incorrect gradient when indexing into a matrix of vectors

[ylxdzsw]

I'm using latest master of AutoGrad.jl in Julia 0.6.3 on Windows and it gives an unexpected gradient when indexing into a matrix of vectors:

julia> a = Matrix{Array{Float32}}(2, 1)
2×1 Array{Array{Float32,N} where N,2}:
 #undef
 #undef

julia> a[1] = [2,3,4]
3-element Array{Int64,1}:
 2
 3
 4

julia> a[2] = [4,5,6]
3-element Array{Int64,1}:
 4
 5
 6

julia> g = grad(x->x[1]' * x[2])
(::gradfun) (generic function with 1 method)

julia> g(a)
2×1 Array{Any,2}:
 Float32[4.0, 5.0, 6.0]
 Float32[2.0, 3.0, 4.0]

julia> g = grad(x->x[1, 1]' * x[2, 1])
(::gradfun) (generic function with 1 method)

julia> g(a)
2×1 Array{Any,2}:
 4.0
 2.0

The two functions are identical in my opinion, but the latter one gives only the gradient of the first element in each vector.

[CarloLucibello]

Also, ideally in both definitions the result should be a 2×1 Array{Array{Float32,N} where N,2} (note that the first definition of g instead is an Array{Any,2})

[denizyuret]

@ylxdzsw fixed in latest master, please test.

@CarloLucibello I cannot do this in general because I want to be able to represent zero gradients for large arrays using nothing. So non-bitstype arrays may get converted to Any arrays.

[ylxdzsw]

Yes it works, thanks.