README.md

DPMM.jl

This repository is a research work on parallel dirichlet process mixture models and clustering on Julia by Ekin Akyürek with supervision of John W. Fischer III.

Getting Started

Demo:

  gm = GridMixture(2)
  X, clabels = rand_with_label(gm,100000)
  fit(X; ncpu=3) # runs parallel split-merge algorithm

Visual Demo (requires OpenGL) :

  gm = GridMixture(2)
  X, clabels = rand_with_label(gm,100000)
  scene = setup_scene(X)
  fit(X; ncpu=3, scene=scene) # visualize parallel split-merge algorithm

For details please see the function documentation

Technical Report

Algorithms

1. Collapsed Gibbs Sampler

labels = fit(X; algorithm=CollapsedAlgorithm) # serial collapsed

2. Quasi-Collapsed Gibbs Sampler

labels = fit(X; algorithm=CollapsedAlgorithm, quasi=true) # quasi & serial collapsed
labels = fit(X; algorithm=CollapsedAlgorithm, quasi=true, ncpu=4) # quasi & parallel collapsed

3. Direct Gibbs Sampler

labels = fit(X; algorithm=DirectAlgorithm) # direct
labels = fit(X; algorithm=DirectAlgorithm ncpu=4) # parallel direct

4. Quasi-Direct Gibbs Sampler

labels = fit(X; algorithm=DirectAlgorithm, quasi=true) # quasi direct gibbs algorithm
labels = fit(X; algorithm=DirectAlgorithm, quasi=true, ncpu=4) # quasi & parallel direct gibbs direct gibbs

5. Split-Merge Gibbs Sampler

labels = fit(X; algorithm=SplitMergeAlgorithm) # split-merge
labels = fit(X; algorithm=SplitMergeAlgorithm, ncpu=4) # parallel split-merge

Parallel Benchmarking

Run below command:

julia --project test/parallel_benchmark.jl  --N 1000000 --K 6 --Kinit 1 --ncpu 4

• Results-I: Time (sec) to run 100 DP-GMM iterations for d=2, N=1e6, K=6.

Code	ncpu=1	ncpu=2	ncpu=4	ncpu=8
C++	76.94	40.57	22.23	13.01
DPMM.jl	75.71	41.54	20.86	12.77
Julia-BNP	1101.97	572.50	345.58	172.30

• Results-II: Time (sec) to run 100 DP-MNMM iterations for d=100, N=1e6, K=6.

Code	ncpu=1	ncpu=2	ncpu=4	ncpu=8
C++	134.25	77.55	40.97	23.60
DPMM.jl	113.131	68.46	45.55	30.79
Julia-BNP	234.40	136.43	87.34	55.10