Problem displaying error message (pretty not defined)

[yha]

Mismatched dimensions on a @cast fail to print the proper error:

julia> x = [rand(2,3) for i=1:4];

julia> @cast y[i,j,k] := x[i,j][k]
ERROR: UndefVarError: pretty not defined

[mcabbott]

Sorry about that. I think this is fixed on master, and that this expressions should work:

(jl_iIumVJ) pkg> add TensorCast#master

julia> using TensorCast

julia> x = [rand(2,3) for i=1:4];

julia> @cast y[i,j,k] := x[i,j][k]
3×4×2 PermutedDimsArray(::Array{Float64,3}, (2, 3, 1)) with eltype Float64:
[:, :, 1]:
 0.899111  0.958207  0.972891   0.0562823
 0.675006  0.353919  0.0110689  0.781209
 0.410219  0.545557  0.370502   0.512076
...

[yha]

I can't make sense of this example. Shouldn't it error, since x should be accessed as x[i][j,k] rather than x[i,j][k]?

[mcabbott]

Oh I missed that. As you say, the it's exactly a problem with the attempt to write a friendly error message.

Tagged version:

julia> @pretty @cast y[i,j,k] := x[i,j][k]
begin
    local reindeer = PermutedDimsArray(red_glue(x, (:, *, *)), (2, 3, 1))
    y = reindeer
end

julia> @cast y[i,j,k] := x[i,j][k]
ERROR: UndefVarError: pretty not defined
Stacktrace:
 [1] gluecodecheck(A::Vector{Matrix{Float64}}, code::Tuple{Colon, typeof(*), typeof(*)})
   @ TensorCast.Fast ~/.julia/packages/TensorCast/rCpV9/src/slice.jl:130
 [2] red_glue(A::Vector{Matrix{Float64}}, code::Tuple{Colon, typeof(*), typeof(*)})
   @ TensorCast.Fast ~/.julia/packages/TensorCast/rCpV9/src/slice.jl:75

Master, after #17, uses a different glueing method, which works for any dimensionality, hence no error.

julia> @pretty @cast y[i,j,k] := x[i,j][k]
begin
    local hyena = PermutedDimsArray(TensorCast.stack_iter(x), (2, 3, 1))
    y = hyena
end

julia> @cast y[i,j,k] := x[i,j][k]
3×4×2 PermutedDimsArray(::Array{Float64, 3}, (2, 3, 1)) with eltype Float64:
...

julia> @cast y[i,j,k] := x[i,j][k]  assert
ERROR: DimensionMismatch("expected a 2-tensor x[i, j]: ndims(x) == 2")

There are many other circumstances in which things will work despite wrong shapes. Originally the macro tried quite hard to produce very short answers, but perhaps it would be better to always include these checks. There is never a guarantee that extra dimensions will be OK, its just an accident.

julia> y = rand(2,5);

julia> @cast z[i] := log(y[i])
2×5 Matrix{Float64}:
 -0.369487  -0.390997  -0.685209   -1.84328  -0.145182
 -0.326173  -0.331384  -0.0361702  -0.80519  -0.254444

julia> @pretty @cast z[i] := log(y[i])
begin
    z = @__dot__(log(y))
end

julia> @cast z[i] := log(y[i]) assert
ERROR: DimensionMismatch("expected a 1-tensor y[i]: ndims(y) == 1")

[mcabbott]

Should be fixed on v0.3.3, and on v0.4.