MRG: Unify _prepare_gain

[larsoner]

Okay @agramfort I had to change a tolerance for one gamma_map test. All other tests should be passing except mne/inverse_sparse/tests/test_mxne_inverse.py:test_mxne_inverse_standard:

>       assert stc_prox.vertices[1][0] in label.vertices
E       IndexError: index 0 is out of bounds for axis 0 with size 0

Then the loose_method='sum' should change to loose_method='svd', but this will break the sphere test. (The orientation only matches to 0.89, not 0.99, and the dipole fit is not 1.7mm away but 30mm :( ).

I'm out of my depth working on the MxNE stuff, so hopefully you find some hacking time soon.

Todo:

• add limit_depth_chs='whiten' to MxNE and MNE
• fix test_mxne_inverse_standard
• set loose_method='svd' and fix the failing sphere model test
• add bias tests for LCMV

Next PR:

• rework compute_depth_prior to operate on forward

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (master@73ea718). Click here to learn what that means.
The diff coverage is 93.85%.

@@            Coverage Diff            @@
##             master    #5984   +/-   ##
=========================================
  Coverage          ?   88.83%           
=========================================
  Files             ?      401           
  Lines             ?    72880           
  Branches          ?    12180           
=========================================
  Hits              ?    64742           
  Misses            ?     5218           
  Partials          ?     2920

[larsoner]

@agramfort okay with you for a backward compat break to refactor compute_depth_prior to take forward instead of gain, gain_info, is_fixed_orient, patch_areas? Given that this function wasn't in the doc and buried in mne.forward.forward.compute_depth_prior I think it's probably fine, esp. if we note it well in the doc. Or should we do a proper deprecation cycle? We could allow gain to be an instance of forward, and have the other options be None and removed in 0.19. Let me know which you prefer.

If you want this refactoring/deprecation to appear in this PR, let me know. Otherwise this can be merged and the compute_depth_prior cleanups can be the next PR.

[larsoner]

FYI this shows the performance improvement for MNE when using whiten-based depth weighting (at least in this case, with this measure for localization bias)

[agramfort]

@agramfort <https://github.com/agramfort> okay with you for a backward compat break to refactor compute_depth_prior to take forward instead of gain, gain_info, is_fixed_orient, patch_areas? Given that this function wasn't in the doc and buried in mne.forward.forward.compute_depth_prior I think it's probably fine, esp. if we note it well in the doc. Or should we do a proper deprecation cycle? We could allow gain to be an instance of forward, and have the other options be None and removed in 0.19. Let me know which you prefer. If you want this refactoring/deprecation to appear in this PR, let me know. Otherwise this can be merged and the compute_depth_prior cleanups can be the next PR.

if this is still relevant I let you judge. Maybe in a next PR if this is pretty much ready.