Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Adding one for structured matrices that preserves type #29777

Merged
merged 8 commits into from
Feb 6, 2019

Conversation

mcognetta
Copy link
Contributor

As with several other PRs, one falls back to generic methods for structured matrices and it does not maintain type (zero does not have these problems).

This PR adds specialized one methods that preserve input type for structured matrices (Diagonal, Bidiagonal, Tridiagonal, SymTridiagonal).

It should be noted that when the container type is a range, it gets promoted to a Vector, which mimics the behavior or zero. Otherwise, the container type is preserved.

julia> D = Diagonal(1:3)
3×3 Diagonal{Int64,UnitRange{Int64}}:
 1  ⋅  ⋅
 ⋅  2  ⋅
 ⋅  ⋅  3

julia> zero(D)
3×3 Diagonal{Int64,Array{Int64,1}}:
 0  ⋅  ⋅
 ⋅  0  ⋅
 ⋅  ⋅  0

julia> one(D)
3×3 Diagonal{Int64,Array{Int64,1}}:
 1  ⋅  ⋅
 ⋅  1  ⋅
 ⋅  ⋅  1

julia> D = Diagonal(sparsevec([1, 2]))
2×2 Diagonal{Int64,SparseVector{Int64,Int64}}:
 1  ⋅
 ⋅  2

julia> zero(D)
2×2 Diagonal{Int64,SparseVector{Int64,Int64}}:
 0  ⋅
 ⋅  0

julia> one(D)
2×2 Diagonal{Int64,SparseVector{Int64,Int64}}:
 1  ⋅
 ⋅  1


Some times:

v1.0

julia> D = Diagonal(rand(10000))
julia> B = Bidiagonal(rand(10000), rand(9999), 'U')
julia> T = Tridiagonal(rand(9999), rand(10000), rand(9999))
julia> S = SymTridiagonal(rand(10000), rand(9999))

julia> @time one(D)
 0.213831 seconds (6 allocations: 762.940 MiB, 10.34% gc time)
julia> @time one(B)
 0.212062 seconds (6 allocations: 762.940 MiB, 9.95% gc time)
julia> @time one(T)
 0.212362 seconds (6 allocations: 762.940 MiB, 10.02% gc time)
julia> @time one(S)
 0.209232 seconds (6 allocations: 762.940 MiB, 9.72% gc time)

julia> one(D)
10000×10000 Array{Float64,2}:
 1.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  …  0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  1.0  0.0  0.0  0.0  0.0  0.0  0.0     0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  1.0  0.0  0.0  0.0  0.0  0.0     0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  1.0  0.0  0.0  0.0  0.0     0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  1.0  0.0  0.0  0.0     0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  1.0  0.0  0.0  …  0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  1.0  0.0     0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  1.0     0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     0.0  0.0  0.0  0.0  0.0  0.0  0.0
 ⋮                        ⋮              ⋱            ⋮                      
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     1.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     0.0  1.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  …  0.0  0.0  1.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     0.0  0.0  0.0  1.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     0.0  0.0  0.0  0.0  1.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     0.0  0.0  0.0  0.0  0.0  1.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     0.0  0.0  0.0  0.0  0.0  0.0  1.0

After PR

julia> D = Diagonal(rand(10000))
julia> B = Bidiagonal(rand(10000), rand(9999), 'U')
julia> T = Tridiagonal(rand(9999), rand(10000), rand(9999))
julia> S = SymTridiagonal(rand(10000), rand(9999))

julia> @time one(D)
 0.000024 seconds (7 allocations: 78.375 KiB)
julia> @time one(B)
 0.000043 seconds (9 allocations: 156.594 KiB)
julia> @time one(T)
 0.000069 seconds (11 allocations: 234.813 KiB)
julia> @time one(S)
 0.000043 seconds (9 allocations: 156.594 KiB)

julia> one(D)
10000×10000 Diagonal{Float64,Array{Float64,1}}:
 1.0   ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅   …   ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅   1.0   ⋅    ⋅    ⋅    ⋅    ⋅    ⋅       ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅   1.0   ⋅    ⋅    ⋅    ⋅    ⋅       ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅   1.0   ⋅    ⋅    ⋅    ⋅       ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅   1.0   ⋅    ⋅    ⋅       ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅   1.0   ⋅    ⋅   …   ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅   1.0   ⋅       ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅   1.0      ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅       ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅       ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅ 
 ⋮                        ⋮              ⋱            ⋮                      
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅       ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅       ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅      1.0   ⋅    ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅       ⋅   1.0   ⋅    ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅   …   ⋅    ⋅   1.0   ⋅    ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅       ⋅    ⋅    ⋅   1.0   ⋅    ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅       ⋅    ⋅    ⋅    ⋅   1.0   ⋅    ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅       ⋅    ⋅    ⋅    ⋅    ⋅   1.0   ⋅ 
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅       ⋅    ⋅    ⋅    ⋅    ⋅    ⋅   1.0

@fredrikekre fredrikekre added linear algebra Linear algebra minor change Marginal behavior change acceptable for a minor release labels Oct 23, 2018
one(A::Diagonal{T}) where T = Diagonal(fill!(similar(A.diag), one(T)))
one(A::Bidiagonal{T}) where T = Bidiagonal(fill!(similar(A.dv), one(T)), zero(A.ev), A.uplo)
one(A::Tridiagonal{T}) where T = Tridiagonal(zero(A.du), fill!(similar(A.d), one(T)), zero(A.dl))
one(A::SymTridiagonal{T}) where T = SymTridiagonal(fill!(similar(A.dv), one(T)), zero(A.ev))
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

similar(foo) in all of these cases is wrong. It should be similar(foo, typeof(one(T)))

The issue is that, if T is a dimensionful type, one returns a dimensionless type.

