Track := variables in resulting chains

[torfjelde]

Currently, := variables are "tracked" as if they are just random variables. This can be annoying in certain cases where we make use of chains obtained from sample calls, e.g. when using generated_quantities, since variables in := expressions aren't treated differently.

Ideally, we should keep track of which variables are actually "deterministic" (i.e. occur in a := expression) vs. "random" (i.e. occur in a ~ expression).

Example:

julia> using DynamicPPL, Distributions

julia> @model function demo()
           x ~ Normal()
           y := x + 100
       end;

julia> model = demo();

julia> varinfo = VarInfo(model);

julia> values_as(varinfo, OrderedDict)
OrderedDict{VarName{:x, typeof(identity)}, Float64} with 1 entry:
  x => 0.66373

julia> values_as_in_model(model, varinfo)
OrderedDict{Any, Any} with 2 entries:
  x => 0.66373
  y => 100.664

Ideally, the values_as_in_model should also keep track of which are considered determinstic and random.

This should be fairly "trivial" to add, since it's mainly an issue of adding the some way to keep track of this information in the tilde calls, e.g. here

DynamicPPL.jl/src/values_as_in_model.jl

Lines 65 to 77 in 1d10278

	function tilde_assume(context::ValuesAsInModelContext, right, vn, vi)
	if is_tracked_value(right)
	value = right.value
	logp = zero(getlogp(vi))
	else
	value, logp, vi = tilde_assume(childcontext(context), right, vn, vi)
	end
	# Save the value.
	push!(context, vn, value)
	# Save the value.
	# Pass on.
	return value, logp, vi
	end

Miiight be a nice issue for you @penelopeysm to have a go at if you want? Should be a nice way to familiarize yourself with DynamicPPL.jl:)

Tag: @mhauru