Adds nlatent and nclass to Categorical

Project.toml

@@ -1,7 +1,7 @@
 name = "GPLikelihoods"
 uuid = "6031954c-0455-49d7-b3b9-3e1c99afaf40"
 authors = ["JuliaGaussianProcesses Team"]
-version = "0.3.1"
+version = "0.3.2"
 
 [deps]
 Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"

src/GPLikelihoods.jl

@@ -28,12 +28,24 @@ export Link,
     ProbitLink,
     NormalCDFLink,
     SoftMaxLink
+export nlatent
 
 # Links
 include("links.jl")
 
 # Likelihoods
 abstract type AbstractLikelihood end
+
+"""
+    nlatent(::AbstractLikelihood)::Int
+
+Returns the number of latent Gaussian processes needed to build the likelihood.
+In other terms the input dimensionality passed to the likelihood from the GP perspective.
+It is typically 1, but for some likelihoods like [`CategoricalLikelihood`](@ref) or 
+[`HeteroscedasticGaussianLikelihood`](@ref) multiple latent GPs are necessary.
+"""
+nlatent(::AbstractLikelihood) = 1 # Default number of latent GPs required is 1
+
 include("likelihoods/bernoulli.jl")
 include("likelihoods/categorical.jl")
 include("likelihoods/gaussian.jl")

src/TestInterface.jl

@@ -4,7 +4,7 @@ using Functors
 using Random
 using Test
 
-function test_interface(rng::AbstractRNG, lik, out_dist, D_in=1; functor_args=())
+function test_interface(rng::AbstractRNG, lik, out_dist, D_in=nlatent(lik); functor_args=())
     N = 10
     T = Float64 # TODO test Float32 as well
     f, fs = if D_in == 1
@@ -53,7 +53,7 @@ samples is correct and if the functor works as intended.
 - `functor_args=()`: a collection of symbols of arguments to match functor parameters with.
 ...
 """
-function test_interface(lik, out_dist, D_in=1; kwargs...)
+function test_interface(lik, out_dist, D_in=nlatent(lik); kwargs...)
     return test_interface(Random.GLOBAL_RNG, lik, out_dist, D_in; kwargs...)
 end
 

src/likelihoods/categorical.jl

@@ -1,7 +1,7 @@
 """
-    CategoricalLikelihood(l=BijectiveSimplexLink(softmax))
+    CategoricalLikelihood(n::Int, l=BijectiveSimplexLink(softmax))
 
-Categorical likelihood is to be used if we assume that the 
+Categorical likelihood with `n` categories is to be used if we assume that the 
 uncertainty associated with the data follows a [Categorical distribution](https://en.wikipedia.org/wiki/Categorical_distribution).
 
 Assuming a distribution with `n` categories:
@@ -27,10 +27,14 @@ For more details, see the end of the section of this [Wikipedia link](https://en
 where it corresponds to Variant 1 and 2.
 """
 struct CategoricalLikelihood{Tl<:AbstractLink} <: AbstractLikelihood
+    n::Int # Number of categories
     invlink::Tl
 end
 
-CategoricalLikelihood(l=BijectiveSimplexLink(softmax)) = CategoricalLikelihood(link(l))
+CategoricalLikelihood(n=BijectiveSimplexLink(softmax)) = CategoricalLikelihood(n, link(l))
+
+nlatent(l::CategoricalLikelihood) = l.n
+nlatent(l::CategoricalLikelihood{<:BijectiveSimplexLink}) = l.n - 1
 
 function (l::CategoricalLikelihood)(f::AbstractVector{<:Real})
     return Categorical(l.invlink(f))

src/likelihoods/gaussian.jl

@@ -42,6 +42,8 @@ end
 
 HeteroscedasticGaussianLikelihood(l=exp) = HeteroscedasticGaussianLikelihood(link(l))
 
+nlatent(::HeteroscedasticGaussianLikelihood) = 2
+
 function (l::HeteroscedasticGaussianLikelihood)(f::AbstractVector{<:Real})
     return Normal(f[1], sqrt(l.invlink(f[2])))
 end

test/likelihoods/categorical.jl

@@ -1,13 +1,15 @@
 @testset "CategoricalLikelihood" begin
-    @test CategoricalLikelihood() isa
+    nclass = 4
+    @test CategoricalLikelihood(nclass) isa
         CategoricalLikelihood{<:GPLikelihoods.BijectiveSimplexLink}
+
+    @test CategoricalLikelihood(nclass, softmax) isa CategoricalLikelihood{SoftMaxLink}
+    @test CategoricalLikelihood(nclass, SoftMaxLink()) isa CategoricalLikelihood{SoftMaxLink}
 
-    @test CategoricalLikelihood(softmax) isa CategoricalLikelihood{SoftMaxLink}
-    @test CategoricalLikelihood(SoftMaxLink()) isa CategoricalLikelihood{SoftMaxLink}
-
-    OUT_DIM = 4
-    lik_bijective = CategoricalLikelihood()
-    test_interface(lik_bijective, Categorical, OUT_DIM)
-    lik_nonbijective = CategoricalLikelihood(softmax)
+    lik_bijective = CategoricalLikelihood(nclass)
+    test_interface(lik_bijective, Categorical)
+    @test nlatent(lik_bijective) == nclass - 1
+    lik_nonbijective = CategoricalLikelihood(nclass, softmax)
     test_interface(lik_nonbijective, Categorical, OUT_DIM)
+    @test nlatent(lik_nonbijective) == nclass
 end

test/likelihoods/gaussian.jl

@@ -10,8 +10,7 @@ end
     end
 
     lik = HeteroscedasticGaussianLikelihood()
-    IN_DIM = 3
-    OUT_DIM = 2 # one for the mean the other for the log-standard deviation
     N = 10
-    test_interface(lik, Normal, 2)
+    test_interface(lik, Normal)
+    @test nlatent(lik) == 2
 end