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

reactivate AD tests: FiniteGP #314

Merged
merged 14 commits into from
Apr 7, 2022
Merged

reactivate AD tests: FiniteGP #314

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
7a102e4
uncomment tests and see what happens
st-- Apr 6, 2022
1c3366b
Apply suggestions from code review
st-- Apr 6, 2022
669c985
EQ -> SqExponentialKernel
st-- Apr 7, 2022
0042246
Merge branch 'st/ad_tests2' of github.com:JuliaGaussianProcesses/Abst…
st-- Apr 7, 2022
b9dc7d4
move tr_Cf_invΣy test
st-- Apr 7, 2022
53f705c
add dense test case
st-- Apr 7, 2022
42ec86c
Update test/sparse_approximations.jl
st-- Apr 7, 2022
6e6e35b
Update test/sparse_approximations.jl
st-- Apr 7, 2022
4000aa6
scaling of GPs not implemented in AbstractGPs
st-- Apr 7, 2022
0de8fa6
comment out broken tests
st-- Apr 7, 2022
0cf0b6b
Merge branch 'st/ad_tests2' of github.com:JuliaGaussianProcesses/Abst…
st-- Apr 7, 2022
5eebf06
remove unneeded/broken tests
st-- Apr 7, 2022
c823efd
add comment
st-- Apr 7, 2022
d479359
Merge branch 'master' into st/ad_tests2
st-- Apr 7, 2022
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
146 changes: 58 additions & 88 deletions test/finite_gp_projection.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,57 +102,29 @@ end
@test mean(abs.(Σ′ - cov(f))) < 1e-2
end
end
# @testset "rand (gradients)" begin
# rng, N, S = MersenneTwister(123456), 10, 3
# x = collect(range(-3.0, stop=3.0, length=N))
# Σy = 1e-12

# # Check that the gradient w.r.t. the samples is correct (single-sample).
# adjoint_test(
# 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), GP(sin, EQ(), GPC())(x, Σy), S),
# randn(rng, N, S),
# x;
# atol=1e-9, rtol=1e-9,
# )
# end
# @testset "tr_Cf_invΣy" begin
# N = 11
# x = collect(range(-3.0, 3.0; length=N))
# @testset "dense" begin
# rng = MersenneTwister(123456)
# A = randn(rng, N, N - 2)
# adjoint_test(
# (x, A)->begin
# f = GP(sin, EQ(), GPC())
# Σy = _to_psd(A)
# C = cholesky(Σy)
# return tr_Cf_invΣy(f(x, Σy), Σy, C)
# end,
# randn(rng), x, A,
# )
# end
# @testset "Diagonal" begin
# rng = MersenneTwister(123456)
# a = 0.01 .* randn(rng, N)
# adjoint_test(
# (x, a)->begin
# f = GP(sin, EQ(), GPC())
# Σy = Diagonal(exp.(a .+ 1))
# C = cholesky(Σy)
# return tr_Cf_invΣy(f(x, Σy), Σy, C)
# end,
# randn(rng), x, a,
# )
# end
# end
@testset "rand (gradients)" begin
rng, N, S = MersenneTwister(123456), 10, 3
x = collect(range(-3.0; stop=3.0, length=N))
Σy = 1e-12

# Check that the gradient w.r.t. the samples is correct (single-sample).
adjoint_test(
x -> rand(MersenneTwister(123456), GP(sin, SqExponentialKernel())(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), GP(sin, SqExponentialKernel())(x, Σy), S),
randn(rng, N, S),
x;
atol=1e-9,
rtol=1e-9,
)
end
@testset "logpdf / loglikelihood" begin
rng = MersenneTwister(123456)
N = 10
Expand All @@ -177,43 +149,41 @@ end
@test logpdf(y, Ŷ) ≈ [logpdf(y, Ŷ[:, n]) for n in 1:S]
@test loglikelihood(y, Ŷ) == sum(logpdf(y, Ŷ))

# # Check gradient of logpdf at mean is zero for `f`.
# adjoint_test(ŷ->logpdf(fx, ŷ), 1, ones(size(ŷ)))
# lp, back = Zygote.pullback(ŷ->logpdf(fx, ŷ), ones(size(ŷ)))
# @test back(randn(rng))[1] == zeros(size(ŷ))

