Gamma-Exponential Process (GEP) Software

Usage overview

After following the installation instructions, the current features of the package are:

• loading a collection of partially observed times series, including lists of points to hold out for evaluation
• running a PMCMC sampler on this data
• after each scan, imputing the held out points using the MAP from the samples collected so far, and computing the reconstruction error.

Example usage

To test the software:

• Go to the root directory of the distribution
• Type:
  java -Xmx2g gep.GEPMain -dataType LOAD_PLAIN_TXT -pathToHeldoutSpec data/synthetic-heldout.txt -flatDataFile data/synthetic-data.txt
• Go in state/execs/. These directories are created by Percy Liang's fig system (http://cs.stanford.edu/~pliang/software/), and the last exec folder (x.exec) contains information about the last run (see below).

Input options

The following command line options are currently supported (you can see the full list by typing java gep.GEPMain -printOptionsAndExit)

• pmcmc.init*: initial values for the parameters/hyperparameters
• N: number of particles in the PMCMC algorithm
• maxMCMCIters: number of MCMC iterations
• saveReconstruction: if set to true, the reconstructed/imputed value will be saved at each iteration (instead of recording only the average errors). The output can be found in the exec folder under the reconstructions directory.
• pathToHeldoutSpec: path to a file that specifies which points to hold out. See data/synthetic-heldout.txt for information on the format of this file.
• flatDataFile: path to the file containing partially observed time series. See data/synthetic-data.txt for information on the format of this file.

Contents of output folder

As mentionned earlier, each folder in state/execs/ contains information about each run. See http://cs.stanford.edu/~pliang/software/ for more information, but here are some examples of what you can find there:

• options.map: options used for this run
• log: a copy of the stdout of this run
• results/*: various statistics on the error, acceptance ratio, nHidden states