Open femtomc opened 1 year ago
Thanks for the kind words!!
I did think about including distributions like VMF and Bingham at one point and would be open to it, the main reason this never happened is that tangent-space Gaussians have been lower-effort and sufficient for my own use cases.
Two things I'm wondering:
jaxlie.SO3.sample_vmf()
? jaxlie.distributions.sample_vmf()
?@brentyi Re (1) -- that's a fair point.
Honestly, I don't know how large a dependency TFP is. I'd really prefer to not have to pull in all of TensorFlow for my own package. This is something I'll have to investigate.
For (2) -- the main functionality that I'm interested in is sampling and scoring. I would be wrapping distributions for use in a probabilistic programming context, so evaluating logpdf
is also valuable to me.
(Also, re -- TFP: I don't fully understand the "substrate" layer. My stuff is all in pure JAX anyways, so I've been reaching for distributions that are also in pure JAX -- I didn't really investigate integration with TFP, although the documentation seems to indicate a high quality for their distributions. If I could somehow figure out how to just depend on distributions + the JAX substrate, it would basically solve my motivating problem)
Just to be fully transparent -- I'm interested in functionality like this: https://github.com/probcomp/GenDirectionalStats.jl
But in JAX. If these distributions seem interesting and in scope, I can likely contribute a PR with my colleague.
On tensorflow_probability
: I don't know much about it either, but the required dependencies seem light. I also just tried and was able to install/use it without installing TensorFlow. (turns out they also emphasize this in an example)
If a contribution to jaxlie
still makes sense despite that and the you have time for a PR with a sensible API + types + tests, I'm definitely happy to review / merge!
Cool! -- I'll play a bit with tfp
, and see if it still makes sense. Thanks for the discussion nonetheless!
Hi -- first off, this package is incredible. Thank you.
Second, have you considered adding further random sampling functionality (for other distributions on spaces of rotations). In the title, I referred to the VMF distribution.
If you think this functionality should live elsewhere, totally fine! If you're curious and open to it, I may setup a PR and move some of the functionality from a package I'm working on to here.