feat: bump narwhals and adapt to support pyarrow (#694)

* refactor all **kwargs into kwargs

* refactor tests

* rerun meta-models

* print

* codebase

* tests

* bump min version

* fix with narwhals branch

* bump narwhals

* list -> numpy constant creation

---------

Co-authored-by: vincent d warmerdam <vincentwarmerdam@gmail.com>

pyproject.toml

@@ -20,7 +20,7 @@ maintainers = [
 ]
 
 dependencies = [
-    "narwhals>=1.0.0",
+    "narwhals>=1.2.0",
     "pandas>=1.1.5",
     "scikit-learn>=1.0",
     "importlib-metadata >= 1.0; python_version < '3.8'",
@@ -62,8 +62,7 @@ docs = [
 ]
 
 test = [
-    "narwhals[polars]",
-    "pyarrow",
+    "narwhals[polars,pyarrow]",
     "pytest>=6.2.5",
     "pytest-xdist>=1.34.0",
     "pytest-cov>=2.6.1",

setup.py

@@ -2,6 +2,7 @@
 """
 setup.py is used to distribute scikit-lego as a package using setuptools and twine
 """
+
 from setuptools import setup
 
 if __name__ == "__main__":

sklego/meta/_grouped_utils.py

@@ -33,10 +33,11 @@ def parse_X_y(X, y, groups, check_X=True, **kwargs) -> nw.DataFrame:
 
     # Convert y and assign it to the frame
     n_samples = X.shape[0]
-    native_space = nw.get_native_namespace(X)
+    y_series = nw.from_dict(
+        data={"tmp": [None] * n_samples if y is None else y}, native_namespace=nw.get_native_namespace(X)
+    )["tmp"]
 
-    y_native = native_space.Series([None] * n_samples) if y is None else native_space.Series(y)
-    return X.with_columns(__sklego_target__=nw.from_native(y_native, allow_series=True))
+    return X.with_columns(__sklego_target__=y_series)
 
 
 def _validate_groups_values(X: nw.DataFrame, groups: List[int] | List[str]) -> None:

sklego/meta/grouped_predictor.py

@@ -123,12 +123,13 @@ def __fit_grouped_estimator(
         if columns is None:
             columns = self._groups
 
+        to_drop = list(set(["__sklego_target__", *columns, *as_list(self.groups)]))
         grouped_estimators = {
             # Fit a clone of the estimators to each group
             (group_name[0] if len(group_name) == 1 else group_name): self.__fit_single_group(
                 group=(group_name[0] if len(group_name) == 1 else group_name),
-                X=nw.to_native(X_grp.drop(["__sklego_target__", *columns, *as_list(self.groups)])),
-                y=(nw.to_native(X_grp.select("__sklego_target__")).to_numpy().reshape(-1) if y is not None else None),
+                X=nw.to_native(X_grp.drop(to_drop)),
+                y=(X_grp.select("__sklego_target__").to_numpy().reshape(-1) if y is not None else None),
             )
             for group_name, X_grp in frame.group_by(columns)
         }
@@ -149,12 +150,12 @@ def __add_shrinkage_column(self, frame, groups=None):
 
         if self.shrinkage is not None and self.use_global_model:
             n_samples = frame.shape[0]
-            native_space = nw.get_native_namespace(frame)
 
             frame = frame.select(
-                nw.from_native(native_space.Series([self._global_col_value] * n_samples), allow_series=True).alias(
-                    self._global_col_name
-                ),
+                nw.from_dict(
+                    data={self._global_col_name: np.full(shape=n_samples, fill_value=self._global_col_value)},
+                    native_namespace=nw.get_native_namespace(frame),
+                )[self._global_col_name],
                 nw.all(),
             )
             groups = [self._global_col_name] if groups is None else [self._global_col_name, *groups]
@@ -285,7 +286,7 @@ def __predict_groups(self, frame: nw.DataFrame, method="predict", groups=None):
                         (group_value[0] if len(group_value) == 1 else group_value),
                         nw.to_native(X_grp.drop(["__sklego_index__", *groups, *as_list(self.groups)])),
                         method=method,
-                    ).set_index(nw.to_native(X_grp["__sklego_index__"]).to_numpy().reshape(-1).astype(int))
+                    ).set_index(X_grp["__sklego_index__"].to_numpy().reshape(-1).astype(int))
                     for group_value, X_grp in frame.group_by(groups)
                 ],
                 axis=0,

sklego/meta/grouped_transformer.py

@@ -164,7 +164,7 @@ def __transform_groups(self, frame: nw.DataFrame):
 
         results = [
             (
-                nw.to_native(X_grp.select("__sklego_index__")).to_numpy().squeeze().astype(int),
+                X_grp.select("__sklego_index__").to_numpy().squeeze().astype(int),
                 self.__transform_single_group(
                     group_name, nw.to_native(X_grp.drop(["__sklego_index__", *self.groups_]))
                 ),

sklego/meta/hierarchical_predictor.py

@@ -281,12 +281,15 @@ def fit(self, X, y=None):
             msg = "Found 0 features, while a minimum of 1 if required."
             raise ValueError(msg)
 
