cannot take derrivative of a scalar function

[ekinakyurek]

I have strange issue, not sure it is a bug. I am sure that loss is a scalar function.

using Knet,LinearAlgebra
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)

julia> J = @diff loss(P,x,y)
ERROR: Only scalar valued functions supported.
Stacktrace:
 [1] error(::String) at ./error.jl:33
 [2] #differentiate#1(::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}, ::Function, ::Function) at /kuacc/users/eakyurek13/.julia/packages/AutoGrad/eAmjh/src/core.jl:61
 [3] differentiate(::Function) at /kuacc/users/eakyurek13/.julia/packages/AutoGrad/eAmjh/src/core.jl:45
 [4] top-level scope at none:0

julia> loss(P,x,y)
0.0036112795041667906

[denizyuret]

In a @diff context, the return value seems to have multiple embedded Results:
R(R(R(8.982629056671427e32)))
and value(ans) does not strip all of them, therefore is not recognized as a Number. Need to see why this is happening, whether it is a valid problem and if so what is the best way to solve it.

[denizyuret]

Since this is an AutoGrad problem, I am moving it to denizyuret/AutoGrad.jl#106 and closing here.