[fix] fix duplicate added initial scores for single-leaf trees (#fixes #4708)

* fix duplicate added initial scores for single-leaf trees

* add test case

* Fix import in Python test

* commit python suggestions
Co-authored-by: Nikita Titov <nekit94-08@mail.ru>

src/boosting/gbdt.cpp

@@ -419,20 +419,15 @@ bool GBDT::TrainOneIter(const score_t* gradients, const score_t* hessians) {
     } else {
       // only add default score one-time
       if (models_.size() < static_cast<size_t>(num_tree_per_iteration_)) {
-        double output = 0.0;
-        if (!class_need_train_[cur_tree_id]) {
-          if (objective_function_ != nullptr) {
-            output = objective_function_->BoostFromScore(cur_tree_id);
-          }
-        } else {
-          output = init_scores[cur_tree_id];
-        }
-        new_tree->AsConstantTree(output);
+        if (objective_function_ != nullptr && !config_->boost_from_average && !train_score_updater_->has_init_score()) {
+          init_scores[cur_tree_id] = ObtainAutomaticInitialScore(objective_function_, cur_tree_id);
           // updates scores
-        train_score_updater_->AddScore(output, cur_tree_id);
+          train_score_updater_->AddScore(init_scores[cur_tree_id], cur_tree_id);
           for (auto& score_updater : valid_score_updater_) {
-          score_updater->AddScore(output, cur_tree_id);
+            score_updater->AddScore(init_scores[cur_tree_id], cur_tree_id);
+          }
         }
+        new_tree->AsConstantTree(init_scores[cur_tree_id]);
       }
     }
     // add model

tests/python_package_test/test_engine.py

@@ -3424,3 +3424,31 @@ def test_pandas_nullable_dtypes():
 
     # test equal predictions
     np.testing.assert_allclose(preds, preds_nullable_dtypes)
+
+
+def test_boost_from_average_with_single_leaf_trees():
+    # test data are taken from bug report
+    # https://github.com/microsoft/LightGBM/issues/4708
+    X = np.array([
+        [1021.0589, 1018.9578],
+        [1023.85754, 1018.7854],
+        [1024.5468, 1018.88513],
+        [1019.02954, 1018.88513],
+        [1016.79926, 1018.88513],
+        [1007.6, 1018.88513]], dtype=np.float32)
+    y = np.array([1023.8, 1024.6, 1024.4, 1023.8, 1022.0, 1014.4], dtype=np.float32)
+    params = {
+        "extra_trees": True,
+        "min_data_in_bin": 1,
+        "extra_seed": 7,
+        "objective": "regression",
+        "verbose": -1,
+        "boost_from_average": True,
+        "min_data_in_leaf": 1,
+    }
+    train_set = lgb.Dataset(X, y)
+    model = lgb.train(params=params, train_set=train_set, num_boost_round=10)
+
+    preds = model.predict(X)
+    mean_preds = np.mean(preds)
+    assert y.min() <= mean_preds <= y.max()