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

Zygote AD failure workarounds & test cleanup #414

Merged
merged 10 commits into from
Dec 18, 2021
Merged

Zygote AD failure workarounds & test cleanup #414

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
7ccfbef
revert test_utils workaround for broken Zygote
st-- Dec 9, 2021
313535e
finer-grained testsets
st-- Dec 16, 2021
11b2bcd
disable broken Zygote AD test for ChainTransform
st-- Dec 18, 2021
feb5b5c
Apply suggestions from code review
st-- Dec 18, 2021
919e445
add missing test cases to test_AD
st-- Dec 18, 2021
7d38977
replace test_FiniteDiff with test_AD(..., :FiniteDiff, ...)
st-- Dec 18, 2021
a4ad008
more explicit test broken
st-- Dec 18, 2021
51c9250
Merge branch 'st/zygote_again' of github.com:JuliaGaussianProcesses/K…
st-- Dec 18, 2021
821503c
remove more duplication
st-- Dec 18, 2021
8b1cddc
Update test/test_utils.jl
st-- Dec 18, 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
176 changes: 67 additions & 109 deletions test/test_utils.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,6 +69,10 @@ function gradient(f, ::Val{:FiniteDiff}, args)
return first(FiniteDifferences.grad(FDM, f, args))
end

function compare_gradient(f, ::Val{:FiniteDiff}, args)
@test_nowarn gradient(f, :FiniteDiff, args)
end

function compare_gradient(f, AD::Symbol, args)
grad_AD = gradient(f, AD, args)
grad_FD = gradient(f, :FiniteDiff, args)
Expand All @@ -88,7 +92,7 @@ testdiagfunction(k::MOKernel, A, B) = sum(kernelmatrix_diag(k, A, B))
function test_ADs(
kernelfunction, args=nothing; ADs=[:Zygote, :ForwardDiff, :ReverseDiff], dims=[3, 3]
)
test_fd = test_FiniteDiff(kernelfunction, args, dims)
test_fd = test_AD(:FiniteDiff, kernelfunction, args, dims)
if !test_fd.anynonpass
for AD in ADs
test_AD(AD, kernelfunction, args, dims)
Expand All @@ -100,7 +104,7 @@ function check_zygote_type_stability(f, args...; ctx=Zygote.Context())
@inferred f(args...)
@inferred Zygote._pullback(ctx, f, args...)
out, pb = Zygote._pullback(ctx, f, args...)
@test_throws ErrorException @inferred pb(out)
@inferred pb(out)
end

function test_ADs(
Expand All @@ -114,70 +118,6 @@ function test_ADs(
end
end

function test_FiniteDiff(kernelfunction, args=nothing, dims=[3, 3])
Copy link
Member Author

Choose a reason for hiding this comment

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

This was doing exactly the same as test_AD (except for testdiagfunction(kernelfunction(p), A, dim) vs testdiagfunction(kernelfunction(p), A, B, dim), which I've now added to test_AD). The only difference was calling compare_gradient vs the @test_warn, so I simply added a compare_gradient(f, ::Val{:FiniteDiff}, args) so we can avoid all this code duplication.

Copy link
Member

Choose a reason for hiding this comment

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

Very nice.

# Init arguments :
k = if args === nothing
kernelfunction()
else
kernelfunction(args)
end
rng = MersenneTwister(42)
@testset "FiniteDifferences" begin
if k isa SimpleKernel
for d in log.([eps(), rand(rng)])
@test_nowarn gradient(:FiniteDiff, [d]) do x
kappa(k, exp(first(x)))
end
end
end
## Testing Kernel Functions
x = rand(rng, dims[1])
y = rand(rng, dims[1])
@test_nowarn gradient(:FiniteDiff, x) do x
k(x, y)
end
if !(args === nothing)
@test_nowarn gradient(:FiniteDiff, args) do p
kernelfunction(p)(x, y)
end
end
## Testing Kernel Matrices
A = rand(rng, dims...)
B = rand(rng, dims...)
for dim in 1:2
@test_nowarn gradient(:FiniteDiff, A) do a
testfunction(k, a, dim)
end
@test_nowarn gradient(:FiniteDiff, A) do a
testfunction(k, a, B, dim)
end
@test_nowarn gradient(:FiniteDiff, B) do b
testfunction(k, A, b, dim)
end
if !(args === nothing)
@test_nowarn gradient(:FiniteDiff, args) do p
testfunction(kernelfunction(p), A, B, dim)
end
end

@test_nowarn gradient(:FiniteDiff, A) do a
testdiagfunction(k, a, dim)
end
@test_nowarn gradient(:FiniteDiff, A) do a
testdiagfunction(k, a, B, dim)
end
@test_nowarn gradient(:FiniteDiff, B) do b
testdiagfunction(k, A, b, dim)
end
if args !== nothing
@test_nowarn gradient(:FiniteDiff, args) do p
testdiagfunction(kernelfunction(p), A, B, dim)
end
end
end
end
end

function test_FiniteDiff(k::MOKernel, dims=(in=3, out=2, obs=3))
rng = MersenneTwister(42)
@testset "FiniteDifferences" begin
Expand Down Expand Up @@ -224,65 +164,83 @@ end

function test_AD(AD::Symbol, kernelfunction, args=nothing, dims=[3, 3])
@testset "$(AD)" begin
# Test kappa function
k = if args === nothing
kernelfunction()
else
kernelfunction(args)
end
rng = MersenneTwister(42)

