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.

Sign up for GitHub

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

Jump to bottom

make to_vec(::Integer) an empty vector #189

Open
wants to merge 6 commits into
base: main
Choose a base branch
from
Open

make to_vec(::Integer) an empty vector #189

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
826ac23
add to_vec(::Integer) implementation
Aug 4, 2021
649779f
fix grad y\bar
Aug 4, 2021
cf30781
add Int64 tests
Aug 4, 2021
212f37e
make sure int become zerotangents
Aug 5, 2021
0152019
bump version
Aug 5, 2021
5993c0b
_ the y
Aug 5, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion Project.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
name = "FiniteDifferences"
uuid = "26cc04aa-876d-5657-8c51-4c34ba976000"
version = "0.12.18"
version = "0.12.19"
Copy link
Member Author

Choose a reason for hiding this comment

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

should this be breaking instead? i think of it as a bug fix

Copy link
Member

Choose a reason for hiding this comment

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

My view is treat it as a bug fix for now -- if it turns out to be super breaking and we later decide it's actually a breaking change, we can always bump the minor version number then.


[deps]
ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
Expand Down
16 changes: 13 additions & 3 deletions src/grad.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,7 +55,7 @@ is defined. Each 2-`Tuple` in `xẋs` contains the value `x` and its tangent `x
function jvp(fdm, f, (x, ẋ)::Tuple{Any, Any})
x_vec, vec_to_x = to_vec(x)
_, vec_to_y = to_vec(f(x))
return vec_to_y(_jvp(fdm, x_vec->to_vec(f(vec_to_x(x_vec)))[1], x_vec, to_vec(ẋ)[1]))
return _int2zero(vec_to_y(_jvp(fdm, x_vec->to_vec(f(vec_to_x(x_vec)))[1], x_vec, to_vec(ẋ)[1])))
end
function jvp(fdm, f, xẋs::Tuple{Any, Any}...)
x, ẋ = collect(zip(xẋs...))
Expand All @@ -70,7 +70,7 @@ Compute an adjoint with any types of arguments `x` for which [`to_vec`](@ref) is
function j′vp(fdm, f, ȳ, x)
x_vec, vec_to_x = to_vec(x)
ȳ_vec, _ = to_vec(ȳ)
return (vec_to_x(_j′vp(fdm, first ∘ to_vec ∘ f ∘ vec_to_x, ȳ_vec, x_vec)), )
return (_int2zero(vec_to_x(_j′vp(fdm, first ∘ to_vec ∘ f ∘ vec_to_x, ȳ_vec, x_vec))), )
end

j′vp(fdm, f, ȳ, xs...) = j′vp(fdm, xs->f(xs...), ȳ, xs)[1]
Expand All @@ -85,4 +85,14 @@ end

Compute the gradient of `f` for any `xs` for which [`to_vec`](@ref) is defined.
"""
grad(fdm, f, xs...) = j′vp(fdm, f, 1, xs...) # `j′vp` with seed of 1
grad(fdm, f, xs...) = j′vp(fdm, f, 1.0, xs...) # `j′vp` with seed of 1

# This deals with the fact that integers are non perturbable
# v, b = to_vec(1);
# v == []
# b(v) == 1
# which means that jvp always returns the integer itself, since [] - [] == []
_int2zero(x) = x
Copy link
Member

Choose a reason for hiding this comment

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

I suspect that this will miss a few cases (e.g. vectors of integers), but probably that's fine for now -- we can revisit later it needs be.

Copy link
Member

Choose a reason for hiding this comment

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

On a more general note: is the need for this function a manifestation of the fact that FiniteDifferences doesn't really know how to handle tangents properly?

_int2zero(x::Tuple) = map(_int2zero, x)
_int2zero(x::NamedTuple) = map(_int2zero, x)
_int2zero(::Integer) = ZeroTangent()
5 changes: 5 additions & 0 deletions src/to_vec.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,11 @@ function to_vec(z::Complex)
return [real(z), imag(z)], Complex_from_vec
end

function to_vec(x::Integer)
Integer_from_vec(v) = x
return Bool[], Integer_from_vec
end

# Base case -- if x is already a Vector{<:Real} there's no conversion necessary.
to_vec(x::Vector{<:Real}) = (x, identity)

Expand Down
65 changes: 34 additions & 31 deletions test/to_vec.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,8 +67,11 @@ function test_to_vec(x::T; check_inferred=true) where {T}
return nothing
end

myrandn(T::Type{<:Number}, args...) = randn(T, args...)
Copy link
Member Author

Choose a reason for hiding this comment

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

i should probably rename this to something sensible.

robust_randn, randn_ints_ok, randn_that_doesnt_hate_ints

Copy link
Member

Choose a reason for hiding this comment

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

I'm not sure this should have randn in the name, since the integer method doesn't really sample from a normal distribution. Perhaps something a bit more verbose like rand_number?

myrandn(T::Type{<:Integer}, args...) = rand(T[-1, 0, 2], args...)

