Fix memoization issue (#1414)

Project.toml

@@ -1,6 +1,6 @@
 name = "Turing"
 uuid = "fce5fe82-541a-59a6-adf8-730c64b5f9a0"
-version = "0.14.3"
+version = "0.14.4"
 
 [deps]
 AbstractMCMC = "80f14c24-f653-4e6a-9b94-39d6b0f70001"

src/core/compat/reversediff.jl

@@ -23,7 +23,7 @@ function gradient_logp(
     context::DynamicPPL.AbstractContext = DynamicPPL.DefaultContext()
 )
     T = typeof(getlogp(vi))
-    
+
     # Specify objective function.
     function f(θ)
         new_vi = VarInfo(vi, sampler, θ)
@@ -46,12 +46,10 @@ end
 @require Memoization = "6fafb56a-5788-4b4e-91ca-c0cea6611c73" @eval begin
     setrdcache(::Val{true}) = RDCache[] = true
     function emptyrdcache()
-        for k in keys(Memoization.caches)
-            if k[1] === typeof(memoized_taperesult)
-                pop!(Memoization.caches, k)
-            end
-        end
+        Memoization.empty_cache!(memoized_taperesult)
+        return
     end
+
     function gradient_logp(
         backend::ReverseDiffAD{true},
         θ::AbstractVector{<:Real},
@@ -61,7 +59,7 @@ end
         context::DynamicPPL.AbstractContext = DynamicPPL.DefaultContext()
     )
         T = typeof(getlogp(vi))
-        
+
         # Specify objective function.
         function f(θ)
             new_vi = VarInfo(vi, sampler, θ)
@@ -81,15 +79,13 @@ end
         f::F
         x::Tx
     end
-    function Memoization._get!(f::Union{Function, Type}, d::IdDict, keys::Tuple{Tuple{RDTapeKey}, Any})
+    function Memoization._get!(f, d::Dict, keys::Tuple{Tuple{RDTapeKey}, Any})
         key = keys[1][1]
-        return Memoization._get!(f, d, (typeof(key.f), typeof(key.x), size(key.x)))
+        return Memoization._get!(f, d, (key.f, typeof(key.x), size(key.x), Threads.threadid()))
     end
     memoized_taperesult(f, x) = memoized_taperesult(RDTapeKey(f, x))
-    Memoization.@memoize function memoized_taperesult(k::RDTapeKey)
+    Memoization.@memoize Dict function memoized_taperesult(k::RDTapeKey)
         return compiledtape(k.f, k.x), GradientResult(k.x)
     end
-    memoized_tape(f, x) = memoized_tape(RDTapeKey(f, x))
-    Memoization.@memoize memoized_tape(k::RDTapeKey) = compiledtape(k.f, k.x)
     compiledtape(f, x) = compile(GradientTape(f, x))
 end

test/core/ad.jl

@@ -276,9 +276,13 @@ _to_cov(B) = B * B' + Matrix(I, size(B)...)
         sample(dir(), HMC(0.01, 1), 1000);
         Turing.setrdcache(true)
         sample(dir(), HMC(0.01, 1), 1000);
-        @test length(Memoization.caches) == 1
+        caches = Memoization.find_caches(Turing.Core.memoized_taperesult)
+        @test length(caches) == 1
+        @test !isempty(first(values(caches)))
         Turing.emptyrdcache()
-        @test length(Memoization.caches) == 0
+        caches = Memoization.find_caches(Turing.Core.memoized_taperesult)
+        @test length(caches) == 1
+        @test isempty(first(values(caches)))
     end
     # FIXME: For some reasons PDMatDistribution AD tests fail with ReverseDiff
     @testset "PDMatDistribution AD" begin
@@ -340,4 +344,24 @@ _to_cov(B) = B * B' + Matrix(I, size(B)...)
         @test H_f == [1.0 0.0; 0.0 1.0]
         @test H_f == H_r
     end
+
+    @testset "memoization: issue #1393" begin
+        Turing.setadbackend(:reversediff)
+        Turing.setrdcache(true)
+
+        @model function demo(data)
+            sigma ~ Uniform(0.0, 20.0)
+            data ~ Normal(0, sigma)
+        end
+
+        N = 1000
+        for i in 1:5
+            d = Normal(0.0, i)
+            data = rand(d, N)
+            chn = sample(demo(data), NUTS(0.65), 1000)
+            @test mean(Array(chn[:sigma])) ≈ std(data) atol=0.5
+        end
+
+        Turing.emptyrdcache()
+    end
 end

test/runtests.jl

@@ -15,13 +15,8 @@ include("test_utils/AllUtils.jl")
         include("core/container.jl")
     end
 
-    test_adbackends = if VERSION >= v"1.2"
-        [:forwarddiff, :tracker, :reversediff]
-    else
-        [:forwarddiff, :tracker]
-    end
     Turing.setrdcache(false)
-    for adbackend in test_adbackends
+    for adbackend in (:forwarddiff, :tracker, :reversediff)
         Turing.setadbackend(adbackend)
         @testset "inference: $adbackend" begin
             @testset "samplers" begin