nsiccha
nsiccha
@yebai: The following "works" (but actually, the likelihood term is not computed correctly). The original failure seems to be due to [Distributions.jl (correctly I must assume) calling `besselix`](https://github.com/JuliaStats/Distributions.jl/blob/abb151c42eb67b4994e21da15e24c65f1cc3dd9f/src/univariate/continuous/vonmises.jl#L29), which I...
@yakir12: the original posterior (with `x = rand(VonMises(0.3, 3), 100)` and the `VonMises` likelihood) samples without issues via Stan. Do note though that in general these kinds of models can...
> the following lines override progress if the user didn't provide a callback (no clue why this is the case): I'd assume that's because the callback my prematurely terminate sampling...
They (including @bob-carpenter) write (emphasis mine): > Remark 1. In the optimized implementation of WALNUTS, the energy error used to assess the validity of a micro trajectory is computed as...
Hm, I guess that _would_ make sense! Any proposals for a name? 😅
Well, I do currently have this implementation below lying around, basically doing what StanLogDensityProblems.jl currently supports but also supports printing the error msg to stdout. I guess one could also...
I did use StanLogDensityProblems.jl as my template, which is why the annotation for the order is slightly weird (like in [the original code](https://github.com/sethaxen/StanLogDensityProblems.jl/blob/28b69a205c793392e72c49e7d24af88215f0885e/src/stanproblem.jl#L39)) 😅 .
I guess this can be easily modified to work as you proposed. Will do so at some uncertain point in the future! Probably when I'm starting a new project such...
> I'd actually drop the nan_on_error field here I was gonna say that the handling of errors in the wrapped LogDensityProblems (as here) would stand in the way of LogDensityProblemsCallbacks...
> I agree it's not ideal we suppress the error message entirely I'm often quite happy with the default! I've just recently added that `silent_nan` handler above, because all the...