Adding sparse finite GP type for integration with Turing

[ElOceanografo]

A continuation/focusing of this discussion on Discourse. Defining the following structure and methods allows using a sparse GP in a Turing model:

struct SparseFiniteGP{T1<:Stheno.FiniteGP, T2<:Stheno.FiniteGP} <: Distributions.AbstractMvNormal
    fobs::T1
    finducing::T2
end

function Stheno.logpdf(f::T, y::AbstractArray{T1, 1}) where {T <: SparseFiniteGP, T1 <: Real}
    return elbo(f.fobs, y, f.finducing)
end

Base.length(f::SparseFiniteGP) = length(f.fobs)
Distributions._rand!(rng::Random._GLOBAL_RNG, f::SparseFiniteGP1, x::Array{T}) where {T<:Real} = rand(f.fobs)

An example model is shown below. Using the SparseFiniteGP speeds up inference dramatically for a simulated dataset with ~1000 observations.

@model function example_model_sparse(x, z, xu)
    f = GP(Matern32(), GPC())
    fsparse = SparseFiniteGP(f(x, 1e-6),  f(xu, 1e-6))
    y ~ fsparse
    λ = exp.(y)
    z .~ Poisson.(λ)
end
chn_sparse = sample(example_model_sparse(x, z, xu), NUTS(), 200)

Some questions to discuss while I put together a PR...

• Any thoughts on the interface or structure of the types?
• Besides rand and logpdf, what other methods does this new type need?
• Where should this code go? In src/abstract_gp.jl?
• What tests are needed--in particular, what comparisons should there be between the sparse and dense GPs?

[willtebbutt]

Thanks for opening this.

Any thoughts on the interface or structure of the types?

The above looks really good -- I prefer the design to the one that I proposed on discourse.

Besides rand and logpdf, what other methods does this new type need?

Yes, there are a few methods of mean and cov that should just fall back to fobs. Basically just look for methods in src/abstract_gp.jl that accept FiniteGPs and implement them :)

Where should this code go? In src/abstract_gp.jl?

We'll have quite a lot of approximate inference stuff one this has been added, so maybe it makes sense to create a new file called src/approximate_inference.jl or something, put this there, and move elbo-related stuff from src/abstract_gp.jl into that file?

What tests are needed--in particular, what comparisons should there be between the sparse and dense GPs?

I would just check that mean, cov, and rand (with identical seeds) produce the same thing in all cases, and that elbo with regular FiniteGPs and logpdf under SparseFiniteGPs produce the same answer.

[ElOceanografo]

Ok, that all sounds good. PR on the way.

[ElOceanografo]

Closed by https://github.com/willtebbutt/Stheno.jl/pull/136!