight_[i - num_init_iteration_] *= (k / (k + config_->learning_rate));
         }
       }
     }

tests/python_package_test/test_engine.py

@@ -480,6 +480,31 @@ class TestEngine(unittest.TestCase):
             self.assertAlmostEqual(l1, mae, places=5)
         os.remove(model_name)
 
+    def test_continue_train_dart(self):
+        X, y = load_boston(True)
+        X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.1, random_state=42)
+        params = {
+            'boosting_type': 'dart',
+            'objective': 'regression',
+            'metric': 'l1',
+            'verbose': -1
+        }
+        lgb_train = lgb.Dataset(X_train, y_train, free_raw_data=False)
+        lgb_eval = lgb.Dataset(X_test, y_test, reference=lgb_train, free_raw_data=False)
+        init_gbm = lgb.train(params, lgb_train, num_boost_round=50)
+        evals_result = {}
+        gbm = lgb.train(params, lgb_train,
+                        num_boost_round=50,
+                        valid_sets=lgb_eval,
+                        verbose_eval=False,
+                        evals_result=evals_result,
+                        init_model=init_gbm)
+        ret = mean_absolute_error(y_test, gbm.predict(X_test))
+        self.assertLess(ret, 3.5)
+        self.assertAlmostEqual(evals_result['valid_0']['l1'][-1], ret, places=5)
+        for l1, mae in zip(evals_result['valid_0']['l1'], evals_result['valid_0']['mae']):
+            self.assertAlmostEqual(l1, mae, places=5)
+
     def test_continue_train_multiclass(self):
         X, y = load_iris(True)
         X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.1, random_state=42)