Extra singleton dimension appears during casting

[prittjam]

using TensorCast, EllipsisNotation

@views project(x) = x ./ x[[end],..]
@views project!(x) = @. x = x / x[[end],..]
euclideanize(x) = project(x)[1:end-1,..]

P = rand(3,4,100)   
@cast u[i,k] := euclideanize(P[:,4,:])[i,k]

The result should be 2 dimensional but it is 3 with a singleton dimension added. Hower, this works as expected,

@cast u[i,k] := Affine.euclideanize(view(P, :, 4, :))[i,k]

[mcabbott]

Thanks, that's a bug. I think this isolates it to mcabbott/LazyStack.jl#15

julia> euclideanize(P[:,4,:])
2×100 Matrix{Float64}:
 1.93102   1.61322  0.727326  0.626134  1.35614   …  1.27111   0.463712  2.29733  2.68379
 0.113882  1.52248  1.03129   0.626155  0.292362     0.839144  0.118597  1.97182  6.95894

julia> @cast u[i,k] := euclideanize(P[:,4,:])[i,k]; summary(u)
"2×100×1 lazystack(::Tuple{Matrix{Float64}}) with eltype Float64"

julia> @pretty @cast u[i,k] := euclideanize(P[:,4,:])[i,k]
begin
    @boundscheck ndims(P) == 3 || throw(ArgumentError("expected a 3-tensor P[:, 4, :]"))
    local dolphin = sliceview(view(P, :, 4, :), (:, :))
    local pig = @__dot__(euclideanize(dolphin))
    local armadillo = lazystack(pig)
    u = armadillo
end

julia> TensorCast.sliceview(view(P, :, 4, :), (:, :)) |> summary
"0-dimensional Array{SubArray{Float64, 2, Array{Float64, 3}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64, Base.Slice{Base.OneTo{Int64}}}, false}, 0}"

julia> TensorCast.sliceview(view(P, :, 4, :), (:, :)) .|> euclideanize |> summary
"2×100 Matrix{Float64}"

julia> TensorCast.sliceview(view(P, :, 4, :), (:, :)) .|> euclideanize |> TensorCast.lazystack |> summary
"2×100×1 lazystack(::Tuple{Matrix{Float64}}) with eltype Float64"

Ideally perhaps @cast would recognise that euclideanize(P[:,4,:]) is a function like rand(2,100) which it should pull out, as it contains no indexing that the macro ought to do anything about. But it does not see this.

[mcabbott]

This is fixed in LazyStack v0.1.1, btw. I leave the issue open as a reminder to add a test here too.