[SohamXYZDev]

Reference issue (if any)

Fixes #14394.

What does this implement/fix?

Adds _soft_import("sklearn", "decoding") before top-level sklearn imports across the mne.decoding submodules (_fixes.py, base.py, ems.py, receptive_field.py, search_light.py, time_delaying_ridge.py, time_frequency.py, and transformer.py) so that accessing mne.decoding functionality without scikit-learn installed raises MNE's informative RuntimeError instead of a raw ModuleNotFoundError.

Also updates mne/tests/test_docstring_parameters.py to handle the RuntimeError from _soft_import when walking mne.decoding in minimal environments without scikit-learn, and adds test coverage in mne/utils/tests/test_check.py.

Additional information

Because ruff (I001) sorts third-party from sklearn... imports before relative from .base import ... imports, placing _soft_import only in base.py would still allow from sklearn... in submodules like transformer.py or search_light.py (as well as ems.py and _fixes.py, which do not import from base.py) to fail with ModuleNotFoundError before base.py is reached. Placing _soft_import("sklearn", "decoding") right before the sklearn imports in each submodule (with # noqa: E402, matching mne/export/_edf_bdf.py and _eeglab.py) ensures every entry point in mne.decoding.__all__ consistently raises the _soft_import error while keeping import mne and import mne.decoding lazy.

I used Gemini 3.8 Flash to help check import ordering across the decoding submodules and draft the unit test, and reviewed and verified all changes locally.

[larsoner]

@tsbinns can you look? If not, I can

[tsbinns]

@larsoner Yeah, can do

[tsbinns]

@SohamXYZDev Thanks for submitting this PR.

Because ruff (I001) sorts third-party from sklearn... imports before relative from .base import ... imports, placing _soft_import only in base.py would still allow from sklearn... in submodules like transformer.py or search_light.py (as well as ems.py and _fixes.py, which do not import from base.py) to fail with ModuleNotFoundError before base.py is reached.

Good point, that does complicate things, but having to noqa each sklearn import line is a bit unwieldy.

At the very least, we could nest the sklearn imports that are being moved to the end between # ruff: disable[E402] and # ruff: enable[E402], which would avoid the need to specify this per-line.

However, I wonder if a cleaner approach is to still try and centralise the check in base.py, with something like

from ..utils import _soft_import  # isort: skip
_soft_import("sklearn", name="mne.decoding")

at the top of base.py, and then in the submodules import .base # isort: skip # noqa: E402 at the top of each file (noqa if it's never used).

For the warnings this would cause, we could either nest the affected imports in # ruff: disable[E402] ... # ruff: enable[E402], or add E402 to the ruff per-file-ignores in pyproject.toml, e.g.:

"mne/decoding/*.py" = [
  "E402",  # Module level import not at top of file
]

@larsoner WDYT about that?

[tsbinns]

IIUC, the ImportError comes from ._fixes.py, but shouldn't that never be encountered since we are adding the _soft_import check at the top of that file?

[SohamXYZDev]

In the minimal / 3.12 CI build (where scikit-learn is not installed), pytest collects doctests using --doctest-modules --doctest-ignore-import-errors.

That pytest flag only ignores errors that inherit from ImportError. Because _soft_import raised a raw RuntimeError, pytest failed during test collection when importing _fixes.py. Making _SoftImportError inherit from both RuntimeError and ImportError preserves MNE's RuntimeError behavior while letting pytest's doctest collector skip it cleanly when scikit-learn is absent.

[larsoner]

Yeah I think centralizing is better, that way if we ever want to change it's easier

[SohamXYZDev]

@tsbinns @larsoner Updated! Rebased on latest main and refactored as suggested:

1. Centralized _soft_import("sklearn", "decoding") at the top of mne/decoding/base.py.
2. Submodules now import base first with # isort: skip.
3. Added "mne/decoding/*.py" = ["E402"] to per-file-ignores in pyproject.toml and stripped the per-line noqa comments.
4. Resolved circular dependencies by keeping MNETransformerMixin in base.py and localizing Vectorizer inside get_coef.

All 1,311 decoding tests pass locally, and the missing-sklearn import checks verified cleanly.

[tsbinns]

Ah, didn't realise this would introduce a circular import conflict. It's preferable if we don't have to move the mixin around.

Perhaps we can have the import check in a new _check.py file to avoid the circular imports. CC @larsoner

[SohamXYZDev]

sounds good to me, happy to move the check into a dedicated _check.py and restore the mixin once confirmed

[larsoner]

Yeah that works for me

[SohamXYZDev]

updated: moved the check into mne/decoding/_check.py and restored the mixin and transformer imports back to transformer.py. nesting checks and tests pass locally.