DpMM

A set of Dirichlet Process Mixture Model (DPMM) sampling-based inference algorithms.

This is research code and builds on the following two papers (please cite them appropriately):

• [1] Jason Chang and John W. Fisher III. Parallel Sampling of DP Mixture Models using Sub-Clusters Splits, NIPS 2013.
• [2] Julian Straub, Jason Chang, Oren Freifeld and John W. Fisher III. Dirichlet Process Mixture Model for Spherical Data, AISTATS 2015.

Dependencies

This code depends on the following other libraries: Eigen3, Boost, CUDA. Optionally OpenMP.

It has been tested under Ubuntu 14.04 with:

• Eigen3 (3.0.5)
• Boost (1.52)
• CUDA (6.5)

It has been tested under Windows 7 with:

• Visual Studio 2012
• Eigen3 (3.2.3)
• Boost (1.57)
• CUDA (6.5)

It has been tested under Mac OS (10.9.4) with:

• clang-600.51
• Eigen3 (3.2.1)
• Boost (1.55)
• CUDA (6.5)
• No OpenMP

Compiling

• Linux:

  Install Eigen3 and Boost

  sudo apt-get install libeigen3-dev libboost-dev 
  

  Install the appropriate CUDA version matching with your nvidia drivers. On our machines we use nvidia-340-dev with libcuda1-340 cuda-6-5 cuda-toolkit-6-5

  Clone this repository and compile the code:

  git clone https://github.com/jstraub/dpMM; cd dpMM; mkdir build; cd
  build; cmake ..; make -j6;
  

Getting Started

After compiling the sampler executable as described above, get started quickly by looking at python/dpmmSampler.py. It simply loads a dataset and runs the sub-cluster split/merge algorithm with different base measures:

python ./python/dpmmSampler.py -i ./data/rndSphereDataIwUncertain.csv -b DpNiwSphereFull -T 400
python ./python/dpmmSampler.py -i ./data/rndSphereDataIwUncertain.csv -b DpNiw -T 400

Where DpNiwSphereFull is for the DP-TGMM [2] and DpNiw for the standard DP-GMM [1]. After finishing the specified number of iterations (via the -T option) the log likelihood as well as the number of clusters over the iterations is shown. Note that the true number of clusters of the data in ./data/rndSphereDataIwUncertain.csv is 30 where each of the clusters has 333 data-points. A groundtruth labeling can be found in ./data/rndSphereDataIwUncertain_gt.lbl. Make sure you compiled the .cpp code beforehand since the python script just wraps the call to dpmmSampler.

Executables

• dpmmSampler: Sampler for Dirichlet process mixture model (DPMM) inference using different algorithms. It is usually preferable to use the python script in python/dpmmSampler.py which wraps around this executable to provide an easier-to-use interface.

Allowed options: -h [ --help ] produce help message --seed arg seed for random number generator -N [ --N ] arg number of input datapoints -D [ --D ] arg number of dimensions of the data -T [ --T ] arg number of sampler iterations -a [ --alpha ] arg alpha parameter of the DP (if single value assumes all alpha_i are the same -K [ --K ] arg number of initial clusters -n [ --nopropose ] flag to disable the propsal of splits and merges -s [ --silhouette ] flag to enable output of silhouett value of the last iteration --shuffle shuffle the data before processing --base arg which base measure to use (StickNiw, DpNiw (DP-GMM), DpNiwSphereFull (DP-TGMM), DpNiwSphere, NiwSphere, DirNiwSphereFull NiwSphereUnifNoise, CrpvMF, DirvMF right now) -p [ --params ] arg parameters of the base measure --brief arg brief parameters of the base measure (ie Delta = deltaI; theta=tones(D) -i [ --input ] arg path to input dataset .csv file (rows: dimensions; cols: different datapoints) -o [ --output ] arg path to output labels .csv file (rows: time; cols: different datapoints) ``` Parameter arguments by model (compare src/dpmmSampler.cpp and python/dpmmSampler.py): - DpNiw (DP-GMM): nu kappa theta0 ... thetaD Delta00 Delta01 ... DeltaDD - DpNiwSphereFull (DP-TGMM): nu Delta00 Delta01 ... Delta(D-1)(D-1)

• generateSphericalData: generate spherical data for synthetic data experiments.

Allowed options: -h [ --help ] produce help message --seed arg seed for random number generator -N [ --N ] arg number of input datapoints -D [ --D ] arg number of dimensions of the data -K [ --K ] arg number of initial clusters -n [ --nu ] arg nu parameter of IW from which sigmas are sampled -a [ --minAngle ] arg min angle between means on sphere -d [ --delta ] arg delta of NIW -o [ --output ] arg path to output labels and data .csv file (rows: time; cols: different datapoints) ```

Collaborators

Julian Straub (jstraub) and Randi Cabezas (rcabezas)