Last active
July 1, 2019 23:44
-
-
Save cscherrer/72062a3b9f264a328f00159d84a61b98 to your computer and use it in GitHub Desktop.
Variational Importance Sampling
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
using Pkg | |
Pkg.add.( | |
["Distributions" | |
, "MonteCarloMeasurements" | |
, "StatsFuns" | |
, "LaTeXStrings" | |
, "Plots"]) | |
using Distributions | |
using MonteCarloMeasurements | |
using StatsFuns | |
# Just some setup so inequalities propagate through particles | |
for rel in [<,>,<=,>=] | |
register_primitive(rel) | |
end | |
function fromObs(x,y) | |
function logp(α,β) | |
ℓ = 0.0 | |
ℓ += logpdf(Normal(0,1), α) | |
ℓ += logpdf(Normal(0,2), β) | |
yhat = α .+ β .* x | |
ℓ += sum(logpdf.(Normal.(yhat, 1), y) ) | |
ℓ | |
end | |
end | |
drawcat(ℓ, k) = [argmax(ℓ + Particles(1000,Gumbel())) for j in 1:k] | |
asmatrix(ps...) = Matrix([ps...])' | |
# Kish's effective sample size, | |
# $n _ { \mathrm { eff } } = \frac { \left( \sum _ { i = 1 } ^ { n } w _ { i } \right) ^ { 2 } } { \sum _ { i = 1 } ^ { n } w _ { i } ^ { 2 } }$ | |
function n_eff(ℓ) | |
logw = ℓ.particles | |
exp(2 * logsumexp(logw) - logsumexp(2 .* logw)) | |
end | |
function f(a,b) | |
# generate data | |
x = rand(Normal(),100) | |
yhat = a .+ b .* x | |
y = rand.(Normal.(yhat, 1)) | |
# generate p | |
logp = fromObs(x,y) | |
runInference(x,y,logp) | |
end | |
function runInference(x,y,logp) | |
N = 1000 | |
# initialize q | |
q = MvNormal(2,100000.0) # Really this would be fit from a sample from the prior | |
α,β = Particles(N,q) | |
m = asmatrix(α,β) | |
ℓ = sum(logp(α,β)) - Particles(logpdf(q,m)) | |
numiters = 60 | |
elbo = Vector{Float64}(undef, numiters) | |
for j in 1:numiters | |
α,β = Particles(N,q) | |
m = asmatrix(α,β) | |
ℓ = logp(α,β) - Particles(logpdf(q,m)) | |
elbo[j] = mean(ℓ) | |
ss = suffstats(MvNormal, m, exp(ℓ - maximum(ℓ)).particles .+ 1/N) | |
q = fit_mle(MvNormal, ss) | |
end | |
(α,β,q,ℓ,elbo) | |
end | |
(α,β,q,ℓ,elbo) = f(3,4) | |
using LaTeXStrings | |
using Plots | |
plot(1:60, -elbo | |
, xlabel="Iteration" | |
, ylabel="Negative ELBO" | |
, legend=false | |
, yscale=:log10) | |
xticks!([0,20,40,60], [L"0",L"20", L"40",L"60"]) | |
yticks!(10 .^ [3,6,9,12], [L"10^3", L"10^6",L"10^9",L"10^{12}"]) | |
savefig("neg-elbo.svg") |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment