Dustin Tran

Good question! The only reason is because it hasn't been implemented yet. :) For the current work, we were only considering dense matrices for various simulation studies; of course, in...

One that came up today: `subvals`. Reading through it, I guess it stands for "substitute values"? In general, some of the shorthand can be difficult to trace (no pun intended).

Another is `primitive`, which wraps `f_raw` followed by `wraps` which wraps `f_wrapped` and is parameterized by `f_raw`. It's not clear why there's nested decorators until you read and grok all...

Unfortunately no. :( I can see if it's possible to get it to converge, and maybe that would indirectly point us to where the problem is.

Is this issue solved yet? I think so but I don't want to close it without confirmation.

Oh, you're right. I mistook my above issue with something else about streams. Yes we still have to do a hack by inputting 0 for the log prob's.

Sure, feel free. Happy to provide feedback if you want to submit a PR while still unfinished.

Yep, I think that makes sense. The simplest version to get started would follow Metropolis-Hastings' `build_update()` but without the Metropolis correction. This would implement vanilla importance sampling.

Hey @JhyOnya ! To let you know, this repo is no longer maintained. We still actively respond on the other relevant research codebases like https://github.com/google/edward2, https://github.com/tensorflow/probability, and https://github.com/google/uncertainty-baselines.

@mattjj, @matthewdhoffman SVAEs require a symbolic algebra system that properly accounts for integral identities (conjugacy relationships). We're working on a separate package doing this with Autograd/NumPy; but note it likely...