Stop doing sparse stuff in this package (#213)

* Bump minor version

* Fix comment

* Allow [email protected]

* Remove elbo from core of package

* Remove elbo-related tests

* Update NEWS.md

* Only bump patch

* Update implementation for SparseGP object

* Remove [email protected] from compat

Author: willtebbutt

NEWS.md

@@ -6,6 +6,12 @@ between versions, and discuss new features.
 If you find a breaking change this is not reported here, please either raise an issue or
 make a PR to ammend this document.
 
+## 0.7.14
+
+AbstractGPs now takes care of everything sparsity-related.
+Consequently, Stheno no longer tests anything ELBO-related, and the functionality you get
+will depend entirely upon which version of AbstractGPs you're using.
+
 ## 0.7.0
 
 ### Breaking changes

Project.toml

@@ -1,6 +1,6 @@
 name = "Stheno"
 uuid = "8188c328-b5d6-583d-959b-9690869a5511"
-version = "0.7.13"
+version = "0.7.14"
 
 [deps]
 AbstractGPs = "99985d1d-32ba-4be9-9821-2ec096f28918"
@@ -17,7 +17,7 @@ Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 ZygoteRules = "700de1a5-db45-46bc-99cf-38207098b444"
 
 [compat]
-AbstractGPs = "0.3.9"
+AbstractGPs = "0.4"
 BlockArrays = "0.15, 0.16"
 ChainRulesCore = "1"
 FillArrays = "0.7, 0.8, 0.9, 0.10, 0.11, 0.12"

src/Stheno.jl

@@ -65,7 +65,6 @@ module Stheno
     include("deprecate.jl")
 
     export wrap, BlockData, GPC, GPPPInput, @gppp
-    export elbo, dtc
     export ∘, select, stretch, periodic, shift
     export cov_diag, mean_and_cov_diag
 end # module

src/composite/cross.jl

@@ -61,7 +61,7 @@ end
 
 
 #
-# Util for multi-process versions of `rand`, `logpdf`, and `elbo`.
+# Build a single FiniteGP from a collection of FiniteGPs.
 #
 
 function finites_to_block(fs::AV{<:FiniteGP})

src/gaussian_process_probabilistic_programme.jl

@@ -177,11 +177,15 @@ f = @gppp let
     f3 = f1 + f2
 end
 
-x = GPPPInput(:f3, randn(5))
+x_local = randn(5)
 
-y = rand(f(x, 0.1))
+x = BlockData(GPPPInput(:f1, x_local), GPPPInput(:f2, x_local), GPPPInput(:f3, x_local))
 
-logpdf(f(x, 0.1), y) ≈ elbo(f(x, 0.1), y, f(x, 1e-9))
+y = rand(f(x, 1e-12))
+
+f1, f2, f3 = split(x, y)
+
+isapprox(f1 + f2, f3; rtol=1e-4)
 
 # output
 

src/sparse_finite_gp.jl

@@ -55,14 +55,14 @@ rand(f::SparseFiniteGP, N::Int) = rand(Random.GLOBAL_RNG, f, N)
 rand(rng::AbstractRNG, f::SparseFiniteGP) = vec(rand(rng, f, 1))
 rand(f::SparseFiniteGP) = vec(rand(f, 1))
 
-elbo(f::SparseFiniteGP, y::AV{<:Real}) = elbo(f.fobs, y, f.finducing)
+elbo(f::SparseFiniteGP, y::AV{<:Real}) = elbo(VFE(f.finducing), f.fobs, y)
 
-logpdf(f::SparseFiniteGP, y::AV{<:Real}) = elbo(f.fobs, y, f.finducing)
+logpdf(f::SparseFiniteGP, y::AV{<:Real}) = elbo(VFE(f.finducing), f.fobs, y)
 
 function logpdf(f::SparseFiniteGP, Y::AbstractMatrix{<:Real})
     return map(y -> logpdf(f, y), eachcol(Y))
 end
 
 function posterior(f::SparseFiniteGP, y::AbstractVector{<:Real})
-    return approx_posterior(AbstractGPs.VFE(), f.fobs, y, f.finducing)
+    return posterior(AbstractGPs.VFE(f.finducing), f.fobs, y)
 end

