fix: SGD results now reproducible

flaml/automl/model.py

@@ -2450,6 +2450,9 @@ def config2params(self, config: dict) -> dict:
 
     def __init__(self, task="binary", **config):
         super().__init__(task, **config)
+        random_seed = self.params.pop("random_seed", config.get("random_seed", 10242048))
+        if "random_state" not in self.params:
+            self.params["random_state"] = random_seed
         if self._task.is_classification():
             self.estimator_class = SGDClassifier
         elif self._task.is_regression():

test/automl/test_classification.py

@@ -435,6 +435,7 @@ def test_sparse_matrix_lr(self):
         # "lrl1",
         "lrl2",
         "rf",
+        "sgd",
         "svc",
         "xgboost",
         "xgb_limitdepth",
@@ -498,6 +499,9 @@ def test_reproducibility_of_classification_models(estimator: str):
         "lrl2",
         "svc",
         "rf",
+        # "sgd" omitted: SGDEstimator wraps the sklearn model with a Normalizer
+        # preprocessing step that this test's helper does not replicate, so the
+        # bare underlying model cannot match the FLAML wrapper's CV result.
         "xgboost",
         "xgb_limitdepth",
     ],

test/automl/test_regression.py

@@ -254,6 +254,7 @@ def test_multioutput():
         "kneighbor",
         "lgbm",
         "rf",
+        "sgd",
         "xgboost",
         "xgb_limitdepth",
     ],
@@ -407,6 +408,9 @@ def test_reproducibility_of_lgbm_regression_model():
         "kneighbor",
         "lgbm",
         "rf",
+        # "sgd" omitted: SGDEstimator wraps the sklearn model with a Normalizer
+        # preprocessing step that this test's helper does not replicate, so the
+        # bare underlying model cannot match the FLAML wrapper's CV result.
         "xgboost",
         "xgb_limitdepth",
     ],