Refactor Approximate Inference (#194)

* First pass at new approx posterior API * Move elbo to `approx_posterior_gp.jl` * Update ELBO tests * Fix elbo tests * Parameterise VFE type * Only pass inducing inputs to VFE * Formatting * Apply suggestions from code review Co-authored-by: willtebbutt <[email protected]> * Explain jitter in docstring * Update regression example * Update concrete features docs * Replace comment * Apply suggestions from code review Co-authored-by: st-- <[email protected]> * Use IrrationalConstants and test type stability * Add IrrationalConstants compat * Regenerate docs Manifests * Refactor again * Formatting * Use identity instead of equality * Bump to 0.4.0 * Bump docs versions Co-authored-by: willtebbutt <[email protected]> Co-authored-by: st-- <[email protected]>
JuliaGaussianProcesses · Aug 20, 2021 · 5630f89 · 5630f89 · willtebbutt · Aug 24, 2021
1 parent b44dd01
commit 5630f89
Show file tree

Hide file tree

Showing 14 changed files with 359 additions and 291 deletions.
diff --git a/Project.toml b/Project.toml
@@ -1,12 +1,13 @@
 name = "AbstractGPs"
 uuid = "99985d1d-32ba-4be9-9821-2ec096f28918"
 authors = ["JuliaGaussianProcesses Team"]
-version = "0.3.10"
+version = "0.4.0"
 
 [deps]
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"
 FillArrays = "1a297f60-69ca-5386-bcde-b61e274b549b"
+IrrationalConstants = "92d709cd-6900-40b7-9082-c6be49f344b6"
 KernelFunctions = "ec8451be-7e33-11e9-00cf-bbf324bd1392"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
@@ -20,6 +21,7 @@ Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 ChainRulesCore = "1"
 Distributions = "0.19, 0.20, 0.21, 0.22, 0.23, 0.24, 0.25"
 FillArrays = "0.7, 0.8, 0.9, 0.10, 0.11, 0.12"
+IrrationalConstants = "0.1"
 KernelFunctions = "0.9, 0.10"
 RecipesBase = "1"
 Reexport = "0.2, 1"

diff --git a/docs/Manifest.toml b/docs/Manifest.toml
@@ -6,10 +6,10 @@ uuid = "a4c015fc-c6ff-483c-b24f-f7ea428134e9"
 version = "0.0.1"
 
 [[AbstractGPs]]
-deps = ["ChainRulesCore", "Distributions", "FillArrays", "KernelFunctions", "LinearAlgebra", "Random", "RecipesBase", "Reexport", "Statistics", "StatsBase", "Test"]
+deps = ["ChainRulesCore", "Distributions", "FillArrays", "IrrationalConstants", "KernelFunctions", "LinearAlgebra", "Random", "RecipesBase", "Reexport", "Statistics", "StatsBase", "Test"]
 path = ".."
 uuid = "99985d1d-32ba-4be9-9821-2ec096f28918"
-version = "0.3.10"
+version = "0.4.0"
 
 [[ArgTools]]
 uuid = "0dad84c5-d112-42e6-8d28-ef12dabb789f"

diff --git a/docs/Project.toml b/docs/Project.toml
@@ -4,5 +4,5 @@ Documenter = "e30172f5-a6a5-5a46-863b-614d45cd2de4"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 
 [compat]
-AbstractGPs = "0.3"
+AbstractGPs = "0.4"
 Documenter = "0.27"
diff --git a/docs/src/concrete_features.md b/docs/src/concrete_features.md
@@ -45,22 +45,21 @@ logpdf(f_posterior(x), y)
 Here, z is a set of pseudo-points. 
 ```julia
 z = randn(rng, 4)
-u = f(z)
 ```
 
-### Evidence Lower BOund (ELBO)
+#### Evidence Lower BOund (ELBO)
 We provide a ready implentation of elbo w.r.t to the pseudo points. We can perform Variational Inference on pseudo-points by maximizing the ELBO term w.r.t pseudo-points `z` and any kernel parameters. For more information, see [examples](https://github.com/JuliaGaussianProcesses/AbstractGPs.jl/tree/master/examples). 
 ```julia
-elbo(fx, y, u)
+elbo(VFE(f(z)), fx, y)
 ```
 
-### Construct the approximate posterior process implied by the VFE approximation.
+#### Construct the approximate posterior process implied by the VFE approximation.
 The optimal pseudo-points obtained above can be used to create a approximate/sparse posterior. This can be used like a regular posterior in many cases.
 ```julia
-f_approx_posterior = approx_posterior(VFE(), fx, y, u)
+f_approx_posterior = posterior(VFE(f(z)), fx, y)
 ```
 
-### An approximate posterior process is yet another `AbstractGP`, so you can do things with it like
+#### An approximate posterior process is yet another `AbstractGP`, so you can do things with it like
 ```julia
 marginals(f_approx_posterior(x))
 ```
@@ -87,12 +86,12 @@ p_p_fx = posterior(p_fx(x[4:10], 0.1), y[4:10])
 ```julia
 Z1 = rand(rng, 4)
 Z2 = rand(rng, 3)
-p_fx = approx_posterior(VFE(), f(x[1:7], 0.1), y[1:7], f(Z))
-u_p_fx = update_approx_posterior(p_fx1, f(x[8:10], 0.1), y[8:10])
+Z = vcat(Z1, Z2)
+p_fx1 = posterior(VFE(f(Z)), f(x[1:7], 0.1), y[1:7])
+u_p_fx1 = update_posterior(p_fx1, f(x[8:10], 0.1), y[8:10])
 ```
 ##### Adding pseudo-points in an sequential fashion
 ```julia
-
-p_fx1 = approx_posterior(VFE(), f(X, 0.1), y, f(Z1))
-u_p_fx1 = update_approx_posterior(p_fx1, f(Z2))
+p_fx2 = posterior(VFE(f(Z1)), f(x, 0.1), y)
+u_p_fx2 = update_posterior(p_fx2, f(Z2))
 ```