Merge pull request #67 from theogf/Additionalfixes

more general fixes

Author: theogf

.travis.yml

@@ -4,20 +4,20 @@ os:
   - linux
   - osx
 julia:
-  - 1.2
-  - 1.4
+  - 1.3
+  - 1
   - nightly
 notifications:
   email: false
 after_success:
   # push coverage results to Coveralls
-  - if [[ $TRAVIS_JULIA_VERSION = 1.4 ]] && [[ $TRAVIS_OS_NAME = linux ]]; then
+  - if [[ $TRAVIS_JULIA_VERSION = 1 ]] && [[ $TRAVIS_OS_NAME = linux ]]; then
       julia -e 'using Pkg; Pkg.add("Coverage"); using Coverage; Coveralls.submit(process_folder())';
     fi
 jobs:
   include:
     - stage: "Documentation"
-      julia: 1.4
+      julia: 1
       os: linux
       script:
         - export DOCUMENTER_DEBUG=true

Project.toml

@@ -1,14 +1,15 @@
 name = "AugmentedGaussianProcesses"
 uuid = "38eea1fd-7d7d-5162-9d08-f89d0f2e271e"
 authors = ["Theo Galy-Fajou <theo.galyfajou@gmail.com>"]
-version = "0.8.5"
+version = "0.9.0"
 
 [deps]
 AdvancedHMC = "0bf59076-c3b1-5ca4-86bd-e02cd72cde3d"
 Clustering = "aaaa29a8-35af-508c-8bc3-b662a17a0fe5"
 DataStructures = "864edb3b-99cc-5e75-8d2d-829cb0a9cfe8"
 Dates = "ade2ca70-3891-5945-98fb-dc099432e06a"
 DeterminantalPointProcesses = "4d968f93-c0cd-4b7f-b189-b034d1a24a0e"
+Distances = "b4f34e82-e78d-54a5-968a-f98e89d6e8f7"
 Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"
 FastGaussQuadrature = "442a2c76-b920-505d-bb47-c5924d526838"
 Flux = "587475ba-b771-5e3f-ad9e-33799f191a9c"
@@ -25,7 +26,6 @@ RecipesBase = "3cdcf5f2-1ef4-517c-9805-6587b60abb01"
 Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
 SimpleTraits = "699a6c99-e7fa-54fc-8d76-47d257e15c1d"
 SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
-StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 StatsBase = "2913bbd2-ae8a-5f71-8c99-4fb6c76f3a91"
 StatsFuns = "4c63d2b9-4356-54db-8cca-17b64c39e42c"
@@ -51,7 +51,7 @@ RecipesBase = "1.0, 1.1"
 Reexport = "0.2"
 SimpleTraits = "0.9"
 SpecialFunctions = "0.9, 0.10"
-StatsBase = "0.32.0, 0.33"
-StatsFuns = "0.8.0, 0.9"
-Zygote = "0.4, 0.5"
-julia = "1.2"
+StatsBase = "0.32, 0.33"
+StatsFuns = "0.8, 0.9"
+Zygote = "0.5"
+julia = "1.3"

README.md

@@ -2,7 +2,7 @@
 
 [![Docs Latest](https://img.shields.io/badge/docs-dev-blue.svg)](https://theogf.github.io/AugmentedGaussianProcesses.jl/dev)
 [![Docs Stable](https://img.shields.io/badge/docs-stable-blue.svg)](https://theogf.github.io/AugmentedGaussianProcesses.jl/stable)
-[![Build Status](https://travis-ci.org/theogf/AugmentedGaussianProcesses.jl.svg?branch=master)](https://travis-ci.org/theogf/AugmentedGaussianProcesses.jl)
+[![Build Status](https://travis-ci.com/theogf/AugmentedGaussianProcesses.jl.svg?branch=master)](https://travis-ci.com/theogf/AugmentedGaussianProcesses.jl)
 [![Coverage Status](https://coveralls.io/repos/github/theogf/AugmentedGaussianProcesses.jl/badge.svg?branch=master)](https://coveralls.io/github/theogf/AugmentedGaussianProcesses.jl?branch=master)
 [![DOI](https://zenodo.org/badge/118922202.svg)](https://zenodo.org/badge/latestdoi/118922202)
 
@@ -59,7 +59,7 @@ AugmentedGaussianProcesses.jl is a Julia package in development for **Data Augme
 
 ## Install the package
 
-The package requires at least [Julia 1.1](https://julialang.org/)
+The package requires at least [Julia 1.3](https://julialang.org/)
 Run `julia`, press `]` and type `add AugmentedGaussianProcesses`, it will install the package and all its dependencies.
 
