Skip to content
master
Go to file
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
cpp
 
 
 
 
 
 
 
 

README.md

Parallel Streaming Wasserstein Barycenters

This repository contains the supporting code for the paper:

Matthew Staib, Sebastian Claici, Justin Solomon, Stefanie Jegelka. Parallel Streaming Wasserstein Barycenters. In Advances in Neural Information Processing Systems 31, 2017.

@inproceedings{staib2017parallel,
 author = {Staib, Matthew and Claici, Sebastian and Solomon, Justin and Jegelka, Stefanie},
 title = {Parallel Streaming {Wasserstein} Barycenters},
 booktitle = {Advances in Neural Information Processing Systems 31},
 year = {2017}
}

Dependencies

Software

Data

Getting started

  1. Edit the Makefile to point to your local copys of the dependencies.
  2. Compile the main barycenter function via make barycenter_mpi.
  3. Then experiments are run by calling barycenter_mpi with various arguments. We include example shell scripts run_vmf_experiment.sh and run_skin_experiment_full.sh. Full explanation of the various parameters can be found below (and also in parse_args.cpp:
e,experiment --  Which experiment to run (skin,vmf,logit,gaussian) (the first two were run for the paper)
s,subsets --  Number of subsets to split into (for WASP)
k,skip --  Number of timesteps between MCMC samples
N,support --  Number of support points
o,outdir --  Output directory for .h5 files
d,saveincrement --  How often to save .h5 files
a,stepsize --  Stepsize for gradient ascent
w,movingwindow --  Width of histogram moving window (or 0 to keep full history)
m,driftrate --  Rate of drift of VMF distributions
b,burniniters --  Number of burn-in iters for MCMC chain
f,fullsampler --  Whether to get samples from the full MCMC chain
p,datapoints --  Number of datapoints to use (for the skin example; useful for testing more quickly)
  1. Plots can then be generated via the included Matlab scripts (be sure to point these scripts to your output directory!). Specifically,
  • skin_wasp_compare compares our stochastic approach to the standard linear programming barycenter algorithm for WASP.
  • plot_cpp_output_vmf produces output images for our Von Mises-Fisher experiments
  • plot_cpp_output_skin gives some example diagnostic plots for the UCI experiments
  • the scripts in scripts produce convergence plots of our barycenter estimate for the UCI experiments

About

Supporting code for "Parallel Streaming Wasserstein Barycenters"

Resources

Releases

No releases published

Packages

No packages published
You can’t perform that action at this time.