Symmetry Boosting

[j-c-c]

This PR adds symmetry boosting to MeanEstimator and WeightedVolumesEstimator.

Boosting is controlled by a flag in the *Estimator's init, which is True (on) by default. Then a symmetry_group is obtained from the provided ImageSource and passed through *Estimator.src_backward() to source.image.im_backward() and finally to image.image.backproject() where boosting of the images and rotations occurs.

Additionally, boosting takes place when computing the Fourier kernel used in mean estimation.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (18fd19c) 88.80% compared to head (3a7e103) 88.82%.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #1072      +/-   ##
===========================================
+ Coverage    88.80%   88.82%   +0.02%     
===========================================
  Files          126      126              
  Lines        12001    12024      +23     
===========================================
+ Hits         10657    10680      +23     
  Misses        1344     1344              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[j-c-c]

@garrettwrong this is ready for you to take a look.

Following up on dev meeting discussion: Implementing the boosting using a loop as opposed to duplicating images (and weights) came with a significant speed up (on laptop) for larger sets of images and high symmetry order. Here are some timings using image duplication with tiling (t_old) and using the loop (t_loop) for various n and symmetric order:

[garrettwrong]

Thanks, this will be nice to have. Can revisit 10081. Very minor changes.

[garrettwrong]

Why apply 1/sym_order here instead of in backproject?

[j-c-c]

No reason other than it made sense to me at the time. This is where we are scaling by number of images and boosting increases the number of images. It also pairs nicely with scaling on line 187.

[garrettwrong]

Thanks for the updates!

I'll try to re-attempt the 10081 gallery example with this branch soon (and the reduction in images that would be induced by boosting as discussed). Depending on what happens there, and with Joakim's review pass, I might ask you to update this PR or make another one with minor updates for the gallery.

[janden]

Looks great!

I feel like I've seen this symmetry group parsing code before. May make sense to have a parsing function (SymmetryGroup.from_string) to encode that logic (including the special case of None → IdentitySymmetryGroup).

[j-c-c]

I feel like I've seen this symmetry group parsing code before. May make sense to have a parsing function (SymmetryGroup.from_string) to encode that logic (including the special case of None → IdentitySymmetryGroup).

Sounds good. Added issue #1083. I'll swing back around and clean that up (in addition to some other symmetry_group issues #1081 I noticed during this PR)

Thanks!