Merge pull request #130 from CTUAvastLab/unthunk

This fixes the problem that Zygote started to use unthunk.

Author: pevnak

src/aggregations/aggregation_stack.jl

@@ -67,8 +67,18 @@ AggregationStack(fs::AbstractAggregation...) = AggregationStack(fs)
 
 Flux.@layer :ignore AggregationStack
 
-function (a::AggregationStack)(x::Maybe{AbstractArray}, bags::AbstractBags, args...)
-    reduce(vcat, (f(x, bags, args...) for f in a.fs))
+# function (a::AggregationStack)(x::Maybe{AbstractArray}, bags::AbstractBags, args...)
+#     reduce(vcat, (f(x, bags, args...) for f in a.fs))
+# end
+
+@generated function (a::AggregationStack{T})(x::Maybe{AbstractArray}, bags::AbstractBags, args...) where {T<:Tuple}
+    l = T.parameters |> length
+    chs = map(1:l) do i
+        :(a.fs[$i](x, bags, args...))
+    end
+    quote
+        vcat($(chs...))
+    end
 end
 
 Flux.@forward AggregationStack.fs Base.getindex, Base.firstindex, Base.lastindex, Base.first,

src/aggregations/aggregations.jl

@@ -22,7 +22,7 @@ abstract type AbstractAggregation end
 @inline _weightsum(ws::AbstractVector, i) = ws[i]
 
 # more stable definitions for r_map and p_map
-ChainRulesCore.rrule(::typeof(softplus), x) = softplus.(x), Δ -> (NoTangent(), Δ .* σ.(x))
+ChainRulesCore.rrule(::typeof(softplus), x) = softplus.(x), Δ -> (NoTangent(), unthunk(Δ) .* σ.(x))
 
 # our definition of type min for Maybe{...} types
 _typemin(t::Type) = typemin(t)

src/aggregations/segmented_lse.jl

@@ -137,13 +137,13 @@ function ChainRulesCore.rrule(::typeof(segmented_lse_forw),
         x::AbstractMatrix, ψ::AbstractVector, r::AbstractVector, bags::AbstractBags)
     M = _lse_precomp(x, r, bags)
     y = _segmented_lse_norm(x, ψ, r, bags, M)
-    grad = Δ -> (NoTangent(), segmented_lse_back(Δ, y, x, ψ, r, bags, M)...)
+    grad = Δ -> (NoTangent(), segmented_lse_back(unthunk(Δ), y, x, ψ, r, bags, M)...)
     y, grad
 end
 
 function ChainRulesCore.rrule(::typeof(segmented_lse_forw),
         x::Missing, ψ::AbstractVector, r::AbstractVector, bags::AbstractBags)
     y = segmented_lse_forw(x, ψ, r, bags)
-    grad = Δ -> (NoTangent(), segmented_lse_back(Δ, x, ψ, bags)...)
+    grad = Δ -> (NoTangent(), segmented_lse_back(unthunk(Δ), x, ψ, bags)...)
     y, grad
 end

src/aggregations/segmented_max.jl

@@ -97,6 +97,6 @@ end
 
 function ChainRulesCore.rrule(::typeof(segmented_max_forw), args...)
     y = segmented_max_forw(args...)
-    grad = Δ -> (NoTangent(), segmented_max_back(Δ, y, args...)...)
+    grad = Δ -> (NoTangent(), segmented_max_back(unthunk(Δ), y, args...)...)
     y, grad
 end

src/aggregations/segmented_mean.jl

@@ -98,6 +98,6 @@ end
 
 function ChainRulesCore.rrule(::typeof(segmented_mean_forw), args...)
     y = segmented_mean_forw(args...)
-    grad = Δ -> (NoTangent(), segmented_mean_back(Δ, y, args...)...)
+    grad = Δ -> (NoTangent(), segmented_mean_back(unthunk(Δ), y, args...)...)
     y, grad
 end

src/aggregations/segmented_pnorm.jl

@@ -141,12 +141,12 @@ end
 function ChainRulesCore.rrule(::typeof(segmented_pnorm_forw), a::AbstractMatrix, ψ, p, bags, w)
     M = _pnorm_precomp(a, bags)
     y = _segmented_pnorm_norm(a, ψ, p, bags, w, M)
-    grad = Δ -> (NoTangent(), segmented_pnorm_back(Δ, y, a, ψ, p, bags, w, M)...)
+    grad = Δ -> (NoTangent(), segmented_pnorm_back(unthunk(Δ), y, a, ψ, p, bags, w, M)...)
     y, grad
 end
 
 function ChainRulesCore.rrule(::typeof(segmented_pnorm_forw), a::Missing, ψ, p, bags, w)
     y = segmented_pnorm_forw(a, ψ, p, bags, w)
-    grad = Δ -> (NoTangent(), segmented_pnorm_back(Δ, y, ψ, bags)...)
+    grad = Δ -> (NoTangent(), segmented_pnorm_back(unthunk(Δ), y, ψ, bags)...)
     y, grad
 end

src/aggregations/segmented_sum.jl

@@ -56,6 +56,7 @@ function segmented_sum_forw(x::AbstractMatrix, ψ::AbstractVector, bags::Abstrac
 end
 
 function segmented_sum_back(Δ, y, x, ψ, bags, w) 
+    Δ = unthunk(Δ)
     dx = zero(x)
     dψ = zero(ψ)
     dw = isnothing(w) ? ZeroTangent() : zero(w)
@@ -96,6 +97,6 @@ end
 
 function ChainRulesCore.rrule(::typeof(segmented_sum_forw), args...)
     y = segmented_sum_forw(args...)
-    grad = Δ -> (NoTangent(), segmented_sum_back(Δ, y, args...)...)
+    grad = Δ -> (NoTangent(), segmented_sum_back(unthunk(Δ), y, args...)...)
     y, grad
 end