Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Improve type inference of MixtureModel #1308

Merged
merged 6 commits into from
May 2, 2021
Merged

Improve type inference of MixtureModel #1308

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
98960e4
Improve type inference of `MixtureModel`
devmotion Apr 21, 2021
6fccef9
Bump version
devmotion Apr 21, 2021
0d2dd90
Remove comment
devmotion Apr 21, 2021
6d48f31
Bump minor version
devmotion Apr 22, 2021
39b7191
Merge branch 'master' into dw/mixture_inference
devmotion Apr 22, 2021
d51a7a6
Merge branch 'master' into dw/mixture_inference
devmotion May 2, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion Project.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
name = "Distributions"
uuid = "31c24e10-a181-5473-b8eb-7969acd0382f"
authors = ["JuliaStats"]
version = "0.24.17"
version = "0.24.18"
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What is the convention in Distributions - is the different type parameter of MixtureModel considered a breaking change?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

OK, I updated the version number.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

But now tests fail since there exists no version of HypothesisTests that is compatible with Distributions 0.25 🤔

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I opened a PR that removes the test dependency on HypothesisTests to fix these test failures: #1311


[deps]
FillArrays = "1a297f60-69ca-5386-bcde-b61e274b549b"
Expand Down
64 changes: 10 additions & 54 deletions src/mixtures/mixturemodel.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,11 +19,11 @@ A mixture of distributions, parametrized on:
* `C` distribution family of the mixture
* `CT` the type for probabilities of the prior
"""
struct MixtureModel{VF<:VariateForm,VS<:ValueSupport,C<:Distribution,CT<:Real} <: AbstractMixtureModel{VF,VS,C}
struct MixtureModel{VF<:VariateForm,VS<:ValueSupport,C<:Distribution,CT<:Categorical} <: AbstractMixtureModel{VF,VS,C}
components::Vector{C}
prior::Categorical{CT}
prior::CT

function MixtureModel{VF,VS,C}(cs::Vector{C}, pri::Categorical{CT}) where {VF,VS,C,CT}
function MixtureModel{VF,VS,C}(cs::Vector{C}, pri::CT) where {VF,VS,C,CT}
length(cs) == ncategories(pri) ||
error("The number of components does not match the length of prior.")
new{VF,VS,C,CT}(cs, pri)
Expand Down Expand Up @@ -171,16 +171,8 @@ minimum(d::MixtureModel) = minimum([minimum(dci) for dci in d.components])
maximum(d::MixtureModel) = maximum([maximum(dci) for dci in d.components])

function mean(d::UnivariateMixture)
K = ncomponents(d)
p = probs(d)
m = 0.0
for i = 1:K
pi = p[i]
if pi > 0.0
c = component(d, i)
m += mean(c) * pi
end
end
m = sum(pi * mean(component(d, i)) for (i, pi) in enumerate(p) if !iszero(pi))
return m
end

Expand Down Expand Up @@ -281,38 +273,22 @@ end
#### Evaluation

function insupport(d::AbstractMixtureModel, x::AbstractVector)
K = ncomponents(d)
p = probs(d)
@inbounds for i in eachindex(p)
pi = p[i]
if pi > 0.0 && insupport(component(d, i), x)
return true
end
end
return false
return any(insupport(component(d, i), x) for (i, pi) in enumerate(p) if !iszero(pi))
end

function _cdf(d::UnivariateMixture, x::Real)
K = ncomponents(d)
p = probs(d)
r = 0.0
@inbounds for i in eachindex(p)
pi = p[i]
if pi > 0.0
c = component(d, i)
r += pi * cdf(c, x)
end
end
r = sum(pi * cdf(component(d, i), x) for (i, pi) in enumerate(p) if !iszero(pi))
return r
end

cdf(d::UnivariateMixture{Continuous}, x::Real) = _cdf(d, x)
cdf(d::UnivariateMixture{Discrete}, x::Integer) = _cdf(d, x)

function _mixpdf1(d::AbstractMixtureModel, x)
ps = probs(d)
cs = components(d)
return sum((ps[i] > 0) * (ps[i] * pdf(cs[i], x)) for i in eachindex(ps))
p = probs(d)
return sum(pi * pdf(component(d, i), x) for (i, pi) in enumerate(p) if !iszero(pi))
end

function _mixpdf!(r::AbstractArray, d::AbstractMixtureModel, x)
Expand Down Expand Up @@ -345,29 +321,9 @@ function _mixlogpdf1(d::AbstractMixtureModel, x)
# m is chosen to be the maximum of logpri[i] + logpdf(cs[i], x)
# such that the argument of exp is in a reasonable range
#

K = ncomponents(d)
p = probs(d)
lp = Vector{eltype(p)}(undef, K)
m = -Inf # m <- the maximum of log(p(cs[i], x)) + log(pri[i])
@inbounds for i in eachindex(p)
pi = p[i]
if pi > 0.0
# lp[i] <- log(p(cs[i], x)) + log(pri[i])
lp_i = logpdf(component(d, i), x) + log(pi)
lp[i] = lp_i
if lp_i > m
m = lp_i
end
end
end
v = 0.0
@inbounds for i = 1:K
if p[i] > 0.0
v += exp(lp[i] - m)
end
end
return m + log(v)
lp = logsumexp(log(pi) + logpdf(component(d, i), x) for (i, pi) in enumerate(p) if !iszero(pi))
return lp
end

function _mixlogpdf!(r::AbstractArray, d::AbstractMixtureModel, x)
Expand Down
46 changes: 23 additions & 23 deletions test/mixture.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,15 +18,15 @@ function test_mixture(g::UnivariateMixture, n::Int, ns::Int,
end