@testset "to_vec" begin
@testset "$T" for T in (Float32, ComplexF32, Float64, ComplexF64)
@testset "$T" for T in (Int64, Float32, ComplexF32, Float64, ComplexF64)
if T == Float64
test_to_vec(1.0)
test_to_vec(1)
Expand All @@ -79,52 +82,52 @@ end
test_to_vec(T[])
test_to_vec(Vector{T}[])
test_to_vec(Matrix{T}[])
test_to_vec(randn(T, 3))
test_to_vec(randn(T, 5, 11))
test_to_vec(randn(T, 13, 17, 19))
test_to_vec(randn(T, 13, 0, 19))
test_to_vec([1.0, randn(T, 2), randn(T, 1), 2.0]; check_inferred=false)
test_to_vec([randn(T, 5, 4, 3), (5, 4, 3), 2.0]; check_inferred=false)
test_to_vec(reshape([1.0, randn(T, 5, 4, 3), randn(T, 4, 3), 2.0], 2, 2); check_inferred=false)
test_to_vec(UpperTriangular(randn(T, 13, 13)))
test_to_vec(Diagonal(randn(T, 7)))
test_to_vec(DummyType(randn(T, 2, 9)))
test_to_vec(myrandn(T, 3))
test_to_vec(myrandn(T, 5, 11))
test_to_vec(myrandn(T, 13, 17, 19))
test_to_vec(myrandn(T, 13, 0, 19))
test_to_vec([1.0, myrandn(T, 2), myrandn(T, 1), 2.0]; check_inferred=false)
test_to_vec([myrandn(T, 5, 4, 3), (5, 4, 3), 2.0]; check_inferred=false)
test_to_vec(reshape([1.0, myrandn(T, 5, 4, 3), myrandn(T, 4, 3), 2.0], 2, 2); check_inferred=false)
test_to_vec(UpperTriangular(myrandn(T, 13, 13)))
test_to_vec(Diagonal(myrandn(T, 7)))
test_to_vec(DummyType(myrandn(T, 2, 9)))
test_to_vec(SVector{2, T}(1.0, 2.0); check_inferred=false)
test_to_vec(SMatrix{2, 2, T}(1.0, 2.0, 3.0, 4.0); check_inferred=false)
test_to_vec(@view randn(T, 10)[1:4]) # SubArray -- Vector
test_to_vec(@view randn(T, 10, 2)[1:4, :]) # SubArray -- Matrix
test_to_vec(@view myrandn(T, 10)[1:4]) # SubArray -- Vector
test_to_vec(@view myrandn(T, 10, 2)[1:4, :]) # SubArray -- Matrix
test_to_vec(Base.ReshapedArray(rand(T, 3, 3), (9,), ()))

@testset "$Op" for Op in (Symmetric, Hermitian)
test_to_vec(Op(randn(T, 11, 11)))
test_to_vec(Op(myrandn(T, 11, 11)))
@testset "$uplo" for uplo in (:L, :U)
A = Op(randn(T, 11, 11), uplo)
A = Op(myrandn(T, 11, 11), uplo)
test_to_vec(A)
x_vec, back = to_vec(A)
@test back(x_vec).uplo == A.uplo
end
end

@testset "$Op" for Op in (Adjoint, Transpose)
test_to_vec(Op(randn(T, 4, 4)))
test_to_vec(Op(randn(T, 6)))
test_to_vec(Op(randn(T, 2, 5)))
test_to_vec(Op(myrandn(T, 4, 4)))
test_to_vec(Op(myrandn(T, 6)))
test_to_vec(Op(myrandn(T, 2, 5)))

