ADTypes Interop (#127)

* ADTypes interop

* Bump patch

Project.toml

@@ -1,9 +1,10 @@
 name = "Tapir"
 uuid = "07d77754-e150-4737-8c94-cd238a1fb45b"
 authors = ["Will Tebbutt, Hong Ge, and contributors"]
-version = "0.2.1"
+version = "0.2.2"
 
 [deps]
+ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 DiffRules = "b552c78f-8df3-52c6-915a-8e097449b14b"
 DiffTests = "de460e47-3fe3-5279-bb4a-814414816d5d"
@@ -25,6 +26,7 @@ TapirLogDensityProblemsADExt = "LogDensityProblemsAD"
 TapirSpecialFunctionsExt = "SpecialFunctions"
 
 [compat]
+ADTypes = "1"
 BenchmarkTools = "1"
 ChainRulesCore = "1"
 DiffRules = "1"
@@ -40,7 +42,7 @@ Setfield = "1"
 SpecialFunctions = "2"
 StableRNGs = "1"
 TemporalGPs = "0.6"
-Turing = "0.31"
+Turing = "0.31.3"
 julia = "1"
 
 [extras]

ext/TapirLogDensityProblemsADExt.jl

@@ -4,38 +4,42 @@
 module TapirLogDensityProblemsADExt
 
 if isdefined(Base, :get_extension)
+    using ADTypes
     using LogDensityProblemsAD: ADGradientWrapper
     import LogDensityProblemsAD: ADgradient, logdensity_and_gradient, dimension, logdensity
     import Tapir
 else
+    using ADTypes
     using ..LogDensityProblemsAD: ADGradientWrapper
     import ..LogDensityProblemsAD: ADgradient, logdensity_and_gradient, dimension, logdensity
     import ..Tapir
 end
 
 struct TapirGradientLogDensity{Trule, L} <: ADGradientWrapper
     rule::Trule
-    l::L
+    ℓ::L
 end
 
-dimension(∇l::TapirGradientLogDensity) = dimension(Tapir.primal(∇l.l))
+dimension(∇l::TapirGradientLogDensity) = dimension(Tapir.primal(∇l.ℓ))
 
 function logdensity(∇l::TapirGradientLogDensity, x::Vector{Float64})
-    return logdensity(Tapir.primal(∇l.l), x)
+    return logdensity(Tapir.primal(∇l.ℓ), x)
 end
 
 """
     ADgradient(Val(:Tapir), ℓ)
 
 Gradient using algorithmic/automatic differentiation via Tapir.
 """
-function ADgradient(::Val{:Tapir}, l)
-    primal_sig = Tuple{typeof(logdensity), typeof(l), Vector{Float64}}
+function ADgradient(::Val{:Tapir}, ℓ)
+    primal_sig = Tuple{typeof(logdensity), typeof(ℓ), Vector{Float64}}
     rule = Tapir.build_rrule(Tapir.TapirInterpreter(), primal_sig)
-    return TapirGradientLogDensity(rule, Tapir.uninit_fcodual(l))
+    return TapirGradientLogDensity(rule, Tapir.uninit_fcodual(ℓ))
 end
 
-Base.show(io::IO, ∇ℓ::TapirGradientLogDensity) = print(io, "Tapir AD wrapper for ", ∇ℓ.ℓ)
+function Base.show(io::IO, ∇ℓ::TapirGradientLogDensity)
+    return print(io, "Tapir AD wrapper for ", Tapir.primal(∇ℓ.ℓ))
+end
 
 # We only test Tapir with `Float64`s at the minute, so make strong assumptions about the
 # types supported in order to prevent silent errors.
@@ -46,10 +50,13 @@ end
 
 function logdensity_and_gradient(∇l::TapirGradientLogDensity, x::Vector{Float64})
     dx = zeros(length(x))
-    y, pb!! = ∇l.rule(Tapir.zero_fcodual(logdensity), ∇l.l, Tapir.CoDual(x, dx))
+    y, pb!! = ∇l.rule(Tapir.zero_fcodual(logdensity), ∇l.ℓ, Tapir.CoDual(x, dx))
     @assert Tapir.primal(y) isa Float64
     pb!!(1.0)
     return Tapir.primal(y), dx
 end
 
+# Interop with ADTypes.
+ADgradient(::ADTypes.AutoTapir, ℓ) = ADgradient(Val(:Tapir), ℓ)
+
 end

src/Tapir.jl

@@ -3,6 +3,7 @@ module Tapir
 const CC = Core.Compiler
 
 using
+    ADTypes,
     DiffRules,
     ExprTools,
     Graphs,

test/integration_testing/logdensityproblemsad_interop.jl

@@ -1,4 +1,4 @@
-using LogDensityProblemsAD
+using ADTypes, LogDensityProblemsAD
 using LogDensityProblemsAD: logdensity_and_gradient, capabilities, dimension, logdensity
 
 # Copied over from LogDensityProblemsAD test suite.
@@ -19,4 +19,6 @@ test_gradient(x) = -2 .* x
         @test isapprox(logdensity_and_gradient(∇l, x)[1], logdensity(TestLogDensity2(), x))
         @test isapprox(logdensity_and_gradient(∇l, x)[2], test_gradient(x))
     end
+
+    @test ADgradient(ADTypes.AutoTapir(), l) isa typeof(∇l)
 end