Copy link
Member

@stevengj stevengj Oct 23, 2018

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You could also use fill(one(T), size(foo)), which might be simpler.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

zero(A.dl) etcetera is wrong for a related reason, because zero returns a dimensionful value. You should instead use fill(zero(one(T)), size(A.dl)) or similar.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You could also use fill(one(T), size(foo)), which might be simpler.

No, this would always return Vector.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Would be good to test the dimensionful case … see the tests with Furlongs in the test/triangular.jl file.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@fredrikekre is correct about (not) using fill(one(T), size(foo)). This would not preserve the underlying container type (as an example: sparsevector).

As for the dimensionless type, using similar should an abstractvector with the given eltype, which then promotes the result of one to the appropriate dimensionful type.

For example:

julia> x = Furlong(3)
Furlong{1,Int64}(3)

julia> D = Diagonal([x, x, x])
3×3 Diagonal{Furlong{1,Int64},Array{Furlong{1,Int64},1}}:
 Furlong{1,Int64}(3)           ⋅                    ⋅
          ⋅           Furlong{1,Int64}(3)           ⋅
          ⋅                    ⋅           Furlong{1,Int64}(3)

julia> one(D)
3×3 Diagonal{Furlong{1,Int64},Array{Furlong{1,Int64},1}}:
 Furlong{1,Int64}(1)           ⋅                    ⋅
          ⋅           Furlong{1,Int64}(1)           ⋅
          ⋅                    ⋅           Furlong{1,Int64}(1)

julia> y = Furlong{2}(3)
Furlong{2,Int64}(3)

julia> E = Diagonal([y, y, y])
3×3 Diagonal{Furlong{2,Int64},Array{Furlong{2,Int64},1}}:
 Furlong{2,Int64}(3)           ⋅                    ⋅
          ⋅           Furlong{2,Int64}(3)           ⋅
          ⋅                    ⋅           Furlong{2,Int64}(3)

julia> one(E)
3×3 Diagonal{Furlong{2,Int64},Array{Furlong{2,Int64},1}}:
 Furlong{2,Int64}(1)           ⋅                    ⋅
          ⋅           Furlong{2,Int64}(1)           ⋅
          ⋅                    ⋅           Furlong{2,Int64}(1)

Is this not the behavior that we want?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

one should return a multiplicative identity, so it should be unitless, e.g.

julia> x = Furlong(3)
Furlong{1,Int64}(3)

julia> one(x)
1

julia> x * one(x) == x
true

and thus,

julia> D = Diagonal([x, x, x])
3×3 Diagonal{Furlong{1,Int64},Array{Furlong{1,Int64},1}}:
 Furlong{1,Int64}(3)           ⋅                    ⋅         
          ⋅           Furlong{1,Int64}(3)           ⋅         
          ⋅                    ⋅           Furlong{1,Int64}(3)

julia> one(D)
3×3 Diagonal{Int64,Array{Int64,1}}:         <-- dimensionless
 1  ⋅  ⋅
 ⋅  1  ⋅
 ⋅  ⋅  1

should be true.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah, I interpreted @stevengj 's comment the opposite way. I will update this and add some tests with types that have a dimension.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This would not preserve the underlying container type (as an example: sparsevector).

Who cares? Why does the underlying container type matter? An array of ones is not sparse anyway.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This way, it would be consistent with zero. Not that I think making it a vector is a bad idea though (especially considering the behavior of zero and one for these matrices backed by ranges).

julia> x=sprand(Float64, 10, 10, .1)
10×10 SparseMatrixCSC{Float64,Int64} with 5 stored entries:
  [6 ,  1]  =  0.999066
  [4 ,  4]  =  0.45223
  [3 ,  5]  =  0.187222
  [6 ,  6]  =  0.87619
  [8 ,  9]  =  0.926811

julia> zero(x)
10×10 SparseMatrixCSC{Float64,Int64} with 5 stored entries:
  [6 ,  1]  =  0.0
  [4 ,  4]  =  0.0
  [3 ,  5]  =  0.0
  [6 ,  6]  =  0.0
  [8 ,  9]  =  0.0

