[maintenance] prepare next release for sklearn 1.8 API breaking changes (#2578)

* Update validation.py

* Update dispatcher.py

* Update dispatcher.py

* Update dispatcher.py

* Update run_test.sh

* Update _pairwise.py

* Update _pairwise.py

* Update _pairwise.py

* Update _pairwise.py

* Update dispatcher.py

* Update run_test.sh

* Update dispatcher.py

* Update _dlpack.py

* Update _common.py

* Update logistic_regression.py

* Update logistic_path.py

* Update _common.py

* attempt to fix

* attempt to fix

* attempt to fix

* add logistic regression to testing

* formatting

* force value of cv

* ugly solution

* remove vestigial newline

* isort fixes

* move imports to respect sklearn versions

* Update setup_sklearn.sh

* Update __init__.py

* Update __init__.py

* Update __init__.py

* Update test_model_builders.py

* formatting

* try to fix logreg testing

* try to fix issues with testing mb

* fix logic again

* try to fix daal4py again

* Update test_logreg.py

* Update test_logreg.py

* formatting

* Update deselected_tests.yaml

* Update logistic_regression.py

* Update setup_sklearn.sh

* changes as requested from reviews

* make suggestion from review

* fix again

* formatting fixg

* false flag on pyright

* Update _common.py

* Update test_patching.py

* Update test_patching.py

* fix para

Author: icfaust

.ci/scripts/select_sklearn_tests.py

@@ -41,7 +41,12 @@ def parse_tests_tree(entry, prefix=""):
     "covariance/tests": "test_covariance.py",
     "decomposition/tests": ["test_pca.py", "test_incremental_pca.py"],
     "ensemble/tests": "test_forest.py",
-    "linear_model/tests": ["test_base.py", "test_coordinate_descent.py", "test_ridge.py"],
+    "linear_model/tests": [
+        "test_base.py",
+        "test_coordinate_descent.py",
+        "test_logistic.py",
+        "test_ridge.py",
+    ],
     "manifold/tests": "test_t_sne.py",
     "metrics/tests": ["test_pairwise.py", "test_ranking.py"],
     "model_selection/tests": ["test_split.py", "test_validation.py"],

daal4py/mb/__init__.py

@@ -53,15 +53,26 @@ def convert_model(model) -> "GBTDAALModel | LogisticDAALModel":
         offers faster prediction methods.
     """
     if isinstance(model, LogisticRegression):
+        # The multi_class keyword is removed in scikit-learn 1.8, and OvR functionality
+        # has been replaced by other estimators. Therefore checking for linear classifiers
+        # only dependent on the solver.
         if model.classes_.shape[0] > 2:
-            if (model.multi_class == "ovr") or (
-                model.multi_class == "auto" and model.solver == "liblinear"
-            ):
-                raise TypeError(
-                    "Supplied 'model' object is a linear classifier, but not multinomial logistic"
-                    " (hint: pass multi_class='multinomial' to 'LogisticRegression')."
-                )
-        elif (model.classes_.shape[0] == 2) and (model.multi_class == "multinomial"):
+            if not hasattr(model, "multi_class"):
+                if model.solver == "liblinear":
+                    raise TypeError(
+                        "Supplied 'model' object is a linear classifier, but not multinomial logistic"
+                    )
+            else:
+                if (model.multi_class == "ovr") or (
+                    model.multi_class == "auto" and model.solver == "liblinear"
+                ):
+                    raise TypeError(
+                        "Supplied 'model' object is a linear classifier, but not multinomial logistic"
+                        " (hint: pass multi_class='multinomial' to 'LogisticRegression')."
+                    )
+        elif (model.classes_.shape[0] == 2) and (
+            getattr(model, "multi_class", "auto") == "multinomial"
+        ):
             raise TypeError(
                 "Supplied 'model' object is not a logistic regressor "
                 "(hint: pass multi_class='auto' to 'LogisticRegression')."

