Align2D refactor/updates

[garrettwrong]

• Refactors Align2D towards subclassing. One branch going off the base class and returning some latent images (say for EM methods) and others which will do basic image stacking via subclassing AveragedAlign2D.
• Adds an adoption of ReddyChatterji (log polar fourier cross corr method for image registration), primarily based from scikit image examples. After estimated rotation registration attemps second round of cross corr to estimate translation.
• Adds BFSReddyChatterji. Brute Force Shifts, performing RC based cross corr technique to rotation estimation after shifting. Returns best shift composed with rotation over all shifts in a radius.

[codecov]

Codecov Report

Merging #538 (d51b0ec) into develop (5e2d49a) will decrease coverage by 0.33%.
The diff coverage is 74.37%.

@@             Coverage Diff             @@
##           develop     #538      +/-   ##
===========================================
- Coverage    87.20%   86.86%   -0.34%     
===========================================
  Files          108      108              
  Lines         8000     8586     +586     
===========================================
+ Hits          6976     7458     +482     
- Misses        1024     1128     +104     

Impacted Files	Coverage Δ
src/aspire/denoising/denoiser_cov2d.py	80.00% <66.66%> (-0.77%)	⬇️
src/aspire/classification/averager2d.py	72.82% <72.82%> (ø)
src/aspire/basis/basis.py	87.83% <80.00%> (-0.74%)	⬇️
src/aspire/classification/rir_class2d.py	95.48% <92.85%> (+1.84%)	⬆️
src/aspire/classification/__init__.py	100.00% <100.00%> (ø)
src/aspire/classification/class2d.py	94.73% <100.00%> (-5.27%)	⬇️
src/aspire/covariance/covar.py	88.79% <100.00%> (ø)
src/aspire/covariance/covar2d.py	94.21% <100.00%> (ø)
src/aspire/denoising/__init__.py	100.00% <100.00%> (ø)
src/aspire/denoising/denoised_src.py	95.83% <100.00%> (+0.18%)	⬆️
... and 6 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 5e2d49a...d51b0ec. Read the comment docs.

[garrettwrong]

Todays commits meld the simulated experimental pipeline and associated minor patches into this PR. It alters the sphinx gallery config to only run the tutorials, but parse the experiments. I also extend the RIR and aligner to admit CWF denoised source for just the classification. Switching do_cov2d in that pipeline example will now create that additional source while using the original source for alignment and averaging. This also fixes a cov3d bug I introduced in the grid updates and works around couple dtype bugs.

I'll try to go through it on another day to look for any typos I missed so far.

[garrettwrong]

Made some style and name changes to the example. Adds an experimental data pipeline and it even runs. Ran (on my laptop) for a small pipeline similar in size to the simulation example... but the reconstruction left a lot to be desired. I'll run it again in a more serious size while the code gets a once over.

[janden]

First pass. Overall, I think this looks good. Big question here is about the class hierarchy.

I haven't gone into implementation details and tests yet. We'll get to that later.

[janden]

Looks great! Just some small comments, then we should be good to go.

[janden]

Should be able to compute an RMSE with the ground truth volume by rotating estimated_volume by Q_mat (Q_mat.T?) and comparing them. Or you might not. It's worth a shot. Otherwise, just leave it.

[garrettwrong]

I can try it. I'll come back to this after the other items. I didn't have any idea that Q_mat would realize the right rotations for that... I'm not sure the project is there yet (good rmse), lots of things to check, but I was reasonably happy to get things like the attached at this stage of the project. First is the original hi res map from EMD 10028, and second is the simulated abinitio model, both as viewed in chimera.

[garrettwrong]

FWIW that sim was gallery/experiments/simulated_abinitio_pipeline.py, if you'd like to run it too. Just requires downloading the map file and putting in the directory.

[janden]

That doesn't look too bad, actually…

So the idea is that Q_mat tells us how the estimated rotations differ from the ground truth rotations (in other words, what rotation needs to be applied to transform one rotation sequence to the other). Thing that might be tricky is that it may include a possible reflection as well. It's definitely a “nice to have” and not a requirement. If it takes more than fifteen minutes, make an issue and we'll come back to it.

[garrettwrong]

Yes, I was quite pleased with it for the maturity of the project and sort of rushing this last feature given the other situations. Lot of code refinement todo, but I think this should get the project down the road quite a ways.

I'll load the saved volume and give it a try.

[garrettwrong]

Just waiting for the local docs and tests to run. I will push it for another round. While that is happening I will i) kick off similated_abinitio_pipeline.py to check it still works, ii) try this Q_mat thing.

[garrettwrong]

Looks like some of the other changes that went in recently have changed the noise level we can preprocess away, so I'll need to update that. (One param line change in simulated experiment, I just need to try a few to see one that is good example).

[garrettwrong]

Simulated abinitio reconstruction looks similar now with the noise var adjustment. Not quite as good as before, but the whitening was working too well before :). Merging.