julia> one(x)
10×10 SparseMatrixCSC{Float64,Int64} with 10 stored entries:
  [1 ,  1]  =  1.0
  [2 ,  2]  =  1.0
  [3 ,  3]  =  1.0
  [4 ,  4]  =  1.0
  [5 ,  5]  =  1.0
  [6 ,  6]  =  1.0
  [7 ,  7]  =  1.0
  [8 ,  8]  =  1.0
  [9 ,  9]  =  1.0
  [10, 10]  =  1.0

julia> UpperTriangular(x)
10×10 UpperTriangular{Float64,SparseMatrixCSC{Float64,Int64}}:
 0.0  0.0  0.0  0.0      0.0       0.0      0.0  0.0  0.0       0.0
  ⋅   0.0  0.0  0.0      0.0       0.0      0.0  0.0  0.0       0.0
  ⋅    ⋅   0.0  0.0      0.187222  0.0      0.0  0.0  0.0       0.0
  ⋅    ⋅    ⋅   0.45223  0.0       0.0      0.0  0.0  0.0       0.0
  ⋅    ⋅    ⋅    ⋅       0.0       0.0      0.0  0.0  0.0       0.0
  ⋅    ⋅    ⋅    ⋅        ⋅        0.87619  0.0  0.0  0.0       0.0
  ⋅    ⋅    ⋅    ⋅        ⋅         ⋅       0.0  0.0  0.0       0.0
  ⋅    ⋅    ⋅    ⋅        ⋅         ⋅        ⋅   0.0  0.926811  0.0
  ⋅    ⋅    ⋅    ⋅        ⋅         ⋅        ⋅    ⋅   0.0       0.0
  ⋅    ⋅    ⋅    ⋅        ⋅         ⋅        ⋅    ⋅    ⋅        0.0

julia> zero(UpperTriangular(x))
10×10 UpperTriangular{Float64,SparseMatrixCSC{Float64,Int64}}:
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
  ⋅   0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
  ⋅    ⋅   0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
  ⋅    ⋅    ⋅   0.0  0.0  0.0  0.0  0.0  0.0  0.0
  ⋅    ⋅    ⋅    ⋅   0.0  0.0  0.0  0.0  0.0  0.0
  ⋅    ⋅    ⋅    ⋅    ⋅   0.0  0.0  0.0  0.0  0.0
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅   0.0  0.0  0.0  0.0
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅   0.0  0.0  0.0
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅   0.0  0.0
  ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅    ⋅   0.0

@chriscoey
Copy link

This is great! Is it ready?

@mcognetta
Copy link
Contributor Author

Sorry, I've been afk working on my thesis. @stevengj approved it but I'd like to know the consensus on preserving container type in these sorts of operations.

@StefanKarpinski
Copy link
Member

Seems great to me! Unlike zeros, which has been deprecated, zero and one for matrices clearly fall into the functional, value-based class of linear algebra APIs and it's unusual to call them and then mutate the result. Given that, I can't see why we wouldn't want this.

@GiggleLiu
Copy link
Contributor

GiggleLiu commented Oct 30, 2018

Shouldn't zero and one always maintain type (#29839)?
Type preserving can make programs more predictable.

@StefanKarpinski
Copy link
Member

No, that's been claimed over there but it's far from concluded.

@StefanKarpinski
Copy link
Member

Also, maintaining the input type is what these definitions do... or were you just supporting the change?

@GiggleLiu
Copy link
Contributor

I have given an example in
#29839
This example shows how type non-preserving can affect program design. This is exactly the bug I encountered in Flux.

Please have a look. @StefanKarpinski

@StefanKarpinski
Copy link
Member

Yeah, I saw it, it's not really definitive.

@mcognetta
Copy link
Contributor Author

Sorry to ping, but I believe this is ready for a final review. @stevengj @fredrikekre @mschauer

@fredrikekre
Copy link
Member

LGTM, but are we allowed to merged such (technically) breaking changes? Will we then revert if it breaks some package?

@StefanKarpinski
Copy link
Member

Yes, mark it as “minor change”. We’ll run PkgEval before releasing 1.1.

@mcognetta
Copy link
Contributor Author

Bump on this one. The CI failure seems erroneous.

@stevengj
Copy link
Member

stevengj commented Feb 6, 2019

Yes, macOS travis failure seems to be a network glitch. (No point in restarting, since nowadays Travis macOS is failing for another reason.)

@stevengj stevengj merged commit fb9f1fd into JuliaLang:master Feb 6, 2019
@mcognetta mcognetta deleted the add_one_structured_matrix branch February 6, 2019 18:02
@StefanKarpinski
Copy link
Member

@mcognetta: would you mind making a PR to add a NEWS item about this?

@StefanKarpinski StefanKarpinski added the needs news A NEWS entry is required for this change label Feb 7, 2019
@mcognetta
Copy link
Contributor Author

Yes, I'll add it shortly.

mcognetta added a commit to mcognetta/julia that referenced this pull request Feb 7, 2019
ararslan pushed a commit that referenced this pull request Feb 7, 2019
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
linear algebra Linear algebra minor change Marginal behavior change acceptable for a minor release needs news A NEWS entry is required for this change
Projects
None yet
Development

Successfully merging this pull request may close these issues.

7 participants