daal4py/sklearn/linear_model/logistic_path.py

@@ -20,15 +20,14 @@
 import scipy.optimize as optimize
 import scipy.sparse as sparse
 import sklearn.linear_model._logistic as logistic_module
-from sklearn.linear_model._sag import sag_solver
-from sklearn.utils import (
-    check_array,
-    check_consistent_length,
-    check_random_state,
-    compute_class_weight,
+from sklearn.linear_model._logistic import _LOGISTIC_SOLVER_CONVERGENCE_MSG
+from sklearn.linear_model._logistic import (
+    LogisticRegression as LogisticRegression_original,
 )
+from sklearn.linear_model._logistic import _check_solver
+from sklearn.utils import check_array, check_consistent_length, check_random_state
 from sklearn.utils.optimize import _check_optimize_result, _newton_cg
-from sklearn.utils.validation import _check_sample_weight, check_is_fitted
+from sklearn.utils.validation import check_is_fitted
 
 import daal4py as d4p
 
@@ -44,35 +43,6 @@
     _daal4py_loss_and_grad,
 )
 
-if sklearn_check_version("1.1"):
-    from sklearn._loss.loss import HalfBinomialLoss, HalfMultinomialLoss
-    from sklearn.linear_model._linear_loss import LinearModelLoss
-    from sklearn.linear_model._logistic import _LOGISTIC_SOLVER_CONVERGENCE_MSG
-    from sklearn.linear_model._logistic import (
-        LogisticRegression as LogisticRegression_original,
-    )
-    from sklearn.linear_model._logistic import (
-        _check_multi_class,
-        _check_solver,
-        _fit_liblinear,
-    )
-else:
-    from sklearn.linear_model._logistic import _LOGISTIC_SOLVER_CONVERGENCE_MSG
-    from sklearn.linear_model._logistic import (
-        LogisticRegression as LogisticRegression_original,
-    )
-    from sklearn.linear_model._logistic import (
-        _check_multi_class,
-        _check_solver,
-        _fit_liblinear,
-        _logistic_grad_hess,
-        _logistic_loss,
-        _logistic_loss_and_grad,
-        _multinomial_grad_hess,
-        _multinomial_loss,
-        _multinomial_loss_grad,
-    )
-
 if sklearn_check_version("1.7.1"):
     from sklearn.utils.fixes import _get_additional_lbfgs_options_dict
 else:
@@ -86,6 +56,25 @@ def _get_additional_lbfgs_options_dict(k, v):
 from sklearn.preprocessing import LabelBinarizer, LabelEncoder
 
 
+# This code is a patch for sklearn 1.8, which is related to https://github.com/scikit-learn/scikit-learn/pull/32073
+# where the multi_class keyword is deprecated and this aspect is removed.
+def _check_multi_class(multi_class, solver, n_classes):
+    """Computes the multi class type, either "multinomial" or "ovr".
+    For `n_classes` > 2 and a solver that supports it, returns "multinomial".
+    For all other cases, in particular binary classification, return "ovr".
+    """
+    if multi_class == "auto":
+        if solver in ("liblinear",):
+            multi_class = "ovr"
+        elif n_classes > 2:
+            multi_class = "multinomial"
+        else:
+            multi_class = "ovr"
+    if multi_class == "multinomial" and solver in ("liblinear",):
+        raise ValueError("Solver %s does not support a multinomial backend." % solver)
+    return multi_class
+
+
 # Code adapted from sklearn.linear_model.logistic version 0.21
 def __logistic_regression_path(
     X,
@@ -110,46 +99,6 @@ def __logistic_regression_path(
     l1_ratio=None,
     n_threads=1,
 ):
-    _patching_status = PatchingConditionsChain(
-        "sklearn.linear_model.LogisticRegression.fit"
-    )
-    _dal_ready = _patching_status.and_conditions(
-        [
-            (
-                solver in ["lbfgs", "newton-cg"],
-                f"'{solver}' solver is not supported. "
-                "Only 'lbfgs' and 'newton-cg' solvers are supported.",
-            ),
-            (not sparse.issparse(X), "X is sparse. Sparse input is not supported."),
-            (sample_weight is None, "Sample weights are not supported."),
-            (class_weight is None, "Class weights are not supported."),
-        ]
-    )
-    if not _dal_ready:
-        _patching_status.write_log()
-        return lr_path_original(
-            X,
-            y,
-            pos_class=pos_class,
-            Cs=Cs,
-            fit_intercept=fit_intercept,
-            max_iter=max_iter,
-            tol=tol,
-            verbose=verbose,
-            solver=solver,
-            coef=coef,
-            class_weight=class_weight,
-            dual=dual,
-            penalty=penalty,
-            intercept_scaling=intercept_scaling,
-            multi_class=multi_class,
-            random_state=random_state,
-            check_input=check_input,
-            max_squared_sum=max_squared_sum,
-            sample_weight=sample_weight,
-            l1_ratio=l1_ratio,
-            **({"n_threads": n_threads} if sklearn_check_version("1.1") else {}),
-        )
 
