**BREAKING** improve structs & constructors

src/AbstractGPs.jl

@@ -59,6 +59,9 @@ include("latent_gp.jl")
 # Plotting utilities.
 include("util/plotting.jl")
 
+# Autodiff utilities
+include("util/ad.jl")
+
 # Testing utilities.
 include("util/TestUtils.jl")
 

src/finite_gp_projection.jl

@@ -4,7 +4,8 @@
 The finite-dimensional projection of the AbstractGP `f` at `x`. Assumed to be observed under
 Gaussian noise with zero mean and covariance matrix `Σy`
 """
-struct FiniteGP{Tf<:AbstractGP,Tx<:AbstractVector,TΣ} <: AbstractMvNormal
+struct FiniteGP{Tf<:AbstractGP,Tx<:AbstractVector,TΣ<:AbstractMatrix{<:Real}} <:
+       AbstractMvNormal
     f::Tf
     x::Tx
     Σy::TΣ
@@ -17,7 +18,11 @@
 const default_σ² = 1e-18
 
 function FiniteGP(f::AbstractGP, x::AbstractVector, σ²::Real=default_σ²)
-    return FiniteGP(f, x, Fill(σ², length(x)))
+    return FiniteGP(f, x, ScalMat(length(x), σ²))
+end
+
+function FiniteGP(f::AbstractGP, x::AbstractVector, σ²::UniformScaling)
+    return FiniteGP(f, x, σ²[1, 1])
 end
 
 ## conversions

src/util/ad.jl

@@ -0,0 +1,6 @@
+import ChainRulesCore: ProjectTo, Tangent
+using PDMats: ScalMat
+
+ProjectTo(x::T) where {T<:ScalMat} = ProjectTo{T}(; dim=x.dim, value=ProjectTo(x.value))
+(pr::ProjectTo{<:ScalMat})(dx::ScalMat) = ScalMat(pr.dim, pr.value(dx.value))
+(pr::ProjectTo{<:ScalMat})(dx::Tangent{<:ScalMat}) = ScalMat(pr.dim, pr.value(dx.value))

test/finite_gp_projection.jl

@@ -205,6 +205,15 @@ end
             first(FiniteDifferences.grad(central_fdm(3, 1), Base.Fix1(logpdf, fx), y))
         @test Distributions.sqmahal!(r, fx, Y) ≈ Distributions.sqmahal(fx, Y)
     end
+
+    @testset "FiniteGP with UniformScaling" begin
+        f = GP(SqExponentialKernel())
+        fx = f(rand(10), 2.0 * I)
+        # for now, just check that it runs
+        _ = mean(fx)
+        _ = mean_and_cov(fx)
+        _ = rand(fx)
+    end
 end
 
 @testset "Docs" begin