Here we provide efficient Rcpp implementation for fitting Bayesian non-parametric mixture models. Specifically, we provide MCMC-based inference for
- Dirichlet Process mixture (DPM)
- Mixture of Finite mixtures (MFM) (Miller and Harrison, 2017)
for the case where mixture components are assumed to follow a Normal-Inverse-Wishart distribution.
We have implemented two MCMC moves:
- The collapsed Gibbs sampler
- The Sequentially Allocated Split-Merge sampler (Dahl 2003)
Our implementation exposes the C++ classes to R via Rcpp modules. All MCMC updates utilise the rank-one updates for the Cholesky decomposition.
Example usage:
library(mixtureModels)
set.seed(1)
N <- 50
mu_list <- list(c(-1, -2), c(1, 1), c(3, -1))
Sigma_list <- list(0.4*diag(2), 0.4*diag(2), 0.4*diag(2))
data <- generate_from_2D_GMM(N, mu_list, Sigma_list)
X <- data$X
n_iter <- 20
plot_freq <- 5
# initialise cluster allocations
z <- rep(1, N)
# initialise mixture model object
m <- mixture(X, z, type = "DPM")
# do MCMC sampling
for(i in 1:100){
m$collapsed_gibbs()
m$split_merge()
}
plot_2D_GMM(m)
For another demonstration see this toy example.