scikit-learn-intelex 2025.4.0__py313-none-manylinux_2_28_x86_64.whl

onedal/covariance/tests/test_incremental_covariance.py

@@ -0,0 +1,190 @@
+# ===============================================================================
+# 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
+import pytest
+from numpy.testing import assert_allclose
+
+from onedal.datatypes import from_table
+from onedal.tests.utils._device_selection import get_queues
+
+
+@pytest.mark.parametrize("queue", get_queues())
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+def test_on_gold_data_unbiased(queue, dtype):
+    from onedal.covariance import IncrementalEmpiricalCovariance
+
+    X = np.array([[0, 1], [0, 1]])
+    X = X.astype(dtype)
+    X_split = np.array_split(X, 2)
+    inccov = IncrementalEmpiricalCovariance()
+
+    for i in range(2):
+        inccov.partial_fit(X_split[i], queue=queue)
+    result = inccov.finalize_fit()
+
+    expected_covariance = np.array([[0, 0], [0, 0]])
+    expected_means = np.array([0, 1])
+
+    assert_allclose(expected_covariance, result.covariance_)
+    assert_allclose(expected_means, result.location_)
+
+    X = np.array([[1, 2], [3, 6]])
+    X = X.astype(dtype)
+    X_split = np.array_split(X, 2)
+    inccov = IncrementalEmpiricalCovariance()
+
+    for i in range(2):
+        inccov.partial_fit(X_split[i], queue=queue)
+    result = inccov.finalize_fit()
+
+    expected_covariance = np.array([[2, 4], [4, 8]])
+    expected_means = np.array([2, 4])
+
+    assert_allclose(expected_covariance, result.covariance_)
+    assert_allclose(expected_means, result.location_)
+
+
+@pytest.mark.parametrize("queue", get_queues())
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+def test_on_gold_data_biased(queue, dtype):
+    from onedal.covariance import IncrementalEmpiricalCovariance
+
+    X = np.array([[0, 1], [0, 1]])
+    X = X.astype(dtype)
+    X_split = np.array_split(X, 2)
+    inccov = IncrementalEmpiricalCovariance(bias=True)
+
+    for i in range(2):
+        inccov.partial_fit(X_split[i], queue=queue)
+    result = inccov.finalize_fit()
+
+    expected_covariance = np.array([[0, 0], [0, 0]])
+    expected_means = np.array([0, 1])
+
+    assert_allclose(expected_covariance, result.covariance_)
+    assert_allclose(expected_means, result.location_)
+
+    X = np.array([[1, 2], [3, 6]])
+    X = X.astype(dtype)
+    X_split = np.array_split(X, 2)
+    inccov = IncrementalEmpiricalCovariance(bias=True)
+
+    for i in range(2):
+        inccov.partial_fit(X_split[i], queue=queue)
+    result = inccov.finalize_fit()
+
+    expected_covariance = np.array([[1, 2], [2, 4]])
+    expected_means = np.array([2, 4])
+
+    assert_allclose(expected_covariance, result.covariance_)
+    assert_allclose(expected_means, result.location_)
+
+
+@pytest.mark.parametrize("queue", get_queues())
+@pytest.mark.parametrize("num_batches", [2, 4, 6, 8, 10])
+@pytest.mark.parametrize("row_count", [100, 1000, 2000])
+@pytest.mark.parametrize("column_count", [10, 100, 200])
+@pytest.mark.parametrize("bias", [True, False])
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+def test_partial_fit_on_random_data(
+    queue, num_batches, row_count, column_count, bias, dtype
+):
