Adapt to scikit-learn 1.6 estimator tag changes

.gitignore

@@ -144,11 +144,13 @@ credentials.csv
 .bloop
 
 # python tests
+*.bin
 demo/**/*.txt
 *.dmatrix
 .hypothesis
 __MACOSX/
 model*.json
+/tests/python/models/models/
 
 # R tests
 *.htm

python-package/pyproject.toml

@@ -63,6 +63,8 @@ disable = [
     "import-error",
     "attribute-defined-outside-init",
     "import-outside-toplevel",
+    "too-few-public-methods",
+    "too-many-ancestors",
     "too-many-nested-blocks",
     "unsubscriptable-object",
     "useless-object-inheritance"

python-package/xgboost/compat.py

@@ -43,6 +43,7 @@ def lazy_isinstance(instance: Any, module: str, name: str) -> bool:
 
 # sklearn
 try:
+    from sklearn import __version__ as _sklearn_version
     from sklearn.base import BaseEstimator as XGBModelBase
     from sklearn.base import ClassifierMixin as XGBClassifierBase
     from sklearn.base import RegressorMixin as XGBRegressorBase
@@ -55,20 +56,43 @@ def lazy_isinstance(instance: Any, module: str, name: str) -> bool:
         from sklearn.cross_validation import KFold as XGBKFold
         from sklearn.cross_validation import StratifiedKFold as XGBStratifiedKFold
 
+    # sklearn.utils Tags types can be imported unconditionally once
+    # xgboost's minimum scikit-learn version is 1.6 or higher
+    try:
+        from sklearn.utils import ClassifierTags as _sklearn_ClassifierTags
+        from sklearn.utils import RegressorTags as _sklearn_RegressorTags
+        from sklearn.utils import Tags as _sklearn_Tags
+    except ImportError:
+        _sklearn_ClassifierTags = object
+        _sklearn_RegressorTags = object
+        _sklearn_Tags = object
+
     SKLEARN_INSTALLED = True
 
 except ImportError:
     SKLEARN_INSTALLED = False
 
     # used for compatibility without sklearn
-    XGBModelBase = object
-    XGBClassifierBase = object
-    XGBRegressorBase = object
-    LabelEncoder = object
+    class XGBModelBase:  # type: ignore[no-redef]
+        """Dummy class for sklearn.base.BaseEstimator."""
+
+    class XGBClassifierBase:  # type: ignore[no-redef]
+        """Dummy class for sklearn.base.ClassifierMixin."""
+
+    class XGBRegressorBase:  # type: ignore[no-redef]
+        """Dummy class for sklearn.base.RegressorMixin."""
+
+    class LabelEncoder:  # type: ignore[no-redef]
+        """Dummy class for sklearn.preprocessing.LabelEncoder."""
 
     XGBKFold = None
     XGBStratifiedKFold = None
 
+    _sklearn_ClassifierTags = object
+    _sklearn_RegressorTags = object
+    _sklearn_Tags = object
+    _sklearn_version = object
+
 
 _logger = logging.getLogger(__name__)
 

python-package/xgboost/core.py

@@ -430,7 +430,7 @@ def c_array(
 def from_array_interface(interface: dict) -> NumpyOrCupy:
     """Convert array interface to numpy or cupy array"""
 
-    class Array:  # pylint: disable=too-few-public-methods
+    class Array:
         """Wrapper type for communicating with numpy and cupy."""
 
         _interface: Optional[dict] = None

python-package/xgboost/dask/__init__.py

@@ -1,7 +1,6 @@
 # pylint: disable=too-many-arguments, too-many-locals
 # pylint: disable=missing-class-docstring, invalid-name
 # pylint: disable=too-many-lines
-# pylint: disable=too-few-public-methods
 """
 Dask extensions for distributed training
 ----------------------------------------

python-package/xgboost/sklearn.py

@@ -35,6 +35,10 @@
     XGBClassifierBase,
     XGBModelBase,
     XGBRegressorBase,
+    _sklearn_ClassifierTags,
+    _sklearn_RegressorTags,
+    _sklearn_Tags,
+    _sklearn_version,
     import_cupy,
 )
 from .config import config_context
@@ -54,7 +58,7 @@
 from .training import train
 
 
-class XGBRankerMixIn:  # pylint: disable=too-few-public-methods
+class XGBRankerMixIn:
     """MixIn for ranking, defines the _estimator_type usually defined in scikit-learn
     base classes.
 
