ENH: Add a script to plot the signal estimated by the GP

[jhlegarreta]

• ENH: Add a script to plot the signal estimated by the GP
• STYLE: Rename the error analysis script to honor better its purpose
• STYLE: Rename the EddyMotionGPR instance to honor better its classname

[jhlegarreta]

pyhon /path/to/eddymotion/scripts/dwi_estimation_plot.py \ 
  60 1000 100 \
  --evals1 0.0015 0.0003 0.0003 \
  --snr 20

gives the below result (with the caveat mentioned in #228 (comment))

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 66.18%. Comparing base (796c501) to head (8835e63).
Report is 9 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #228   +/-   ##
=======================================
  Coverage   66.18%   66.18%           
=======================================
  Files          19       19           
  Lines         905      905           
  Branches      112      112           
=======================================
  Hits          599      599           
  Misses        264      264           
  Partials       42       42           

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

[jhlegarreta]

After 9358298 the script is able to simulate accept one, or two or three crossing fibers, e.g.

One fiber:

pyhon /path/to/eddymotion/scripts/dwi_estimation_plot.py \ 
  60 1000 100 \
  --angles 0 0 \
  --evals 0.0015 0.0003 0.0003 \
  --snr 20

Two fibers:

pyhon /path/to/eddymotion/scripts/dwi_estimation_plot.py \ 
  60 1000 100 \ 
  --angles 0 0 90 0 \
  --evals 0.0015 0.0003 0.0003 0.0015 0.0003 0.0003 \
  --snr 20

Three fibers:

pyhon /path/to/eddymotion/scripts/dwi_estimation_plot.py \ 
  60 1000 100 \
  --angles 0 0 90 0 90 90 \
  --evals 0.0015 0.0003 0.0003 0.0015 0.0003 0.0003 0.0015 0.0003 0.0003 \
  --snr 20

Plot for two fibers (made the colormap solid and changed the opacity value only):

Plot for three fibers:

[jhlegarreta]

A few notes:

• I have tested the error analysis and plot scripts and they work as intended (or at least as they were working prior to this commit).

  • The std dev is monotonically increasing only after the 10-fold CV in my experiments (happens in the main branch as well: @oesteban you may want to check this; if we conclude that there is something wrong, I'd rather try to fix that in a separate PR).
  

• Pending to create the predicted surface on more vectors than those used to generate the synthetic signal (xref ENH: Add a script to plot the signal estimated by the GP #228 (comment)). That requires reformatting the script, as we do not save the gp parameters or the fitted object in the error analysis script.

  • We may be able to pickle the EddyMotionGPR instance, e.g.
    https://stackoverflow.com/questions/56107259/how-to-save-a-trained-model-by-scikit-learn

    However, it is unclear to me at this point whether that fits in this script, as we are doing a CV and do not have access to the models. scikit-learn's cross-validate with return_estimator=True returns the estimators but at this point it looks like best would be a separate script where we train on all available data and predict on a set of independent directions and save the estimator and predictions. Thoughts are welcome.

• I would set the seed in dwi_gp_estimation_error_analysis.py to a fixed value for the sake of reproducibility. @oesteban do you prefer to keep the randomness through a flag to the script?

• We may want to remove the evals parameter from the dwi_gp_estimation_error_analysis.py script, and rely entirely on random values, as it is not feasible to provide N different values for each voxels, and even less so for multi-fiber cases (xref PR ENH: Allow simulating multi-fiber voxels #247) It will be easier to handle towards generalizing this to a synthetic dataset with voxels having multiple fibers.

• Checking the overlap between/merging eddymotion.testing.simulations.create_random_polar_coordinates and eddymotion.testing.simulations.create_random_polar_angles is left for a separate PR.