ENH: Add GP estimation notebook

[jhlegarreta]

Add GP estimation notebook.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 66.24%. Comparing base (795a9b7) to head (cbff151).
Report is 9 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #255   +/-   ##
=======================================
  Coverage   66.24%   66.24%           
=======================================
  Files          19       19           
  Lines         942      942           
  Branches      119      119           
=======================================
  Hits          624      624           
  Misses        274      274           
  Partials       44       44           

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

[jhlegarreta]

A few notes:

• In my CV I am only allowing two values; I've set it to 5 locally, but takes a while, so for now leaving 2 in the notebook.
• Andersson2015 use lambda [119, 1.15, 186] (p. 175) for [lambda, a, sigma^2]. The values I've used are those in https://github.com/nipreps/eddymotion/blob/main/docs/notebooks/dmri_covariance.ipynb. I have tried with lambda 100 without an improvement in results.

[jhlegarreta]

For the records, the issues I had when computing the GP and fixed by 03b3e54, they were due to these two lines arranging the data in a different way:

dwi_mask = np.repeat(brain_mask[..., np.newaxis], shell_data.shape[-1], axis=-1)
y = shell_data[dwi_mask]
y_train = y.reshape((X_train.shape[0], -1))

wrt to:

y = shell_data[brain_mask]
y_train = y.T

(in the example above I am pretending all data is used for the GP fit to remove any effect of the DWI train_test_mask)

Thanks Óscar, for having taken another, simpler approach in 03b3e54 that allowed to get around the issue.

Ready to be merged.