David Widmann

Can we add at least tests for every new function and type and fix the CI errors? And I think it would be nice to see as well what actually...

Oh it seems maybe @torfjelde has already fixed the problems in https://github.com/TuringLang/DynamicPPL.jl/pull/360/commits/0f9765bda684b27202982cf95d11e8de07304f62?

I was just looking at https://github.com/TuringLang/Turing.jl/blob/master/src/stdlib/distributions.jl for completely unrelated reasons, and discovered [some definitions of `Bijectors.logpdf_with_trans(::NoDist, x, t)`](https://github.com/TuringLang/Turing.jl/blob/9f482f3c65c52485c16f511b56f622198aee3d2d/src/stdlib/distributions.jl#L248-L253) :open_mouth: Regardless of whether they are useful etc., this seems like one...

> As of right now, this would require re-instantiating the model between every call to AbstractMCMC.step. This is only possible if observations are provided as arguments to the model which...

By the way, at this stage I think this should be addressed in Turing rather than in DynamicPPL by defining a custom context for SGHMC and SGLD, similar to the...

> One issue with both approaches: neither handles the case where the full data set is to large to have in memory, and ideally you'd lazily load the batch. An...

> Surely this could fall under a more general "filter VarName" context which simply as a filtering function specified by the user? This filter could be random, deterministic, etc. In...

> Hmm, I guess it might be better to put it there. You mean "for now" while we are in a bit of an experimentation phase? Yes, I think we...

We need some structure or algorithm and can't leave this only up to the user. In contrast to when you implement the log probability function yourself, the Turing `@model` does...

There's some explanation here: https://turing.ml/dev/docs/for-developers/compiler The main thing is just that one can provide a context when executing a model, and this context is forwarded to all `observe` and `assume`...