Vincent Roulet
Vincent Roulet
I think the most important point is that the method ``make_perturbed_fun`` does not require normalized log probabilities hence the advantage of highlighting that with the ``unnormalized_log_prob`` name. This can at...
Sorry I had not seen #1213. So before considering merging #1213, we should create a proper class or simply make these distributions some simple NamedTuples with sample, and unnormalized_log_prob as...
@carlosgmartin sorry for the delay. - So I'd prefer to remove the classes and use functions. - Having indeed one function that both samples and provides the unnormalized log prob...
I am surprised that the issue could be the installation of packages. To me what slows down the tests a lot is (i) pytype, (ii) doc building and doctests. However,...
> Wouldn't setting `bias_correction_v=False` not skip applying bias b2 entirely now? > > Is the point to not use the bias at all or change the order of the operations...
Thanks for pointing this out! Would you be willing to contribute and fix this?
Totally fine, we really appreciate the effort!
Yes, that would be great!
Thank you again @carlosgmartin Why not adding this algorithm to mpax? Optax is more oriented deep learning, you would have much less visibility here than in mpax I think.
Hello @GeophyAI, It depends on the optimizer. We cannot help without a concrete example of what you are trying to do.