@@ -79,7 +83,7 @@ def _can_use_qdm(tree_method: Optional[str], device: Optional[str]) -> bool:
     return tree_method in ("hist", "gpu_hist", None, "auto") and not_sycl
 
 
-class _SklObjWProto(Protocol):  # pylint: disable=too-few-public-methods
+class _SklObjWProto(Protocol):
     def __call__(
         self,
         y_true: ArrayLike,
@@ -805,6 +809,41 @@ def _more_tags(self) -> Dict[str, bool]:
             tags["non_deterministic"] = True
         return tags
 
+    @staticmethod
+    def _update_sklearn_tags_from_dict(
+        *,
+        tags: _sklearn_Tags,
+        tags_dict: Dict[str, bool],
+    ) -> _sklearn_Tags:
+        """Update ``sklearn.utils.Tags`` inherited from ``scikit-learn`` base classes.
+
+        ``scikit-learn`` 1.6 introduced a dataclass-based interface for estimator tags.
+        ref: https://github.com/scikit-learn/scikit-learn/pull/29677
+
+        This method handles updating that instance based on the values in ``self._more_tags()``.
+        """
+        tags.non_deterministic = tags_dict.get("non_deterministic", False)
+        tags.no_validation = tags_dict["no_validation"]
+        tags.input_tags.allow_nan = tags_dict["allow_nan"]
+        return tags
+
+    def __sklearn_tags__(self) -> _sklearn_Tags:
+        # XGBModelBase.__sklearn_tags__() cannot be called unconditionally,
+        # because that method isn't defined for scikit-learn<1.6
+        if not hasattr(XGBModelBase, "__sklearn_tags__"):
+            err_msg = (
+                "__sklearn_tags__() should not be called when using scikit-learn<1.6. "
+                f"Detected version: {_sklearn_version}"
+            )
+            raise AttributeError(err_msg)
+
+        # take whatever tags are provided by BaseEstimator, then modify
+        # them with XGBoost-specific values
+        return self._update_sklearn_tags_from_dict(
+            tags=XGBModelBase.__sklearn_tags__(self),  # pylint: disable=no-member
+            tags_dict=self._more_tags(),
+        )
+
     def __sklearn_is_fitted__(self) -> bool:
         return hasattr(self, "_Booster")
 
@@ -898,13 +937,30 @@ def get_params(self, deep: bool = True) -> Dict[str, Any]:
         """Get parameters."""
         # Based on: https://stackoverflow.com/questions/59248211
         # The basic flow in `get_params` is:
-        # 0. Return parameters in subclass first, by using inspect.
-        # 1. Return parameters in `XGBModel` (the base class).
+        # 0. Return parameters in subclass (self.__class__) first, by using inspect.
+        # 1. Return parameters in all parent classes (especially `XGBModel`).
         # 2. Return whatever in `**kwargs`.
         # 3. Merge them.
+        #
+        # This needs to accommodate being called recursively in the following
+        # inheritance graphs (and similar for classification and ranking):
+        #
+        #   XGBRFRegressor -> XGBRegressor -> XGBModel -> BaseEstimator
+        #                     XGBRegressor -> XGBModel -> BaseEstimator
+        #                                     XGBModel -> BaseEstimator
+        #
         params = super().get_params(deep)
         cp = copy.copy(self)
-        cp.__class__ = cp.__class__.__bases__[0]
+        # if the immediate parent is a mixin, skip it (mixins don't define get_params())
+        if cp.__class__.__bases__[0] in (
+            XGBClassifierBase,
+            XGBRankerMixIn,
+            XGBRegressorBase,
+        ):
+            cp.__class__ = cp.__class__.__bases__[1]
+        # otherwise, run get_params() from the immediate parent class
+        else:
+            cp.__class__ = cp.__class__.__bases__[0]
         params.update(cp.__class__.get_params(cp, deep))
         # if kwargs is a dict, update params accordingly
         if hasattr(self, "kwargs") and isinstance(self.kwargs, dict):
@@ -1481,7 +1537,7 @@ def _cls_predict_proba(n_classes: int, prediction: PredtT, vstack: Callable) ->
         Number of boosting rounds.
 """,
 )
-class XGBClassifier(XGBModel, XGBClassifierBase):
+class XGBClassifier(XGBClassifierBase, XGBModel):
     # pylint: disable=missing-docstring,invalid-name,too-many-instance-attributes
     @_deprecate_positional_args
     def __init__(
@@ -1497,6 +1553,15 @@ def _more_tags(self) -> Dict[str, bool]:
         tags["multilabel"] = True
         return tags
 
