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

Deprecate implicit obsdim keyword argument and improve tests #264

Merged
merged 9 commits into from
Feb 1, 2022
Merged

Deprecate implicit obsdim keyword argument and improve tests #264

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
98ca012
Deprecate implicit `obsdim` keyword argument and improve tests
devmotion Jan 31, 2022
59004c0
Fix format
devmotion Jan 31, 2022
6f27050
Remove kwargs
devmotion Jan 31, 2022
5706cbb
Fix typo
devmotion Jan 31, 2022
7db260f
Fix test
devmotion Jan 31, 2022
d40d73f
Fix format
devmotion Jan 31, 2022
50488b7
More general approach
devmotion Feb 1, 2022
47e5851
Merge branch 'master' into dw/obsdim
devmotion Feb 1, 2022
e79cfc8
Update Project.toml
devmotion Feb 1, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
11 changes: 5 additions & 6 deletions src/finite_gp_projection.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,12 +20,6 @@ function FiniteGP(f::AbstractGP, x::AbstractVector, σ²::Real=default_σ²)
return FiniteGP(f, x, Fill(σ², length(x)))
end

function FiniteGP(
f::AbstractGP, X::AbstractMatrix, σ²=default_σ²; obsdim::Int=KernelFunctions.defaultobs
)
return FiniteGP(f, KernelFunctions.vec_of_vecs(X; obsdim=obsdim), σ²)
end

## conversions
Base.convert(::Type{MvNormal}, f::FiniteGP) = MvNormal(mean_and_cov(f)...)
function Base.convert(::Type{MvNormal{T}}, f::FiniteGP) where {T}
Expand All @@ -36,6 +30,11 @@ end
Base.length(f::FiniteGP) = length(f.x)

(f::AbstractGP)(x...) = FiniteGP(f, x...)
Copy link
Member

Choose a reason for hiding this comment

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

Should it be

Suggested change
(f::AbstractGP)(x...) = FiniteGP(f, x...)
(f::AbstractGP)(x...; kwargs...) = FiniteGP(f, x...; kwargs...)

or is it done automatically?

Copy link
Member Author

Choose a reason for hiding this comment

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

No, it isn't done automatically (and hence one can't change obsdim with the official API - but this is not tested either...). However, since FiniteGP does not accept keyword arguments it is not really needed currently after applying the fixes and dispatches in this PR.

function (f::AbstractGP)(
X::AbstractMatrix, args...; obsdim::Union{Int,Nothing}=KernelFunctions.defaultobs
Copy link
Member

Choose a reason for hiding this comment

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

why the args... instead of sigma^2 here?

Copy link
Member Author

Choose a reason for hiding this comment

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

To capture all possible arguments (if someone decides to create some other internal FiniteGP constructors) and to not have to worry about default values of some arguments. Similar to what we do in KernelFunctions, with this PR we just forward everything to the vector case and handle arguments, default values, etc. there.

)
return f(KernelFunctions.vec_of_vecs(X; obsdim=obsdim), args...)
end

"""
mean(fx::FiniteGP)
Expand Down
83 changes: 47 additions & 36 deletions test/finite_gp_projection.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,8 @@ end
f = GP(sin, SqExponentialKernel())
x = randn(10)
Σy = generate_noise_matrix(Random.GLOBAL_RNG, 10)
fx = FiniteGP(f, x, Σy)
fx = f(x, Σy)
@test fx isa AbstractGPs.FiniteGP

dist = @inferred(convert(MvNormal, fx))
@test dist isa MvNormal{Float64}
Expand All @@ -28,12 +29,18 @@ end
σ² = 1e-3
Xmat = randn(rng, N, N′)
f = GP(sin, SqExponentialKernel())
fx, fx′ = FiniteGP(f, x, Σy), FiniteGP(f, x′, Σy′)
fx, fx′ = f(x, Σy), f(x′, Σy′)
@test fx isa AbstractGPs.FiniteGP
@test fx′ isa AbstractGPs.FiniteGP

for (x, obsdim) in ((RowVecs(Xmat), 1), (ColVecs(Xmat), 2))
@test f(x) isa AbstractGPs.FiniteGP
@test f(x, σ²) isa AbstractGPs.FiniteGP
@test f(Xmat; obsdim=obsdim) == f(x)
@test f(Xmat, σ²; obsdim=obsdim) == f(x, σ²)
end
@test @test_deprecated(f(Xmat)) == f(ColVecs(Xmat))

