[BUG] `cdist_soft_dtw_normalized` fails unexpectedly when time series panels have different number of instances about tslearn HOT 3 CLOSED

@YannCabanes could you have a look?

The current implementation of cdist_soft_dtw_normalized is wrong in the case where dataset2 is provided (in this case, the self-distances are not stored on the diagonal but have to be computed explicitly). This bug seems to have existed for a while (it was already present in v0.3 at least).

from tslearn.

OK, I gave it a try in #476, let's see what tests have to say...

from tslearn.

@rtavenar, thanks!

Quick question on tests - do you have systematic contract tests for distances?
Odd that it did not seem to cover such a central case (dataset2 is non-null)

What I mean is a test that automatically gets all distances, then substitutes common cases for dataset, dataset2 for it. We do that in sktime - using skbase (https://github.com/sktime/skbase) makes it relatively easy. That's how we caught it, just "run everything for every distance" and something broke.

Happy to help setting this up in tslearn if you'd like? I have not looked at your tests though - in case you're already doing this, please disregard my comment.

from tslearn.