FIX: CTF Projector comp

[bloyl]

Closes #5124

When computing ssp projection operators, include the meg_ref channels for data objects that include compensation channels (i.e CTF data), otherwise creating epochs raises an exception.

[bloyl]

A more general approach would be to modify pick_types to take a comp keyword that would include channels listed in info['comps']. That would obviously have wider ramifications.

[larsoner]

you can just do ref_meg = len(my_info['comps']) > 0; picks = pick_types(..., ref_meg=ref_meg).

Then we'll need a test that 1) proj can actually be calculated for CTF data (this should fail on master), and 2) that the proj do not contain ref channels.

[bloyl]

For the test do you know anything about the files in MNE-testing-data/CTF? Some of them (testdata_ctf.ds for example) have some eeg channels but I don't know if any of them EOG or ECG.

…

On Thu, Apr 12, 2018 at 10:40 AM Eric Larson ***@***.***> wrote: ***@***.**** commented on this pull request. ------------------------------ In mne/preprocessing/ssp.py <#5125 (comment)>: > @@ -183,8 +183,14 @@ def _compute_exg_proj(mode, raw, raw_event, tmin, tmax, _safe_del_key(flat, 'eog') # exclude bad channels from projection - picks = pick_types(my_info, meg=True, eeg=True, eog=True, ref_meg=False, - exclude='bads') + if len(my_info['comps']) > 0: + # Keep reference channels if compensation channels are present + picks = pick_types(my_info, meg=True, eeg=True, eog=True, ref_meg=True, + exclude='bads') + else: + picks = pick_types(my_info, meg=True, eeg=True, eog=True, ref_meg=False, + exclude='bads') you can just do ref_meg = len(my_info['comps']) > 0; picks = pick_types(..., ref_meg=ref_meg). Then we'll need a test that 1) proj can actually be calculated for CTF data (this should fail on master), and 2) that the proj do not contain ref channels. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#5125 (review)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AFU7nfmulhPm5Bh542m78_Ak0HUxJaPZks5tn2dcgaJpZM4TRxn2> .

[larsoner]

I don't know if there are ECG or EOG, but you can use any MEG/EEG channel as a surrogate if necessary (should be an option in the proj computation function). It doesn't matter if the projectors are correct here, just that they can be computed.

[codecov]

Codecov Report

Merging #5125 into master will increase coverage by 0.08%.
The diff coverage is 93.33%.

@@            Coverage Diff             @@
##           master    #5125      +/-   ##
==========================================
+ Coverage   88.06%   88.15%   +0.08%     
==========================================
  Files         356      357       +1     
  Lines       65303    65805     +502     
  Branches    11128    11198      +70     
==========================================
+ Hits        57511    58008     +497     
+ Misses       4963     4959       -4     
- Partials     2829     2838       +9

[bloyl]

In writing the test I've bumped into something I think is strange.

on master I only get this exception when I use average=True. This means that calls like

epochs = Epochs(raw, events, None, tmin, tmax, baseline=None, preload=True,
                             picks=picks, reject=reject, flat=flat, proj=True)

Don't actually call pick_channels or _pick_drop_channels until the actual data array is read, even if preload=True. Is this the expected behaviour?

[larsoner]

That sounds like a bug. Perhaps the comps check needs to be replicated in this other code path, too. It's likely that there are some paths like Epochs(..., picks=picks) that don't actually use the pick_channels / _pick_drop_channels path for efficiency reasons.

[bloyl]

What is your advice for this PR?
I say I add the test and then we can merge this and I'll open another issue for the epochs stuff.
Hows that sounds?

[larsoner]

Sure, fixing one bug at a time can simplify things a bit

[larsoner]

don't forget the @testing.requires_testing_data decorator here

[larsoner]

With pytest it's better to do now assert len(projs) == ... directly, it gives nicer error messages than assert_true

[larsoner]

For consistency "Test to show that ..."

[larsoner]

Can you check that none of the compensation channels are in the projs?

[bloyl]

I added a check, although I think the best way would be would be something like

for p in projs:
    if _bad_chans_comp(raw.info, pick_ch_names)[0]:
         raise RuntimeError('something')

with _bad_chans_comp from #5133

Once this is merged I'll add an additional check to that PR.

[larsoner]

Thanks @bloyl