scikit-learn-intelex 2024.5.0__py310-none-manylinux1_x86_64.whl → 2024.6.0__py310-none-manylinux1_x86_64.whl

sklearnex/svm/_common.py

@@ -15,15 +15,20 @@
 # ==============================================================================
 
 from abc import ABC
+from numbers import Number, Real
 
 import numpy as np
+from scipy import sparse as sp
+from sklearn.base import BaseEstimator, ClassifierMixin
 from sklearn.calibration import CalibratedClassifierCV
+from sklearn.metrics import r2_score
 from sklearn.model_selection import StratifiedKFold
 from sklearn.preprocessing import LabelEncoder
 
 from daal4py.sklearn._utils import sklearn_check_version
-from onedal.utils import _column_or_1d
+from onedal.utils import _check_array, _check_X_y, _column_or_1d
 
+from .._config import config_context, get_config
 from .._utils import PatchingConditionsChain
 
 
@@ -51,7 +56,8 @@ def set_intercept(self, value):
             del self._onedal_estimator._onedal_model
 
 
-class BaseSVM(ABC):
+class BaseSVM(BaseEstimator, ABC):
+
     def _onedal_gpu_supported(self, method_name, *data):
         patching_status = PatchingConditionsChain(f"sklearn.{method_name}")
         patching_status.and_conditions([(False, "GPU offloading is not supported.")])
@@ -74,7 +80,7 @@ class BaseSVM(ABC):
             )
             return patching_status
         inference_methods = (
-            ["predict"]
+            ["predict", "score"]
             if class_name.endswith("R")
             else ["predict", "predict_proba", "decision_function", "score"]
         )
@@ -85,6 +91,130 @@ class BaseSVM(ABC):
             return patching_status
         raise RuntimeError(f"Unknown method {method_name} in {class_name}")
 
