Pierre Ablin
Pierre Ablin
Hi, Unfortunately it is not available as such. In picardo (i.e. picard with ortho=True), the extended version is automatically used to mimic the behavior of fastica. If I have some...
Picard minimizes the same cost function as the aforementioned algorithms, but due to non-convexity of the problem, it might very well find another solution (it is a general problem with...
Hello @adrianjav ! Do you have a MWE to reproduce this behavior? There is definitely a bug somewhere because with this safe step size the distance to the manifold should...
Declaring k rotations at a time should not be a problem, computations are batched. In the code snippet above, `self.rotation` is not orthogonal; you should project it to make it...
It seems like a bug, but it's hard without running some code to know where it comes from. If that is not too much trouble, could you save the last...
Hey @adrianjav , sorry for the lack of responsiveness... I now have a new PR #4 that should hopefully fix this problem!
We should rather use https://jax.readthedocs.io/en/latest/_autosummary/jax.flatten_util.ravel_pytree.html to flatten every pytree instead of rewriting all algorithms.
Note: this is a WIP!
Actually seems like https://github.com/ott-jax/ott/commit/db9f58a13755901e877ececa6d2388dd9d62d6b1 makes this useless, right? If yes, we can close.