+    def __sklearn_tags__(self) -> _sklearn_Tags:
+        tags = XGBModel.__sklearn_tags__(self)
+        tags.estimator_type = "classifier"
+        tags_dict = self._more_tags()
+        tags.classifier_tags = _sklearn_ClassifierTags(
+            multi_label=tags_dict["multilabel"]
+        )
+        return tags
+
     @_deprecate_positional_args
     def fit(
         self,
@@ -1769,7 +1834,7 @@ def fit(
     "Implementation of the scikit-learn API for XGBoost regression.",
     ["estimators", "model", "objective"],
 )
-class XGBRegressor(XGBModel, XGBRegressorBase):
+class XGBRegressor(XGBRegressorBase, XGBModel):
     # pylint: disable=missing-docstring
     @_deprecate_positional_args
     def __init__(
@@ -1783,6 +1848,15 @@ def _more_tags(self) -> Dict[str, bool]:
         tags["multioutput_only"] = False
         return tags
 
+    def __sklearn_tags__(self) -> _sklearn_Tags:
+        tags = XGBModel.__sklearn_tags__(self)
+        tags.estimator_type = "regressor"
+        tags_dict = self._more_tags()
+        tags.regressor_tags = _sklearn_RegressorTags()
+        tags.target_tags.multi_output = tags_dict["multioutput"]
+        tags.target_tags.single_output = not tags_dict["multioutput_only"]
+        return tags
+
 
 @xgboost_model_doc(
     "scikit-learn API for XGBoost random forest regression.",
@@ -1910,7 +1984,7 @@ def _get_qid(
         `qid` can be a special column of input `X` instead of a separated parameter, see
         :py:meth:`fit` for more info.""",
 )
-class XGBRanker(XGBModel, XGBRankerMixIn):
+class XGBRanker(XGBRankerMixIn, XGBModel):
     # pylint: disable=missing-docstring,too-many-arguments,invalid-name
     @_deprecate_positional_args
     def __init__(self, *, objective: str = "rank:ndcg", **kwargs: Any):

python-package/xgboost/spark/core.py

@@ -2,8 +2,8 @@
 
 import base64
 
-# pylint: disable=fixme, too-many-ancestors, protected-access, no-member, invalid-name
-# pylint: disable=too-few-public-methods, too-many-lines, too-many-branches
+# pylint: disable=fixme, protected-access, no-member, invalid-name
+# pylint: disable=too-many-lines, too-many-branches
 import json
 import logging
 import os

python-package/xgboost/spark/estimator.py

@@ -1,7 +1,6 @@
 """Xgboost pyspark integration submodule for estimator API."""
 
