David Widmann

Yes, this is the plan (and #95 already makes use of ChainRules).

I guess the problem is that the the samples in `vi` are saved as `Real`, and hence rerunning leads to the error since the evaluation of the logpdf in Bijectors...

https://github.com/TuringLang/Bijectors.jl/blob/1f3b581afe04f690bd93fba9edd88735cc1fc140/src/Bijectors.jl#L124 is actually not defined for `x >= 1 - eps` mathematically even though it works with Distributions (and incorrect for any `x > 0`). Maybe one should apply the...

More generally speaking, I'm wondering if for sampling and optimization in the unconstrained space we could use a `rand_trans` function that generates samples in the transformed space directly to avoid...

Copy from https://github.com/TuringLang/DistributionsAD.jl/issues/145#issuecomment-853794894: > The main issue was fixed upstream, only some interesting issue due to `Zygote._zero` with `ChainRulesCore.NotImplemented` for matrices of Skellam variates is left: https://github.com/TuringLang/DistributionsAD.jl/pull/172

No, Skellam is still broken with Zygote in one test.

Are you sure? I fixed these in https://github.com/TuringLang/Bijectors.jl/pull/182, I thought.

Ah, it seems you updated the tests before merging the master branch. I updated this issue here when I fixed the tests in https://github.com/TuringLang/Bijectors.jl/pull/182, so the remaining error is still...

Additionally we might want to: - Define the bijector on the Cholesky factor directly - Map it to a parameter vector instead of a matrix

I opened a PR over at Distributions (https://github.com/JuliaStats/Distributions.jl/pull/1144) to add support for single samples (and basically anything you could get from `rand`) to `loglikelihood`. A similar approach in Bijectors would...