Optimal way to flatten array

General Usage
1

I have an array of array

y = [ones(Float32, 3) for i=1:3]
sum(abs2, y)

which I can’t apply abs2 to.
I need it to compute my loss function

Now I use custom function

function flat_this(x)
    arr = ones(0)
    for i=1:length(x)
        arr = vcat(arr, x[i])
    end
    return arr
end

But I doubt it’s optimal and fast enough

2

reduce(vcat, y)

or in this case, to avoid an intermediary Array

sum(i -> sum(abs2,i), y)

3 Likes
3

Or another alternative:
sum(abs2,Iterators.flatten(y))

3 Likes
4

Iterators.flatten doesn’t allocate because it’s lazy but it’s slower (I have no idea why)

julia> y = [rand(100) for i in 1:100];

julia> @btime sum(abs2, reduce(vcat, $y))
  7.200 μs (2 allocations: 78.20 KiB)
3333.509067117525

julia> @btime sum(i -> sum(abs2,i), $y)
  1.400 μs (0 allocations: 0 bytes)
3333.5090671175253

julia> @btime sum(abs2,Iterators.flatten($y))
  9.600 μs (0 allocations: 0 bytes)
3333.5090671175276
3 Likes
5

Another way to flatten y, quite inefficient compared to reduce(vcat,y) above, though:

[yi for yr in y for yi in yr]
1 Like
6

Thanks for replies