How to perform explicit array broadcasting like np.broadcast_to/torch.expand in Julia?

General Usage
, , ,
4

From my vague knowledge of python, I think its a will actually be a solid array, not an array of arrays. (Or, something more like Julia’s Matrix, than like Vector{Vector}.) This will be more efficient than making arrays of small arrays, unless (as suggested) you use StaticArrays.

julia> a = [[0,0], [0,1], [0, 1]];

julia> b = [[0,0], [0,1], [1, 0], [1, 0]];

julia> as = reduce(hcat, a); bs = reduce(hcat, b)
2×4 Matrix{Int64}:
 0  0  1  1
 0  1  0  0

julia> sqrt.(dropdims(sum((as .- reshape(bs, size(bs,1), 1, size(bs,2))).^2, dims=1), dims=1))
3×4 Matrix{Float64}:
 0.0  1.0  1.0      1.0
 1.0  0.0  1.41421  1.41421
 1.0  0.0  1.41421  1.41421

# other tricks for how to write this:

julia> const newaxis = [CartesianIndex{0}()];

julia> view(bs, :, newaxis, :) = reshape(bs, size(bs,1), 1, size(bs,2))
true

julia> using TensorCast

julia> sqrt.(@reduce tmp[i,j] := sum(x) (as[x,i] - bs[x,j])^2)
3×4 Matrix{Float64}:
 0.0  1.0  1.0      1.0
 1.0  0.0  1.41421  1.41421
 1.0  0.0  1.41421  1.41421

Is this true? (I mean as a real question, not snark!) Julia’s broadcasting will always (at present) materialise the big array, before reducing (i.e. the 2×3×4 Array{Int64, 3} which is the argument of sum.) But things like reshape or view(bs, :, newaxis, :) will not not make a copy.

3 Likes