David Widmann
David Widmann
It would be interesting to understand why this is happening, it is definitely not intended.
I noticed that sampling from these models works fine (without the KeyError) in the Julia REPL ([code](https://gist.github.com/devmotion/1fde42d174c6f53c9f7f749aaf4d6706)).
Note that you have to be careful to not share the tape though: https://github.com/TuringLang/Turing.jl/issues/1412#issuecomment-698160681
IMO it would be great to make the setup more modular and less "magic". I'm not sure though if one should be able to specify the tape as part of...
Can you check if it still fails with the latest release (0.15.12)? I assume it does but it's good to check it first. In general, I assume the problem is...
> but the second example also uses a vector of missing and it does work: The type of `x` is not relevant here - it only matters what values are...
> So maybe it's a problem with Distributions.jl (or Turing should call disallowmissing before loglikelihood)? Maybe one could relax the type constraints in Distributions (AFAIK currently they require a `AbstractMatrix{
> I think the correct solution here is just to pass in a single missing rather than a vector. Yes, this is what one can do, as pointed out in...
I use pinned environments for all of my literate scripts, and therefore I would be very interested if one could use specific environments. Maybe one could use the same behaviour...
http://www.cs.toronto.edu/pub/reports/na/adjoint_paperNSERC.pdf