scikit-learn-intelex 2025.0.0__py311-none-manylinux_2_28_x86_64.whl

sklearnex/tests/_utils.py

@@ -0,0 +1,328 @@
+# ==============================================================================
+# 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.
+# ==============================================================================
+
+from functools import partial
+from inspect import getattr_static, isclass, signature
+
+import numpy as np
+from scipy import sparse as sp
+from sklearn import clone
+from sklearn.base import (
+    BaseEstimator,
+    ClassifierMixin,
+    ClusterMixin,
+    OutlierMixin,
+    RegressorMixin,
+    TransformerMixin,
+)
+from sklearn.datasets import load_diabetes, load_iris
+from sklearn.neighbors._base import KNeighborsMixin
+
+from onedal.tests.utils._dataframes_support import _convert_to_dataframe
+from sklearnex import get_patch_map, patch_sklearn, sklearn_is_patched, unpatch_sklearn
+from sklearnex.linear_model import LogisticRegression
+from sklearnex.neighbors import (
+    KNeighborsClassifier,
+    KNeighborsRegressor,
+    LocalOutlierFactor,
+    NearestNeighbors,
+)
+from sklearnex.svm import SVC, NuSVC
+
+
+def _load_all_models(with_sklearnex=True, estimator=True):
+    """Convert sklearnex patch_map into a dictionary of estimators or functions
+
+    Parameters
+    ----------
+    with_sklearnex: bool (default=True)
+        Discover estimators and methods with sklearnex patching enabled (True)
+        or disabled (False) from the sklearnex patch_map
+
+    estimator: bool (default=True)
+        yield estimators (True) or functions (False)
+
+    Returns
+    -------
+    dict: {name:estimator}
+        estimator is a class or function from sklearn or sklearnex
+    """
+    # insure that patch state is correct as dictated by patch_sklearn boolean
+    # and return it to the previous state no matter what occurs.
+    already_patched_map = sklearn_is_patched(return_map=True)
+    already_patched = any(already_patched_map.values())
+    try:
+        if with_sklearnex:
+            patch_sklearn()
+        elif already_patched:
+            unpatch_sklearn()
+
+        models = {}
+        for patch_infos in get_patch_map().values():
+            candidate = getattr(patch_infos[0][0][0], patch_infos[0][0][1], None)
+            if candidate is not None and isclass(candidate) == estimator:
+                if not estimator or issubclass(candidate, BaseEstimator):
+                    models[patch_infos[0][0][1]] = candidate
+    finally:
+        if with_sklearnex:
+            unpatch_sklearn()
+        # both branches are now in an unpatched state, repatch as necessary
+        if already_patched:
+            patch_sklearn(name=[i for i in already_patched_map if already_patched_map[i]])
+
+    return models
+
+
+PATCHED_MODELS = _load_all_models(with_sklearnex=True)
+UNPATCHED_MODELS = _load_all_models(with_sklearnex=False)
+
+PATCHED_FUNCTIONS = _load_all_models(with_sklearnex=True, estimator=False)
+UNPATCHED_FUNCTIONS = _load_all_models(with_sklearnex=False, estimator=False)
+
+mixin_map = [
+    [
+        ClassifierMixin,
+        ["decision_function", "predict", "predict_proba", "predict_log_proba", "score"],
+        "classification",
+    ],
+    [RegressorMixin, ["predict", "score"], "regression"],
+    [ClusterMixin, ["fit_predict"], "classification"],
+    [TransformerMixin, ["fit_transform", "transform", "score"], "classification"],
+    [OutlierMixin, ["fit_predict", "predict"], "classification"],
+    [KNeighborsMixin, ["kneighbors"], None],
+]
+
+
+class _sklearn_clone_dict(dict):
+    """Special dict type for returning state-free sklearn/sklearnex estimators
+    with the same parameters"""
+
+    def __getitem__(self, key):
+        return clone(super().__getitem__(key))
+
+
+# Special dictionary of sklearnex estimators which must be specifically tested, this
+# could be because of supported non-default parameters, blocked support via sklearn's
+# 'available_if' decorator, or not being a native sklearn estimator (i.e. those not in
+# the default PATCHED_MODELS dictionary)
+SPECIAL_INSTANCES = _sklearn_clone_dict(
+    {
+        str(i): i
+        for i in [
+            LocalOutlierFactor(novelty=True),
+            SVC(probability=True),
+            NuSVC(probability=True),
+            KNeighborsClassifier(algorithm="brute"),
+            KNeighborsRegressor(algorithm="brute"),
+            NearestNeighbors(algorithm="brute"),
+            LogisticRegression(solver="newton-cg"),
+        ]
+    }
+)
+
+
+def gen_models_info(algorithms, required_inputs=["X", "y"]):
+    """Generate estimator-attribute pairs for pytest test collection.
+
+    Parameters
+    ----------