# Ensure that if an `AbstractVector` is `Adjoint`ed, then the reconstructed
# version also contains an `AbstractVector`, rather than an `AbstractMatrix`
# whose 2nd dimension is of size 1.
@testset "Vector" begin
x_vec, back = to_vec(Op(randn(T, 5)))
x_vec, back = to_vec(Op(myrandn(T, 5)))
@test parent(back(x_vec)) isa AbstractVector
end
end

@testset "PermutedDimsArray" begin
test_to_vec(PermutedDimsArray(randn(T, 3, 1), (2, 1)))
test_to_vec(PermutedDimsArray(randn(T, 4, 2, 3), (3, 1, 2)))
test_to_vec(PermutedDimsArray(myrandn(T, 3, 1), (2, 1)))
test_to_vec(PermutedDimsArray(myrandn(T, 4, 2, 3), (3, 1, 2)))
test_to_vec(
PermutedDimsArray(
[randn(T, 3) for _ in 1:3, _ in 1:2, _ in 1:4], (2, 1, 3),
[myrandn(T, 3) for _ in 1:3, _ in 1:2, _ in 1:4], (2, 1, 3),
),
)
end
Expand All @@ -133,7 +136,7 @@ end
# (100, 100) is needed to test for the NaNs that can appear in the
# qr(M).T matrix
for dims in [(7, 3), (100, 100)]
M = randn(T, dims...)
M = myrandn(T, dims...)
P = M * M' + I # Positive definite matrix
test_to_vec(svd(M))
test_to_vec(cholesky(P))
Expand Down Expand Up @@ -172,25 +175,25 @@ end

@testset "Tuples" begin
test_to_vec((5, 4))
test_to_vec((5, randn(T, 5)); check_inferred = VERSION ≥ v"1.2") # broken on Julia 1.6.0, fixed on 1.6.1
test_to_vec((randn(T, 4), randn(T, 4, 3, 2), 1); check_inferred=false)
test_to_vec((5, randn(T, 4, 3, 2), UpperTriangular(randn(T, 4, 4)), 2.5); check_inferred = VERSION ≥ v"1.2") # broken on Julia 1.6.0, fixed on 1.6.1
test_to_vec(((6, 5), 3, randn(T, 3, 2, 0, 1)); check_inferred=false)
test_to_vec((DummyType(randn(T, 2, 7)), DummyType(randn(T, 3, 9))))
test_to_vec((DummyType(randn(T, 3, 2)), randn(T, 11, 8)))
test_to_vec((5, myrandn(T, 5)); check_inferred = VERSION ≥ v"1.2") # broken on Julia 1.6.0, fixed on 1.6.1
test_to_vec((myrandn(T, 4), myrandn(T, 4, 3, 2), 1); check_inferred=false)
test_to_vec((5, myrandn(T, 4, 3, 2), UpperTriangular(myrandn(T, 4, 4)), 2.5); check_inferred = VERSION ≥ v"1.2") # broken on Julia 1.6.0, fixed on 1.6.1
test_to_vec(((6, 5), 3, myrandn(T, 3, 2, 0, 1)); check_inferred=false)
test_to_vec((DummyType(myrandn(T, 2, 7)), DummyType(myrandn(T, 3, 9))))
test_to_vec((DummyType(myrandn(T, 3, 2)), myrandn(T, 11, 8)))
end
@testset "NamedTuple" begin
if T == Float64
test_to_vec((a=5, b=randn(10, 11), c=(5, 4, 3)); check_inferred = VERSION ≥ v"1.2")
else
test_to_vec((a=3 + 2im, b=randn(T, 10, 11), c=(5+im, 2-im, 1+im)); check_inferred = VERSION ≥ v"1.2")
test_to_vec((a=3 + 2im, b=myrandn(T, 10, 11), c=(5+im, 2-im, 1+im)); check_inferred = VERSION ≥ v"1.2")
end
end
@testset "Dictionary" begin
if T == Float64
test_to_vec(Dict(:a=>5, :b=>randn(10, 11), :c=>(5, 4, 3)); check_inferred=false)
else
test_to_vec(Dict(:a=>3 + 2im, :b=>randn(T, 10, 11), :c=>(5+im, 2-im, 1+im)); check_inferred=false)
test_to_vec(Dict(:a=>3 + 2im, :b=>myrandn(T, 10, 11), :c=>(5+im, 2-im, 1+im)); check_inferred=false)
end
end
end
Expand Down