Skip to content

Latest commit

 

History

History
83 lines (59 loc) · 2.51 KB

README.md

File metadata and controls

83 lines (59 loc) · 2.51 KB
Raw

Deep Exponential Family

Reference

Deep Exponential Families by Rajesh Ranganath, Linpeng Tang, Laurent Charlin, and David M. Blei, AISTATS 2015.

Requirements

  • armadillo
  • boost 1.55
  • OpenMP
  • GSL
  • g++ >= 4.7

Instructions to Build and Run

Configuring: ./waf configure

Building: ./waf build (binary is build/def_main)

Running: def reads its options from a config file and from the command line. ./build/def_main --help shows the command line options.

We give a full example below (including a sample config file) of a def running a dataset of wikipedia articles.

Input Format for Text

The header:

n_examples n_words

Followed by n_examples examples, each has two lines:

example_ind example_words
word_ind0 word_count0 word_ind1 word_count1 ...

Comprehensive Example

  1. Data. We have pre-processed a corpus of wikipedia articles containing 1000 train articles, 500 validation and, 500 test articles. The data are available in folder wikpedia

  2. Configuration file. The configuration file will be read by def and contains all options regarding the model to be trained (e.g., number of layers, size of layers, distribution of global and local variables). Example configuration for the wikipedia dataset is available in folder wikipedia. Note that this file will look for the dataset in a directory specified by the WIKIPEDIA_DEF environment variable.

  3. Running. Here is an example invocation:

cd deep-exponential-families
# define environment variable used in def_wikipedia_50_25_10.ini
export WIKIPEDIA_DEF=`pwd`/wikipedia
./build/def_main --v=3 --folder=experiments/def_wikipedia --algo=rmsprop --rho=.2 --samples=64 --max_examples=1000000 --model=wikipedia/def_wikipedia_50_25_10.ini --batch=10000 --batch_order=rand --threads=5 --test_interval=5 --iter=2000

The above settings (including the values provided in the configuration file) are the ones we used in the paper. We have found these settings to be useful across several datasets.

Topic visualization

We show some of the tools that we have used to explore the def fits in this DEF IPython Notebook.