+    algorithms : iterable (list, tuple, 1D array-like object)
+        Iterable of valid sklearnex estimators or keys from PATCHED_MODELS
+
+    required_inputs : list, tuple of strings or None
+        list of required args/kwargs for callable attribute (only non-private,
+        non-BaseEstimator attributes).  Only one must be present, None
+        signifies taking all non-private attribues, callable or not.
+
+    Returns
+    -------
+    list of 2-element tuples: (estimator, string)
+        Returns a list of valid methods or attributes without "fit"
+    """
+    output = []
+    for estimator in algorithms:
+
+        if estimator in PATCHED_MODELS:
+            est = PATCHED_MODELS[estimator]
+        elif isinstance(algorithms[estimator], BaseEstimator):
+            est = algorithms[estimator].__class__
+        else:
+            raise KeyError(f"Unrecognized sklearnex estimator: {estimator}")
+
+        # remove BaseEstimator methods (get_params, set_params)
+        candidates = set(dir(est)) - set(dir(BaseEstimator))
+        # remove private methods
+        candidates = set([attr for attr in candidates if not attr.startswith("_")])
+        # required to enable other methods
+        candidates = candidates - {"fit"}
+
+        # allow only callable methods with any of the required inputs
+        if required_inputs:
+            methods = []
+            for attr in candidates:
+                attribute = getattr_static(est, attr)
+                if callable(attribute):
+                    params = signature(attribute).parameters
+                    if any([inp in params for inp in required_inputs]):
+                        methods += [attr]
+        else:
+            methods = candidates
+
+        output += (
+            [(estimator, method) for method in methods]
+            if methods
+            else [(estimator, None)]
+        )
+
+    # In the case that no methods are available, set method to None.
+    # This will allow estimators without mixins to still test the fit
+    # method in various tests.
+    return output
+
+
+def call_method(estimator, method, X, y, **kwargs):
+    """Generalized interface to call most sklearn estimator methods
+
+    Parameters
+    ----------
+    estimator : sklearn or sklearnex estimator instance
+
+    method: string
+        Valid callable method to estimator
+
+    X: array-like
+        data
+
+    y: array-like (for 'score', 'partial-fit', and 'path')
+        X-dependent data
+
+    **kwargs: keyword dict
+        keyword arguments to estimator.method
+
+    Returns
+    -------
+    return value from estimator.method
+    """
+    # useful for repository wide testing
+    if method == "inverse_transform":
+        # PCA's inverse_transform takes (n_samples, n_components)
+        data = (
+            (X[:, : estimator.n_components_],)
+            if X.shape[1] != estimator.n_components_
+            else (X,)
+        )
+    elif method not in ["score", "partial_fit", "path"]:
+        data = (X,)
+    else:
+        data = (X, y)
+    return getattr(estimator, method)(*data, **kwargs)
+
+
+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. Private method.
+    dataset = "classification"
+    estimator = est.__class__ if isinstance(est, BaseEstimator) else est
+
+    for mixin, _, data in mixin_map:
+        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,
+):
+    """Generate dataset for pytest testing.
+
+    Parameters
+    ----------
+    est : sklearn or sklearnex estimator class
+        Must inherit an sklearn Mixin or sklearn's BaseEstimator
+
+    dataset: dataset dict
+        Dictionary with keys "classification" and/or "regression"
+        Value must be a list of object which yield X, y array
+        objects when called, ideally using a lambda or
+        functools.partial.
+
+    sparse: bool (default False)
+        Convert X data to a scipy.sparse csr_matrix format.
+
+    queue: SYCL queue or None
+        Queue necessary for device offloading following the
+        SYCL 2020 standard, usually generated by dpctl.
+
+    target_df: string or None
+        dataframe type for returned dataset, as dictated by
+        onedal's _convert_to_dataframe.
+
+    dtype: numpy dtype or None
+       target datatype for returned datasets (see DTYPES).
+
+    Returns
+    -------
+    list of 2-element list X,y: (array-like, array-like)
+        list of datasets for analysis
+    """
+    dataset_type = _gen_dataset_type(est)
+    output = []
+    # load data
+    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 = [
+    np.int8,
+    np.int16,
+    np.int32,
+    np.int64,
+    np.float16,
+    np.float32,
+    np.float64,
+    np.uint8,
+    np.uint16,
+    np.uint32,
+    np.uint64,
+]

