check_estimators check_estimators_pickle fails on Prescott architecture

[DerWeh]

Describe the bug

On machines with the Prescott architecture, tests using check_estimator unexpectedly fail due to #23994, which forces aligned=True. To my mind, this makes little sense as an estimator cannot be aligned which only makes sense for simple arrays.

I have no idea what the intent of #23994, is. Maybe replacing

if has_prescott_openblas:
    aligned = True

by

if isinstance(data, np.ndarray) and data.flags.aligned and has_prescott_openblas:
    aligned = True

is sufficient to fix this? But as I said, I don't really know what the point of that PR was.

Steps/Code to Reproduce

from sklearn.ensemble import RandomForestClassifier
from sklearn.utils.estimator_checks import check_estimator

for estimator, check in check_estimator(RandomForestClassifier(), generate_only=True):
    if "check_estimators_pickle" not in check.func.__name__:
         continue
    check(estimator)

Expected Results

Passes

Actual Results

On the Prescott architecture, this results in the following error:

ValueError                                Traceback (most recent call last)
Cell In[2], line 13
      9 if "check_estimators_pickle" not in check.func.__name__:
     11      continue
---> 13 check(estimator)

File ~/micromamba/envs/RandomForestClassifier/lib/python3.10/site-packages/sklearn/utils/_testing.py:156, in _IgnoreWarnings.__call__.<locals>.wrapper(*args, **kwargs)
    154 with warnings.catch_warnings():
    155     warnings.simplefilter("ignore", self.category)
--> 156     return fn(*args, **kwargs)

File ~/micromamba/envs/RandomForestClassifier/lib/python3.10/site-packages/sklearn/utils/estimator_checks.py:2007, in check_estimators_pickle(name, estimator_orig, readonly_memmap)
   2004 estimator.fit(X, y)
   2006 if readonly_memmap:
-> 2007     unpickled_estimator = create_memmap_backed_data(estimator)
   2008 else:
   2009     # pickle and unpickle!
   2010     pickled_estimator = pickle.dumps(estimator)

File ~/micromamba/envs/RandomForestClassifier/lib/python3.10/site-packages/sklearn/utils/_testing.py:510, in create_memmap_backed_data(data, mmap_mode, return_folder, aligned)
    507     aligned = True
    509 if aligned:
--> 510     memmap_backed_data = _create_aligned_memmap_backed_arrays(
    511         data, mmap_mode, temp_folder
    512     )
    513 else:
    514     filename = op.join(temp_folder, "data.pkl")

File ~/micromamba/envs/RandomForestClassifier/lib/python3.10/site-packages/sklearn/utils/_testing.py:476, in _create_aligned_memmap_backed_arrays(data, mmap_mode, folder)
    466 if isinstance(data, Sequence) and all(
    467     isinstance(each, np.ndarray) for each in data
    468 ):
    469     return [
    470         _create_memmap_backed_array(
    471             array, op.join(folder, f"data{index}.dat"), mmap_mode
    472         )
    473         for index, array in enumerate(data)
    474     ]
--> 476 raise ValueError(
    477     "When creating aligned memmap-backed arrays, input must be a single array or a"
    478     " sequence of arrays"
    479 )

ValueError: When creating aligned memmap-backed arrays, input must be a single array or a sequence of arrays

Versions

System:
    python: 3.10.13 | packaged by conda-forge | (main, Oct 26 2023, 18:07:37) [GCC 12.3.0]
executable: ~/micromamba/envs/RandomForestClassifier/bin/python3.10
   machine: Linux-5.15.0-88-generic-x86_64-with-glibc2.31

Python dependencies:
      sklearn: 1.3.2
          pip: 23.3.1
   setuptools: 68.2.2
        numpy: 1.26.1
        scipy: 1.11.3
       Cython: 3.0.5
       pandas: None
   matplotlib: None
       joblib: 1.3.2
threadpoolctl: 3.2.0

Built with OpenMP: True

threadpoolctl info:
       user_api: openmp
   internal_api: openmp
    num_threads: 2
         prefix: libgomp
       filepath: ~/micromamba/envs/RandomForestClassifier/lib/python3.10/site-packages/scikit_learn.libs/libgomp-a34b3233.so.1.0.0
        version: None

       user_api: blas
   internal_api: openblas
    num_threads: 2
         prefix: libopenblas
       filepath: ~/micromamba/envs/RandomForestClassifier/lib/python3.10/site-packages/numpy.libs/libopenblas64_p-r0-0cf96a72.3.23.dev.so
        version: 0.3.23.dev
threading_layer: pthreads
   architecture: Prescott

       user_api: blas
   internal_api: openblas
    num_threads: 2
         prefix: libopenblas
       filepath: ~/micromamba/envs/RandomForestClassifier/lib/python3.10/site-packages/scipy.libs/libopenblasp-r0-23e5df77.3.21.dev.so
        version: 0.3.21.dev
threading_layer: pthreads
   architecture: Prescott

[glemaitre]

Are you sure that you have a Prescott architecture. It is a old Intel Pentium D.
If this is not the case then you probably have the same issue as reported in #27613.

It should be solved in #27614 that is merged in main and in OpenBLAS the issue linked with: OpenMathLib/OpenBLAS#4268

[glemaitre]

@DerWeh You can install the nightly build to check that we really solve the problem. But I am almost sure this is the case.

[DerWeh]

I forgot the drop the filter for open issues, so I overlooked #27614...

@glemaitre Sorry, I don't know how to check the architecture. But sklearn.utils.fixes.threadpool_info shows "Prescott". The machine in question is actually a CI server, so I have no access to install the nightly build.

However, the error can be reproduced on any architecture by monkey patching sklearn.utils.fixes.threadpool_info. The following code fails on any architecture:

# BEGIN: monkey patch architecture
import sklearn.utils.fixes


threadpool_info = sklearn.utils.fixes.threadpool_info


def hide_prescott():
    """Something is wrong in sklearn for Prescott architecture, let's hide it."""
    result = threadpool_info()
    for dict_ in result:
        if dict_["internal_api"] == "openblas":
            dict_["architecture"] = "Prescott"
    return result


sklearn.utils.fixes.threadpool_info = hide_prescott
# END: monkey patch architecture


from sklearn.ensemble import RandomForestClassifier
from sklearn.utils.estimator_checks import check_estimator


for estimator, check in check_estimator(RandomForestClassifier(), generate_only=True):
    if "check_estimators_pickle" not in check.func.__name__:
        continue
    check(estimator)

But as you say, on the nightly build is passes. Sorry for the trouble.