+    def _compute_gamma_sigma(self, X):
+        # only run extended conversion if kernel is not linear
+        # set to a value = 1.0, so gamma will always be passed to
+        # the onedal estimator as a float type
+        if self.kernel == "linear":
+            return 1.0
+
+        if isinstance(self.gamma, str):
+            if self.gamma == "scale":
+                if sp.issparse(X):
+                    # var = E[X^2] - E[X]^2
+                    X_sc = (X.multiply(X)).mean() - (X.mean()) ** 2
+                else:
+                    X_sc = X.var()
+                _gamma = 1.0 / (X.shape[1] * X_sc) if X_sc != 0 else 1.0
+            elif self.gamma == "auto":
+                _gamma = 1.0 / X.shape[1]
+            else:
+                raise ValueError(
+                    "When 'gamma' is a string, it should be either 'scale' or "
+                    "'auto'. Got '{}' instead.".format(self.gamma)
+                )
+        else:
+            if sklearn_check_version("1.1") and not sklearn_check_version("1.2"):
+                if isinstance(self.gamma, Real):
+                    if self.gamma <= 0:
+                        msg = (
+                            f"gamma value must be > 0; {self.gamma!r} is invalid. Use"
+                            " a positive number or use 'auto' to set gamma to a"
+                            " value of 1 / n_features."
+                        )
+                        raise ValueError(msg)
+                    _gamma = self.gamma
+                else:
+                    msg = (
+                        "The gamma value should be set to 'scale', 'auto' or a"
+                        f" positive float value. {self.gamma!r} is not a valid option"
+                    )
+                    raise ValueError(msg)
+            else:
+                _gamma = self.gamma
+        return _gamma
+
+    def _onedal_fit_checks(self, X, y, sample_weight=None):
+        if hasattr(self, "decision_function_shape"):
+            if self.decision_function_shape not in ("ovr", "ovo", None):
+                raise ValueError(
+                    f"decision_function_shape must be either 'ovr' or 'ovo', "
+                    f"got {self.decision_function_shape}."
+                )
+
+        if y is None:
+            if self._get_tags()["requires_y"]:
+                raise ValueError(
+                    f"This {self.__class__.__name__} estimator "
+                    f"requires y to be passed, but the target y is None."
+                )
+        # using onedal _check_X_y to insure X and y are contiguous
+        # finite check occurs in onedal estimator
+        X, y = _check_X_y(
+            X,
+            y,
+            dtype=[np.float64, np.float32],
+            force_all_finite=False,
+            accept_sparse="csr",
+        )
+        y = self._validate_targets(y)
+        sample_weight = self._get_sample_weight(X, y, sample_weight)
+        return X, y, sample_weight
+
+    def _get_sample_weight(self, X, y, sample_weight):
+        n_samples = X.shape[0]
+        dtype = X.dtype
+        if n_samples == 1:
+            raise ValueError("n_samples=1")
+
+        sample_weight = np.ascontiguousarray(
+            [] if sample_weight is None else sample_weight, dtype=np.float64
+        )
+
+        sample_weight_count = sample_weight.shape[0]
+        if sample_weight_count != 0 and sample_weight_count != n_samples:
+            raise ValueError(
+                "sample_weight and X have incompatible shapes: "
+                "%r vs %r\n"
+                "Note: Sparse matrices cannot be indexed w/"
+                "boolean masks (use `indices=True` in CV)."
+                % (len(sample_weight), X.shape)
+            )
+
+        if sample_weight_count == 0:
+            if not isinstance(self, ClassifierMixin) or self.class_weight_ is None:
+                return None
+            sample_weight = np.ones(n_samples, dtype=dtype)
+        elif isinstance(sample_weight, Number):
+            sample_weight = np.full(n_samples, sample_weight, dtype=dtype)
+        else:
+            sample_weight = _check_array(
+                sample_weight,
+                accept_sparse=False,
+                ensure_2d=False,
+                dtype=dtype,
+                order="C",
+            )
+            if sample_weight.ndim != 1:
+                raise ValueError("Sample weights must be 1D array or scalar")
+
+            if sample_weight.shape != (n_samples,):
+                raise ValueError(
+                    "sample_weight.shape == {}, expected {}!".format(
+                        sample_weight.shape, (n_samples,)
+                    )
+                )
+
+        if np.all(sample_weight <= 0):
+            if "nusvc" in self.__module__:
+                raise ValueError("negative dimensions are not allowed")
+            else:
+                raise ValueError(
+                    "Invalid input - all samples have zero or negative weights."
+                )
+
+        return sample_weight
+
 
 class BaseSVC(BaseSVM):
     def _compute_balanced_class_weight(self, y):
@@ -105,22 +235,31 @@ class BaseSVC(BaseSVM):
         params["decision_function_shape"] = "ovr"
         clf_base = self.__class__(**params)
 
-        try:
-            n_splits = 5
-            n_jobs = n_splits if queue is None or queue.sycl_device.is_cpu else 1
-            cv = StratifiedKFold(
-                n_splits=n_splits, shuffle=True, random_state=self.random_state
-            )
-            self.clf_prob = CalibratedClassifierCV(
-                clf_base, ensemble=False, cv=cv, method="sigmoid", n_jobs=n_jobs
-            )
-            self.clf_prob.fit(X, y, sample_weight)
-        except ValueError:
-            clf_base = clf_base.fit(X, y, sample_weight)
-            self.clf_prob = CalibratedClassifierCV(
-                clf_base, cv="prefit", method="sigmoid"
-            )
-            self.clf_prob.fit(X, y, sample_weight)
+        # We use stock metaestimators below, so the only way
+        # to pass a queue is using config_context.
+        cfg = get_config()
+        cfg["target_offload"] = queue
+        with config_context(**cfg):
+            try:
+                n_splits = 5
+                n_jobs = n_splits if queue is None or queue.sycl_device.is_cpu else 1
+                cv = StratifiedKFold(
+                    n_splits=n_splits, shuffle=True, random_state=self.random_state
+                )
+                self.clf_prob = CalibratedClassifierCV(
+                    clf_base,
+                    ensemble=False,
+                    cv=cv,
+                    method="sigmoid",
+                )
+                self.clf_prob.fit(X, y, sample_weight)
+
+            except ValueError:
+                clf_base = clf_base.fit(X, y, sample_weight)
+                self.clf_prob = CalibratedClassifierCV(
+                    clf_base, cv="prefit", method="sigmoid"
+                )
+                self.clf_prob.fit(X, y, sample_weight)
 