@test FiniteGP(f, Xmat; obsdim=1) == FiniteGP(f, RowVecs(Xmat))
@test FiniteGP(f, Xmat; obsdim=2) == FiniteGP(f, ColVecs(Xmat))
@test FiniteGP(f, Xmat, σ²; obsdim=1) == FiniteGP(f, RowVecs(Xmat), σ²)
@test FiniteGP(f, Xmat, σ²; obsdim=2) == FiniteGP(f, ColVecs(Xmat), σ²)
@test mean(fx) == mean(f, x)
@test cov(fx) == cov(f, x)
@test var(fx) == diag(cov(fx))
Expand All @@ -57,7 +64,8 @@ end
N = 10
x = randn(rng, N)
Σy = generate_noise_matrix(rng, N)
fx = FiniteGP(GP(1, SqExponentialKernel()), x, Σy)
fx = GP(1, SqExponentialKernel())(x, Σy)
@test fx isa AbstractGPs.FiniteGP

# Check that samples are the correct size.
@test length(rand(rng, fx)) == length(x)
Expand All @@ -79,7 +87,8 @@ end
m0 = 1
S = 100_000
x = range(-3.0, 3.0; length=N)
f = FiniteGP(GP(m0, SqExponentialKernel()), x, 1e-12)
f = GP(m0, SqExponentialKernel())(x, 1e-12)
@test f isa AbstractGPs.FiniteGP

# Check mean + covariance estimates approximately converge for single-GP sampling.
f̂1 = rand(rng, f, S)
Expand All @@ -100,15 +109,15 @@ end

# # Check that the gradient w.r.t. the samples is correct (single-sample).
# adjoint_test(
# x->rand(MersenneTwister(123456), FiniteGP(GP(sin, EQ(), GPC()), x, Σy)),
# x->rand(MersenneTwister(123456), GP(sin, EQ(), GPC())(x, Σy)),
# randn(rng, N),
# x;
# atol=1e-9, rtol=1e-9,
# )

# # Check that the gradient w.r.t. the samples is correct (multisample).
# adjoint_test(
# x->rand(MersenneTwister(123456), FiniteGP(GP(sin, EQ(), GPC()), x, Σy), S),
# x->rand(MersenneTwister(123456), GP(sin, EQ(), GPC())(x, Σy), S),
# randn(rng, N, S),
# x;
# atol=1e-9, rtol=1e-9,
Expand All @@ -125,7 +134,7 @@ end
# f = GP(sin, EQ(), GPC())
# Σy = _to_psd(A)
# C = cholesky(Σy)
# return tr_Cf_invΣy(FiniteGP(f, x, Σy), Σy, C)
# return tr_Cf_invΣy(f(x, Σy), Σy, C)
# end,
# randn(rng), x, A,
# )
Expand All @@ -138,7 +147,7 @@ end
# f = GP(sin, EQ(), GPC())
# Σy = Diagonal(exp.(a .+ 1))
# C = cholesky(Σy)
# return tr_Cf_invΣy(FiniteGP(f, x, Σy), Σy, C)
# return tr_Cf_invΣy(f(x, Σy), Σy, C)
# end,
# randn(rng), x, a,
# )
Expand All @@ -151,8 +160,10 @@ end
σ = 1e-1
x = collect(range(-3.0; stop=3.0, length=N))
f = GP(1, SqExponentialKernel())
fx = FiniteGP(f, x, 0)
y = FiniteGP(f, x, σ^2)
fx = f(x, 0)
@test fx isa AbstractGPs.FiniteGP
y = f(x, σ^2)
@test y isa AbstractGPs.FiniteGP
ŷ = rand(rng, y)

