David Widmann
David Widmann
It seems the adjoint definitions have to be updated to the ChainRules-syntax? In particular, I guess `nothing` should be replaced with the ChainRules-equivalents.
The `cholesky(Hermitian(...))` issue should probably be fixed by https://github.com/FluxML/Zygote.jl/pull/1114 on the master branch.
It seems the issue was reintroduced in https://github.com/hlissner/doom-emacs/commit/48658d2136c3ae08fcf6ea0e94804d2a2cc1455d
Seems reasonable to add definitions whenever it makes sense, and fall back to the generic definition otherwise. This came up also in https://github.com/TuringLang/MCMCChains.jl/pull/354.
In MCMCChains/MCMCDiagnosticTools, we use a bit special setup, due to our application and for performance reasons, so I'm not completely sure how generally useful it is. The implementation there, and...
As said, in the literature the use of `n` (and not `n - k`) is recommended for MCMC analysis. Surely not everyone agrees - but the nice thing is that...
Not sure what you want to say unfortunately. We implement exactly the version from BDA.
> I'm saying that the version from BDA uses `n-k` instead of `n`. Sure, but that's what's implemented in MCMCDiagnosticTools. I don't understand where the problem is.
Stan uses the FFT approach, so any other comparison will be flawed.
> The columns/rows of a matrix are easy to interchange or confuse, and a user not paying close attention to the documentation might mistakenly believe the function is broadcasted rowwise....