give up and use AbstractArray

src/backend/mxnet/graph.jl

@@ -90,7 +90,7 @@ end
 
 register(ctx::Context, node) = node
 
-function var(ctx::Context, p::Union{Flux.Param{<:AArray},AArray,AlterParam})
+function var(ctx::Context, p::Union{Flux.Param{<:AbstractArray},AbstractArray,AlterParam})
   id = gensym()
   ctx[:params][id] = p
   return mx.Variable(id)

src/backend/tensorflow/graph.jl

@@ -62,7 +62,7 @@ end
 interp(ctx, c::Conv2D, x) =
   nn.conv2d(x, interp(ctx, constant(c.filter)), [1,c.stride...,1], "VALID")
 
-param(ctx, p::Flux.Param{<:AArray}) =
+param(ctx, p::Flux.Param{<:AbstractArray}) =
   haskey(ctx[:params], p) ?
      ctx[:params][p] :
     (ctx[:params][p] =

src/compiler/code.jl

@@ -34,7 +34,7 @@ function build_type(T, params)
   end
   if any(x->isexpr(x, Symbol), params)
     push!(ex.args,
-      :($T($(map(x->isexpr(x, Symbol) ? :($x::AArray) : x, params)...)) =
+      :($T($(map(x->isexpr(x, Symbol) ? :($x::AbstractArray) : x, params)...)) =
           $T($(map(x->isexpr(x, Symbol) ? :(param($x)) : namify(x), params)...))))
   end
   ex

src/utils.jl

@@ -2,8 +2,6 @@ call(f, xs...) = f(xs...)
 
 # Arrays
 
-const AArray = AbstractArray
-
 initn(dims...) = randn(dims...)/100
 
 unsqueeze(xs, dim = 1) = reshape(xs, (size(xs)[1:dim-1]..., 1, size(xs)[dim:end]...))