The inverse solutions had some bad EEG channels removed, whereas the trialwise data didn't. When using mne.minimum_norm.apply_inverse the extra data channels are silently dropped. When using the new faster code introduced in #276, the first time the code was run, it would drop the channels and save the combined inverse solution. The second time the code was run, it would load the combined inverse solution and then attempt to matrix-multiply without first dropping the mismatched channels. So it worked in our tests using the first codepath but then didn't work in production when the second codepath was taken 😅
Now fixed.
Fixes #300
Addresses #302 but arguably does not fix it enough
Haha, that was DOOZY!
The inverse solutions had some bad EEG channels removed, whereas the trialwise data didn't. When using
mne.minimum_norm.apply_inversethe extra data channels are silently dropped. When using the new faster code introduced in #276, the first time the code was run, it would drop the channels and save the combined inverse solution. The second time the code was run, it would load the combined inverse solution and then attempt to matrix-multiply without first dropping the mismatched channels. So it worked in our tests using the first codepath but then didn't work in production when the second codepath was taken 😅Now fixed.
Fixes #300 Addresses #302 but arguably does not fix it enough