     # Comment 2025-08-04: this file might have dead code paths from unsupported solvers.
     # It appears to have initially been a copy-paste of scikit-learn with a few additions
@@ -269,7 +218,6 @@ def __logistic_regression_path(
             func = _daal4py_loss_
             grad = _daal4py_grad_
             hess = _daal4py_grad_hess_
-        warm_start_sag = {"coef": w0.T}
     else:
         target = y_bin
         if solver == "lbfgs":
@@ -280,7 +228,6 @@ def __logistic_regression_path(
             func = _daal4py_loss_
             grad = _daal4py_grad_
             hess = _daal4py_grad_hess_
-        warm_start_sag = {"coef": np.expand_dims(w0, axis=1)}
 
     coefs = list()
     n_iter = np.zeros(len(Cs), dtype=np.int32)
@@ -385,8 +332,6 @@ def _func_(x, *args):
         for i, ci in enumerate(coefs):
             coefs[i] = np.delete(ci, 0, axis=-1)
 
-    _patching_status.write_log()
-
     return np.array(coefs), np.array(Cs), n_iter
 
 
@@ -427,20 +372,21 @@ def daal4py_predict(self, X, resultsToEvaluate):
         f"sklearn.linear_model.LogisticRegression.{_function_name}"
     )
     if _function_name != "predict":
+        multi_class = getattr(self, "multi_class", "auto")
         _patching_status.and_conditions(
             [
                 (
                     self.classes_.size == 2
-                    or logistic_module._check_multi_class(
-                        self.multi_class if self.multi_class != "deprecated" else "auto",
+                    or _check_multi_class(
+                        multi_class if multi_class != "deprecated" else "auto",
                         self.solver,
                         self.classes_.size,
                     )
                     != "ovr",
                     f"selected multiclass option is not supported for n_classes > 2.",
                 ),
                 (
-                    not (self.classes_.size == 2 and self.multi_class == "multinomial"),
+                    not (self.classes_.size == 2 and multi_class == "multinomial"),
                     "multi_class='multinomial' not supported with binary data",
                 ),
             ],
@@ -502,52 +448,35 @@ def daal4py_predict(self, X, resultsToEvaluate):
         return LogisticRegression_original.predict_log_proba(self, X)
 
