Proposal for MH only works correctly with subtypes of MvNormal

[torfjelde]

MH is implemented in a way where even if it receives a AdvancedMH.RandomWalkProposal it will still hit the "general" propose!! that assumes the proposal is a NamedTuple, i.e.

Turing.jl/src/mcmc/mh.jl

Lines 377 to 383 in 3a315ce

	function propose!!(
	rng::AbstractRNG,
	vi::AbstractVarInfo,
	model::Model,
	spl::Sampler{<:MH},
	proposal
	)

instead of the one that is hit if we're a subtype of MvNormal

Turing.jl/src/mcmc/mh.jl

Lines 410 to 417 in 3a315ce

	# Make a proposal if we DO have a covariance proposal matrix.
	function propose!!(
	rng::AbstractRNG,
	vi::AbstractVarInfo,
	model::Model,
	spl::Sampler{<:MH},
	proposal::AdvancedMH.RandomWalkProposal{issymmetric,<:MvNormal}
	) where {issymmetric}

For example, if I want to use a filldist(..., n) as a proposal instead of product_distribution(Fill(...), n), stuff breaks:

julia> using Turing, AdvancedMH, LinearAlgebra

julia> @model demo() = x ~ MvNormal(zeros(3), I)
demo (generic function with 2 methods)

julia> sample(
           demo(),
           MH(AdvancedMH.RandomWalkProposal(filldist(Normal(), 3))),
           10
       )
Sampling 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| Time: 0:00:00
ERROR: MethodError: no method matching _dist_tuple(::RandomWalkProposal{…}, ::DynamicPPL.TypedVarInfo{…}, ::@NamedTuple{…})

Closest candidates are:
  _dist_tuple(::NamedTuple{propnames}, ::DynamicPPL.VarInfo, ::NamedTuple{names}) where {names, propnames}
   @ Turing ~/.julia/packages/Turing/I5L1s/src/mcmc/mh.jl:335

Stacktrace:
  [1] dist_val_tuple(spl::DynamicPPL.Sampler{MH{(), RandomWalkProposal{…}}}, vi::DynamicPPL.TypedVarInfo{@NamedTuple{x::DynamicPPL.Metadata{…}}, Float64})
    @ Turing.Inference ~/.julia/packages/Turing/I5L1s/src/mcmc/mh.jl:317
  [2] propose!!(rng::Random.TaskLocalRNG, vi::DynamicPPL.TypedVarInfo{…}, model::DynamicPPL.Model{…}, spl::DynamicPPL.Sampler{…}, proposal::RandomWalkProposal{…})
    @ Turing.Inference ~/.julia/packages/Turing/I5L1s/src/mcmc/mh.jl:385
  [3] step(rng::Random.TaskLocalRNG, model::DynamicPPL.Model{…}, spl::DynamicPPL.Sampler{…}, vi::DynamicPPL.TypedVarInfo{…}; kwargs::@Kwargs{})
    @ Turing.Inference ~/.julia/packages/Turing/I5L1s/src/mcmc/mh.jl:466
  [4] step(rng::Random.TaskLocalRNG, model::DynamicPPL.Model{…}, spl::DynamicPPL.Sampler{…}, vi::DynamicPPL.TypedVarInfo{…})
    @ Turing.Inference ~/.julia/packages/Turing/I5L1s/src/mcmc/mh.jl:456

IIUC we should just relax

Turing.jl/src/mcmc/mh.jl

Lines 410 to 417 in 3a315ce

	# Make a proposal if we DO have a covariance proposal matrix.
	function propose!!(
	rng::AbstractRNG,
	vi::AbstractVarInfo,
	model::Model,
	spl::Sampler{<:MH},
	proposal::AdvancedMH.RandomWalkProposal{issymmetric,<:MvNormal}
	) where {issymmetric}

to be any AdvancedMH.RandomWalkProposal 🤷