scikit-learn · thomasjpfan · Jul 26, 2022 · Jul 25, 2022 · Jul 25, 2022 · Jul 25, 2022
diff --git a/sklearn/utils/_testing.py b/sklearn/utils/_testing.py
@@ -22,6 +22,7 @@
 import re
 import contextlib
 from collections.abc import Iterable
+from collections.abc import Sequence
 
 import scipy as sp
 from functools import wraps
@@ -60,6 +61,7 @@
     check_is_fitted,
     check_X_y,
 )
+from sklearn.utils.fixes import threadpool_info
 
 
 __all__ = [
@@ -602,6 +604,38 @@ def __exit__(self, exc_type, exc_val, exc_tb):
         _delete_folder(self.temp_folder)
 
 
+def _create_memmap_backed_array(array, filename, mmap_mode):
+    # https://numpy.org/doc/stable/reference/generated/numpy.memmap.html
+    fp = np.memmap(filename, dtype=array.dtype, mode="w+", shape=array.shape)
+    fp[:] = array[:]  # write array to memmap array
+    fp.flush()
+    memmap_backed_array = np.memmap(
+        filename, dtype=array.dtype, mode=mmap_mode, shape=array.shape
+    )
+    return memmap_backed_array
+
+
+def _create_aligned_memmap_backed_arrays(data, mmap_mode, folder):
+    if isinstance(data, np.ndarray):
+        filename = op.join(folder, "data.dat")
+        return _create_memmap_backed_array(data, filename, mmap_mode)
+
+    if isinstance(data, Sequence) and all(
+        isinstance(each, np.ndarray) for each in data
+    ):
+        return [
+            _create_memmap_backed_array(
+                array, op.join(folder, f"data{index}.dat"), mmap_mode
+            )
+            for index, array in enumerate(data)
+        ]
+
+    raise ValueError(
+        "When creating aligned memmap-backed arrays, input must be a single array or a"
+        " sequence of arrays"
+    )
+
+
 def create_memmap_backed_data(data, mmap_mode="r", return_folder=False, aligned=False):
     """
     Parameters
@@ -616,18 +650,23 @@ def create_memmap_backed_data(data, mmap_mode="r", return_folder=False, aligned=
     """
     temp_folder = tempfile.mkdtemp(prefix="sklearn_testing_")
     atexit.register(functools.partial(_delete_folder, temp_folder, warn=True))
+    # OpenBLAS is known to segfault with unaligned data on the Prescott
+    # architecture so force aligned=True on Prescott. For more details, see:
+    # https://github.com/scipy/scipy/issues/14886
+    has_prescott_openblas = any(
+        True
+        for info in threadpool_info()
+        if info["internal_api"] == "openblas"
+        # Prudently assume Prescott might be the architecture if it is unknown.
+        and info.get("architecture", "prescott").lower() == "prescott"
+    )
+    if has_prescott_openblas:
+        aligned = True
+
     if aligned:
-        if isinstance(data, np.ndarray) and data.flags.aligned:
-            # https://numpy.org/doc/stable/reference/generated/numpy.memmap.html
-            filename = op.join(temp_folder, "data.dat")
-            fp = np.memmap(filename, dtype=data.dtype, mode="w+", shape=data.shape)
-            fp[:] = data[:]  # write data to memmap array
-            fp.flush()
-            memmap_backed_data = np.memmap(
-                filename, dtype=data.dtype, mode=mmap_mode, shape=data.shape
-            )
-        else:
-            raise ValueError("If aligned=True, input must be a single numpy array.")
+        memmap_backed_data = _create_aligned_memmap_backed_arrays(
+            data, mmap_mode, temp_folder
+        )
     else:
         filename = op.join(temp_folder, "data.pkl")
         joblib.dump(data, filename)

diff --git a/sklearn/utils/estimator_checks.py b/sklearn/utils/estimator_checks.py
@@ -54,7 +54,6 @@
 from ..model_selection import ShuffleSplit
 from ..model_selection._validation import _safe_split
 from ..metrics.pairwise import rbf_kernel, linear_kernel, pairwise_distances
-from ..utils.fixes import threadpool_info
 from ..utils.fixes import sp_version
 from ..utils.fixes import parse_version
 from ..utils.validation import check_is_fitted
@@ -2120,22 +2119,6 @@ def check_classifiers_one_label(name, classifier_orig):
         assert_array_equal(classifier.predict(X_test), y, err_msg=error_string_predict)
 
 
-def _create_memmap_backed_data(numpy_arrays):
-    # OpenBLAS is known to segfault with unaligned data on the Prescott architecture
-    # See: https://github.com/scipy/scipy/issues/14886
-    has_prescott_openblas = any(
-        True
-        for info in threadpool_info()
-        if info["internal_api"] == "openblas"
-        # Prudently assume Prescott might be the architecture if it is unknown.
-        and info.get("architecture", "prescott").lower() == "prescott"
-    )
-    return [
-        create_memmap_backed_data(array, aligned=has_prescott_openblas)
-        for array in numpy_arrays
-    ]
-
-
 @ignore_warnings  # Warnings are raised by decision function
 def check_classifiers_train(
     name, classifier_orig, readonly_memmap=False, X_dtype="float64"
@@ -2153,7 +2136,7 @@ def check_classifiers_train(
         X_b -= X_b.min()
 
     if readonly_memmap:
-        X_m, y_m, X_b, y_b = _create_memmap_backed_data([X_m, y_m, X_b, y_b])
+        X_m, y_m, X_b, y_b = create_memmap_backed_data([X_m, y_m, X_b, y_b])
 
     problems = [(X_b, y_b)]
     tags = _safe_tags(classifier_orig)
@@ -2814,7 +2797,7 @@ def check_regressors_train(
         y_ = y
 
     if readonly_memmap:
-        X, y, y_ = _create_memmap_backed_data([X, y, y_])
+        X, y, y_ = create_memmap_backed_data([X, y, y_])
 
     if not hasattr(regressor, "alphas") and hasattr(regressor, "alpha"):
         # linear regressors need to set alpha, but not generalized CV ones

diff --git a/sklearn/utils/tests/test_testing.py b/sklearn/utils/tests/test_testing.py
@@ -702,16 +702,19 @@ def test_create_memmap_backed_data(monkeypatch, aligned):
     assert registration_counter.nb_calls == 3
 
     input_list = [input_array, input_array + 1, input_array + 2]
-    if aligned:
-        with pytest.raises(
-            ValueError, match="If aligned=True, input must be a single numpy array."
-        ):
-            create_memmap_backed_data(input_list, aligned=True)
-    else:
-        mmap_data_list = create_memmap_backed_data(input_list, aligned=False)
-        for input_array, data in zip(input_list, mmap_data_list):
-            check_memmap(input_array, data)
-        assert registration_counter.nb_calls == 4
+    mmap_data_list = create_memmap_backed_data(input_list, aligned=aligned)
+    for input_array, data in zip(input_list, mmap_data_list):
+        check_memmap(input_array, data)
+    assert registration_counter.nb_calls == 4
+
+    with pytest.raises(
+        ValueError,
+        match=(
+            "When creating aligned memmap-backed arrays, input must be a single array"
+            " or a sequence of arrays"
+        ),
+    ):
+        create_memmap_backed_data([input_array, "not-an-array"], aligned=True)
 
 
 @pytest.mark.parametrize("dtype", [np.float32, np.float64, np.int32, np.int64])