-        native_space = nw.get_native_namespace(X)
-
+        native_namespace = nw.get_native_namespace(X)
+        target_series = nw.from_dict({self._TARGET_NAME: y}, native_namespace=native_namespace)[self._TARGET_NAME]
+        global_series = nw.from_dict({self._GLOBAL_NAME: np.ones(n_samples)}, native_namespace=native_namespace)[
+            self._GLOBAL_NAME
+        ]
         frame = X.with_columns(
             **{
-                self._TARGET_NAME: nw.from_native(native_space.Series(y), allow_series=True),
-                self._GLOBAL_NAME: nw.from_native(native_space.Series([1] * n_samples), allow_series=True),
+                self._TARGET_NAME: target_series,
+                self._GLOBAL_NAME: global_series,
             }
         ).pipe(self.__validate_frame)
 
@@ -318,11 +321,14 @@ def _predict_estimators(self, X, method_name):
             X = nw.from_native(pd.DataFrame(X))
 
         n_samples = X.shape[0]
-        native_space = nw.get_native_namespace(X)
+        native_namespace = nw.get_native_namespace(X)
+        global_series = nw.from_dict({self._GLOBAL_NAME: np.ones(n_samples)}, native_namespace=native_namespace)[
+            self._GLOBAL_NAME
+        ]
 
         frame = X.with_columns(
             **{
-                self._GLOBAL_NAME: nw.from_native(native_space.Series([1] * n_samples), allow_series=True),
+                self._GLOBAL_NAME: global_series,
                 self._INDEX_NAME: np.arange(n_samples),
             }
         ).pipe(self.__validate_frame)
@@ -340,7 +346,7 @@ def _predict_estimators(self, X, method_name):
 
         for level_idx, grp_names in enumerate(self.fitted_levels_):
             for grp_values, grp_frame in frame.group_by(grp_names):
-                grp_idx = nw.to_native(grp_frame.select(self._INDEX_NAME)).to_numpy().reshape(-1)
+                grp_idx = grp_frame.select(self._INDEX_NAME).to_numpy().reshape(-1)
 
                 _estimator, _level = _get_estimator(
                     estimators=self.estimators_,

sklego/model_selection.py

@@ -263,11 +263,13 @@ def update_split_info(indices, j, part, summary):
 
         j = 0
         for i in self.split(nw.to_native(X)):
-            update_split_info(native_namespace.Series(i[0]), j, "train", summary)
-            update_split_info(native_namespace.Series(i[1]), j, "valid", summary)
+            train_info = nw.to_native(nw.from_dict({"tmp": i[0]}, native_namespace=native_namespace)["tmp"])
+            valid_info = nw.to_native(nw.from_dict({"tmp": i[1]}, native_namespace=native_namespace)["tmp"])
+            update_split_info(train_info, j, "train", summary)
+            update_split_info(valid_info, j, "valid", summary)
             j = j + 1
 
-        result = nw.from_native(native_namespace.DataFrame(summary))
+        result = nw.from_dict(summary, native_namespace=native_namespace)
         result = nw.maybe_set_index(result, ["fold", "part"])
         return nw.to_native(result)
 

tests/test_meta/test_grouped_predictor.py

@@ -1,8 +1,10 @@
 from contextlib import nullcontext as does_not_raise
 
+import narwhals.stable.v1 as nw
 import numpy as np
 import pandas as pd
 import polars as pl
+import pyarrow as pa
 import pytest
 from sklearn.dummy import DummyRegressor
 from sklearn.impute import SimpleImputer
@@ -60,32 +62,32 @@ def random_xy_grouped_clf_different_classes(request):
 
 
 @pytest.mark.parametrize("groups, expected", [("diet", {1, 2, 3, 4}), ("chick", set(range(1, 50 + 1)))])
-@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame])
+@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame, pa.table])
 def test_chickweight_keys(groups, expected, frame_func):
-    df = frame_func(load_chicken(as_frame=True))
+    df = nw.from_native(frame_func(load_chicken(as_frame=True).to_dict(orient="list")))
     mod = GroupedPredictor(estimator=LinearRegression(), groups=groups)
-    mod.fit(df[["time", groups]], df["weight"])
+    mod.fit(nw.to_native(df.select("time", groups)), nw.to_native(df["weight"]))
     assert set(mod.estimators_.keys()) == expected
 