sklearnex/tests/_utils_spmd.py

@@ -0,0 +1,198 @@
+# ==============================================================================
+# 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 numpy as np
+from numpy.testing import assert_allclose
+from sklearn.datasets import make_blobs, make_classification, make_regression
+from sklearn.model_selection import train_test_split
+
+from onedal.tests.utils._dataframes_support import _as_numpy
+
+try:
+    import dpctl
+    from dpctl import SyclQueue
+    from mpi4py import MPI
+
+    mpi_libs_available = True
+    gpu_is_available = dpctl.has_gpu_devices()
+except (ImportError, ModuleNotFoundError):
+    mpi_libs_available = False
+
+_mpi_libs_and_gpu_available = mpi_libs_available and gpu_is_available
+
+
+def _get_local_tensor(full_data):
+    """Splits data across ranks.
+
+    Called on each rank to extract the subset of data assigned to that rank.
+
+    Args:
+        full_data (numpy or dpctl array): The entire set of data
+
+    Returns:
+        local_data (numpy or dpctl array): The subset of data used by the rank
+    """
+
+    # create sycl queue and gather communicator details
+    q = SyclQueue("gpu")
+    comm = MPI.COMM_WORLD
+    rank = comm.Get_rank()
+    size = comm.Get_size()
+
+    # divide data across ranks and move to dpt tensor
+    data_rows = full_data.shape[0]
+    local_start = rank * data_rows // size
+    local_end = (1 + rank) * data_rows // size
+    local_data = full_data[local_start:local_end]
+
+    return local_data
+
+
+def _generate_regression_data(n_samples, n_features, dtype=np.float64, random_state=42):
+    # Generates regression data and divides between train and test
+    X, y = make_regression(
+        n_samples=n_samples, n_features=n_features, random_state=random_state
+    )
+    X = X.astype(dtype)
+    y = y.astype(dtype)
+    X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=random_state)
+    return X_train, X_test, y_train, y_test
+
+
+def _generate_classification_data(
+    n_samples, n_features, n_classes=2, dtype=np.float64, random_state=42
+):
+    # Generates classification data and divides between train and test
+    X, y = make_classification(
+        n_samples=n_samples,
+        n_features=n_features,
+        n_classes=n_classes,
+        n_informative=int(0.5 * n_classes + 1),
+        random_state=random_state,
+    )
+    X = X.astype(dtype)
+    y = y.astype(dtype)
+    X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=random_state)
+    return X_train, X_test, y_train, y_test
+
+
+def _generate_statistic_data(
+    n_samples, n_features=None, dtype=np.float64, random_state=42
+):
+    # Generates statistical data
+    gen = np.random.default_rng(random_state)
+    data = gen.uniform(
+        low=-0.3,
+        high=+0.7,
+        size=(n_samples, n_features) if n_features is not None else (n_samples,),
+    ).astype(dtype)
+    return data
+
+
+def _generate_clustering_data(
+    n_samples, n_features, centers=None, dtype=np.float64, random_state=42
+):
+    # Generates clustering data and divides between train and test
+    X, _ = make_blobs(
+        n_samples=n_samples,
+        centers=centers,
+        n_features=n_features,
+        random_state=random_state,
+    )
+    X = X.astype(dtype)
+    X_train, X_test = train_test_split(X, random_state=random_state)
+    return X_train, X_test
+
+
+def _spmd_assert_allclose(spmd_result, batch_result, **kwargs):
+    """Calls assert_allclose on spmd and batch results.
+
+    Called on each rank to compare the spmd result specific to that rank and
+    subset of batch result that corresponds to that rank.
+
+    Args:
+        spmd_result (numpy or dpctl array): The result for the subset of data on the rank the function is called from, computed by the spmd estimator
+        batch_result (numpy array): The result for all data, computed by the batch estimator
+
