[ENH] Let cuDF handle input types for label encoder. (rapidsai#5783)

cuDF handles more types than the label encoder currently does (like torch tensor). This PR delegates the type checking to cuDF.

- Let cuDF handle input types for label encoder.
- Small cleanups.

Authors:
  - Jiaming Yuan (https://github.com/trivialfis)
  - Dante Gama Dessavre (https://github.com/dantegd)

Approvers:
  - Dante Gama Dessavre (https://github.com/dantegd)

URL: rapidsai#5783

python/cuml/_thirdparty/sklearn/utils/validation.py

@@ -225,6 +225,8 @@ def check_is_fitted(estimator, attributes=None, *, msg=None, all_or_any=all):
         if not isinstance(attributes, (list, tuple)):
             attributes = [attributes]
         attrs = all_or_any([hasattr(estimator, attr) for attr in attributes])
+    elif hasattr(estimator, "__sklearn_is_fitted__"):
+        attrs = estimator.__sklearn_is_fitted__()
     else:
         attrs = [v for v in vars(estimator)
                  if v.endswith("_") and not v.startswith("__")]

python/cuml/preprocessing/LabelEncoder.py

@@ -1,5 +1,5 @@
 #
-# Copyright (c) 2019-2023, NVIDIA CORPORATION.
+# Copyright (c) 2019-2024, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -14,16 +14,27 @@
 # limitations under the License.
 #
 
-from cuml.common.exceptions import NotFittedError
-from cuml.internals.safe_imports import cpu_only_import_from
-from cuml import Base
-from cuml.internals.safe_imports import cpu_only_import
-from cuml.internals.safe_imports import gpu_only_import
+from typing import TYPE_CHECKING
 
-cudf = gpu_only_import("cudf")
-cp = gpu_only_import("cupy")
-np = cpu_only_import("numpy")
-pdSeries = cpu_only_import_from("pandas", "Series")
+from cuml import Base
+from cuml._thirdparty.sklearn.utils.validation import check_is_fitted
+from cuml.common.exceptions import NotFittedError
+from cuml.internals.safe_imports import (
+    cpu_only_import,
+    cpu_only_import_from,
+    gpu_only_import,
+)
+
+if TYPE_CHECKING:
+    import cudf
+    import cupy as cp
+    import numpy as np
+    from pandas import Series as pdSeries
+else:
+    cudf = gpu_only_import("cudf")
+    cp = gpu_only_import("cupy")
+    np = cpu_only_import("numpy")
+    pdSeries = cpu_only_import_from("pandas", "Series")
 
 
 class LabelEncoder(Base):
@@ -125,7 +136,7 @@ def __init__(
         handle=None,
         verbose=False,
         output_type=None,
-    ):
+    ) -> None:
 
         super().__init__(
             handle=handle, verbose=verbose, output_type=output_type
@@ -136,13 +147,8 @@ def __init__(
         self._fitted: bool = False
         self.handle_unknown = handle_unknown
 
-    def _check_is_fitted(self):
-        if not self._fitted:
-            msg = (
-                "This LabelEncoder instance is not fitted yet. Call 'fit' "
-                "with appropriate arguments before using this estimator."
-            )
-            raise NotFittedError(msg)
+    def __sklearn_is_fitted__(self) -> bool:
+        return self.classes_ is not None
 
     def _validate_keywords(self):
         if self.handle_unknown not in ("error", "ignore"):
@@ -174,17 +180,13 @@ def fit(self, y, _classes=None):
         self._validate_keywords()
 
         if _classes is None:
-            y = (
-                self._to_cudf_series(y)
-                .drop_duplicates()
-                .sort_values(ignore_index=True)
-            )  # dedupe and sort
+            # dedupe and sort
+            y = cudf.Series(y).drop_duplicates().sort_values(ignore_index=True)
             self.classes_ = y
         else:
             self.classes_ = _classes
 
         self.dtype = y.dtype if y.dtype != cp.dtype("O") else str
-        self._fitted = True
         return self
 
     def transform(self, y) -> cudf.Series:
@@ -211,11 +213,9 @@ def transform(self, y) -> cudf.Series:
         KeyError
             if a category appears that was not seen in `fit`
         """
-        y = self._to_cudf_series(y)
+        check_is_fitted(self)
 
-        self._check_is_fitted()
-
-        y = y.astype("category")
+        y = cudf.Series(y, dtype="category")
 
         encoded = y.cat.set_categories(self.classes_)._column.codes
         encoded = cudf.Series(encoded, index=y.index)
@@ -233,13 +233,12 @@ def fit_transform(self, y, z=None) -> cudf.Series:
         `LabelEncoder().fit(y).transform(y)`
         """
 
