wip #106

src/core.jl

@@ -22,7 +22,7 @@ mutable struct Result{T} <: Tracked{T}
     args
     kwargs
     Result{T}(v,f,a,k) where {T} = new(v,f,a,k)
-    Result{T}(v,f,a,k) where {T<:Value} = error("Result cannot take $T as arg.")
+    # Result{T}(v,f,a,k) where {T<:Value} = error("Result cannot take $T as arg.") # See #106
 end
 
 # value() should give a regular (non-Value) result regardless of recursion
@@ -132,6 +132,9 @@ function track(f, args, kwargs, bcasted)
     end
 end
 
+Result(v::T, f, args, kwargs) where {T<:Tracked} = v
+Result(v::T, f, args, kwargs) where {T<:Bcasted} = error("Result cannot take $T as arg")
+
 function Result(v::T, f, args, kwargs) where {T}
     record!(t::Tape, v::Tracked) = (n = get(t.dict, v, nothing); n === nothing ? record!(t, Node(v)) : n)
     record!(t::Tape, n::Node) = (t.dict[n.Value] = n; pushfirst!(t.list, n); n)

test/core.jl

@@ -46,4 +46,26 @@ using Statistics
     # Double broadcasting
     x = Param([1.,2.]); f3(x)=sin(x); f4(x)=sin.(x)
     @test grad((@diff sum(f3.(x))), x) == grad((@diff sum(f4.(x))), x) == grad((@diff sum(f4(x))), x)
+
+    # Issue #106: Double Result
+    h(x) = exp(-x); h′(x,y) = -y
+    𝓁(x,y) = sum(abs2,x-y)/2
+    function neural_net(mparams, input; h=h, h′=h′, N=length(mparams))
+        δ = [];
+        X = Any[input];
+        for i=1:N
+            x = sum(mparams[i] .* [X[i],1])
+            y = h.(x)
+            push!(δ, h′.(x,y))
+            push!(X,y)
+        end
+        return X,δ
+    end
+    mparams =[[randn(),randn()] for i=1:3]
+    P = Param(mparams)
+    loss(P,x,y)= 𝓁(neural_net(P,x)[1][end],y)
+    x,y=randn(),randn()
+    J = @diff loss(P,x,y)
+    @test isa(J, AutoGrad.Tape)
+    @test_broken @gcheck loss(P,x,y)
 end

test/gradcheck.jl

@@ -151,7 +151,7 @@ function gcheck(f, x...; kw=(), nsample=10, verbose=1, rtol=0.05, atol=0.01, del
     y = @diff gcsum(f, x...; kw...)
     if !isa(y, Tape); @warn("Output independent of params"); return true; end
     f0 = value(y)
-    ps = Param[ n.Value for n in y if isa(n.Value, Param) ]
+    ps = Param[ n.Value for n in y.list if isa(n.Value, Param) ]
     if isempty(ps); @error("Cannot find any params"); end
     vs = value.(ps)
     gs = (p->grad(y,p)).(ps)

test/header.jl

@@ -1,2 +1,2 @@
 using AutoGrad, Test
-using AutoGrad: gradcheck, randcheck
+using AutoGrad: gradcheck, randcheck, gcheck, @gcheck