Move towards `np.random.default_rng`

[sappelhoff]

We are still using numpy's old random API in many places. It would be good to migrate to the recommended np.random.default_rng.

This might turn out to be difficult, given our many tests that need to remain compatible with matlab, etc.

xrefs:

• https://github.com/a-hurst/matprep_artifacts/tree/master
• API: RandomState deprecation, default_rng, and SPEC7 mne-tools/mne-python#9233

cc @a-hurst

[viranovskaya]

Hi! I’d be interested in looking into this migration.

My current work involves MNE-Python EEG preprocessing and reproducible, test-based workflows. I would start by identifying the remaining legacy random calls and migrate them incrementally, while preserving deterministic behavior and MATLAB compatibility.

Given the reproducibility constraints, I’d keep the first PR narrowly scoped. If this approach makes sense and the issue is still available, I’d be happy to prepare it.

[sappelhoff]

@larsoner what are the plans of moving towards default_rng in mne?

[larsoner]

Nothing beyond what's in mne-tools/mne-python#9233 ... so far I've been trying to use it in new code at least. We should probably make a PR to switch as many usages as we can at some point.

[sappelhoff]

Ol thanks for the update!

@viranovskaya thanks for reaching out, but we will not switch in pyprep before mne switches :)

[viranovskaya]

Thanks for clarifying — I’ll keep an eye on the MNE issue!

[sappelhoff]

• note, there has been some work in Move toward determinism mne-tools/mne-python#14090