trappmartin
commented
4 years ago You might want to do something like this instead.
@model dp_gmm_crp_delayed(y) = begin
nobs = length(y)
alpha ~ Gamma(1, 0.1) # mean = a*b
rpm = DirichletProcess(alpha)
# Base measure.
H_mu = Normal(0, 3)
H_sigma = InverseGamma(2, 3)
# Latent assignment.
z = tzeros(Int, nobs)
n = tzeros(Int, nobs)
K = 0
mu = tzeros(Float64, nobs)
sigma = tzeros(Float64, nobs)
for i in 1:nobs
z[i] ~ ChineseRestaurantProcess(rpm, n)
n[z[i]] += 1
if z[i] > K
mu[z[i]] ~ H_mu
sigma[z[i]] ~ H_sigma
K += 1
end
y[i] ~ Normal(mu[z[i]], sqrt(sigma[z[i]]))
end
endThis might contain a bug as I haven't tested it.
In general, it is beneficial to delay sampling of RVs for SMC and PG, as in this code. I wasn't aware of this myself, Hong told me a while ago.
xukai92
mohamed82008
luiarthur
yebai
I simulated some data and fit the following model to the data. I get poor inference. i.e., the results don't really match up with the simulation truth. Particularly, I tend to learn 4 to 5 clusters when my data has 4 almost-equally-sized clusters, but there are usually only 1 or 2 big (dominating) clusters in the posterior inference.
So, my first question is am I abusing the API? In my DP gaussian mixtures of location and scale, I use a base measure (H) which is Normal x InverseGamma (two independent distributions) for the location (mu) and scale (sigma). Am I doing this correctly? (It runs, but I suspect I'm doing something outside the intended use.)
My second question is, if the model is implemented correctly, what might be the cause for poor inference? Admittedly, I'm not familiar with
SMC/PC. But does increasing the number of particles generally lead to better inference?Here is the complete notebook.