scikit-learn · ogrisel · Sep 11, 2024 · Jul 12, 2024 · Aug 22, 2024 · Aug 22, 2024
diff --git a/sklearn/utils/_array_api.py b/sklearn/utils/_array_api.py
@@ -10,6 +10,7 @@
 
 import numpy
 import scipy
+import scipy.sparse as sp
 import scipy.special as special
 
 from .._config import get_config
@@ -162,7 +163,9 @@ def device(*array_list, remove_none=True, remove_types=(str,)):
         *array_list, remove_none=remove_none, remove_types=remove_types
     )
 
-    # Note that _remove_non_arrays ensures that array_list is not empty.
+    if not array_list:
+        return None
+
     device_ = _single_array_device(array_list[0])
 
     # Note: here we cannot simply use a Python `set` as it requires
@@ -451,6 +454,8 @@ def _remove_non_arrays(*arrays, remove_none=True, remove_types=(str,)):
 
     Raise ValueError if no arrays are left after filtering.
 
+    Sparse arrays are always filtered out.
+
     Parameters
     ----------
     *arrays : array objects
@@ -465,7 +470,8 @@ def _remove_non_arrays(*arrays, remove_none=True, remove_types=(str,)):
     Returns
     -------
     filtered_arrays : list
-        List of arrays with None and typoe
+        List of arrays filtered as requested. An empty list is returned if no input
+        passes the filters.
     """
     filtered_arrays = []
     remove_types = tuple(remove_types)
@@ -474,14 +480,10 @@ def _remove_non_arrays(*arrays, remove_none=True, remove_types=(str,)):
             continue
         if isinstance(array, remove_types):
             continue
+        if sp.issparse(array):
+            continue
         filtered_arrays.append(array)
 
-    if not filtered_arrays:
-        raise ValueError(
-            f"At least one input array expected after filtering with {remove_none=}, "
-            f"remove_types=[{', '.join(t.__name__ for t in remove_types)}]. Got none. "
-            f"Original types: [{', '.join(type(a).__name__ for a in arrays)}]."
-        )
     return filtered_arrays
 
 
@@ -491,6 +493,8 @@ def get_namespace(*arrays, remove_none=True, remove_types=(str,), xp=None):
     Introspect `arrays` arguments and return their common Array API compatible
     namespace object, if any.
 
+    Note that sparse arrays are filtered by default.
+
     See: https://numpy.org/neps/nep-0047-array-api-standard.html
 
     If `arrays` are regular numpy arrays, an instance of the `_NumPyAPIWrapper`
@@ -509,6 +513,9 @@ def get_namespace(*arrays, remove_none=True, remove_types=(str,), xp=None):
     always returned irrespective of the fact that arrays implement the
     `__array_namespace__` protocol or not.
 
+    Note that if no arrays pass the set filters, ``_NUMPY_API_WRAPPER_INSTANCE, False``
+    is returned.
+
     Parameters
     ----------
     *arrays : array objects
@@ -546,9 +553,14 @@ def get_namespace(*arrays, remove_none=True, remove_types=(str,), xp=None):
         return xp, True
 
     arrays = _remove_non_arrays(
-        *arrays, remove_none=remove_none, remove_types=remove_types
+        *arrays,
+        remove_none=remove_none,
+        remove_types=remove_types,
     )
 
+    if not arrays:
+        return _NUMPY_API_WRAPPER_INSTANCE, False
+
     _check_array_api_dispatch(array_api_dispatch)
 
     # array-api-compat is a required dependency of scikit-learn only when
@@ -591,20 +603,19 @@ def get_namespace_and_device(*array_list, remove_none=True, remove_types=(str,))
         `device` object (see the "Device Support" section of the array API spec).
     """
     array_list = _remove_non_arrays(
-        *array_list, remove_none=remove_none, remove_types=remove_types
+        *array_list,
+        remove_none=remove_none,
+        remove_types=remove_types,
     )
 
     skip_remove_kwargs = dict(remove_none=False, remove_types=[])
 
