Skip to content

JuliaArrays/StackViews.jl

Folders and files

NameName
Last commit message
Last commit date

Latest commit

Β 

History

18 Commits
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation

StackViews

Build Status Coverage

StackViews provides only one array type: StackView. There are multiple ways to understand StackView:

  • inverse of eachslice
  • cat variant
  • view object
  • lazy version of repeat special case

StackView as the inverse of eachslice

StackView can be seen as the inverse of eachslice:

julia> using StackViews

julia> X = rand(100, 100);

julia> StackView(collect(eachslice(X, dims=1)), 1) == X
true

StackView as a variant of cat

StackView works very similar to cat and its vcat/hcat/hvcat variants in Base.

julia> using StackViews

julia> A = reshape(collect(1:6), 2, 3)
2Γ—3 Matrix{Int64}:
 1  3  5
 2  4  6

julia> B = reshape(collect(7:12), 2, 3)
2Γ—3 Matrix{Int64}:
 7   9  11
 8  10  12

julia> StackView(A, B; dims=3) # mostly equivalent to `cat(A, B, dims=3)`
2Γ—3Γ—2 StackView{Int64, 3, 3, ...}:
[:, :, 1] =
 1  3  5
 2  4  6

[:, :, 2] =
 7   9  11
 8  10  12

Unlike cats, StackView always creats new dimension:

julia> StackView(A, B; dims=1) # `cat(A, B, dims=1)` outputs 4Γ—3 Matrix
2Γ—2Γ—3 StackView{Int64, 3, 1, ...}:
[:, :, 1] =
 1  2
 7  8

[:, :, 2] =
 3   4
 9  10

[:, :, 3] =
  5   6
 11  12

julia> StackView(A, B; dims=2) # `cat(A, B, dims=2)` outputs 2Γ—6 Matrix
2Γ—2Γ—3 StackView{Int64, 3, 2, ...}:
[:, :, 1] =
 1  7
 2  8

[:, :, 2] =
 3   9
 4  10

[:, :, 3] =
 5  11
 6  12

StackView as a view object

Without StackView, people use reshape + cat for the previous examples:

julia> StackView(A, B, dims=1) == cat(map(x->reshape(x, 1, axes(x)...), (A, B))...; dims=1)
true

but StackView is only a view and thus don't create any memory:

frames = [rand(1000, 1000) for _ in 1:100];
@btime StackView($frames, Val(1)); # 143.905 ns (0 allocations: 0 bytes)
@btime cat(map(x->reshape(x, 1, axes(x)...), $frames)...; dims=1); # 1.127 s (1119 allocations: 763.06 MiB)

Of course, since it is a view, if you modify it, original arrays get modified, too:

A = [1, 2, 3, 4]
B = [5, 6, 7, 8]
sv = StackView(A, B)
fill!(sv, -1) # A and B are modified
A == B == fill(-1, 4) # true

If indexed in contiguous memeory order, it has almost zero overhead for getindex and setindex!:

function arrsum_cart(A::AbstractArray)
    rst = zero(eltype(A))
    @inbounds for I in CartesianIndices(A)
        rst += A[I]
    end
    return rst
end

As = StackView(frames);
Ac = cat(map(x->reshape(x, axes(x)..., 1), frames)...; dims=3);
As == Ac # true

@btime arrsum_cart($As); # 122.703 ms (0 allocations: 0 bytes)
@btime arrsum_cart($Ac); # 123.813 ms (0 allocations: 0 bytes)

StackView as a lazy version of repeat special case

StackView allows you to stack the same array object multiple times, which makes a special version of repeat when there's only one none-1 repeat count:

A = rand(1000, 1000);
n = 100;
StackView([A for _ in 1:n]) == repeat(A, ntuple(_->1, ndims(A))..., n) # true
@btime StackView([$A for _ in 1:$n]); # 403.156 ns (2 allocations: 1.75 KiB)
@btime repeat($A, ntuple(_->1, ndims($A))..., $n) # 590.043 ms (4 allocations: 762.94 MiB)

More examples

When arrays are of different types and sizes, StackView just kills cats:

julia> using StackViews, PaddedViews

julia> A = collect(reshape(1:8, 2, 4));

julia> B = collect(reshape(9:16, 4, 2));

julia> StackView(paddedviews(-1, A, B), 3)
4Γ—4Γ—2 StackView{Int64, 3, 3, ...}:
[:, :, 1] =
  1   3   5   7
  2   4   6   8
 -1  -1  -1  -1
 -1  -1  -1  -1

[:, :, 2] =
  9  13  -1  -1
 10  14  -1  -1
 11  15  -1  -1
 12  16  -1  -1

julia> StackView(sym_paddedviews(-1, A, B), 3)
4Γ—4Γ—2 StackView{Int64, 3, 3, ...):
[:, :, 1] =
 -1  -1  -1  -1
  1   3   5   7
  2   4   6   8
 -1  -1  -1  -1

[:, :, 2] =
 -1   9  13  -1
 -1  10  14  -1
 -1  11  15  -1
 -1  12  16  -1

There is some mind work here but by chaining more views you can get some interesting result:

using PaddedViews, StackViews
using ImageCore, ImageShow, TestImages, ColorVectorSpace

toucan = testimage("toucan") # 150Γ—162 RGBA image
moon = testimage("moon") # 256Γ—256 Gray image

# equivalently, you can just use `mosaic(toucan, moon; nrow=1)`
mosaicview(StackView(sym_paddedviews(zero(RGB), toucan, moon)); nrow=1)

Other similar packages

  • There's a lot of overlap between this and LazyStack.jl; I didn't notice its existance when I wrote this.