     def _save_attributes(self):
         self.support_vectors_ = self._onedal_estimator.support_vectors_
@@ -129,7 +268,8 @@ class BaseSVC(BaseSVM):
         self.dual_coef_ = self._onedal_estimator.dual_coef_
         self.shape_fit_ = self._onedal_estimator.class_weight_
         self.classes_ = self._onedal_estimator.classes_
-        self.class_weight_ = self._onedal_estimator.class_weight_
+        if isinstance(self, ClassifierMixin) or not sklearn_check_version("1.2"):
+            self.class_weight_ = self._onedal_estimator.class_weight_
         self.support_ = self._onedal_estimator.support_
 
         self._intercept_ = self._onedal_estimator.intercept_
@@ -183,3 +323,8 @@ class BaseSVR(BaseSVM):
 
         if sklearn_check_version("1.1"):
             self.n_iter_ = self._onedal_estimator.n_iter_
+
+    def _onedal_score(self, X, y, sample_weight=None, queue=None):
+        return r2_score(
+            y, self._onedal_predict(X, queue=queue), sample_weight=sample_weight
+        )

sklearnex/svm/nusvc.py

@@ -83,6 +83,17 @@ class NuSVC(sklearn_NuSVC, BaseSVC):
     def fit(self, X, y, sample_weight=None):
         if sklearn_check_version("1.2"):
             self._validate_params()
+        elif self.nu <= 0 or self.nu > 1:
+            # else if added to correct issues with
+            # sklearn tests:
+            # svm/tests/test_sparse.py::test_error
+            # svm/tests/test_svm.py::test_bad_input
+            # for sklearn versions < 1.2 (i.e. without
+            # validate_params parameter checking)
+            # Without this, a segmentation fault with
+            # Windows fatal exception: access violation
+            # occurs
+            raise ValueError("nu <= 0 or nu > 1")
         if sklearn_check_version("1.0"):
             self._check_feature_names(X, reset=True)
         dispatch(
@@ -94,7 +105,7 @@ class NuSVC(sklearn_NuSVC, BaseSVC):
             },
             X,
             y,
-            sample_weight,
+            sample_weight=sample_weight,
         )
 
         return self
@@ -242,12 +253,31 @@ class NuSVC(sklearn_NuSVC, BaseSVC):
 
     decision_function.__doc__ = sklearn_NuSVC.decision_function.__doc__
 
+    def _get_sample_weight(self, X, y, sample_weight=None):
+        sample_weight = super()._get_sample_weight(X, y, sample_weight)
+        if sample_weight is None:
+            return sample_weight
+
+        weight_per_class = [
+            np.sum(sample_weight[y == class_label]) for class_label in np.unique(y)
+        ]
+
+        for i in range(len(weight_per_class)):
+            for j in range(i + 1, len(weight_per_class)):
+                if self.nu * (weight_per_class[i] + weight_per_class[j]) / 2 > min(
+                    weight_per_class[i], weight_per_class[j]
+                ):
+                    raise ValueError("specified nu is infeasible")
+
+        return sample_weight
+
     def _onedal_fit(self, X, y, sample_weight=None, queue=None):
+        X, _, weights = self._onedal_fit_checks(X, y, sample_weight)
         onedal_params = {
             "nu": self.nu,
             "kernel": self.kernel,
             "degree": self.degree,
-            "gamma": self.gamma,
+            "gamma": self._compute_gamma_sigma(X),
             "coef0": self.coef0,
             "tol": self.tol,
             "shrinking": self.shrinking,
@@ -259,10 +289,16 @@ class NuSVC(sklearn_NuSVC, BaseSVC):
         }
 
         self._onedal_estimator = onedal_NuSVC(**onedal_params)
-        self._onedal_estimator.fit(X, y, sample_weight, queue=queue)
+        self._onedal_estimator.fit(X, y, weights, queue=queue)
 