 
-def logistic_regression_path(
-    X,
-    y,
-    pos_class=None,
-    Cs=10,
-    fit_intercept=True,
-    max_iter=100,
-    tol=1e-4,
-    verbose=0,
-    solver="lbfgs",
-    coef=None,
-    class_weight=None,
-    dual=False,
-    penalty="l2",
-    intercept_scaling=1.0,
-    multi_class="auto",
-    random_state=None,
-    check_input=True,
-    max_squared_sum=None,
-    sample_weight=None,
-    l1_ratio=None,
-    n_threads=1,
-):
-    return __logistic_regression_path(
-        X,
-        y,
-        pos_class=pos_class,
-        Cs=Cs,
-        fit_intercept=fit_intercept,
-        max_iter=max_iter,
-        tol=tol,
-        verbose=verbose,
-        solver=solver,
-        coef=coef,
-        class_weight=class_weight,
-        dual=dual,
-        penalty=penalty,
-        intercept_scaling=intercept_scaling,
-        multi_class=multi_class,
-        random_state=random_state,
-        check_input=check_input,
-        max_squared_sum=max_squared_sum,
-        sample_weight=sample_weight,
-        l1_ratio=l1_ratio,
-        n_threads=n_threads,
+def logistic_regression_path(*args, **kwargs):
+
+    _patching_status = PatchingConditionsChain(
+        "sklearn.linear_model.LogisticRegression.fit"
+    )
+    _dal_ready = _patching_status.and_conditions(
+        [
+            (
+                kwargs["solver"] in ["lbfgs", "newton-cg"],
+                f"'{kwargs['solver']}' solver is not supported. "
+                "Only 'lbfgs' and 'newton-cg' solvers are supported.",
+            ),
+            (not sparse.issparse(args[0]), "X is sparse. Sparse input is not supported."),
+            (kwargs["sample_weight"] is None, "Sample weights are not supported."),
+            (kwargs["class_weight"] is None, "Class weights are not supported."),
+        ]
     )
+    if not _dal_ready:
+        _patching_status.write_log()
+        return lr_path_original(*args, **kwargs)
+
+    if sklearn_check_version("1.8"):
+        kwargs.pop("classes", None)
+        res = __logistic_regression_path(*(args[:2]), **kwargs)
+    else:
+        res = __logistic_regression_path(*args, **kwargs)
+
+    _patching_status.write_log()
+    return res
 
 
 @control_n_jobs(

daal4py/sklearn/monkeypatch/tests/_models_info.py

@@ -30,7 +30,7 @@
 from sklearn.neighbors import KNeighborsClassifier, KNeighborsRegressor, NearestNeighbors
 from sklearn.svm import SVC
 
-from daal4py.sklearn._utils import daal_check_version
+from daal4py.sklearn._utils import daal_check_version, sklearn_check_version
 
 MODELS_INFO = [
     {
@@ -84,7 +84,10 @@
         "dataset": "classifier",
     },
     {
-        "model": LogisticRegression(max_iter=100, multi_class="multinomial"),
+        "model": LogisticRegression(
+            max_iter=100,
+            **({} if sklearn_check_version("1.8") else {"multi_class": "multinomial"})
+        ),
         "methods": [
             "decision_function",
             "predict",

daal4py/sklearn/utils/validation.py

@@ -280,9 +280,9 @@ def _daal_check_array(
     array_converted : object
         The converted and validated array.
     """
-    if force_all_finite not in (True, False, "allow-nan"):
+    if force_all_finite not in (True, False, "allow-nan", None):
         raise ValueError(
-            'force_all_finite should be a bool or "allow-nan"'
+            'force_all_finite should be a bool, None, or "allow-nan"'
             ". Got {!r} instead".format(force_all_finite)
         )
 

deselected_tests.yaml

@@ -457,6 +457,9 @@ public:
   # single-threaded computation is used which is a frequent case for CI
   - decomposition/tests/test_pca.py::test_pca_dtype_preservation[full] <1.5
 
+  # Convergence failure due to different settings and cost function implementation
+  - linear_model/tests/test_logistic.py::test_logistic_regression_path_convergence_fail
+
   # --------------------------------------------------------
   # The following tests currently fail with GPU offloading
 gpu:

onedal/datatypes/_dlpack.py

@@ -45,11 +45,14 @@ def dlpack_to_numpy(obj):
             raise TypeError(f"cannot move {type(obj)} to cpu")
 
     # convert to numpy
-    if hasattr(obj, "__array__"):
-        # ``copy`` param for the ``asarray`` is not set.
-        # The object is copied only if needed
-        obj = np.asarray(obj)
-    else:
+    try:
+        # Some frameworks implement an __array__ method just to
+        # throw a RuntimeError when used (array_api_strict, dpctl),
+        # or a TypeError (array_api-strict) rather than an AttributeError
+        # therefore a try catch is necessary (logic is essentially a
+        # getattr call + some)
+        obj = obj.__array__()
+    except (AttributeError, RuntimeError, TypeError):
         # requires numpy 1.23
         try:
             obj = np.from_dlpack(obj)