TIF_E41253394/venv/Lib/site-packages/sklearn/tests/test_build.py

99 lines
4.1 KiB
Python

import importlib
import inspect
import os
import pkgutil
import textwrap
import pytest
import sklearn
from sklearn import __version__
from sklearn.utils._openmp_helpers import _openmp_parallelism_enabled
@pytest.mark.thread_unsafe # import side-effects
def test_extension_type_module():
"""Check that Cython extension types have a correct ``__module__``.
When a subpackage containing Cython extension types has a misconfigured
``meson.build`` (e.g. missing ``__init__.py`` in its Cython tree), Cython
cannot detect the package hierarchy and sets ``__module__`` to just the
submodule name (e.g. ``'_loss'``) instead of the fully qualified
``'sklearn._loss._loss'``. This breaks downstream tools like skops that
rely on ``__module__`` for serialization.
"""
sklearn_path = [os.path.dirname(sklearn.__file__)]
failures = []
for _, modname, ispkg in pkgutil.walk_packages(
path=sklearn_path, prefix="sklearn.", onerror=lambda _: None
):
# Packages are directories, not modules that can hold extension
# types. ``tests``, ``externals`` (vendored third-party code),
# ``_build_utils`` (build-time helpers that import ``Cython``, which
# is not installed in the wheel test environment) and ``_cyutility`` (that is a
# Cython shared utility module whose classes report a runtime-level module path)
# are out of scope for this check. The latter is needed for the pyodide
# platform in particular.
if (
ispkg
or ".tests." in modname
or ".externals." in modname
or "._build_utils." in modname
or "._cyutility" in modname
):
continue
mod = importlib.import_module(modname)
mod_file = getattr(mod, "__file__", "") or ""
# Only compiled extension modules can produce the misconfigured
# ``__module__`` this test guards against. Pure-Python modules get
# the correct ``__module__`` from the import system by construction.
if not mod_file.endswith((".so", ".pyd")):
continue
for name, cls in inspect.getmembers(mod, inspect.isclass):
try:
cls_file = inspect.getfile(cls)
except TypeError: # pragma: no cover
# Raised for built-in types (``object``, stdlib C types) that
# have no source file — they were not defined in ``mod``.
continue # pragma: no cover
# Skip classes imported into ``mod`` from elsewhere (e.g. numpy,
# scipy, or another sklearn module). Only classes whose source
# file *is* this extension's .so are candidates for the bug.
if cls_file != mod_file:
continue
if cls.__module__ != modname:
failures.append( # pragma: no cover
f"{modname}.{name}.__module__ == {cls.__module__!r}, "
f"expected {modname!r}"
)
assert not failures, "Extension types with incorrect __module__:\n" + "\n".join(
failures
)
def test_openmp_parallelism_enabled():
# Check that sklearn is built with OpenMP-based parallelism enabled.
# This test can be skipped by setting the environment variable
# ``SKLEARN_SKIP_OPENMP_TEST``.
if os.getenv("SKLEARN_SKIP_OPENMP_TEST"):
pytest.skip("test explicitly skipped (SKLEARN_SKIP_OPENMP_TEST)")
base_url = "dev" if __version__.endswith(".dev0") else "stable"
err_msg = textwrap.dedent(
"""
This test fails because scikit-learn has been built without OpenMP.
This is not recommended since some estimators will run in sequential
mode instead of leveraging thread-based parallelism.
You can find instructions to build scikit-learn with OpenMP at this
address:
https://scikit-learn.org/{}/developers/advanced_installation.html
You can skip this test by setting the environment variable
SKLEARN_SKIP_OPENMP_TEST to any value.
"""
).format(base_url)
assert _openmp_parallelism_enabled(), err_msg