-# pylint: disable=too-many-ancestors
-# pylint: disable=fixme, too-many-ancestors, protected-access, no-member, invalid-name
+# pylint: disable=fixme, protected-access, no-member, invalid-name
 # pylint: disable=unused-argument, too-many-locals
 
 import warnings

python-package/xgboost/spark/params.py

@@ -2,7 +2,6 @@
 
 from typing import Dict
 
-# pylint: disable=too-few-public-methods
 from pyspark.ml.param import TypeConverters
 from pyspark.ml.param.shared import Param, Params
 

python-package/xgboost/spark/utils.py

@@ -47,7 +47,7 @@ def _get_default_params_from_func(
     return filtered_params_dict
 
 
-class CommunicatorContext(CCtx):  # pylint: disable=too-few-public-methods
+class CommunicatorContext(CCtx):
     """Context with PySpark specific task ID."""
 
     def __init__(self, context: BarrierTaskContext, **args: CollArgsVals) -> None:

python-package/xgboost/testing/data.py

@@ -566,7 +566,7 @@ def is_binary(self) -> bool:
         return self.max_rel == 1
 
 
-class PBM:  # pylint: disable=too-few-public-methods
+class PBM:
     """Simulate click data with position bias model. There are other models available in
     `ULTRA <https://github.com/ULTR-Community/ULTRA.git>`_ like the cascading model.
 

tests/python/test_with_sklearn.py

@@ -2,6 +2,7 @@
 import os
 import pickle
 import random
+import re
 import tempfile
 import warnings
 from typing import Callable, Optional
@@ -1517,7 +1518,7 @@ def test_tags() -> None:
         assert tags["multioutput"] is True
         assert tags["multioutput_only"] is False
 
-    for clf in [xgb.XGBClassifier()]:
+    for clf in [xgb.XGBClassifier(), xgb.XGBRFClassifier()]:
         tags = clf._more_tags()
         assert "multioutput" not in tags
         assert tags["multilabel"] is True
@@ -1526,6 +1527,58 @@ def test_tags() -> None:
     assert "multioutput" not in tags
 
 
+# the try-excepts in this test should be removed once xgboost's
+# minimum supported scikit-learn version is at least 1.6
+def test_sklearn_tags():
+
+    def _assert_has_xgbmodel_tags(tags):
+        # values set by XGBModel.__sklearn_tags__()
+        assert tags.non_deterministic is False
+        assert tags.no_validation is True
+        assert tags.input_tags.allow_nan is True
+
+    for reg in [xgb.XGBRegressor(), xgb.XGBRFRegressor()]:
+        try:
+            # if no AttributeError was thrown, we must be using scikit-learn>=1.6,
+            # and so the actual effects of __sklearn_tags__() should be tested
+            tags = reg.__sklearn_tags__()
+            _assert_has_xgbmodel_tags(tags)
+            # regressor-specific values
+            assert tags.estimator_type == "regressor"
+            assert tags.regressor_tags is not None
+            assert tags.classifier_tags is None
+            assert tags.target_tags.multi_output is True
+            assert tags.target_tags.single_output is True
+        except AttributeError as err:
+            # only the exact error we expected to be raised should be raised
+            assert bool(re.search(r"__sklearn_tags__.* should not be called", str(err)))
+
+    for clf in [xgb.XGBClassifier(), xgb.XGBRFClassifier()]:
+        try:
+            # if no AttributeError was thrown, we must be using scikit-learn>=1.6,
+            # and so the actual effects of __sklearn_tags__() should be tested
+            tags = clf.__sklearn_tags__()
+            _assert_has_xgbmodel_tags(tags)
+            # classifier-specific values
+            assert tags.estimator_type == "classifier"
+            assert tags.regressor_tags is None
+            assert tags.classifier_tags is not None
+            assert tags.classifier_tags.multi_label is True
+        except AttributeError as err:
+            # only the exact error we expected to be raised should be raised
+            assert bool(re.search(r"__sklearn_tags__.* should not be called", str(err)))
+
+    for rnk in [xgb.XGBRanker(),]:
+        try:
+            # if no AttributeError was thrown, we must be using scikit-learn>=1.6,
+            # and so the actual effects of __sklearn_tags__() should be tested
+            tags = rnk.__sklearn_tags__()
+            _assert_has_xgbmodel_tags(tags)
+        except AttributeError as err:
+            # only the exact error we expected to be raised should be raised
+            assert bool(re.search(r"__sklearn_tags__.* should not be called", str(err)))
+
+
 def test_doc_link() -> None:
     for est in [
         xgb.XGBRegressor(),