Conv with non homogenous array eltypes gives confusing error message

[femtomc]

ERROR: LoadError: MethodError: no method matching Flux.Conv(::Array{Float64, 4}, ::Vector{Float32}, ::typeof(identity); stride=1, pad=0, dilation=1)
Closest candidates are:
  Flux.Conv(::F, ::A, ::V, ::Tuple{Vararg{Int64, N}}, ::Tuple{Vararg{Int64, M}}, ::Tuple{Vararg{Int64, N}}) where {N, M, F, A, V} at /home/mccoy/.julia/packages/Flux/q3zeA/src/layers/conv.jl:71 got unsupported keyword arguments "stride", "pad", "dilation"
  Flux.Conv(::AbstractArray{T, N}, ::Union{AbstractVector{T}, Flux.Zeros}, ::Any; stride, pad, dilation) where {T, N} at /home/mccoy/.julia/packages/Flux/q3zeA/src/layers/conv.jl:105
  Flux.Conv(::Tuple{Vararg{Integer, N}}, ::Pair{var"#s150", var"#s149"} where var"#s149"<:Integer where var"#s150"<:Integer, ::Any; init, stride, pad, dilation, weight, bias) where N at /home/mccoy/.julia/packages/Flux/q3zeA/src/layers/conv.jl:132

when using

Conv((5, 5), 10 => 20, identity; init = glorot_uniform64),

[CarloLucibello]

Conv was expecting homogeneous weights and bias (e.g. Float32). We should relax this or give some informative error

[gxyd]

(Correct me if I'm wrong) I was thinking about the two, and from https://docs.julialang.org/en/v1/manual/conversion-and-promotion/#Conversion-vs.-Construction . If we do relax this, that would mean a conversion from Float32 to Float64 of a (possibly) large array to take place, which might encourage its use that way.

Better go with simply throwing an informative error?

[DhairyaLGandhi]

I feel like the error message can be improved to not be a footgun, but other than that erring on the side of caution is good. Of course none of the machinery expects the same eltypes.

The second method in the stack trace also points at the missing eltypes