Allow concatenating BaseRaw objects

[cbrnr]

As reported in this forum post, mne.concatenate_raws() only works if all objects share the exact same type. In particular, this means that it is not possible to concatenate Raw (an alias for RawFIF) and RawArray objects, and other types like RawEDF, RawEEGLAB, etc. do not work either. However, since all concrete types inherit from BaseRaw, I think we could relax the compatibility check accordingly.

The following example fails on main, but works on this branch:

import numpy as np

import mne

sfreq = 100.0
ch_names = ["EEG 001", "EEG 002"]
ch_types = ["eeg"] * 2
info = mne.create_info(ch_names=ch_names, sfreq=sfreq, ch_types=ch_types)
data = np.random.randn(len(ch_names), 1000)

raw_array = mne.io.RawArray(data, info)
raw_array.save("temp_raw.fif", overwrite=True)
raw_fiff = mne.io.read_raw_fif("temp_raw.fif", preload=True)

mne.concatenate_raws([raw_fiff, raw_array])

[hoechenberger]

I didn't review any of the changes and don't know if there's any side effects to be expected; but purely from a user's POV, I'd say we should definitely try to support this

[sappelhoff]

I also think we should support this. Potential issues could arise from concatenating Raw types that support different attributes. For example, the BrainVision Raw supports impedances, but other Raws might not.

[larsoner]

At a bare minimum we would have to ensure all data are preloaded. And as @sappelhoff points out to keep things simple the resulting class should probably be RawArray since IIRC it doesn't have any extra attributes like impedance and such. (We could later maybe think about how to combine/keep them but I'd rather instead move toward removing them entirely.)

[cbrnr]

@larsoner I implemented your suggestions, I think this should be ready for review now.

[larsoner]

I'm a bit surprised you need this. It seems like the logic in raws[0].append(raws[1:], preload) can/should be smart enough to handle this case. Did you try to make this work?

In other words, I would expect something of the form:

if mixed_types:
    raws = list(raws)  # local copy of list
    raws[0] = RawArray(raws[0]._data, raws[0].info, first_samp=raws[0].first_samp)
    preload = True
raws[0].append(raws[1:], preload)

after the events code to be potentially much less memory intensive (by a factor approaching 2), and also fewer lines and easier to follow here

[cbrnr]

Hmm, you may be right. I pushed a fix, but I think I still have to preload raws[0] in order to use raws[0]._data, right? Also, I think we need to copy the annotations.

[larsoner]

Yeah you're right, you'd either need to preload that one or use .get_data. And yes probably annotations too

[cbrnr]

I mean, I'm starting to doubt if this is really worth it and if people should just do the conversion manually... WDYT?

[larsoner]

Well there are some slightly tricky bits that people can silently get wrong, like the first_samp plus annotations copying. So I think it's probably still worthwhile

[cbrnr]

OK, I think I implemented all of your suggestions.

[larsoner]

Already some failures

https://github.com/mne-tools/mne-python/actions/runs/24842637783/job/72721124939?pr=13263#step:17:4375

[larsoner]

LGTM +1 for merge once CIs come back green!