Compiler refactor 2.0 (#613)

* compiler refactor proof of concept

* rename compiler to model_info

* improve compiler readibility and make it more compact

* invokelatest -> calling model

* respond to Martin's comments

* fix ambiguity

* fix missing data case and make more compact

* fix broken compiler.jl/explicit_ret.jl test

* fix sampling from prior

* make compiler more compact

* fix update_vars!

* fix tests

* fix mh.jl tests

* fix Sampler API

* uncomment commented tests

* fix seed in test/mh.jl/mh_cons.jl

* make pvars and dvars type parameters in CallableModel

* make data a field in CallableModel

* fix #544 - allows default value of data vars when treated as params

* add support for passing data as kwargs to outer function

* model::Function -> model

* CallableModel -> Model

* compiler hygiene and organization

* docstrings and some renaming

* respond to Will's comments

* some more style issues

* add test for #544

* comment

* model -> model::Model

src/Turing.jl

@@ -112,6 +112,7 @@ mutable struct Sampler{T<:InferenceAlgorithm} <: AbstractSampler
   alg   ::  T
   info  ::  Dict{Symbol, Any}         # sampler infomation
 end
+Sampler(alg, model) = Sampler(alg)
 
 # mutable struct HMCState{T<:Real}
 #     epsilon  :: T

src/core/ad.jl

@@ -30,7 +30,7 @@ getADtype(::Type{<:Hamiltonian{AD}}) where {AD} = AD
 gradient(
     θ::AbstractVector{<:Real},
     vi::VarInfo,
-    model::Function,
+    model::Model,
     sampler::Union{Nothing, Sampler}=nothing,
 )
 
@@ -40,7 +40,7 @@ Computes the gradient of the log joint of `θ` for the model specified by
 @generated function gradient(
     θ::AbstractVector{<:Real},
     vi::VarInfo,
-    model::Function,
+    model::Model,
     sampler::TS=nothing,
 ) where {TS <: Union{Nothing, Sampler}}
     if TS == Nothing
@@ -68,7 +68,7 @@ end
 gradient_forward(
     θ::AbstractVector{<:Real},
     vi::VarInfo,
-    model::Function,
+    model::Model,
     spl::Union{Nothing, Sampler}=nothing,
     chunk_size::Int=CHUNKSIZE[],
 )
@@ -79,7 +79,7 @@ using forwards-mode AD from ForwardDiff.jl.
 function gradient_forward(
     θ::AbstractVector{<:Real},
     vi::VarInfo,
-    model::Function,
+    model::Model,
     sampler::Union{Nothing, Sampler}=nothing,
     ::Val{chunk_size}=Val(CHUNKSIZE[]),
 ) where chunk_size
@@ -111,7 +111,7 @@ end
 gradient_reverse(
     θ::AbstractVector{<:Real},
     vi::VarInfo,
-    model::Function,
+    model::Model,
     sampler::Union{Nothing, Sampler}=nothing,
 )
 
@@ -121,7 +121,7 @@ Computes the gradient of the log joint of `θ` for the model specified by
 function gradient_reverse(
     θ::AbstractVector{<:Real},
     vi::Turing.VarInfo,
-    model::Function,
+    model::Model,
     sampler::Union{Nothing, Sampler}=nothing,
 )
     vals_old, logp_old = copy(vi.vals), copy(vi.logp)