if k isa SimpleKernel
for d in log.([eps(), rand(rng)])
compare_gradient(AD, [d]) do x
kappa(k, exp(x[1]))
@testset "kappa function" begin
for d in log.([eps(), rand(rng)])
compare_gradient(AD, [d]) do x
kappa(k, exp(x[1]))
end
end
end
end
# Testing kernel evaluations
x = rand(rng, dims[1])
y = rand(rng, dims[1])
compare_gradient(AD, x) do x
k(x, y)
end
compare_gradient(AD, y) do y
k(x, y)
end
if !(args === nothing)
compare_gradient(AD, args) do p
kernelfunction(p)(x, y)
end
end
# Testing kernel matrices
A = rand(rng, dims...)
B = rand(rng, dims...)
for dim in 1:2
compare_gradient(AD, A) do a
testfunction(k, a, dim)
end
compare_gradient(AD, A) do a
testfunction(k, a, B, dim)

@testset "kernel evaluations" begin
x = rand(rng, dims[1])
y = rand(rng, dims[1])
@testset "first argument" begin
compare_gradient(AD, x) do x
k(x, y)
end
end
compare_gradient(AD, B) do b
testfunction(k, A, b, dim)
@testset "second argument" begin
compare_gradient(AD, y) do y
k(x, y)
end
end
if !(args === nothing)
compare_gradient(AD, args) do p
testfunction(kernelfunction(p), A, dim)
@testset "hyperparameters" begin
compare_gradient(AD, args) do p
kernelfunction(p)(x, y)
end
end
end
end

compare_gradient(AD, A) do a
testdiagfunction(k, a, dim)
end
compare_gradient(AD, A) do a
testdiagfunction(k, a, B, dim)
end
compare_gradient(AD, B) do b
testdiagfunction(k, A, b, dim)
end
if args !== nothing
compare_gradient(AD, args) do p
testdiagfunction(kernelfunction(p), A, dim)
Comment on lines -274 to -285
Copy link
Member Author

Choose a reason for hiding this comment

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

this was doing exactly the same for testdiagfunction as the code above for testfunction, so I've unified it with a for loop over the two functions.

@testset "kernel matrices" begin
A = rand(rng, dims...)
B = rand(rng, dims...)
for dim in 1:2
compare_gradient(AD, A) do a
testfunction(k, a, dim)
end
compare_gradient(AD, A) do a
testfunction(k, a, B, dim)
end
compare_gradient(AD, B) do b
testfunction(k, A, b, dim)
end
if !(args === nothing)
compare_gradient(AD, args) do p
testfunction(kernelfunction(p), A, dim)
end
compare_gradient(AD, args) do p
testfunction(kernelfunction(p), A, B, dim)
end
end

compare_gradient(AD, A) do a
testdiagfunction(k, a, dim)
end
compare_gradient(AD, A) do a
testdiagfunction(k, a, B, dim)
end
compare_gradient(AD, B) do b
testdiagfunction(k, A, b, dim)
end
if args !== nothing
compare_gradient(AD, args) do p
testdiagfunction(kernelfunction(p), A, dim)
end
compare_gradient(AD, args) do p
testdiagfunction(kernelfunction(p), A, B, dim)
end
end
end
end
Copy link
Member Author

Choose a reason for hiding this comment

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

@rossviljoen doing the changes above I spotted that you overloaded test_AD() for MOKernel - this worries me: it seems like all the "fallback" AD tests above are not run. It also seems to suggest that we can't substitute an MOKernel where we need an arbitrary Kernel (which would I believe violate Liskov substitution principle). Could you comment pls? thanks:)

Copy link
Contributor

Choose a reason for hiding this comment

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

I don't think that was me?

Copy link
Member Author

Choose a reason for hiding this comment

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

Ah, apologies! Now I'm not even sure anymore why I assumed it was you 😔 @willtebbutt / @thomasgudjonwright ?

Copy link
Member

Choose a reason for hiding this comment

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

Hmm don't think it was me. A lot of this code was @theogf I think?

Copy link
Member

Choose a reason for hiding this comment

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

I did the original compare gradient but someone else extended it to MOKernel

Copy link
Member

Choose a reason for hiding this comment

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

Wasn't me! Could always look back at the commit history :)

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've checked git blame in the first place... it was #263 (@david-vicente).

This to me seems like a case where overloading the same method isn't really the right thing to do (because it suggests they do the same thing - it's a kernel AD check in either case - but they don't actually seem to check the same things)... now that i've got y'all in a thread, what is your opinion on these ?

Copy link
Member Author

Choose a reason for hiding this comment

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

Separated out into #416

Expand Down
4 changes: 3 additions & 1 deletion test/transform/chaintransform.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,8 @@
@test repr(tp ∘ tf) == "Chain of 2 transforms:\n\t - $(tf) |> $(tp)"
test_ADs(
x -> SEKernel() ∘ (ScaleTransform(exp(x[1])) ∘ ARDTransform(exp.(x[2:4]))),
randn(rng, 4),
randn(rng, 4);
ADs=[:ForwardDiff, :ReverseDiff], # explicitly pass ADs to exclude :Zygote
)
@test_broken "test_AD of chain transform is currently broken in Zygote, see GitHub issue #263"
end