+    Raises:
+        AssertionError: If all results are not adequately close.
+    """
+
+    # extract chunk from batch result to match with local spmd result
+    local_batch_result = _get_local_tensor(batch_result)
+
+    assert_allclose(_as_numpy(spmd_result), _as_numpy(local_batch_result), **kwargs)
+
+
+def _assert_unordered_allclose(spmd_result, batch_result, localize=False, **kwargs):
+    """Checks if rows in spmd and batch results are aligned, even if not in the same order.
+
+    Called to verify correct unordered results are present. Useful to check KMeans centers
+    or KNN neighbors, where order does not matter. Sorts inputs to handle unordering. Also
+    capable of handling localization.
+
+    Args:
+        spmd_result (numpy or dpctl array): Result computed by the spmd estimator
+        batch_result (numpy array): Result computed by batch estimator
+        localize (bool): Whether of not spmd result is specific to the rank, in which case batch result needs to be localized
+
+    Raises:
+        AssertionError: If results do not match.
+    """
+    np_spmd_result = _as_numpy(spmd_result)
+
+    sorted_spmd_result = np_spmd_result[
+        np.argsort(np.linalg.norm(np_spmd_result, axis=1))
+    ]
+    if localize:
+        local_batch_result = _get_local_tensor(batch_result)
+        sorted_batch_result = local_batch_result[
+            np.argsort(np.linalg.norm(local_batch_result, axis=1))
+        ]
+    else:
+        sorted_batch_result = batch_result[
+            np.argsort(np.linalg.norm(batch_result, axis=1))
+        ]
+
+    assert_allclose(sorted_spmd_result, sorted_batch_result, **kwargs)
+
+
+def _assert_kmeans_labels_allclose(
+    spmd_labels, batch_labels, spmd_centers, batch_centers, **kwargs
+):
+    """Checks if labels for spmd and batch results are aligned, even cluster indices don't match.
+
+    Called to verify labels are assigned the same way on spmd and batch. Uses raw labels (which
+    may not match) to identify cluster center and ensure results match.
+
+    Args:
+        spmd_labels (numpy or dpctl array): The labels for the subset of data on the rank the function is called from, computed by the spmd estimator
+        batch_labels (numpy array): The labels for all data, computed by the batch estimator
+        spmd_centers (numpy or dpctl array): Centers computed by the spmd estimator
+        batch_centers (numpy array): Centers computed by batch estimator
+
+    Raises:
+        AssertionError: If clusters are not correctly assigned.
+    """
+
+    np_spmd_labels = _as_numpy(spmd_labels)
+    np_spmd_centers = _as_numpy(spmd_centers)
+    local_batch_labels = _get_local_tensor(batch_labels)
+    assert_allclose(
+        np_spmd_centers[np_spmd_labels],
+        batch_centers[local_batch_labels],
+        **kwargs,
+    )

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}"

sklearnex/tests/test_config.py

@@ -0,0 +1,43 @@
+# ==============================================================================
+# Copyright 2021 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 sklearn
+
+import onedal
+import sklearnex
+
+
+def test_get_config_contains_sklearn_params():
+    skex_config = sklearnex.get_config()
+    sk_config = sklearn.get_config()
+
+    assert all(value in skex_config.keys() for value in sk_config.keys())
+
+
+def test_set_config_works():
+    default_config = sklearnex.get_config()
+    sklearnex.set_config(
+        assume_finite=True, target_offload="cpu:0", allow_fallback_to_host=True
+    )
+
+    config = sklearnex.get_config()
+    onedal_config = onedal._config._get_config()
+    assert config["target_offload"] == "cpu:0"
+    assert config["allow_fallback_to_host"]
+    assert config["assume_finite"]
+    assert onedal_config["target_offload"] == "cpu:0"
+    assert onedal_config["allow_fallback_to_host"]
+    sklearnex.set_config(**default_config)