test/abstract_gp.jl

@@ -73,85 +73,6 @@ end
             atol=1e-9, rtol=1e-9,
         )
     end
-    @testset "logpdf / elbo / dtc" begin
-        rng, N, S, σ, gpc = MersenneTwister(123456), 10, 11, 1e-1, GPC()
-        x = collect(range(-3.0, stop=3.0, length=N))
-        f = wrap(GP(1, SEKernel()), gpc)
-        fx, y = FiniteGP(f, x, 0), FiniteGP(f, x, σ^2)
-        ŷ = rand(rng, y)
-
-        # Check that logpdf returns the correct type and roughly agrees with Distributions.
-        @test logpdf(y, ŷ) isa Real
-        @test logpdf(y, ŷ) ≈ logpdf(MvNormal(Vector(mean(y)), cov(y)), ŷ)
-
-        # Check that multi-sample logpdf returns the correct type and is consistent with
-        # single-sample logpdf
-        Ŷ = rand(rng, y, S)
-        @test logpdf(y, Ŷ) isa Vector{Float64}
-        @test logpdf(y, Ŷ) ≈ [logpdf(y, Ŷ[:, n]) for n in 1:S]
-
-        # 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(FiniteGP(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)))),
-            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)
-
-        # 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);
-            atol=1e-8, rtol=1e-8,
-        )
-        adjoint_test(
-            α->sum(logpdf(FiniteGP(α * f, x, 1e-1), Ŷ)), l̄, randn(rng);
-            atol=1e-8, rtol=1e-8,
-        )
-
-        # Ensure that the elbo is close to the logpdf when appropriate.
-        @test elbo(y, ŷ, fx) isa Real
-        @test elbo(y, ŷ, fx) ≈ logpdf(y, ŷ)
-        @test elbo(y, ŷ, y) < logpdf(y, ŷ)
-        @test elbo(y, ŷ, FiniteGP(f, x, 2 * σ^2)) < elbo(y, ŷ, y)
-
-        # Check adjoint w.r.t. elbo is correct.
-        adjoint_test(
-            (x, ŷ, σ)->elbo(FiniteGP(f, x, σ^2), ŷ, FiniteGP(f, x, 0)),
-            randn(rng), x, ŷ, σ;
-            atol=1e-6, rtol=1e-6,
-        )
-
-        # Ensure that the dtc is close to the logpdf when appropriate.
-        @test dtc(y, ŷ, fx) isa Real
-        @test dtc(y, ŷ, fx) ≈ logpdf(y, ŷ)
-
-        # Check adjoint w.r.t. dtc is correct.
-        adjoint_test(
-            (x, ŷ, σ)->dtc(FiniteGP(f, x, σ^2), ŷ, FiniteGP(f, x, 0)),
-            randn(rng), x, ŷ, σ;
-            atol=1e-6, rtol=1e-6,
-        )
-    end
     @testset "Type Stability - $T" for T in [Float64, Float32]
         rng = MersenneTwister(123456)
         x = randn(rng, T, 123)
@@ -164,6 +85,5 @@ end
         y = rand(rng, fx)
         @test y isa Vector{T}
         @test logpdf(fx, y) isa T
-        @test elbo(fx, y, u) isa T
     end
 end

test/composite/addition.jl

@@ -89,56 +89,3 @@
         )
     end
 end
-
-# # θ = Dict(:l1=>0.5, :l2=>2.3);
-# x, z = collect(range(-5.0, 5.0; length=512)), collect(range(-5.0, 5.0; length=128));
-# y = rand(GP(sin, SqExponentialKernel(), GPC())(x, 0.1));
-
-# foo_logpdf = (x, y) -> begin
-#     gpc = GPC()
-#     f = GP(sin, SqExponentialKernel(), gpc)
-#     return logpdf(f(x, 0.1), y)
-# end
-
-# foo_elbo = (x, y, z) -> begin
-#     f = GP(0, SqExponentialKernel(), GPC())
-#     return elbo(f(x, 0.1), y, f(z, 0.001))
-# end
-
-# @benchmark foo_logpdf($x,  $y)
-# @benchmark Zygote.pullback(foo_logpdf, $x, $y)
-
-# let
-#     z, back = Zygote.pullback(foo_logpdf, x, y)
-#     @benchmark $back($(randn()))
-# end
-
-# let
-#     foo = function(x, y)
-#         fx = GP(0, SqExponentialKernel(), GPC())(x, 0.1)
-#         C = cholesky(Symmetric(cov(fx)))
-#         return logdet(C) + Xt_invA_X(C, y)
-#     end
-#     display(@benchmark Zygote.pullback($foo, $x, $y)) 
-#     z_pw, back_pw = Zygote.pullback(foo, x, y)
-#     @benchmark $back_pw(randn())
-# end
-
-
-# @benchmark foo_elbo($x, $y, $z)
-# @benchmark Zygote.pullback(foo_elbo, $x, $y, $z)
-
-# let
-#     L, back = Zygote.pullback(foo_elbo, x, y, z)
-#     @benchmark $back($L)
-# end
-
-
-# θ->begin
-#     gpc = GPC()
-#     f1 = GP(sin, SqExponentialKernel(l=θ[:l1]), gpc)
-#     f2 = GP(cos, SqExponentialKernel(l=θ[:l2]), gpc)
-#     f3 = f1 + f2
-#     return f3, f3
-# end,
-# 13, 11,