# Check that logpdf returns the correct type and roughly agrees with Distributions.
Expand All @@ -179,28 +190,28 @@ end
# # Check that gradient w.r.t. inputs is approximately correct for `f`.
# x, l̄ = randn(rng, N), randn(rng)
# adjoint_test(
# x->logpdf(FiniteGP(f, x, 1e-3), ones(size(x))),
# x->logpdf(f(x, 1e-3), ones(size(x))),
# l̄, collect(x);
# atol=1e-8, rtol=1e-8,
# )
# adjoint_test(
# x->sum(logpdf(FiniteGP(f, x, 1e-3), ones(size(Ŷ)))),
# x->sum(logpdf(f(x, 1e-3), ones(size(Ŷ)))),
# l̄, collect(x);
# atol=1e-8, rtol=1e-8,
# )

# # Check that the gradient w.r.t. the noise is approximately correct for `f`.
# σ_ = randn(rng)
# adjoint_test((σ_, ŷ)->logpdf(FiniteGP(f, x, exp(σ_)), ŷ), l̄, σ_, ŷ)
# adjoint_test((σ_, Ŷ)->sum(logpdf(FiniteGP(f, x, exp(σ_)), Ŷ)), l̄, σ_, Ŷ)
# adjoint_test((σ_, ŷ)->logpdf(f(x, exp(σ_)), ŷ), l̄, σ_, ŷ)
# adjoint_test((σ_, Ŷ)->sum(logpdf(f(x, exp(σ_)), Ŷ)), l̄, σ_, Ŷ)

# # Check that the gradient w.r.t. a scaling of the GP works.
# adjoint_test(
# α->logpdf(FiniteGP(α * f, x, 1e-1), ŷ), l̄, randn(rng);
# α->logpdf((α * f)(x, 1e-1), ŷ), l̄, randn(rng);
# atol=1e-8, rtol=1e-8,
# )
# adjoint_test(
# α->sum(logpdf(FiniteGP(α * f, x, 1e-1), Ŷ)), l̄, randn(rng);
# α->sum(logpdf((α * f)(x, 1e-1), Ŷ)), l̄, randn(rng);
# atol=1e-8, rtol=1e-8,
# )
end
Expand Down Expand Up @@ -235,7 +246,7 @@ end
end

@testset "Docs" begin
docstring = string(Docs.doc(logpdf, Tuple{FiniteGP,Vector{Float64}}))
docstring = string(Docs.doc(logpdf, Tuple{AbstractGPs.FiniteGP,Vector{Float64}}))
@test occursin("logpdf(f::FiniteGP, y::AbstractVecOrMat{<:Real})", docstring)
end

Expand All @@ -257,35 +268,35 @@ end
# # Test gradient w.r.t. random sampling.
# N = length(x)
# adjoint_test(
# (x, isp_σ)->rand(_rng(), FiniteGP(f, x, exp(isp_σ)^2)),
# (x, isp_σ)->rand(_rng(), f(x, exp(isp_σ)^2)),
# randn(rng, N),
# x,
# isp_σ,;
# atol=atol, rtol=rtol,
# )
# adjoint_test(
# (x, isp_σ)->rand(_rng(), FiniteGP(f, x, exp(isp_σ)^2), 11),
# (x, isp_σ)->rand(_rng(), f(x, exp(isp_σ)^2), 11),
# randn(rng, N, 11),
# x,
# isp_σ,;
# atol=atol, rtol=rtol,
# )

