Fix type inference of eltype(vi, spl) (Turing#2151) (#568)

* Fix type inference of `eltype(vi, spl)` (Turing#2151)

* Update Project.toml

* Add test

* Apply suggestions from code review

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* Fix test

---------

Co-authored-by: Tor Erlend Fjelde <[email protected]>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

Project.toml

@@ -1,7 +1,6 @@
 name = "DynamicPPL"
 uuid = "366bfd00-2699-11ea-058f-f148b4cae6d8"
-
-version = "0.24.4"
+version = "0.24.5"
 
 [deps]
 AbstractMCMC = "80f14c24-f653-4e6a-9b94-39d6b0f70001"

src/abstract_varinfo.jl

@@ -363,7 +363,14 @@ Determine the default `eltype` of the values returned by `vi[spl]`.
     This method is considered legacy, and is likely to be deprecated in the future.
 """
 function Base.eltype(vi::AbstractVarInfo, spl::Union{AbstractSampler,SampleFromPrior})
-    return eltype(Core.Compiler.return_type(getindex, Tuple{typeof(vi),typeof(spl)}))
+    T = Base.promote_op(getindex, typeof(vi), typeof(spl))
+    if T === Union{}
+        # In this case `getindex(vi, spl)` errors
+        # Let us throw a more descriptive error message
+        # Ref https://github.com/TuringLang/Turing.jl/issues/2151
+        return eltype(vi[spl])
+    end
+    return eltype(T)
 end
 
 # TODO: Should relax constraints on `vns` to be `AbstractVector{<:Any}` and just try to convert

src/varinfo.jl

@@ -1444,7 +1444,7 @@ function getindex(vi::TypedVarInfo, spl::Sampler)
     # Gets the ranges as a NamedTuple
     ranges = _getranges(vi, spl)
     # Calling getfield(ranges, f) gives all the indices in `vals` of the `vn`s with symbol `f` sampled by `spl` in `vi`
-    return vcat(_getindex(vi.metadata, ranges)...)
+    return reduce(vcat, _getindex(vi.metadata, ranges))
 end
 # Recursively builds a tuple of the `vals` of all the symbols
 @generated function _getindex(metadata, ranges::NamedTuple{names}) where {names}

test/turing/Project.toml

@@ -2,10 +2,12 @@
 DynamicPPL = "366bfd00-2699-11ea-058f-f148b4cae6d8"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
+ReverseDiff = "37e2e3b7-166d-5795-8a7a-e32c996b4267"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 Turing = "fce5fe82-541a-59a6-adf8-730c64b5f9a0"
 
 [compat]
 DynamicPPL = "0.24"
+ReverseDiff = "1.15"
 Turing = "0.30"
 julia = "1.7"

test/turing/runtests.jl

@@ -1,6 +1,7 @@
 using DynamicPPL
 using Turing
 using LinearAlgebra
+using ReverseDiff
 
 using Random
 using Test

test/turing/varinfo.jl

@@ -311,4 +311,35 @@
         @test vi.metadata.w.gids[1] == Set([hmc.selector])
         @test vi.metadata.u.gids[1] == Set([hmc.selector]) =#
     end
+
+    @testset "Turing#2151: eltype(vi, spl)" begin
+        # build data
+        t = 1:0.05:8
+        σ = 0.3
+        y = @. rand(sin(t) + Normal(0, σ))
+
+        @model function state_space(y, TT, ::Type{T}=Float64) where {T}
+            # Priors 
+            α ~ Normal(y[1], 0.001)
+            τ ~ Exponential(1)
+            η ~ filldist(Normal(0, 1), TT - 1)
+            σ ~ Exponential(1)
+
+            # create latent variable
+            x = Vector{T}(undef, TT)
+            x[1] = α
+            for t in 2:TT
+                x[t] = x[t - 1] + η[t - 1] * τ
+            end
+
+            # measurement model
+            y ~ MvNormal(x, σ^2 * I)
+
+            return x
+        end
+
+        n = 10
+        model = state_space(y, length(t))
+        @test size(sample(model, NUTS(; adtype=AutoReverseDiff(true)), n), 1) == n
+    end
 end