朱莉娅函数调用的歧义

ambuigity on function call in julia

我有这个错误

ERROR: MethodError: vcat(::Array{Real,2}, ::TrackedArray{…,Array{Float32,2}}) is ambiguous. Candidates:
  vcat(364::AbstractArray, x::Union{TrackedArray, TrackedReal}, xs::Union{Number, AbstractArray}...) in Tracker at C:\Users\Henri\.julia\packages\TrackerwcYJ\src\lib\array.jl:167
  vcat(A::Union{AbstractArray{T,2}, AbstractArray{T,1}} where T...) in Base at abstractarray.jl:1296
Possible fix, define
  vcat(::Union{AbstractArray{T,2}, AbstractArray{T,1}} where T, ::Union{TrackedArray{T,1,A} where A<:AbstractArray{T,1} where T, TrackedArray{T,2,A} where A<:AbstractArray{T,2} where T}, ::Vararg{Union{AbstractArray{T,2}, AbstractArray{T,1}} where T,N} where N)

告诉我两个 vcat() 函数不明确。我想使用 Base.vcat() 函数,但明确使用它会引发相同的错误。这是为什么 ?错误抛出的这个 "possible fix" 是什么?

此外,当我手动调用 REPL 中的每一行时,没有抛出任何错误。我不明白这种行为。只有当 vcat() 在另一个函数内部调用的函数中时才会发生这种情况。就像我下面的例子。

这是重现错误的代码:

using Flux

function loss(a, b, net, net2)
    net2(vcat(net(a),a))

end

function test()    
    opt = ADAM()
    net = Chain(Dense(3,3))
    net2 = Chain(Dense(6,1))
    L(a, b) = loss(a, b, net, net2)

    data = tuple(rand(3,1), rand(3,1))
    xs = Flux.params(net)
    gs = Tracker.gradient(() -> L(data...), xs)
    Tracker.update!(opt, xs, gs)
end

正如在 Henri.D 的评论中提到的,我们已经设法通过小心处理 a 的类型来修复它,它是 Float64Arrayrand 返回的默认类型,而 net(a) 返回 Float32TrackedArray 并且无法使用 a.

vcat

I've managed to fix vcat by changing your loss function with this: net2(vcat(net(a),Float32.(a))) because vcat couldn't concatenate as net(a) was a Float32 Array and a a Float64 one. Then L(data...) is a TrackedArray of 1 element whereas I think you need a Float32 that's why I finally replace loss function by net2(vcat(net(a),Float32.(a)))[1]