test/composite/test_util.jl

@@ -3,7 +3,7 @@
 
 Some basic consistency checks for the function `f(θ)::Tuple{GP, GP}`. Mainly just checks
 that Zygote works properly for `f`, and correctly derives the gradients w.r.t. `θ` for
-`rand`, `logpdf`, `elbo` when considering the f.d.d.s `f(x, Σ)` and observations `y`, where
+`rand`, `logpdf`, when considering the f.d.d.s `f(x, Σ)` and observations `y`, where
 `Σ = _to_psd(A)`. The first output of `f` will be the GP sampled from and whose `logpdf`
 will be computed, while the second will be used as the process for the pseudo-points, whose
 inputs are `z`.
@@ -46,7 +46,7 @@ function check_consistency(rng::AbstractRNG, θ, f, x::AV, y::AV, A, z::AV, B)
 
 
     # #
-    # # rand / logpdf / elbo tests
+    # # rand / logpdf tests
     # #
 
     # # Check that the gradient w.r.t. the samples is correct (single-sample).
@@ -69,19 +69,8 @@ function check_consistency(rng::AbstractRNG, θ, f, x::AV, y::AV, A, z::AV, B)
     #     rtol=1e-4, atol=1e-4,
     # )
 
-    # # Check adjoint for elbo.
-    # adjoint_test(
-    #     (ϴ, x, A, y, z, B)->begin
-    #         fx, uz = h(θ, x, A, z, B)
-    #         return elbo(fx, y, uz)
-    #     end,
-    #     randn(rng), θ, x, A, y, z, B;
-    #     rtol=1e-4, atol=1e-4,
-    # )
-
-
     # #
-    # # multi-process rand / logpdf / elbo tests - this stuff won't work for anything if
+    # # multi-process rand / logpdf tests - this stuff won't work for anything if
     # # cross-related functionality doesn't work properly
     # #
 
@@ -120,15 +109,6 @@ function check_consistency(rng::AbstractRNG, θ, f, x::AV, y::AV, A, z::AV, B)
     #     randn(rng), θ, x, A, y;
     #     rtol=1e-4, atol=1e-4,
     # )
-
-    # adjoint_test(
-    #     (ϴ, x, A, y, z, B)->begin
-    #         fx, uz = h(θ, x, A, z, B)
-    #         return elbo([fx, fx], [y, y], [uz, uz])
-    #     end,
-    #     randn(rng), θ, x, A, y, z, B;
-    #     rtol=1e-4, atol=1e-4,
-    # )
 end
 
 function standard_1D_dense_test(rng::AbstractRNG, θ, f, x::AV, z::AV)

test/sparse_finite_gp.jl

@@ -30,12 +30,12 @@
         fxu = SparseFiniteGP(f(x, σ), f(xu, σu))
         y = rand(MersenneTwister(12345), fxu)
 
-        fpost1 = approx_posterior(VFE(), fxu.fobs, y, fxu.finducing)
+        fpost1 = posterior(VFE(fxu.finducing), fxu.fobs, y)
         fpost2 = posterior(fxu, y)
 
         @test marginals(fpost1(x)) == marginals(fpost2(x))
         @test elbo(fxu, y) == logpdf(fxu, y)
-        @test logpdf(fxu, y) == elbo(fxu.fobs, y, fxu.finducing)
+        @test logpdf(fxu, y) == elbo(VFE(fxu.finducing),fxu.fobs, y)
         yy = rand(fxu, 10)
         @test all(logpdf(fx, yy) .> logpdf(fxu, yy))
     end