 ## Use the package

docs/examples/multiclassgp.jl

@@ -68,7 +68,7 @@ end;
 
 function plot_contour(model, σ)
     n_grid = 100
-    global pred_proba, pred, x, y = compute_grid(model, n_grid);
+    pred_proba, pred, x, y = compute_grid(model, n_grid);
     colors = reshape(
             [
                 RGB([pred_proba[model.likelihood.ind_mapping[j]][i] for j in 1:n_class]...)

src/ar_predict.jl

@@ -21,7 +21,7 @@ ks = []
     @assert length(y_past) == m.nTask
     @assert all(length.(y_past).>= p)
     @assert all(m.nDim .== p)
-    global Xtest = [reshape(y[(end-p+1):end],1,:) for y in y_past]
+    Xtest = [reshape(y[(end-p+1):end],1,:) for y in y_past]
     y_new = [zeros(T, n) for _ in 1:m.nTask]
     for i in 1:n
         setindex!.(y_new, first.(first.(first(_predict_f(m, Xtest, covf = false)))), i)

src/data/datacontainer.jl

@@ -14,20 +14,20 @@ struct DataContainer{
     TY<:AbstractVector,
 } <: AbstractDataContainer
     X::TX # Feature vectors
-    y::TY # Output (-1,1 for classification, real for regression, matrix for multiclass)
+    y::TY # Output (-1,1 for classification, real for regression, vector{vector} for multiclass)
     nSamples::Int # Number of samples
     nDim::Int # Number of features per sample
 end
 
 function wrap_data(X::TX, y::TY) where {TX, TY<:AbstractVector{<:Real}}
-    size(y, 1) == size(X, 1) || error("There is not the same number of samples in X ($(length(TX))) and y ($(size(y, 1)))")
+    length(y) == length(X) || error("There is not the same number of samples in X ($(length(X))) and y ($(length(y)))")
     Tx = eltype(first(X))
     Ty = eltype(first(y))
     return DataContainer{Tx, TX, Ty, TY}(X, y, length(X), length(first(X)))
 end
 
 function wrap_data(X::TX, y::TY) where {TX, TY<:AbstractVector}
-    size(first(y), 1) == size(X, 1) || error("There is not the same number of samples in X ($(length(TX))) and y ($(size(y, 1)))")
+    all(length.(y) .== length(X)) || error("There is not the same number of samples in X ($(length(X))) and y ($(length.(y))))")
     Tx = eltype(first(X))
     Ty = eltype(first(y))
     return DataContainer{Tx, TX, Ty, TY}(X, y, length(X), length(first(X)))
@@ -46,7 +46,7 @@ struct MODataContainer{
 end
 
 function wrap_modata(X::TX, y::TY) where {TX, TY<:AbstractVector}
-    all(size.(y, 1) .== size(X, 1)) || error("There is not the same number of samples in X ($(length(TX))) and y ($(size(y, 1)))")
+    all(length.(y) .== length(X)) || error("There is not the same number of samples in X ($(length(X))) and y ($(length.(y))))")
     Tx = eltype(first(X))
     return MODataContainer{Tx, TX, TY}(X, y, length(X), length(first(X)), length(y))
 end

src/data/utils.jl

@@ -2,6 +2,8 @@ view_x(d::AbstractDataContainer, indices) = view(d.X, indices)
 view_y(l::Likelihood, y::AbstractVector, i::AbstractVector) = view(y, i)
 view_y(l::Likelihood, d::AbstractDataContainer, i::AbstractVector) = view_y(l, output(d), i)
 view_y(l::Likelihood, d::MODataContainer, i::AbstractVector) = view_y.(l, output(d), Ref(i))
+view_y(l::AbstractVector{<:Likelihood}, d::MODataContainer, i::AbstractVector) = view_y.(l, output(d), Ref(i))
+
 