+    from onedal.covariance import IncrementalEmpiricalCovariance
+
+    seed = 77
+    gen = np.random.default_rng(seed)
+    X = gen.uniform(low=-0.3, high=+0.7, size=(row_count, column_count))
+    X = X.astype(dtype)
+    X_split = np.array_split(X, num_batches)
+    inccov = IncrementalEmpiricalCovariance(bias=bias)
+
+    for i in range(num_batches):
+        inccov.partial_fit(X_split[i], queue=queue)
+    result = inccov.finalize_fit()
+
+    expected_covariance = np.cov(X.T, bias=bias)
+    expected_means = np.mean(X, axis=0)
+
+    assert_allclose(expected_covariance, result.covariance_, atol=1e-6)
+    assert_allclose(expected_means, result.location_, atol=1e-6)
+
+
+@pytest.mark.parametrize("queue", get_queues())
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+def test_incremental_estimator_pickle(queue, dtype):
+    import pickle
+
+    from onedal.covariance import IncrementalEmpiricalCovariance
+
+    inccov = IncrementalEmpiricalCovariance()
+
+    # Check that estimator can be serialized without any data.
+    dump = pickle.dumps(inccov)
+    inccov_loaded = pickle.loads(dump)
+    seed = 77
+    gen = np.random.default_rng(seed)
+    X = gen.uniform(low=-0.3, high=+0.7, size=(10, 10))
+    X = X.astype(dtype)
+    X_split = np.array_split(X, 2)
+    inccov.partial_fit(X_split[0], queue=queue)
+    inccov_loaded.partial_fit(X_split[0], queue=queue)
+
+    assert inccov._need_to_finalize == True
+    assert inccov_loaded._need_to_finalize == True
+
+    # Check that estimator can be serialized after partial_fit call.
+    dump = pickle.dumps(inccov)
+    inccov_loaded = pickle.loads(dump)
+
+    assert inccov._need_to_finalize == False
+    # Finalize is called during serialization to make sure partial results are finalized correctly.
+    assert inccov_loaded._need_to_finalize == False
+
+    partial_n_rows = from_table(inccov._partial_result.partial_n_rows)
+    partial_n_rows_loaded = from_table(inccov_loaded._partial_result.partial_n_rows)
+    assert_allclose(partial_n_rows, partial_n_rows_loaded)
+
+    partial_crossproduct = from_table(inccov._partial_result.partial_crossproduct)
+    partial_crossproduct_loaded = from_table(
+        inccov_loaded._partial_result.partial_crossproduct
+    )
+    assert_allclose(partial_crossproduct, partial_crossproduct_loaded)
+
+    partial_sums = from_table(inccov._partial_result.partial_sums)
+    partial_sums_loaded = from_table(inccov_loaded._partial_result.partial_sums)
+    assert_allclose(partial_sums, partial_sums_loaded)
+
+    inccov.partial_fit(X_split[1], queue=queue)
+    inccov_loaded.partial_fit(X_split[1], queue=queue)
+    assert inccov._need_to_finalize == True
+    assert inccov_loaded._need_to_finalize == True
+
+    dump = pickle.dumps(inccov_loaded)
+    inccov_loaded = pickle.loads(dump)
+
+    assert inccov._need_to_finalize == True
+    assert inccov_loaded._need_to_finalize == False
+
+    inccov.finalize_fit()
+    inccov_loaded.finalize_fit()
+
+    # Check that finalized estimator can be serialized.
+    dump = pickle.dumps(inccov_loaded)
+    inccov_loaded = pickle.loads(dump)
+
+    assert_allclose(inccov.location_, inccov_loaded.location_, atol=1e-6)
+    assert_allclose(inccov.covariance_, inccov_loaded.covariance_, atol=1e-6)

