FIX: Broken 4D resampling

[oesteban]

Summary

Resolves several issues in 4D transformations, happening both in "serialized" mode (iterating over 3D volumes) as well as "all-at-once" (interpolating on 4D coordinates). The PR was initiated with codex, generating a new test case that reproduced the problem (it's a small image with only one voxel set to 1, and the transform is a translation along the i-axis of 1 voxel per timepoint (i.e., the voxel moves from one extreme to the other along X in space).

Issues found:

• The serialized mode was not properly initiating the interpolation targets. Their calculation has been made more general and works for both operation modes (serialized, all-at-once).
• The all-at-once was generating nonsensical targets, reshapings were just fitting the shape of the targets array to what map_coordinates wants, but did not make any sense.

Both issues have been now resolved and the new test exercises both.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 94.81%. Comparing base (82f58c1) to head (4e159c2).
Report is 7 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #247      +/-   ##
==========================================
- Coverage   94.88%   94.81%   -0.08%     
==========================================
  Files          16       16              
  Lines        1859     1872      +13     
  Branches      243      242       -1     
==========================================
+ Hits         1764     1775      +11     
- Misses         77       80       +3     
+ Partials       18       17       -1     

Flag	Coverage Δ
unittests	?

Flags with carried forward coverage won't be shown. Click here to find out more.

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[oesteban]

@jmarabotto If you are still curious one year in, this is how this reshaping that took so much of your time works:

nitransforms/nitransforms/resampling.py

Lines 323 to 330 in f1efba1

	n_time, n_dim, n_vox = targets.shape
	# Reshape to (3, n_time x n_vox)
	ijk_targets = np.rollaxis(targets, 0, 2).reshape((n_dim, -1))
	time_row = np.repeat(np.arange(n_time), n_vox)[None, :]
	# Now targets is (4, n_vox x n_time), with indexes (t, i, j, k)
	# t is the slowest-changing axis, so we put it first
	targets = np.vstack((time_row, ijk_targets))