 # Verify that the data is self-consistent and consistent with the likelihood ##
 function check_data!(

src/examples/gpregression.md/GP Regression.md

@@ -1,5 +1,7 @@
-"""Compute the KL Divergence between the GP Prior and the variational distribution"""
-GaussianKL(model::AbstractGP) = sum(broadcast(GaussianKL, model.f, Zviews(model)))
+"""
+    KL Divergence between the GP Prior and the variational distribution
+"""
+GaussianKL(model::AbstractGP) = mapreduce(GaussianKL, +, model.f, Zviews(model))
 
 GaussianKL(gp::AbstractLatent, X::AbstractVector) = GaussianKL(mean(gp), pr_mean(gp, X), cov(gp), pr_cov(gp))
 
@@ -13,15 +15,17 @@ function GaussianKL(
     0.5 * (logdet(K) - logdet(Σ) + tr(K \ Σ) + invquad(K, μ - μ₀) - length(μ))
 end
 
-extraKL(model::AbstractGP{T}) where {T} = zero(T)
+extraKL(::AbstractGP{T}) where {T} = zero(T)
 
-"""Return the extra KL term containing the divergence with the GP at time t and t+1"""
+"""
+    Extra KL term containing the divergence with the GP at time t and t+1
+"""
 function extraKL(model::OnlineSVGP{T}) where {T}
     KLₐ = zero(T)
     for gp in model.f
         κₐμ = gp.κₐ * mean(gp)
         KLₐ += gp.prev𝓛ₐ
-        KLₐ += -0.5 *  sum(opt_trace.([gp.invDₐ], [gp.K̃ₐ, gp.κₐ * cov(gp) * transpose(gp.κₐ)]))
+        KLₐ += -0.5 *  sum(opt_trace.(Ref(gp.invDₐ), [gp.K̃ₐ, gp.κₐ * cov(gp) * transpose(gp.κₐ)]))
         KLₐ += dot(gp.prevη₁, κₐμ) - 0.5 * dot(κₐμ, gp.invDₐ * κₐμ)
     end
     return KLₐ
@@ -57,21 +61,22 @@ function PoissonKL(
 end
 
 
-"""KL(q(ω)||p(ω)), where q(ω) = PG(b,c) and p(ω) = PG(b,0). θ = 𝑬[ω]"""
+"""
+    KL(q(ω)||p(ω)), where q(ω) = PG(b,c) and p(ω) = PG(b,0). θ = 𝑬[ω]
+"""
 function PolyaGammaKL(b, c, θ)
     dot(b, logcosh.(0.5 * c)) - 0.5 * dot(abs2.(c), θ)
 end
 
-
 """
     Entropy of GIG variables with parameters a,b and p and omitting the derivative d/dpK_p cf <https://en.wikipedia.org/wiki/Generalized_inverse_Gaussian_distribution#Entropy>
 """
 function GIGEntropy(a, b, p)
-    sqrtab = sqrt.(a .* b)
-    return sum(0.5 * log.(a ./ b)) +
-           sum(log.(2 * besselk.(p, sqrtab))) +
+    sqrt_ab = sqrt.(a .* b)
+    return 0.5 * (sum(log, a) - sum(log, b)) +
+           mapreduce((p, s) -> log(2 * besselk(p, s), +, p, sqrt_ab)) +
            sum(
-               0.5 * sqrtab ./ besselk.(p, sqrtab) .*
-               (besselk.(p + 1, sqrtab) + besselk.(p - 1, sqrtab)),
+               0.5 * sqrt_ab ./ besselk.(p, sqrt_ab) .*
+               (besselk.(p + 1, sqrt_ab) + besselk.(p - 1, sqrt_ab)),
            )
 end