-        y = self._to_cudf_series(y)
+        y = cudf.Series(y)
         self.dtype = y.dtype if y.dtype != cp.dtype("O") else str
 
         y = y.astype("category")
         self.classes_ = y._column.categories
 
-        self._fitted = True
         return cudf.Series(y._column.codes, index=y.index)
 
     def inverse_transform(self, y: cudf.Series) -> cudf.Series:
@@ -258,9 +257,9 @@ def inverse_transform(self, y: cudf.Series) -> cudf.Series:
             Reverted labels
         """
         # check LabelEncoder is fitted
-        self._check_is_fitted()
+        check_is_fitted(self)
         # check input type is cudf.Series
-        y = self._to_cudf_series(y)
+        y = cudf.Series(y)
 
         # check if ord_label out of bound
         ord_label = y.unique()
@@ -285,20 +284,3 @@ def get_param_names(self):
         return super().get_param_names() + [
             "handle_unknown",
         ]
-
-    def _to_cudf_series(self, y):
-        if isinstance(y, pdSeries):
-            y = cudf.from_pandas(y)
-        elif isinstance(y, cp.ndarray):
-            y = cudf.Series(y)
-        elif isinstance(y, np.ndarray):
-            y = cudf.Series(y)
-        elif not isinstance(y, cudf.Series):
-            msg = (
-                "input should be either 'cupy.ndarray'"
-                " or 'numpy.ndarray' or 'pandas.Series',"
-                " or 'cudf.Series'"
-                "got {0}.".format(type(y))
-            )
-            raise TypeError(msg)
-        return y

python/cuml/tests/dask/test_dask_label_encoder.py

@@ -11,12 +11,13 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from cuml.common.exceptions import NotFittedError
 import pytest
-from cuml.internals.safe_imports import cpu_only_import
+
 import cuml
+from cuml._thirdparty.sklearn.utils.validation import check_is_fitted
+from cuml.common.exceptions import NotFittedError
 from cuml.dask.preprocessing.LabelEncoder import LabelEncoder
-from cuml.internals.safe_imports import gpu_only_import
+from cuml.internals.safe_imports import cpu_only_import, gpu_only_import
 
 cudf = gpu_only_import("cudf")
 np = cpu_only_import("numpy")
@@ -51,7 +52,7 @@ def test_labelencoder_transform(length, cardinality, client):
     tmp = cudf.Series(np.random.choice(cardinality, (length,)))
     df = dask_cudf.from_cudf(tmp, npartitions=len(client.has_what()))
     le = LabelEncoder().fit(df)
-    assert le._fitted
+    check_is_fitted(le)
 
     encoded = le.transform(df)
 
@@ -69,7 +70,7 @@ def test_labelencoder_unseen(client):
         npartitions=len(client.has_what()),
     )
     le = LabelEncoder().fit(df)
-    assert le._fitted
+    check_is_fitted(le)
 
     with pytest.raises(KeyError):
         tmp = dask_cudf.from_cudf(
@@ -141,7 +142,7 @@ def test_inverse_transform(
         le.fit_transform(orig_label)
     else:
         le.fit(orig_label)
-    assert le._fitted is True
+    check_is_fitted(le)
 
     # test if inverse_transform is correct
     reverted = le.inverse_transform(ord_label)
@@ -175,7 +176,7 @@ def test_empty_input(empty, ord_label, client):
     ord_label = dask_cudf.from_cudf(ord_label, npartitions=n_workers)
     le = LabelEncoder()
     le.fit(empty)
-    assert le._fitted is True
+    check_is_fitted(le)
 
     # test if correctly raies ValueError
     with pytest.raises(ValueError, match="y contains previously unseen label"):
@@ -184,7 +185,7 @@ def test_empty_input(empty, ord_label, client):
     # check fit_transform()
     le = LabelEncoder()
     transformed = le.fit_transform(empty).compute()
-    assert le._fitted is True
+    check_is_fitted(le)
     assert len(transformed) == 0
 
 

python/cuml/tests/test_label_encoder.py

@@ -12,12 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from cuml.common.exceptions import NotFittedError
 import pytest
-from cuml.internals.safe_imports import cpu_only_import
+
+from cuml._thirdparty.sklearn.utils.validation import check_is_fitted
+from cuml.common.exceptions import NotFittedError
+from cuml.internals.safe_imports import cpu_only_import, gpu_only_import
 from cuml.preprocessing.LabelEncoder import LabelEncoder
-from cuml.internals.safe_imports import gpu_only_import
 