# # Check that gradient of logpdf at mean is zero for `y`.
# adjoint_test(ŷ->logpdf(y, ŷ), 1, ones(size(ŷ)))
# lp, back = Zygote.pullback(ŷ->logpdf(y, ŷ), ones(size(ŷ)))
# @test back(randn(rng))[1] == zeros(size(ŷ))

# # Check that gradient w.r.t. inputs is approximately correct for `f`.
# x, l̄ = randn(rng, N), randn(rng)
# adjoint_test(
# x->logpdf(f(x, 1e-3), ones(size(x))),
# l̄, collect(x);
# atol=1e-8, rtol=1e-8,
# )
# adjoint_test(
# 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(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((α * f)(x, 1e-1), ŷ), l̄, randn(rng);
# atol=1e-8, rtol=1e-8,
# )
# adjoint_test(
# α->sum(logpdf((α * f)(x, 1e-1), Ŷ)), l̄, randn(rng);
# atol=1e-8, rtol=1e-8,
# )
# Check gradient of logpdf at mean is zero for `f`.
adjoint_test(ŷ -> logpdf(fx, ŷ), 1, ones(size(ŷ)))
lp, back = Zygote.pullback(ŷ -> logpdf(fx, ŷ), ones(size(ŷ)))
@test back(randn(rng))[1] == zeros(size(ŷ))

# Check that gradient of logpdf at mean is zero for `y`.
adjoint_test(ŷ -> logpdf(y, ŷ), 1, ones(size(ŷ)))
lp, back = Zygote.pullback(ŷ -> logpdf(y, ŷ), ones(size(ŷ)))
@test back(randn(rng))[1] == zeros(size(ŷ))

# Check that gradient w.r.t. inputs is approximately correct for `f`.
x, l̄ = randn(rng, N), randn(rng)
adjoint_test(
x -> logpdf(f(x, 1e-3), ones(size(x))), l̄, collect(x); atol=1e-8, rtol=1e-8
)
adjoint_test(
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(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.
st-- marked this conversation as resolved.
Show resolved Hide resolved
adjoint_test(
α -> logpdf((α * f)(x, 1e-1), ŷ), l̄, randn(rng); atol=1e-8, rtol=1e-8
)
adjoint_test(
α -> sum(logpdf((α * f)(x, 1e-1), Ŷ)), l̄, randn(rng); atol=1e-8, rtol=1e-8
)
end
@testset "Type Stability - $T" for T in [Float64, Float32]
rng = MersenneTwister(123456)
Expand Down
32 changes: 31 additions & 1 deletion test/sparse_approximations.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -122,11 +122,41 @@ end
σ² = 1.234
Σy_Diagonal = Diagonal(Fill(σ², N))
Σy_ScalMat = ScalMat(N, σ²)
Σy_dense = _to_psd(randn(N, N - 2))
f = GP(SqExponentialKernel())

for Σy in (Σy_Diagonal, Σy_ScalMat)
for Σy in (Σy_Diagonal, Σy_ScalMat, Σy_dense)
fx = f(x, Σy)
Cf = cov(f, x)
@test AbstractGPs.tr_Cf_invΣy(fx, Σy) ≈ tr(Cf / Matrix(Σy))
end

@testset "adjoint_test" begin
N = 11
x = collect(range(-3.0, 3.0; length=N))
@testset "dense" begin
rng = MersenneTwister(123456)
A = randn(rng, N, N - 2)
adjoint_test(
(x, A) -> begin
f = GP(sin, SqExponentialKernel())
Σy = _to_psd(A)
C = cholesky(Σy)
return AbstractGPs.tr_Cf_invΣy(f(x, Σy), Σy, C)
end, randn(rng), x, A
st-- marked this conversation as resolved.
Show resolved Hide resolved
)
end
@testset "Diagonal" begin
rng = MersenneTwister(123456)
a = 0.01 .* randn(rng, N)
adjoint_test(
(x, a) -> begin
f = GP(sin, SqExponentialKernel())
Σy = Diagonal(exp.(a .+ 1))
C = cholesky(Σy)
return AbstractGPs.tr_Cf_invΣy(f(x, Σy), Σy, C)
end, randn(rng), x, a
st-- marked this conversation as resolved.
Show resolved Hide resolved
)
end
end
end