Simulation `eval_coords` bug

[j-c-c]

When running the notebook in gallery/tutorials/cov3d_simulation.py using a LegacyVolume with C>2 volumes, an error is raised when performing Simulation.eval_coords:

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
Cell In[7], line 22
     17 # Assign the cluster centroids to the different images. Since we expect a discrete distribution of volumes
     18 # (and therefore of coordinates), we assign the centroid coordinate to each image that belongs to that cluster.
     19 # Evaluate the coordinates estimated
     21 clustered_coords_est = centers[vol_idx]
---> 22 coords_perf = sim.eval_coords(mean_est, eigs_est_trunc, clustered_coords_est)

File ~/Work/ASPIRE-python.simulation_default_volume_811/src/aspire/source/simulation.py:438, in Simulation.eval_coords(self, mean_vol, eig_vols, coords_est)
    435 # 0-indexed states vector
    436 states = self.states - 1
--> 438 coords_true = coords_true[states]
    439 res_norms = res_norms[states]
    440 res_inners = res_inners[:, states]

IndexError: index 2 is out of bounds for axis 0 with size 2

In this case, the states indices are being applied to the wrong axis of coords_true. This can be fixed by setting coords_true = coords_true[:, states], but this does not work for C=2 (likely due to the squeeze() used in `Simulation.vol_coords').

[j-c-c]

One additional observation that was made while tinkering with this notebook is that scipy.cluster.vq.kmeans2 returns the correct clusters, but sometimes with incorrect indices. For example, vol_idx returned by kmeans2 might be

vol_idx = [0, 0, 1, 2, 2]

while sim.states - 1 is

sim.states - 1 = [1, 1, 2, 0, 0]

[garrettwrong]

If you believe that the label of the clustering is not deterministic then we just need a small algo that can effectively check that a clustering matches another clustering. You can do this by adding corresponding values to a dict on first occurrence. When you encounter again, check the mapping holds. If you make to the end with no conflicts and both sets are same length, I think you're good.

[j-c-c]

This issue has been partially resolved by using sklearn.metrics.adjusted_rand_score to find the accuracy of the kmeans clustering (#873).

The tutorial will still only run for C=2.

[j-c-c]

Closing. Resolved by #1058.