The Pathfinder algorithm looks like a promising way to find starting points for HMC (and maybe NUTS). I saw some talk on Twitter about implementing it, and have found this example implementation (GitHub - miclegr/pathfinder: JAX implementation of "Pathfinder: Parallel quasi-Newton variational inference"), but I can’t find anything r.e. whether this is an active feature being developed.
If not → how would I go about adding it?
I don’t think it’s being developed, but might be
a good addition to GitHub - pymc-devs/pymc-experimental
In terms of how to go about it, you can have a look at how different initialization strategies are implemented starting from here: pymc/sampling.py at 9a066e75dab62f92a94f30064ecd62b436c88186 · pymc-devs/pymc · GitHub
Did you mean pymc-experimental?
Yes, thanks! I fixed the link