K = ncomponents(g)
pr = probs(g)
pr = @inferred(probs(g))
@assert length(pr) == K

# mean
mu = 0.0
for k = 1:K
mu += pr[k] * mean(component(g, k))
end
@test mean(g) ≈ mu
@test @inferred(mean(g)) ≈ mu

# evaluation of cdf
cf = zeros(T, n)
Expand All @@ -38,7 +38,7 @@ function test_mixture(g::UnivariateMixture, n::Int, ns::Int,
end

for i = 1:n
@test cdf(g, X[i]) ≈ cf[i]
@test @inferred(cdf(g, X[i])) ≈ cf[i]
end
@test cdf.(g, X) ≈ cf

Expand All @@ -58,16 +58,16 @@ function test_mixture(g::UnivariateMixture, n::Int, ns::Int,
mix_lp0 = log.(mix_p0)

for i = 1:n
@test pdf(g, X[i]) ≈ mix_p0[i]
@test logpdf(g, X[i]) ≈ mix_lp0[i]
@test componentwise_pdf(g, X[i]) ≈ vec(P0[i,:])
@test componentwise_logpdf(g, X[i]) ≈ vec(LP0[i,:])
@test @inferred(pdf(g, X[i])) ≈ mix_p0[i]
@test @inferred(logpdf(g, X[i])) ≈ mix_lp0[i]
@test @inferred(componentwise_pdf(g, X[i])) ≈ vec(P0[i,:])
@test @inferred(componentwise_logpdf(g, X[i])) ≈ vec(LP0[i,:])
end

@test pdf.(g, X) ≈ mix_p0
@test logpdf.(g, X) ≈ mix_lp0
@test componentwise_pdf(g, X) ≈ P0
@test componentwise_logpdf(g, X) ≈ LP0
@test @inferred(map(Base.Fix1(pdf, g), X)) ≈ mix_p0
@test @inferred(map(Base.Fix1(logpdf, g), X)) ≈ mix_lp0
@test @inferred(componentwise_pdf(g, X)) ≈ P0
@test @inferred(componentwise_logpdf(g, X)) ≈ LP0

# sampling
if (T <: AbstractFloat)
Expand All @@ -94,15 +94,15 @@ function test_mixture(g::MultivariateMixture, n::Int, ns::Int,
end

K = ncomponents(g)
pr = probs(g)
pr = @inferred(probs(g))
@assert length(pr) == K

# mean
mu = zeros(length(g))
for k = 1:K
mu .+= pr[k] .* mean(component(g, k))
end
@test mean(g) ≈ mu
@test @inferred(mean(g)) ≈ mu

# evaluation
P0 = zeros(n, K)
Expand All @@ -121,20 +121,20 @@ function test_mixture(g::MultivariateMixture, n::Int, ns::Int,

for i = 1:n
x_i = X[:,i]
@test pdf(g, x_i) ≈ mix_p0[i]
@test logpdf(g, x_i) ≈ mix_lp0[i]
@test componentwise_pdf(g, x_i) ≈ vec(P0[i,:])
@test componentwise_logpdf(g, x_i) ≈ vec(LP0[i,:])
@test @inferred(pdf(g, x_i)) ≈ mix_p0[i]
@test @inferred(logpdf(g, x_i)) ≈ mix_lp0[i]
@test @inferred(componentwise_pdf(g, x_i)) ≈ vec(P0[i,:])
@test @inferred(componentwise_logpdf(g, x_i)) ≈ vec(LP0[i,:])
end
#=
@show g
@show size(X)
@show size(mix_p0)
=#
@test pdf(g, X) ≈ mix_p0
@test logpdf(g, X) ≈ mix_lp0
@test componentwise_pdf(g, X) ≈ P0
@test componentwise_logpdf(g, X) ≈ LP0
@test @inferred(pdf(g, X)) ≈ mix_p0
@test @inferred(logpdf(g, X)) ≈ mix_lp0
@test @inferred(componentwise_pdf(g, X)) ≈ P0
@test @inferred(componentwise_logpdf(g, X)) ≈ LP0

# sampling
if ismissing(rng)
Expand Down Expand Up @@ -172,8 +172,8 @@ end
"rand(rng, ...)" => MersenneTwister(123))

@testset "Testing UnivariateMixture" begin
g_u = MixtureModel(Normal, [(0.0, 1.0), (2.0, 1.0), (-4.0, 1.5)], [0.2, 0.5, 0.3])
@test isa(g_u, MixtureModel{Univariate, Continuous, Normal})
g_u = MixtureModel(Normal{Float64}, [(0.0, 1.0), (2.0, 1.0), (-4.0, 1.5)], [0.2, 0.5, 0.3])
@test isa(g_u, MixtureModel{Univariate,Continuous,<:Normal})
@test ncomponents(g_u) == 3
test_mixture(g_u, 1000, 10^6, rng)
test_params(g_u)
Expand Down