 
-@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame])
+@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame, pa.table])
 def test_chickweight_can_do_fallback(frame_func):
-    df = frame_func(load_chicken(as_frame=True))
+    df = nw.from_native(frame_func(load_chicken(as_frame=True).to_dict(orient="list")))
     mod = GroupedPredictor(estimator=LinearRegression(), groups="diet")
-    mod.fit(df[["time", "diet"]], df["weight"])
+    mod.fit(nw.to_native(df.select("time", "diet")), nw.to_native(df["weight"]))
     assert set(mod.estimators_.keys()) == {1, 2, 3, 4}
     to_predict = pd.DataFrame({"time": [21, 21], "diet": [5, 6]})
     assert mod.predict(to_predict).shape == (2,)
     assert mod.predict(to_predict)[0] == mod.predict(to_predict)[1]
 
 
-@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame])
+@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame, pa.table])
 def test_chickweight_can_do_fallback_proba(frame_func):
-    df = frame_func(load_chicken(as_frame=True))
+    df = nw.from_native(frame_func(load_chicken(as_frame=True).to_dict(orient="list")))
 
-    y = np.where(df["weight"] > df["weight"].mean(), 1, 0)
+    y = nw.to_native((df["weight"] > df["weight"].mean()).cast(nw.Int32))
     mod = GroupedPredictor(estimator=LogisticRegression(), groups="diet")
-    mod.fit(df[["time", "diet"]], y)
+    mod.fit(nw.to_native(df.select("time", "diet")), y)
     assert set(mod.estimators_.keys()) == {1, 2, 3, 4}
 
     to_predict = pd.DataFrame({"time": [21, 21], "diet": [5, 6]})

tests/test_meta/test_grouped_transformer.py

@@ -3,6 +3,7 @@
 import numpy as np
 import pandas as pd
 import polars as pl
+import pyarrow as pa
 import pytest
 import sklearn
 from sklearn import clone
@@ -264,9 +265,9 @@ def test_array_with_strings():
     transformer.fit_transform(X)
 
 
-@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame])
+@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame, pa.table])
 def test_df(penguins_df, frame_func):
-    penguins_df = frame_func(penguins_df)
+    penguins_df = frame_func(penguins_df.to_dict(orient="list"))
     meta = GroupedTransformer(StandardScaler(), groups=["island", "sex"])
 
     transformed = meta.fit_transform(penguins_df)
@@ -275,14 +276,14 @@ def test_df(penguins_df, frame_func):
     assert transformed.shape == (penguins_df.shape[0], penguins_df.shape[1] - 2)
 
 
-@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame])
+@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame, pa.table])
 def test_df_missing_group(penguins_df, frame_func):
     meta = GroupedTransformer(StandardScaler(), groups=["island", "sex"])
 
     # Otherwise the fixture is changed
     X = penguins_df.copy()
     X.loc[0, "island"] = None
-    X = frame_func(X)
+    X = frame_func(X.to_dict(orient="list"))
     with pytest.raises(ValueError):
         meta.fit_transform(X)
 
@@ -309,29 +310,31 @@ def test_grouping_column_not_in_array(penguins):
         meta.fit_transform(X[:, :3])
 
 
-@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame])
+@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame, pa.table])
 def test_grouping_column_not_in_df(penguins_df, frame_func):
     meta = GroupedTransformer(StandardScaler(), groups=["island", "unexisting_column"])
 
     # This should raise ValueError
     with pytest.raises(ValueError):
-        meta.fit_transform(frame_func(penguins_df))
+        meta.fit_transform(frame_func(penguins_df.to_dict(orient="list")))
 
 
-@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame])
+@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame, pa.table])
 def test_no_grouping(penguins_df, frame_func):
-    penguins_numeric = frame_func(penguins_df[["bill_length_mm", "bill_depth_mm", "flipper_length_mm", "body_mass_g"]])
+    penguins_numeric = frame_func(
+        penguins_df[["bill_length_mm", "bill_depth_mm", "flipper_length_mm", "body_mass_g"]].to_dict(orient="list")
+    )
 
     meta = GroupedTransformer(StandardScaler(), groups=None)
     nonmeta = StandardScaler()
 
     assert (meta.fit_transform(penguins_numeric) == nonmeta.fit_transform(penguins_numeric)).all()
 