         if self.probability:
-            self._fit_proba(X, y, sample_weight, queue=queue)
+            self._fit_proba(
+                X,
+                y,
+                sample_weight=sample_weight,
+                queue=queue,
+            )
+
         self._save_attributes()
 
     def _onedal_predict(self, X, queue=None):

sklearnex/svm/nusvr.py

@@ -65,6 +65,17 @@ class NuSVR(sklearn_NuSVR, BaseSVR):
     def fit(self, X, y, sample_weight=None):
         if sklearn_check_version("1.2"):
             self._validate_params()
+        elif self.nu <= 0 or self.nu > 1:
+            # else if added to correct issues with
+            # sklearn tests:
+            # svm/tests/test_sparse.py::test_error
+            # svm/tests/test_svm.py::test_bad_input
+            # for sklearn versions < 1.2 (i.e. without
+            # validate_params parameter checking)
+            # Without this, a segmentation fault with
+            # Windows fatal exception: access violation
+            # occurs
+            raise ValueError("nu <= 0 or nu > 1")
         if sklearn_check_version("1.0"):
             self._check_feature_names(X, reset=True)
         dispatch(
@@ -76,7 +87,7 @@ class NuSVR(sklearn_NuSVR, BaseSVR):
             },
             X,
             y,
-            sample_weight,
+            sample_weight=sample_weight,
         )
         return self
 
@@ -94,13 +105,30 @@ class NuSVR(sklearn_NuSVR, BaseSVR):
             X,
         )
 
+    @wrap_output_data
+    def score(self, X, y, sample_weight=None):
+        if sklearn_check_version("1.0"):
+            self._check_feature_names(X, reset=False)
+        return dispatch(
+            self,
+            "score",
+            {
+                "onedal": self.__class__._onedal_score,
+                "sklearn": sklearn_NuSVR.score,
+            },
+            X,
+            y,
+            sample_weight=sample_weight,
+        )
+
     def _onedal_fit(self, X, y, sample_weight=None, queue=None):
+        X, _, sample_weight = self._onedal_fit_checks(X, y, sample_weight)
         onedal_params = {
             "C": self.C,
             "nu": self.nu,
             "kernel": self.kernel,
             "degree": self.degree,
-            "gamma": self.gamma,
+            "gamma": self._compute_gamma_sigma(X),
             "coef0": self.coef0,
             "tol": self.tol,
             "shrinking": self.shrinking,
@@ -117,3 +145,4 @@ class NuSVR(sklearn_NuSVR, BaseSVR):
 
     fit.__doc__ = sklearn_NuSVR.fit.__doc__
     predict.__doc__ = sklearn_NuSVR.predict.__doc__
+    score.__doc__ = sklearn_NuSVR.score.__doc__

sklearnex/svm/svc.py

@@ -85,6 +85,17 @@ class SVC(sklearn_SVC, BaseSVC):
     def fit(self, X, y, sample_weight=None):
         if sklearn_check_version("1.2"):
             self._validate_params()
+        elif self.C <= 0:
+            # else if added to correct issues with
+            # sklearn tests:
+            # svm/tests/test_sparse.py::test_error
+            # svm/tests/test_svm.py::test_bad_input
+            # for sklearn versions < 1.2 (i.e. without
+            # validate_params parameter checking)
+            # Without this, a segmentation fault with
+            # Windows fatal exception: access violation
+            # occurs
+            raise ValueError("C <= 0")
         if sklearn_check_version("1.0"):
             self._check_feature_names(X, reset=True)
         dispatch(
@@ -96,8 +107,9 @@ class SVC(sklearn_SVC, BaseSVC):
             },
             X,
             y,
-            sample_weight,
+            sample_weight=sample_weight,
         )
+
         return self
 
     @wrap_output_data
@@ -270,12 +282,30 @@ class SVC(sklearn_SVC, BaseSVC):
             return patching_status
         raise RuntimeError(f"Unknown method {method_name} in {class_name}")
 