onedal/datatypes/__init__.py

@@ -0,0 +1,19 @@
+# ==============================================================================
+# 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.
+# ==============================================================================
+
+from ._data_conversion import from_table, to_table
+
+__all__ = ["from_table", "to_table"]

onedal/datatypes/_data_conversion.py

@@ -0,0 +1,121 @@
+# ==============================================================================
+# 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 warnings
+
+import numpy as np
+
+from onedal import _backend, _is_dpc_backend
+
+
+def _apply_and_pass(func, *args, **kwargs):
+    if len(args) == 1:
+        return func(args[0], **kwargs)
+    return tuple(map(lambda arg: func(arg, **kwargs), args))
+
+
+def _convert_one_to_table(arg, queue=None):
+    # All inputs for table conversion must be array-like or sparse, not scalars
+    return _backend.to_table(np.atleast_2d(arg) if np.isscalar(arg) else arg, queue)
+
+
+def to_table(*args, queue=None):
+    """Create oneDAL tables from scalars and/or arrays.
+
+    Note: this implementation can be used with scipy.sparse, numpy ndarrays,
+    DPCTL/DPNP usm_ndarrays and scalars. Tables will use pointers to the
+    original array data. Scalars and non-contiguous arrays will be copies.
+    Arrays may be modified in-place by oneDAL during computation. This works
+    for data located on CPU and SYCL-enabled Intel GPUs. Each array may only
+    be of a single datatype (i.e. each must be homogeneous).
+
+    Parameters
+    ----------
+    *args : {scalar, numpy array, sycl_usm_ndarray, csr_matrix, or csr_array}
+        arg1, arg2... The arrays should be given as arguments.
+
+    Returns
+    -------
+    tables: {oneDAL homogeneous tables}
+    """
+    return _apply_and_pass(_convert_one_to_table, *args, queue=queue)
+
+
+if _is_dpc_backend:
+
+    try:
+        # try/catch is used here instead of dpep_helpers because
+        # of circular import issues of _data_conversion.py and
+        # utils/validation.py. This is a temporary fix until the
+        # issue with dpnp is addressed, at which point this can
+        # be removed entirely.
+        import dpnp
+
+        def _table_to_array(table, xp=None):
+            # By default DPNP ndarray created with a copy.
+            # TODO:
+            # investigate why dpnp.array(table, copy=False) doesn't work.
+            # Work around with using dpctl.tensor.asarray.
+            if xp == dpnp:
+                return dpnp.array(dpnp.dpctl.tensor.asarray(table), copy=False)
+            else:
+                return xp.asarray(table)
+
+    except ImportError:
+
+        def _table_to_array(table, xp=None):
+            return xp.asarray(table)
+
+    from ..common._policy import _HostInteropPolicy
+
+    def convert_one_from_table(table, sycl_queue=None, sua_iface=None, xp=None):
+        # Currently only `__sycl_usm_array_interface__` protocol used to
+        # convert into dpnp/dpctl tensors.
+        if sua_iface:
+            if (
+                sycl_queue
+                and sycl_queue.sycl_device.is_cpu
+                and table.__sycl_usm_array_interface__["syclobj"] is None
+            ):
+                # oneDAL returns tables with None sycl queue for CPU sycl queue inputs.
+                # This workaround is necessary for the functional preservation
+                # of the compute-follows-data execution.
+                # Host tables first converted into numpy.narrays and then to array from xp
+                # namespace.
+                return xp.asarray(
+                    _backend.from_table(table), usm_type="device", sycl_queue=sycl_queue
+                )
+            else:
+                return _table_to_array(table, xp=xp)
+
+        return _backend.from_table(table)
+
+else:
+
+    def convert_one_from_table(table, sycl_queue=None, sua_iface=None, xp=None):
+        # Currently only `__sycl_usm_array_interface__` protocol used to
+        # convert into dpnp/dpctl tensors.
+        if sua_iface:
+            raise RuntimeError(
+                "SYCL usm array conversion from table requires the DPC backend"
+            )
+        return _backend.from_table(table)
+
+
+def from_table(*args, sycl_queue=None, sua_iface=None, xp=None):
+    return _apply_and_pass(
+        convert_one_from_table, *args, sycl_queue=sycl_queue, sua_iface=sua_iface, xp=xp
+    )

