Model call: no method matching `NoDist`

[phipsgabler]

On DPPL 0.8.0:

julia> @model function test5(x) 
           μ ~ MvNormal(fill(0, 2), 2.0)
           z = Vector{Int}(undef, length(x))
           z .~ Categorical.(fill([0.5, 0.5], length(x)))
           for i in 1:length(x)
               x[i] ~ Normal(μ[z[i]], 0.1)
           end
       end
ModelGen{var"###generator#506",(:x,),(),Tuple{}}(##generator#506, NamedTuple())

julia> test5([1, 1, -1])(VarInfo(), SampleFromPrior(), LikelihoodContext())
ERROR: MethodError: no method matching NoDist(::Array{DiscreteNonParametric{Int64,Float64,Base.OneTo{Int64},Array{Float64,1}},1})
Closest candidates are:
  NoDist(::NamedDist) at /home/philipp/.julia/packages/DynamicPPL/QgcLg/src/distribution_wrappers.jl:31
  NoDist(::Td) where {variate, support, Td<:Distribution{variate,support}} at /home/philipp/.julia/packages/DynamicPPL/QgcLg/src/distribution_wrappers.jl:29
Stacktrace:
 [1] dot_tilde(::Random._GLOBAL_RNG, ::LikelihoodContext{Nothing}, ::SampleFromPrior, ::Array{DiscreteNonParametric{Int64,Float64,Base.OneTo{Int64},Array{Float64,1}},1}, ::Array{Int64,1}, ::VarName{:z,Tuple{}}, ::Tuple{}, ::VarInfo{DynamicPPL.Metadata{Dict{VarName,Int64},Array{Distribution,1},Array{VarName,1},Array{Real,1},Array{Set{DynamicPPL.Selector},1}},Float64}) at /home/philipp/.julia/packages/DynamicPPL/QgcLg/src/context_implementations.jl:180
 [2] dot_tilde_assume(::Random._GLOBAL_RNG, ::LikelihoodContext{Nothing}, ::SampleFromPrior, ::Array{DiscreteNonParametric{Int64,Float64,Base.OneTo{Int64},Array{Float64,1}},1}, ::Array{Int64,1}, ::VarName{:z,Tuple{}}, ::Tuple{}, ::VarInfo{DynamicPPL.Metadata{Dict{VarName,Int64},Array{Distribution,1},Array{VarName,1},Array{Real,1},Array{Set{DynamicPPL.Selector},1}},Float64}) at /home/philipp/.julia/packages/DynamicPPL/QgcLg/src/context_implementations.jl:215
 [3] macro expansion at ./REPL[88]:4 [inlined]
 [4] ##evaluator#505(::Random._GLOBAL_RNG, ::Model{var"###evaluator#505",(:x,),Tuple{Array{Int64,1}},(),ModelGen{var"###generator#506",(:x,),(),Tuple{}}}, ::VarInfo{DynamicPPL.Metadata{Dict{VarName,Int64},Array{Distribution,1},Array{VarName,1},Array{Real,1},Array{Set{DynamicPPL.Selector},1}},Float64}, ::SampleFromPrior, ::LikelihoodContext{Nothing}) at /home/philipp/.julia/packages/DynamicPPL/QgcLg/src/compiler.jl:356
 [5] evaluate_threadunsafe(::Random._GLOBAL_RNG, ::Model{var"###evaluator#505",(:x,),Tuple{Array{Int64,1}},(),ModelGen{var"###generator#506",(:x,),(),Tuple{}}}, ::VarInfo{DynamicPPL.Metadata{Dict{VarName,Int64},Array{Distribution,1},Array{VarName,1},Array{Real,1},Array{Set{DynamicPPL.Selector},1}},Float64}, ::SampleFromPrior, ::LikelihoodContext{Nothing}) at /home/philipp/.julia/packages/DynamicPPL/QgcLg/src/model.jl:157
 [6] (::Model{var"###evaluator#505",(:x,),Tuple{Array{Int64,1}},(),ModelGen{var"###generator#506",(:x,),(),Tuple{}}})(::Random._GLOBAL_RNG, ::VarInfo{DynamicPPL.Metadata{Dict{VarName,Int64},Array{Distribution,1},Array{VarName,1},Array{Real,1},Array{Set{DynamicPPL.Selector},1}},Float64}, ::SampleFromPrior, ::LikelihoodContext{Nothing}) at /home/philipp/.julia/packages/DynamicPPL/QgcLg/src/model.jl:136
 [7] (::Model{var"###evaluator#505",(:x,),Tuple{Array{Int64,1}},(),ModelGen{var"###generator#506",(:x,),(),Tuple{}}})(::VarInfo{DynamicPPL.Metadata{Dict{VarName,Int64},Array{Distribution,1},Array{VarName,1},Array{Real,1},Array{Set{DynamicPPL.Selector},1}},Float64}, ::SampleFromPrior, ::Vararg{Any,N} where N) at /home/philipp/.julia/packages/DynamicPPL/QgcLg/src/model.jl:126
 [8] top-level scope at REPL[90]:1

[devmotion]

I guess it has never worked since it seems we have never supported constructing a NoDist with an array of distributions and the affected line has not been touched for quite some time. Probably we should just replace NoDist(dist) in

DynamicPPL.jl/src/context_implementations.jl

Line 180 in e85f821

return _dot_tilde(rng, sampler, NoDist(dist), left, vns, vi)

with NoDist.(dist). A bit annoying that we have to allocate a new array there, but I don't think it can be avoided. We shouldn't propagate a distribution instead of an array of distributions there just for LikelihoodContext, I guess, since that might dispatch to a different downstream implementation, and also the implementation of NoDist would become very messy otherwise.

[devmotion]

Fixed by #136.