+    def _get_sample_weight(self, X, y, sample_weight=None):
+        sample_weight = super()._get_sample_weight(X, y, sample_weight)
+        if sample_weight is None:
+            return sample_weight
+
+        if np.any(sample_weight <= 0) and len(np.unique(y[sample_weight > 0])) != len(
+            self.classes_
+        ):
+            raise ValueError(
+                "Invalid input - all samples with positive weights "
+                "belong to the same class"
+                if sklearn_check_version("1.2")
+                else "Invalid input - all samples with positive weights "
+                "have the same label."
+            )
+        return sample_weight
+
     def _onedal_fit(self, X, y, sample_weight=None, queue=None):
+        X, _, weights = self._onedal_fit_checks(X, y, sample_weight)
         onedal_params = {
             "C": self.C,
             "kernel": self.kernel,
             "degree": self.degree,
-            "gamma": self.gamma,
+            "gamma": self._compute_gamma_sigma(X),
             "coef0": self.coef0,
             "tol": self.tol,
             "shrinking": self.shrinking,
@@ -287,10 +317,16 @@ class SVC(sklearn_SVC, BaseSVC):
         }
 
         self._onedal_estimator = onedal_SVC(**onedal_params)
-        self._onedal_estimator.fit(X, y, sample_weight, queue=queue)
+        self._onedal_estimator.fit(X, y, weights, queue=queue)
 
         if self.probability:
-            self._fit_proba(X, y, sample_weight, queue=queue)
+            self._fit_proba(
+                X,
+                y,
+                sample_weight=sample_weight,
+                queue=queue,
+            )
+
         self._save_attributes()
 
     def _onedal_predict(self, X, queue=None):

sklearnex/svm/svr.py

@@ -65,6 +65,17 @@ class SVR(sklearn_SVR, BaseSVR):
     def fit(self, X, y, sample_weight=None):
         if sklearn_check_version("1.2"):
             self._validate_params()
+        elif self.C <= 0:
+            # else if added to correct issues with
+            # sklearn tests:
+            # svm/tests/test_sparse.py::test_error
+            # svm/tests/test_svm.py::test_bad_input
+            # for sklearn versions < 1.2 (i.e. without
+            # validate_params parameter checking)
+            # Without this, a segmentation fault with
+            # Windows fatal exception: access violation
+            # occurs
+            raise ValueError("C <= 0")
         if sklearn_check_version("1.0"):
             self._check_feature_names(X, reset=True)
         dispatch(
@@ -76,7 +87,7 @@ class SVR(sklearn_SVR, BaseSVR):
             },
             X,
             y,
-            sample_weight,
+            sample_weight=sample_weight,
         )
 
         return self
@@ -95,13 +106,30 @@ class SVR(sklearn_SVR, BaseSVR):
             X,
         )
 
+    @wrap_output_data
+    def score(self, X, y, sample_weight=None):
+        if sklearn_check_version("1.0"):
+            self._check_feature_names(X, reset=False)
+        return dispatch(
+            self,
+            "score",
+            {
+                "onedal": self.__class__._onedal_score,
+                "sklearn": sklearn_SVR.score,
+            },
+            X,
+            y,
+            sample_weight=sample_weight,
+        )
+
     def _onedal_fit(self, X, y, sample_weight=None, queue=None):
+        X, _, sample_weight = self._onedal_fit_checks(X, y, sample_weight)
         onedal_params = {
             "C": self.C,
             "epsilon": self.epsilon,
             "kernel": self.kernel,
             "degree": self.degree,
-            "gamma": self.gamma,
+            "gamma": self._compute_gamma_sigma(X),
             "coef0": self.coef0,
             "tol": self.tol,
             "shrinking": self.shrinking,
@@ -118,3 +146,4 @@ class SVR(sklearn_SVR, BaseSVR):
 
     fit.__doc__ = sklearn_SVR.fit.__doc__
     predict.__doc__ = sklearn_SVR.predict.__doc__
+    score.__doc__ = sklearn_SVR.score.__doc__

sklearnex/tests/_utils.py

@@ -14,9 +14,11 @@
 # limitations under the License.
 # ==============================================================================
 
