WIP fix: fix bug in thresholder

[jolars]

Something is not working correctly with the thresholder function. It doesn't converge for some examples, such as this one.

cd_primal_old is the old thresholder function, which works correctly at least for this example.

These are some change to help weed out what the problem is.

[jolars]

I see this:

Epoch: 12, loss: 0.7776163967877969, gap: 7.14e-03
between clusters
between clusters
between clusters, outside loop
between clusters
between clusters, outside loop
j:  4
cluster_indices:  [0 1 2 3 9 8 7 6 5 4]
c:  [0.90675302 0.47672382 0.0469308  0.0207867  0.00389932]
c[j]:  0.003899315241586314
x: 0.03978030871681798
beta_tilde:  0.03978030871681798 , beta_tilde_old: 0.0038103252608220756
ind_old:  4
ind_new:  4 , ind_new_old:  4

Looks like the index of the new cluster returned from the thresholder is wrong. beta_tilde also disagrees between the new and old version.

[Klopfe]

I see this:

Epoch: 12, loss: 0.7776163967877969, gap: 7.14e-03
between clusters
between clusters
between clusters, outside loop
between clusters
between clusters, outside loop
j:  4
cluster_indices:  [0 1 2 3 9 8 7 6 5 4]
c:  [0.90675302 0.47672382 0.0469308  0.0207867  0.00389932]
c[j]:  0.003899315241586314
x: 0.03978030871681798
beta_tilde:  0.03978030871681798 , beta_tilde_old: 0.0038103252608220756
ind_old:  4
ind_new:  4 , ind_new_old:  4

Looks like the index of the new cluster returned from the thresholder is wrong. beta_tilde also disagrees between the new and old version.

I'll look into it. Thanks for spotting it.

[jolars]

Alright, I think I have this sorted out now.

This had me stumped for so long that I finally rewrote the function from scratch. It is a little more verbose than our previous version but should be as fast (or even a bit faster since we avoid forming a vector of indexes which we did previously and also because we avoid checking the zero cluster each time, i.e. summing over all lambdas for the zero cluster, which I realize can actually be expensive if the zero cluster is large).

Let me know if you think this looks fine and I'll clean it up and replace our version.

[Klopfe]

Looks great to me! Let's keep the newest version and clean up the files for linter tests. Do you want me to do that @jolars? Thanks a lot @jolars!