onedal/datatypes/tests/common.py

@@ -0,0 +1,126 @@
+# ===============================================================================
+# 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 onedal.utils._dpep_helpers import dpctl_available, dpnp_available
+
+if dpnp_available:
+    import dpnp
+
+if dpctl_available:
+    import dpctl
+    from dpctl.tensor import usm_ndarray
+
+    def _get_sycl_queue(syclobj):
+        if hasattr(syclobj, "_get_capsule"):
+            return dpctl.SyclQueue(syclobj._get_capsule())
+        else:
+            return dpctl.SyclQueue(syclobj)
+
+    def _assert_tensor_attr(actual, desired, order):
+        """Check attributes of two given USM tensors."""
+        is_usm_tensor = (
+            lambda x: dpnp_available
+            and isinstance(x, dpnp.ndarray)
+            or isinstance(x, usm_ndarray)
+        )
+        assert is_usm_tensor(actual)
+        assert is_usm_tensor(desired)
+        # dpctl.tensor is the dpnp.ndarrays's core tensor structure along
+        # with advanced device management. Convert dpnp to dpctl.tensor with zero copy.
+        get_tensor = lambda x: (
+            x.get_array() if dpnp_available and isinstance(x, dpnp.ndarray) else x
+        )
+        # Now DPCtl tensors
+        actual = get_tensor(actual)
+        desired = get_tensor(desired)
+
+        assert actual.shape == desired.shape
+        assert actual.strides == desired.strides
+        assert actual.dtype == desired.dtype
+        if order == "F":
+            assert actual.flags.f_contiguous
+            assert desired.flags.f_contiguous
+            assert actual.flags.f_contiguous == desired.flags.f_contiguous
+        else:
+            assert actual.flags.c_contiguous
+            assert desired.flags.c_contiguous
+            assert actual.flags.c_contiguous == desired.flags.c_contiguous
+        assert actual.flags == desired.flags
+        assert actual.sycl_queue == desired.sycl_queue
+        # TODO:
+        # check better way to check usm ptrs.
+        assert actual.usm_data._pointer == desired.usm_data._pointer
+
+    def _assert_sua_iface_fields(
+        actual, desired, skip_syclobj=False, skip_data_0=False, skip_data_1=False
+    ):
+        """Check attributes of two given reprsesentations of
+        USM allocations `__sycl_usm_array_interface__`.
+
+        For full documentation about `__sycl_usm_array_interface__` refer
+        https://intelpython.github.io/dpctl/latest/api_reference/dpctl/sycl_usm_array_interface.html.
+
+        Parameters
+        ----------
+        actual : dict, __sycl_usm_array_interface__
+        desired : dict, __sycl_usm_array_interface__
+        skip_syclobj : bool, default=False
+            If True, check for __sycl_usm_array_interface__["syclobj"]
+            will be skipped.
+        skip_data_0 : bool, default=False
+            If True, check for __sycl_usm_array_interface__["data"][0]
+            will be skipped.
+        skip_data_1 : bool, default=False
+            If True, check for __sycl_usm_array_interface__["data"][1]
+            will be skipped.
+        """
+        assert hasattr(actual, "__sycl_usm_array_interface__")
+        assert hasattr(desired, "__sycl_usm_array_interface__")
+        actual_sua_iface = actual.__sycl_usm_array_interface__
+        desired_sua_iface = desired.__sycl_usm_array_interface__
+        # data: A 2-tuple whose first element is a Python integer encoding
+        # USM pointer value. The second entry in the tuple is a read-only flag
+        # (True means the data area is read-only).
+        if not skip_data_0:
+            assert actual_sua_iface["data"][0] == desired_sua_iface["data"][0]
+        if not skip_data_1:
+            assert actual_sua_iface["data"][1] == desired_sua_iface["data"][1]
+        # shape: a tuple of integers describing dimensions of an N-dimensional array.
+        # Reformating shapes for check cases (r,) vs (r,1). Contiguous flattened array
+        # shape (r,) becoming (r,1) just for the check, since oneDAL supports only (r,1)
+        # for 1-D arrays. In code after from_table conversion for 1-D expected outputs
+        # xp.ravel or reshape(-1) is used.
+        get_shape_if_1d = lambda shape: (shape[0], 1) if len(shape) == 1 else shape
+        actual_shape = get_shape_if_1d(actual_sua_iface["shape"])
+        desired_shape = get_shape_if_1d(desired_sua_iface["shape"])
+        assert actual_shape == desired_shape
+        # strides: An optional tuple of integers describing number of array elements
+        # needed to jump to the next array element in the corresponding dimensions.
+        if not actual_sua_iface["strides"] and not desired_sua_iface["strides"]:
+            # None to indicate a C-style contiguous 1D array.
+            # onedal4py constructs __sycl_usm_array_interface__["strides"] with
+            # real values.
+            assert actual_sua_iface["strides"] == desired_sua_iface["strides"]
+        # versions: Version of the interface.
+        assert actual_sua_iface["version"] == desired_sua_iface["version"]
+        # typestr: a string encoding elemental data type of the array.
+        assert actual_sua_iface["typestr"] == desired_sua_iface["typestr"]
+        # syclobj: Python object from which SYCL context to which represented USM
+        # allocation is bound.
+        if not skip_syclobj and dpctl_available:
+            actual_sycl_queue = _get_sycl_queue(actual_sua_iface["syclobj"])
+            desired_sycl_queue = _get_sycl_queue(desired_sua_iface["syclobj"])
+            assert actual_sycl_queue == desired_sycl_queue