# # Check that gradient w.r.t. logpdf is correct.
# y, l̄ = rand(rng, FiniteGP(f, x, exp(isp_σ))), randn(rng)
# y, l̄ = rand(rng, f(x, exp(isp_σ))), randn(rng)
# adjoint_test(
# (x, isp_σ, y)->logpdf(FiniteGP(f, x, exp(isp_σ)), y),
# (x, isp_σ, y)->logpdf(f(x, exp(isp_σ)), y),
# l̄, x, isp_σ, y;
# atol=atol, rtol=rtol,
# )

# # Check that elbo is tight-ish when it's meant to be.
# fx, yx = FiniteGP(f, x, 1e-9), FiniteGP(f, x, exp(isp_σ))
# fx, yx = f(x, 1e-9), f(x, exp(isp_σ))
# @test isapprox(elbo(yx, y, fx), logpdf(yx, y); atol=1e-6, rtol=1e-6)

# # Check that gradient w.r.t. elbo is correct.
# adjoint_test(
# (x, ŷ, isp_σ)->elbo(FiniteGP(f, x, exp(isp_σ)), ŷ, FiniteGP(f, x, 1e-9)),
# (x, ŷ, isp_σ)->elbo(f(x, exp(isp_σ)), ŷ, f(x, 1e-9)),
# randn(rng), x, y, isp_σ;
# atol=1e-6, rtol=1e-6,
# )
Expand Down Expand Up @@ -333,31 +344,31 @@ end
# Smat = Matrix(S)

# f = GP(cos, EQ(), GPC())
# y = rand(FiniteGP(f, x, S))
# y = rand(f(x, S))

# @test logpdf(FiniteGP(f, x, S), y) ≈ logpdf(FiniteGP(f, x, Smat), y)
# @test logpdf(f(x, S), y) ≈ logpdf(f(x, Smat), y)
# adjoint_test(
# (x, S, y)->logpdf(FiniteGP(f, x, S), y), randn(rng), x, Smat, y;
# (x, S, y)->logpdf(f(x, S), y), randn(rng), x, Smat, y;
# atol=1e-6, rtol=1e-6,
# )
# adjoint_test(
# (x, A1, A2, y)->logpdf(FiniteGP(f, x, block_diagonal([A1 * A1' + I, A2 * A2' + I])), y),
# (x, A1, A2, y)->logpdf(f(x, block_diagonal([A1 * A1' + I, A2 * A2' + I])), y),
# randn(rng), x, As[1], As[2], y;
# atol=1e-6, rtol=1e-6
# )

# @test elbo(FiniteGP(f, x, Smat), y, FiniteGP(f, x)) ≈ logpdf(FiniteGP(f, x, Smat), y)
# @test elbo(FiniteGP(f, x, S), y, FiniteGP(f, x)) ≈
# elbo(FiniteGP(f, x, Smat), y, FiniteGP(f, x))
# @test elbo(f(x, Smat), y, f(x)) ≈ logpdf(f(x, Smat), y)
# @test elbo(f(x, S), y, f(x)) ≈
# elbo(f(x, Smat), y, f(x))
# adjoint_test(
# (x, A, y)->elbo(FiniteGP(f, x, _to_psd(A)), y, FiniteGP(f, x)),
# (x, A, y)->elbo(f(x, _to_psd(A)), y, f(x)),
# randn(rng), x, randn(rng, sum(Ns), sum(Ns)), y;
# atol=1e-6, rtol=1e-6,
# )
# adjoint_test(
# (x, A1, A2, y) -> begin
# S = block_diagonal([A1 * A1' + I, A2 * A2' + I])
# return elbo(FiniteGP(f, x, S), y, FiniteGP(f, x))
# return elbo(f(x, S), y, f(x))
# end,
# randn(rng), x, As[1], As[2], y;
# atol=1e-6, rtol=1e-6,
Expand Down
1 change: 0 additions & 1 deletion test/runtests.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,6 @@ using AbstractGPs
using AbstractGPs:
AbstractGP,
MeanFunction,
FiniteGP,
ConstMean,
GP,
ZeroMean,
Expand Down