     xp, is_array_api = get_namespace(*array_list, **skip_remove_kwargs)
+    arrays_device = device(*array_list, **skip_remove_kwargs)
     if is_array_api:
-        return (
-            xp,
-            is_array_api,
-            device(*array_list, **skip_remove_kwargs),
-        )
+        return xp, is_array_api, arrays_device
     else:
-        return xp, False, None
+        return xp, False, arrays_device
 
 
 def _expit(X, xp=None):

diff --git a/sklearn/utils/tests/test_array_api.py b/sklearn/utils/tests/test_array_api.py
@@ -1,5 +1,4 @@
 import os
-import re
 from functools import partial
 
 import numpy
@@ -31,6 +30,7 @@
     yield_namespace_device_dtype_combinations,
 )
 from sklearn.utils._testing import (
+    SkipTest,
     _array_api_for_tests,
     assert_array_equal,
     skip_if_array_api_compat_not_configured,
@@ -238,20 +238,10 @@ def test_average_raises_with_invalid_parameters(
         _average(array_in, axis=axis, weights=weights)
 
 
-def test_device_raises_if_no_input():
-    err_msg = re.escape(
-        "At least one input array expected after filtering with remove_none=True, "
-        "remove_types=[str]. Got none. Original types: []."
-    )
-    with pytest.raises(ValueError, match=err_msg):
-        device()
+def test_device_none_if_no_input():
+    assert device() is None
 
-    err_msg = re.escape(
-        "At least one input array expected after filtering with remove_none=True, "
-        "remove_types=[str]. Got none. Original types: [NoneType, str]."
-    )
-    with pytest.raises(ValueError, match=err_msg):
-        device(None, "name")
+    assert device(None, "name") is None
 
 
 def test_device_inspection():
@@ -540,13 +530,13 @@ def test_get_namespace_and_device():
     some_numpy_array = numpy.arange(3)
 
     # When dispatch is disabled, get_namespace_and_device should return the
-    # default NumPy wrapper namespace and no device. Our code will handle such
+    # default NumPy wrapper namespace and "cpu" device. Our code will handle such
     # inputs via the usual __array__ interface without attempting to dispatch
     # via the array API.
     namespace, is_array_api, device = get_namespace_and_device(some_torch_tensor)
     assert namespace is get_namespace(some_numpy_array)[0]
     assert not is_array_api
-    assert device is None
+    assert device.type == "cpu"
 
     # Otherwise, expose the torch namespace and device via array API compat
     # wrapper.
@@ -605,3 +595,17 @@ def test_fill_or_add_to_diagonal(array_namespace, device_, dtype_name, wrap):
     _fill_or_add_to_diagonal(array_xp, value=1, xp=xp, add_value=False, wrap=wrap)
     numpy.fill_diagonal(array_np, val=1, wrap=wrap)
     assert_array_equal(_convert_to_numpy(array_xp, xp=xp), array_np)
+
+
+@pytest.mark.parametrize("csr_container", CSR_CONTAINERS)
+@pytest.mark.parametrize("dispatch", [True, False])
+def test_sparse_device(csr_container, dispatch):
+    a, b = csr_container(numpy.array([[1]])), csr_container(numpy.array([[2]]))
+    try:
+        with config_context(array_api_dispatch=dispatch):
+            assert device(a, b) is None
+            assert device(a, numpy.array([1])) == "cpu"
+            assert get_namespace_and_device(a, b)[2] is None
+            assert get_namespace_and_device(a, numpy.array([1]))[2] == "cpu"
+    except ImportError:
+        raise SkipTest("array_api_compat is not installed")
diff --git a/sklearn/utils/tests/test_validation.py b/sklearn/utils/tests/test_validation.py
@@ -2200,6 +2200,16 @@ def test_check_array_writeable_df():
     assert out.flags.writeable
 
 
+@skip_if_array_api_compat_not_configured
+def test_check_array_on_sparse_inputs_with_array_api_enabled():
+    X_sp = sp.csr_array([[0, 1, 0], [1, 0, 1]])
+    with config_context(array_api_dispatch=True):
+        assert sp.issparse(check_array(X_sp, accept_sparse=True))
+
+        with pytest.raises(TypeError):
+            check_array(X_sp)
+
+
 # TODO(1.8): remove
 def test_force_all_finite_rename_warning():
     X = np.random.uniform(size=(10, 10))