Merge pull request #513 from TuringLang/ref/compiler_rebased

Refactoring of compiler

REQUIRE

@@ -8,6 +8,7 @@ MCMCChain 0.1.1
 Libtask 0.1.1
 Flux 0.6.7
 Stan
+MacroTools
 
 ProgressMeter 0.6.0
 BinaryProvider 0.4.0

src/Turing.jl

@@ -19,6 +19,7 @@ using LinearAlgebra
 using ProgressMeter
 using Markdown
 using Libtask
+using MacroTools
 
 #  @init @require Stan="682df890-35be-576f-97d0-3d8c8b33a550" begin
 using Stan
@@ -37,6 +38,7 @@ import MCMCChain: AbstractChains, Chains
 const ADBACKEND = Ref(:reverse_diff)
 setadbackend(backend_sym) = begin
   @assert backend_sym == :forward_diff || backend_sym == :reverse_diff
+  backend_sym == :forward_diff && CHUNKSIZE[] == 0 && setchunksize(40)
   ADBACKEND[] = backend_sym
 end
 
@@ -72,7 +74,7 @@ const CACHERANGES = 0b01
 
 abstract type InferenceAlgorithm end
 abstract type Hamiltonian <: InferenceAlgorithm end
-
+abstract type AbstractSampler end
 """
     Sampler{T}
 
@@ -85,11 +87,21 @@ An implementation of an algorithm should include the following:
 Turing translates models to chunks that call the modelling functions at specified points. The dispatch is based on the value of a `sampler` variable. To include a new inference algorithm implements the requirements mentioned above in a separate file,
 then include that file at the end of this one.
 """
-mutable struct Sampler{T<:InferenceAlgorithm}
+mutable struct Sampler{T<:InferenceAlgorithm} <: AbstractSampler
   alg   ::  T
   info  ::  Dict{Symbol, Any}         # sampler infomation
 end
 
+"""
+Robust initialization method for model parameters in Hamiltonian samplers.
+"""
+struct HamiltonianRobustInit <: AbstractSampler end
+struct SampleFromPrior <: AbstractSampler end
+
+# This can be removed when all `spl=nothing` is replaced with
+#   `spl=SampleFromPrior`
+const AnySampler = Union{Nothing, AbstractSampler}
+
 include("utilities/helper.jl")
 include("utilities/transform.jl")
 include("core/varinfo.jl")  # core internal variable container

src/core/ad.jl

@@ -23,7 +23,7 @@ function gradient_forward(
   # Define function to compute log joint.
   function f(θ)
       vi[spl] = θ
-      return -runmodel(model, vi, spl).logp
+      return -runmodel!(model, vi, spl).logp
   end
 
   # Set chunk size and do ForwardMode.
@@ -57,7 +57,7 @@ function gradient_reverse(
   # Specify objective function.
   function f(θ)
     vi[spl] = θ
-    return -runmodel(model, vi, spl).logp
+    return -runmodel!(model, vi, spl).logp
   end
 
   # Compute forward and reverse passes.