[enh] add array API enabled compute_class_weight function for use in classifier estimators

sklearnex/utils/class_weight.py

@@ -0,0 +1,103 @@
+# ==============================================================================
+# Copyright contributors to the oneDAL Project
+#
+# 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 sklearn.preprocessing import LabelEncoder as _sklearn_LabelEncoder
+
+from daal4py.sklearn._utils import sklearn_check_version
+
+from ._array_api import get_namespace
+from .validation import _check_sample_weight
+
+if not sklearn_check_version("1.7"):
+    from sklearn.utils.class_weight import (
+        compute_class_weight as _sklearn_compute_class_weight,
+    )
+
+    def compute_class_weight(class_weight, *, classes, y, sample_weight=None):
+        return _sklearn_compute_class_weight(class_weight, classes=classes, y=y)
+
+else:
+    from sklearn.utils.class_weight import compute_class_weight
+
+
+def _compute_class_weight(class_weight, *, classes, y, sample_weight=None):
+    # this duplicates sklearn code in order to enable it for array API.
+    # Note for the use of LabelEncoder this is only valid for sklearn
+    # versions >= 1.6.
+    xp, is_array_api_compliant = get_namespace(classes, y, sample_weight)
+
+    if not is_array_api_compliant:
+        # use the sklearn version for standard use.
+        return compute_class_weight(class_weight, classes, y, sample_weight=sample_weight)
+
+    sety = xp.unique_values(y)
+    setclasses = xp.unique_values(classes)
+    if sety.shape[0] != xp.unique_values(xp.concat((sety, setclasses))).shape[0]:
+        raise ValueError("classes should include all valid labels that can be in y")
+    if class_weight is None or len(class_weight) == 0:
+        # uniform class weights
+        weight = xp.ones((classes.shape[0],), dtype=xp.float64, device=classes.device)
+    elif class_weight == "balanced":
+        if not sklearn_check_version("1.6"):
+            raise RuntimeError(
+                "array API support with 'balanced' keyword not supported for sklearn <1.6"
+            )
+        # Find the weight of each class as present in y.
+        le = _sklearn_LabelEncoder()
+        y_ind = le.fit_transform(y)
+        if not all([item in le.classes_ for item in classes]):
+            raise ValueError("classes should have valid labels that are in y")
+
+        sample_weight = _check_sample_weight(sample_weight, y)
+        # scikit-learn implementation uses numpy.bincount, which does a combined
+        # min and max search, only erroring when a value < 0. Replicating this
+        # exactly via array API would cause another O(n) evaluation (by doing
+        # min and max separately). However this check can be removed due to the
+        # nature of the LabelEncoder. Therefore only the maximum is found, and
+        # then core logic of bincount is replicated:
+        # https://github.com/numpy/numpy/blob/main/numpy/_core/src/multiarray/compiled_base.c
+        weighted_class_counts = xp.zeros(
+            (xp.max(y_ind) + 1,), dtype=sample_weight.dtype, device=y.device
+        )
+
+        # use a more GPU-friendly summation approach for collecting weighted_class_counts
+        for w_idx in range(weighted_class_counts.shape[0]):
+            weighted_class_counts[w_idx] = xp.sum(sample_weight[y_ind == w_idx])
+
+        recip_freq = xp.sum(weighted_class_counts) / (
+            le.classes_.shape[0] * weighted_class_counts
+        )
+
+        weight = xp.take(recip_freq, le.transform(classes))
+    else:
+        # user-defined dictionary
+        weight = xp.ones((classes.shape[0],), dtype=xp.float64, device=classes.device)
+        unweighted_classes = []
+        for i, c in enumerate(classes):
+            if (fc := float(c)) in class_weight:
+                # array API has only numeric datatypes, convert to float for generality
+                # complex values should never be observed by this function
+                weight[i] = class_weight[fc]
+            else:
+                unweighted_classes.append(c)
+
+        n_weighted_classes = classes.shape[0] - len(unweighted_classes)
+        if unweighted_classes and n_weighted_classes != len(class_weight):
+            raise ValueError(
+                f"The classes, {unweighted_classes}, are not in" " class_weight"
+            )
+
+    return weight

sklearnex/utils/tests/test_class_weight.py

@@ -0,0 +1,69 @@
+# ==============================================================================
+# Copyright contributors to the oneDAL Project
+#
+# 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
+import pytest
+from sklearn.datasets import load_iris
+
+from daal4py.sklearn._utils import sklearn_check_version
+from onedal.tests.utils._dataframes_support import (
+    _as_numpy,
+    _convert_to_dataframe,
+    get_dataframes_and_queues,
+)
+from sklearnex import config_context
+from sklearnex.utils.class_weight import _compute_class_weight
+from sklearnex.utils.class_weight import compute_class_weight as sk_compute_class_weight
+
+
+@pytest.mark.skipif(not sklearn_check_version("1.6"), reason="lacks array API support")
+@pytest.mark.parametrize("class_weight", [None, "balanced", "ramp"])
+@pytest.mark.parametrize("dataframe,queue", get_dataframes_and_queues("array_api,dpctl"))
+def test_compute_class_weight_array_api(class_weight, dataframe, queue):
+    # This verifies that array_api functionality matches sklearn
+
+    _, y = load_iris(return_X_y=True)
+    classes = np.unique(y)
+
+    y_xp = _convert_to_dataframe(y, target_df=dataframe, device=queue)
+    classes_xp = _convert_to_dataframe(classes, target_df=dataframe, device=queue)
+
+    rng = np.random.default_rng(seed=42)
+
+    # support of sample_weights added in sklearn 1.7
+    set_sample_weight = class_weight == "balanced" and sklearn_check_version("1.7")
+
+    sample_weight = rng.random(y.shape).astype(np.float64) if set_sample_weight else None
+
+    if class_weight == "ramp":
+        class_weight = {int(i): int(i) for i in np.unique(y)}
+
+    weight_np = sk_compute_class_weight(
+        class_weight, classes=classes, y=y, sample_weight=sample_weight
+    )
+
+    if set_sample_weight:
+        sample_weight = _convert_to_dataframe(
+            sample_weight, target_df=dataframe, device=queue
+        )
+
+    # evaluate custom sklearnex array API functionality
+    with config_context(array_api_dispatch=True):
+        weight_xp = _compute_class_weight(
+            class_weight, classes=classes_xp, y=y_xp, sample_weight=sample_weight
+        )
+
+    np.testing.assert_allclose(_as_numpy(weight_xp), weight_np)