Wrong type of output?

[andreasvarga]

The following horizontal concatenation produces the expected result:

julia> [1 rand(3)' rand(1,2) 2]
1×7 Array{Float64,2}:
 1.0  0.277368  0.788148  0.388695  0.495806  0.293525  2.0

However, I wonder if the following one has the correct output type:

julia> [1 rand(3)' 2]
1×5 Adjoint{Float64,Array{Float64,2}}:
 1.0  0.551156  0.578134  0.821385  2.0

I see no reason why the output should be of type Adjoint !

[mgkuhn]

Note that rand(3)' is also of type Adjoint. If you want Julia to actually execute the transpose (i.e., copying the matrix data in memory), rather than just putting a “lazy wrapper” type around rand(3) to remember to permute the dimensions here, then you can collect the result:

julia> collect([1 rand(3)' 2])
1×5 Array{Float64,2}:
 1.0  0.441081  0.621996  0.349487  2.0
julia> [1 collect(rand(3)') 2]
1×5 Array{Float64,2}:
 1.0  0.346152  0.274838  0.194685  2.0

(But I would only do so after having confirmed that this has a performance advantage.)

[mbauman]

It's not a performance thing in this particular case; it's a behavior thing. Adjoint vectors have some special behaviors like rand(3)' * rand(3) evaluating to a scalar instead of a 1-vector or 1x1 matrix. So it makes sense to preserve that in some cases. Note that we don't preserve the wrapper in the case of a adjoint matrix column:

julia> [1 rand(3, 1)' 2]
1×5 Matrix{Float64}:
 1.0  0.866621  0.70354  0.0792592  2.0

I had no idea that cat had these smarts, but it makes sense to me (and may just fall out of similar or some such).

[KristofferC]

I had no idea that cat had these smarts, but it makes sense to me (and may just fall out of similar or some such).

I think it is explicitly handled:

julia/stdlib/LinearAlgebra/src/adjtrans.jl

Line 259 in 502c039

hcat(avs::Union{Number,AdjointAbsVec}...) = _adjoint_hcat(avs...)

julia/stdlib/LinearAlgebra/src/adjtrans.jl

Line 261 in 502c039

_adjoint_hcat(avs::Union{Number,AdjointAbsVec}...) = adjoint(vcat(map(adjoint, avs)...))