[python-package] preserve params when copying Booster (fixes #5539)

python-package/lightgbm/basic.py

@@ -3250,6 +3250,9 @@ def __init__(
                             'to create Booster instance')
         self.params = params
 
+        # ensure params are updated on the C++ side
+        self.reset_parameter(params)
+
     def __del__(self) -> None:
         try:
             if self._network:
@@ -3267,7 +3270,7 @@ def __copy__(self) -> "Booster":
 
     def __deepcopy__(self, _) -> "Booster":
         model_str = self.model_to_string(num_iteration=-1)
-        return Booster(model_str=model_str)
+        return Booster(model_str=model_str, params=self.params)
 
     def __getstate__(self) -> Dict[str, Any]:
         this = self.__dict__.copy()

src/boosting/gbdt.cpp

@@ -795,7 +795,10 @@ void GBDT::ResetConfig(const Config* config) {
 
   boosting_on_gpu_ = objective_function_ != nullptr && objective_function_->IsCUDAObjective() &&
                     !data_sample_strategy_->IsHessianChange();  // for sample strategy with Hessian change, fall back to boosting on CPU
-  tree_learner_->ResetBoostingOnGPU(boosting_on_gpu_);
+
+  if (tree_learner_ != nullptr) {
+    tree_learner_->ResetBoostingOnGPU(boosting_on_gpu_);
+  }
 
   if (train_data_ != nullptr) {
     data_sample_strategy_->ResetSampleConfig(new_config.get(), false);

tests/c_api_test/test_.py

@@ -210,6 +210,9 @@ def test_booster():
         c_str('model.txt'),
         ctypes.byref(num_total_model),
         ctypes.byref(booster2))
+    LIB.LGBM_BoosterResetParameter(
+        booster2,
+        c_str("app=binary metric=auc num_leaves=29 verbose=0"))
     data = np.loadtxt(str(binary_example_dir / 'binary.test'), dtype=np.float64)
     mat = data[:, 1:]
     preb = np.empty(mat.shape[0], dtype=np.float64)

tests/python_package_test/test_basic.py

@@ -823,3 +823,29 @@ def test_feature_names_are_set_correctly_when_no_feature_names_passed_into_Datas
         data=np.random.randn(100, 3),
     )
     assert ds.construct().feature_name == ["Column_0", "Column_1", "Column_2"]
+
+
+def test_booster_deepcopy_preserves_parameters():
+    bst = lgb.train(
+        params={'num_leaves': 5, 'verbosity': -1},
+        num_boost_round=2,
+        train_set=lgb.Dataset(np.random.rand(100, 2))
+    )
+    bst2 = copy.deepcopy(bst)
+    assert bst2.params == bst.params
+    assert bst.params["num_leaves"] == 5
+    assert bst.params["verbosity"] == -1
+
+
+def test_booster_params_kwarg_overrides_params_from_model_string():
+    bst = lgb.train(
+        params={'num_leaves': 5, 'learning_rate': 0.708, 'verbosity': -1},
+        num_boost_round=2,
+        train_set=lgb.Dataset(np.random.rand(100, 2))
+    )
+    bst2 = lgb.Booster(
+        params={'num_leaves': 7},
+        model_str=bst.model_to_string()
+    )
+    assert bst2.params["num_leaves"] == 7
+    assert bst2.params["learning_rate"] == 0.708