+from functools import partial
 from inspect import isclass
 
 import numpy as np
+from scipy import sparse as sp
 from sklearn import clone
 from sklearn.base import (
     BaseEstimator,
@@ -116,8 +118,8 @@ def gen_models_info(algorithms):
 
         if i in PATCHED_MODELS:
             est = PATCHED_MODELS[i]
-        elif i in SPECIAL_INSTANCES:
-            est = SPECIAL_INSTANCES[i].__class__
+        elif isinstance(algorithms[i], BaseEstimator):
+            est = algorithms[i].__class__
         else:
             raise KeyError(f"Unrecognized sklearnex estimator: {i}")
 
@@ -138,24 +140,54 @@ def gen_models_info(algorithms):
     return output
 
 
-def gen_dataset(estimator, queue=None, target_df=None, dtype=np.float64):
-    dataset = None
-    name = estimator.__class__.__name__
-    est = PATCHED_MODELS[name]
+def gen_dataset_type(est):
+    # est should be an estimator or estimator class
+    # dataset initialized to classification, but will be swapped
+    # for other types as necessary
+    dataset = "classification"
+    estimator = est.__class__ if isinstance(est, BaseEstimator) else est
+
     for mixin, _, data in mixin_map:
-        if issubclass(est, mixin) and data is not None:
+        if issubclass(estimator, mixin) and data is not None:
             dataset = data
+    return dataset
+
+
+_dataset_dict = {
+    "classification": [partial(load_iris, return_X_y=True)],
+    "regression": [partial(load_diabetes, return_X_y=True)],
+}
+
+
+def gen_dataset(
+    est,
+    datasets=_dataset_dict,
+    sparse=False,
+    queue=None,
+    target_df=None,
+    dtype=None,
+):
+    dataset_type = gen_dataset_type(est)
+    output = []
     # load data
-    if dataset == "classification" or dataset is None:
-        X, y = load_iris(return_X_y=True)
-    elif dataset == "regression":
-        X, y = load_diabetes(return_X_y=True)
-    else:
-        raise ValueError("Unknown dataset type")
-
-    X = _convert_to_dataframe(X, sycl_queue=queue, target_df=target_df, dtype=dtype)
-    y = _convert_to_dataframe(y, sycl_queue=queue, target_df=target_df, dtype=dtype)
-    return X, y
+    flag = dtype is None
+
+    for func in datasets[dataset_type]:
+        X, y = func()
+        if flag:
+            dtype = X.dtype if hasattr(X, "dtype") else np.float64
+
+        if sparse:
+            X = sp.csr_matrix(X)
+        else:
+            X = _convert_to_dataframe(
+                X, sycl_queue=queue, target_df=target_df, dtype=dtype
+            )
+            y = _convert_to_dataframe(
+                y, sycl_queue=queue, target_df=target_df, dtype=dtype
+            )
+        output += [[X, y]]
+    return output
 
 
 DTYPES = [

sklearnex/tests/test_common.py

@@ -0,0 +1,54 @@
+# ==============================================================================
+# Copyright 2024 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+# ==============================================================================
+
+import os
+from glob import glob
+
+import pytest
+
+ALLOWED_LOCATIONS = [
+    "_config.py",
+    "_device_offload.py",
+    "test",
+    "svc.py",
+    "svm" + os.sep + "_common.py",
+]
+
+
+def test_target_offload_ban():
+    """This test blocks the use of target_offload in
+    in sklearnex files. Offloading computation to devices
+    via target_offload should only occur externally, and not
+    within the architecture of the sklearnex classes. This
+    is for clarity, traceability and maintainability.
+    """
+    from sklearnex import __file__ as loc
+
+    path = loc.replace("__init__.py", "")
+    files = [y for x in os.walk(path) for y in glob(os.path.join(x[0], "*.py"))]
+
+    output = []
+
+    for f in files:
+        if open(f, "r").read().find("target_offload") != -1:
+            output += [f.replace(path, "sklearnex" + os.sep)]
+
+    # remove this file from the list
+    for allowed in ALLOWED_LOCATIONS:
+        output = [i for i in output if allowed not in i]
+
+    output = "\n".join(output)
+    assert output == "", f"sklearn versioning is occuring in: \n{output}"