willtebbutt

#
# This uses the Nabla.jl-style interception mechanism whereby
# we wrap things that are to be differentiated w.r.t. in a
# thin wrapper. There are lots of thing that you can't
# propoagate derivative information through with this kind of
# approach without quite a lot of extra machinery, but the
# examples at the bottom do work.
#

using ChainRules, Cassette

willtebbutt

using Revise

using AbstractGPs
using BenchmarkTools
using CUDA
using KernelFunctions
using LinearAlgebra
using Random

using AbstractGPs: AbstractGP, FiniteGP, ZeroMean

willtebbutt

using AbstractGPs, KernelFunctions

# Generate toy data.
num_dims_in = 5
num_dims_out = 4
num_obs = 100
X = randn(num_obs, num_dims_in)
Y = randn(num_obs, num_dims_out)

# Convert to format required for AbstractGPs / KernelFunctions.

willtebbutt

using Pkg
Pkg.activate(; temp=true)
pkg"add AdvancedHMC, LogDensityProblems, LinearAlgebra, Plots, Random, Zygote, Statistics"
using AdvancedHMC, LogDensityProblems, LinearAlgebra, Plots, Random, Zygote, Statistics

# Define the target distribution using the `LogDensityProblem` interface
struct LogTargetDensity
    dim::Int
end
LogDensityProblems.logdensity(::LogTargetDensity, θ) = -sum(abs2, θ) / 2