Xianda Sun

🎉

In the long run, we should just move some of these interface functions to `AbstractMCMC` (https://github.com/TuringLang/AbstractMCMC.jl/pull/144). Then we only need update `SliceSampling` up to the `AbstractMCMC` interface. For now, I...

A `varname` is concretized dynamically using the size information in the scope the concretization is carried out. One of the point is the `varname` will adapt to changing sizes. e.g....

regarding how we should convert `varname` to `string`, I think ideally, the conversion should be (1) easy to read (2) a bijection. (by bijection I mean: `varname(string(vn)) == vn` for...

> say that roundtrip deserialisation need only work for concretised varnames yeah, that would make sense to me. also `x[:]` with `ConcretizedSlice` and `x[1:10]` (the range doesn't matter here) should...

if all we want is serialization, then yeah, storing the full thing would be the right way to go. but do we want to use this "string-ify" also for the...

I think the tradeoff between the new dependency and speed of deserialize is worth it, so the new implementation is good with me. This is fantastic, Penny! > This way...

Just started a PR at https://github.com/JuliaObjects/Accessors.jl/pull/146. For equality comparisons between `ComposedOptic`, I kept the original definition. @jw3126 can you clarity why do we need `circ` (or `opcompose`) in this case?

Yes, perfectly.

An alternative approach I can envision is fully adopting `LogDensityFunction` in Turing through the `ExternalSampler` interface, but this might requires much more serious work encapsulating the `InferenceAlgorithm`s. (Do we have...