[Mne_analysis] tf_mixed_norm applied to epochs?

[Alexandre Gramfort]

> Would it be possible to separate the TF-MxNE inverse operator from the data to which it is applied? This would be similar to the L-1 MNE inverse, where "make_inverse_operator" is separate from "apply_inverse"? In which case the localization could be performed on average evoked data, and then the solution could be applied to epochs?

well we could do this, which would amount to run a least square on the activeset
(active sources + TF coefs) but the TF-MxNE solver like any non L2 solver does
not work this way. The solution is a non-linear operation computed on the
data at hand. How skilled are you with Python / MNE-Python to experiment
with this?

> P.S. tf_mixed_norm appears to discard the sign of the timecourses? I.e. they are always non-negative? Can the sign be retained?

we should add a pick_ori parameter to be able to keep the normal component only.
Otherwise you can keep the sign if you disable the loose parameter or
use a fixed
orientation constraint.

Alex