MCMC inference algorithms for normalized random measure mixtures
- Marginal Gibbs Sampler
- Split-Merge Sampler
- Tree-guided MCMC (NIPS 2015, Paper: http://arxiv.org/abs/1511.05650)
Coded & tested with Microsoft Visual Studio 2013 on Windows machine
Requires Eigen library (http://eigen.tuxfamily.org)
Usage: type nrmm data output bm nrm init sampler params after building
- data: data name (e.g., toy, 10k, nips)
- output: output folder name
- bm: base measure, 0 for NormalWishart and 1 for Multinomial-Dirichlet
- nrm: NRM, 0 for DP and 1 for NGGP
- init: initialization option, 0 for exact IBHC and 1 for noisy IBHC (see experimental section of the paper)
- sampler: sampler, 0 for Gibbs, 1 for split-merge and 2 for TGMCMC
- params: parameters for samplers
- nrmm data output bm nrm init 0 subset et_thres
- nrmm data output bm nrm init 1 subset et_thres
- subset for subset size (see paper) and et_thres for total running time
- nrmm data output bm nrm init 2 num_sm depth et_thres
- num_sm for the parameter G and depth for the parameter D in the paper
Demo: after building in Relase mode, put sample_script.bat in Relase folder and run to produce results for toy dataset.
To see log-likelihood traces, run display_results.m in results folder.
The nips data was accquired from https://archive.ics.uci.edu/ml/datasets/Bag+of+Words.