+pd = cpu_only_import("pandas")
 cudf = gpu_only_import("cudf")
 np = cpu_only_import("numpy")
 cp = gpu_only_import("cupy")
@@ -46,7 +48,7 @@ def test_labelencoder_transform(length, cardinality):
     """Try fitting and then encoding a small subset of the df"""
     df = cudf.Series(np.random.choice(cardinality, (length,)))
     le = LabelEncoder().fit(df)
-    assert le._fitted
+    check_is_fitted(le)
 
     subset = df.iloc[0 : df.shape[0] // 2]
     encoded = le.transform(subset)
@@ -62,7 +64,7 @@ def test_labelencoder_unseen():
     """Try encoding a value that was not present during fitting"""
     df = cudf.Series(np.random.choice(10, (10,)))
     le = LabelEncoder().fit(df)
-    assert le._fitted
+    check_is_fitted(le)
 
     with pytest.raises(KeyError):
         le.transform(cudf.Series([-1]))
@@ -72,7 +74,7 @@ def test_labelencoder_unfitted():
     """Try calling `.transform()` without fitting first"""
     df = cudf.Series(np.random.choice(10, (10,)))
     le = LabelEncoder()
-    assert not le._fitted
+    assert not le.__sklearn_is_fitted__()
 
     with pytest.raises(NotFittedError):
         le.transform(df)
@@ -117,7 +119,7 @@ def test_inverse_transform(
         le.fit_transform(orig_label)
     else:
         le.fit(orig_label)
-    assert le._fitted is True
+    check_is_fitted(le)
 
     # test if inverse_transform is correct
     reverted = le.inverse_transform(ord_label)
@@ -132,7 +134,7 @@ def test_unfitted_inverse_transform():
     """Try calling `.inverse_transform()` without fitting first"""
     df = cudf.Series(np.random.choice(10, (10,)))
     le = LabelEncoder()
-    assert not le._fitted
+    assert not le.__sklearn_is_fitted__()
 
     with pytest.raises(NotFittedError):
         le.transform(df)
@@ -145,7 +147,7 @@ def test_empty_input(empty, ord_label):
     # prepare LabelEncoder
     le = LabelEncoder()
     le.fit(empty)
-    assert le._fitted is True
+    check_is_fitted(le)
 
     # test if correctly raies ValueError
     with pytest.raises(ValueError, match="y contains previously unseen label"):
@@ -154,7 +156,7 @@ def test_empty_input(empty, ord_label):
     # check fit_transform()
     le = LabelEncoder()
     transformed = le.fit_transform(empty)
-    assert le._fitted is True
+    check_is_fitted(le)
     assert len(transformed) == 0
 
 
@@ -187,18 +189,29 @@ def _array_to_similarity_mat(x):
 
 @pytest.mark.parametrize("length", [10, 1000])
 @pytest.mark.parametrize("cardinality", [5, 10, 50])
-@pytest.mark.parametrize("dtype", ["cupy", "numpy"])
-def test_labelencoder_fit_transform_cupy_numpy(length, cardinality, dtype):
-    """Try encoding the cupy array"""
+@pytest.mark.parametrize("dtype", ["cupy", "numpy", "pd"])
+def test_labelencoder_fit_transform_cupy_numpy_pd(length, cardinality, dtype):
+    """Try encoding with various types"""
     x = cp.random.choice(cardinality, (length,))
+    # to series
     if dtype == "numpy":
         x = x.get()
+    elif dtype == "pd":
+        x = pd.Series(x.get())
     encoded = LabelEncoder().fit_transform(x)
 
-    x_arr = _array_to_similarity_mat(x)
+    if dtype == "pd":
+        x_arr = _df_to_similarity_mat(x)
+    else:
+        x_arr = _array_to_similarity_mat(x)
+
     encoded_arr = _array_to_similarity_mat(encoded.values)
-    if dtype == "numpy":
+
+    # to array
+    if dtype == "numpy" or dtype == "pd":
         encoded_arr = encoded_arr.get()
+    if dtype == "pd":
+        x = x.to_numpy()
     assert ((encoded_arr == encoded_arr.T) == (x == x_arr.T)).all()
 
 
@@ -229,7 +242,7 @@ def test_inverse_transform_cupy_numpy(
         le.fit_transform(orig_label)
     else:
         le.fit(orig_label)
-    assert le._fitted is True
+    check_is_fitted(le)
 
     # test if inverse_transform is correct
     reverted = le.inverse_transform(ord_label)