 
-@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame])
+@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame, pa.table])
 def test_with_y(penguins_df, frame_func):
-    X = frame_func(penguins_df.drop(columns=["sex"]))
-    y = penguins_df["sex"]
+    X = frame_func(penguins_df.drop(columns=["sex"]).to_dict(orient="list"))
+    y = penguins_df["sex"].to_numpy()
 
     meta = GroupedTransformer(StandardScaler(), groups="island")
 

tests/test_meta/test_hierarchical_predictor.py

@@ -5,6 +5,7 @@
 import numpy as np
 import pandas as pd
 import polars as pl
+import pyarrow as pa
 import pytest
 from sklearn import clone
 from sklearn.datasets import make_classification, make_regression
@@ -17,7 +18,7 @@
 
 from sklego.meta import HierarchicalClassifier, HierarchicalRegressor
 
-frame_funcs = [pd.DataFrame, pl.DataFrame]
+frame_funcs = [pd.DataFrame, pl.DataFrame, pa.table]
 
 
 @parametrize_with_checks([HierarchicalRegressor(estimator=LinearRegression(), groups=0)])
@@ -60,14 +61,14 @@ def make_hierarchical_dataset(task, frame_func=pd.DataFrame):
     else:
         raise ValueError("Invalid task")
 
-    X_ = (
+    X_ = frame_func(
         pd.DataFrame(X, columns=[f"x_{i}" for i in range(X.shape[1])])
         .assign(
             g_0=1,
             g_1=["A"] * (n_samples // 2) + ["B"] * (n_samples // 2),
             g_2=["X"] * (n_samples // 4) + ["Y"] * (n_samples // 2) + ["Z"] * (n_samples // 4),
         )
-        .pipe(frame_func)
+        .to_dict(orient="list")
     )
     groups = ["g_0", "g_1", "g_2"]
 
@@ -129,7 +130,7 @@ def test_fit_predict(meta_cls, base_estimator, task, fallback_method, shrinkage,
     """Tests that the model can be fit and predict with different configurations of fallback and shrinkage methods if
     X to predict contains same groups as X used to fit.
     """
-    X, y, groups = make_hierarchical_dataset(task, frame_func=frame_funcs[randint(0, 1)])
+    X, y, groups = make_hierarchical_dataset(task, frame_func=frame_funcs[randint(0, 2)])
 
     meta_model = meta_cls(
         estimator=base_estimator,
@@ -159,10 +160,10 @@ def test_fallback(meta_cls, base_estimator, task, fallback_method, context):
     """Tests that the model fails or not when predicting with different fallback methods if X to predict contains
     unseen group values.
     """
-    X, y, groups = make_hierarchical_dataset(task, frame_func=frame_funcs[randint(0, 1)])
+    X, y, groups = make_hierarchical_dataset(task, frame_func=frame_funcs[randint(0, 2)])
 
     meta_model = meta_cls(estimator=base_estimator, groups=groups, fallback_method=fallback_method).fit(X, y)
-    X[groups] = np.ones((X.shape[0], len(groups))) * -1  # Shortcut assignment that works both in pandas and polars
+    X = nw.to_native(nw.from_native(X).with_columns(**{g: nw.lit(-1) for g in groups}))
 
     with context:
         meta_model.predict(X)
@@ -190,7 +191,7 @@ def test_shrinkage(meta_cls, base_estimator, task, metric, shrinkage, kwargs):
     """Tests that the model performance is better than the base estimator when predicting with different shrinkage
     methods.
     """
-    X, y, groups = make_hierarchical_dataset(task, frame_func=frame_funcs[randint(0, 1)])
+    X, y, groups = make_hierarchical_dataset(task, frame_func=frame_funcs[randint(0, 2)])
 
     X_ = nw.from_native(X).drop(groups).pipe(nw.to_native)
     meta_model = meta_cls(

tests/test_meta/test_regression_outlier.py

@@ -1,6 +1,7 @@
 import numpy as np
 import pandas as pd
 import polars as pl
+import pyarrow as pa
 import pytest
 from sklearn.linear_model import LinearRegression, LogisticRegression
 from sklearn.utils.estimator_checks import parametrize_with_checks
@@ -37,7 +38,7 @@ def test_obvious_example():
         assert preds[i] == -1
 
 
-@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame])
+@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame, pa.table])
 def test_obvious_example_dataframe(frame_func):
     # generate random data for illustrative example
     np.random.seed(42)
@@ -54,7 +55,7 @@ def test_obvious_example_dataframe(frame_func):
         assert preds[i] == -1
 
 
-@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame])
+@pytest.mark.parametrize("frame_func", [pd.DataFrame, pl.DataFrame, pa.table])
 def test_raises_error(frame_func):
     # generate random data for illustrative example
     np.random.seed(42)