Repeat example from the docs is broken

[mcabbott]

julia> M = Array(reshape(1:12, 3,4))
3×4 Matrix{Int64}:
 1  4  7  10
 2  5  8  11
 3  6  9  12

julia> @cast R[r,(n,c)] := M[r,c]^2  (n in 1:3)
ERROR: LoadError: index n appears only on the left

julia> V = [10,20,30];

julia> @cast _[i,j] := V[i]   (j in 1:4)  # simpler version
ERROR: LoadError: index j appears only on the left

julia> @cast _[i,j] := V[i] + 0j   (j in 1:4)  # this works!
3×4 Matrix{Int64}:
 10  10  10  10
 20  20  20  20
 30  30  30  30

[RomeoV]

Any progress on this?
I'm running into a similar problem.

x = rand(3, 4)
@test repeat(x, 1, 2) == @cast _[h, (w, 2)] :=  x[h, w]
# or
@test repeat(x, 1, 2) == @cast _[h, (w, j)] :=  x[h, w] (j in 1:2)

(Both fail).
Perhaps the check for "only on the left" should be removed? I think the first version is even nicer though, and is also supported by python's einops library.

[mcabbott]

Sorry, I haven't yet looked into this. Not so hard, I hope.

[RomeoV]

Great. I'm comparing TensorCast.jl to the recent paper published on einops (https://openreview.net/pdf?id=oapKSVM2bcj).
If this is fixed, all examples in Listing 1 can be done with TensorCast.jl and I can provide a pull request with the appropriate tests / examples.

[mcabbott]

That sounds great.

One or twice I've tried to borrow documentation from einops... I see there's an ancient notebook currently committed, but not linked from anywhere. Would be nice to do more. And more tests never hurt either.

Some investigation into this repeat problem:

julia> R = randn(3, 3*4);  # write in-place, easier:

julia> @cast R[r,(n,c)] = M[r,c]^2  # (n in 1:3)  # n can be inferred
ERROR: LoadError: index n appears only on the left
Stacktrace:
 [1] checkallseen ...
 [2] _macro(exone::Expr, extwo::Nothing, exthree::Nothing; call::TensorCast.CallInfo, dict::Dict{Any, Any})
   @ TensorCast ~/.julia/dev/TensorCast/src/macro.jl:199

# comment out check on line 199 and it works:

julia> @cast R[r,(n,c)] = M[r,c]^2  # (n in 1:3)
3×3×4 Array{Float64, 3}:
[:, :, 1] =
 1.0  1.0  1.0
 4.0  4.0  4.0
...

julia> R  # line above returns the wrong thing, reshaped rather than R itself:
3×12 Matrix{Float64}:
 1.0  1.0  1.0  16.0  16.0  16.0  49.0  49.0  49.0  100.0  100.0  100.0
 4.0  4.0  4.0  25.0  25.0  25.0  64.0  64.0  64.0  121.0  121.0  121.0
 9.0  9.0  9.0  36.0  36.0  36.0  81.0  81.0  81.0  144.0  144.0  144.0

# out-of-place again:

julia> @cast R[r,(n,c)] := M[r,c]^2  (n in 1:3)
ERROR: DimensionMismatch: new dimensions (3, 12) must be consistent with array size 12

julia> @pretty @cast R[r,(n,c)] := M[r,c]^2  (n in 1:3)
begin
    @boundscheck ndims(M) == 2 || throw(ArgumentError("expected a 2-tensor M[r, c]"))
    local (ax_c, ax_n, ax_r) = (axes(M, 2), OneTo(3), axes(M, 1))
    local spider = transmute(M, Val((1, nothing, 2)))
    R = reshape(@__dot__(spider ^ 2), (ax_r, star(ax_n, ax_c)))
end

julia> @cast R[r,(n,c)] := M[r,c]^2 + 0n  (n in 1:3)
3×12 Matrix{Int64}:
 1  1  1  16  16  16  49  49  49  100  100  100
 4  4  4  25  25  25  64  64  64  121  121  121
 9  9  9  36  36  36  81  81  81  144  144  144

To make the new dimension, it needs to do something like .+ 0 .* transpose(ax_n) in the broadcast (as in the last expression). There must be, or have been, logic for this, somewhere...

According to git blame I added this to docs here between 0.4.0 and 0.4.1, but it doesn't run on either of those versions.

Ah now I found a branch: master...repeat

[mcabbott]

Note BTW that @cast _[h, (w, 2)] := x[h, w] will probably never work, but would ideally have a better error.

This is because @cast _[h] := x[h, 1] is already the first column, a constant index. Plausibly @cast _[h, j] := x[h, (j, 1)] (j in 1:2) should then extract some subset of the columns.

The way to specify sizes used to be something like @cast _[h, (w, j:2)] := x[h, w], or else j:2 after the expression, notation chosen to be compact. But this was removed when everything was upgraded to allow offsets everywhere.