Hierarchical Dirichlet Process (with Split-Merge Operations)

(C) Copyright 2010, Chong Wang and David Blei written by Chong Wang, [email protected].

---

This is a C++ implementation of hierarchical Dirichlet process for topic modeling.

The split-merge algorithm is preliminary.

Note that this code requires the Gnu Scientific Library, http://www.gnu.org/software/gsl/

---

TABLE OF CONTENTS

A. COMPILING

B. POSTERIOR INFERENCE

C. INFERENCE ON NEW DATA

D. PARAMETER SETTINGS

E. PRINTING TOPICS

---

A. COMPILING

Type "make" in a shell. Make sure the GSL is installed. You may need to change the Makefile a bit.

B. TRANING (POSTERIOR INFERENCE)

The following shows an example of performing posterior inference on a set of documents

hdp --train_data data --directory my_model_dir 

Data format:

--data points to a file where each line is of the form (the LDA-C format):

[M] [term_1]:[count] [term_2]:[count] ...  [term_N]:[count]

where [M] is the number of unique terms in the document, and the [count] associated with each term is how many times that term appeared in the document.

The sampler will produce some files in the --directory,

• *-topics.dat: the word counts for each topic, with each line as a topic

• *.bin: the binary model file used for inference on new data.

• state.log: various information to monitor the Markov chain.

• *-word-assignments.dat: print each word's assignment to the topic and the table, which is in R-friendly format d w z t

  d: document id w: word id z: topic index t: table index (only for document level. If you only analyze the topics, this is irrelevant.)

More parameter settings, run hdp --help

C. INFERENCE ON NEW DATA

To perform inference on a different set of data (in the same format as before), run:

hdp --test_data  data --directory my_model_dir 

where --directory is the saving directory from training.

The sampler will produce some files in the --directory,

• test-*-topics.dat: the word counts for each topic, with each line as a topic
• test*-word-assignments.dat: print each word's assignment to the topic and the table, which is in R-friendly format.
• test.log: various information to monitor the Markov chain.
• test-*.bin: the binary model file used for inference on newer data.

More parameter settings, run hdp --help

D. PARAMETER SETTINGS

The meaning of the parameters is the same as in the in the following paper

Y. Teh, M. Jordan, M. Beal, and D. Blei. Hierarchical Dirichlet processes. Journal of the American Statistical Association, 2006. 101[476]:1566-1581

E. PRINTING TOPICS

A R script (print.topics.R) is included to print topics. Make sure it is executable. (chmod +x print.topics.R) For example,

print.topics.R mode-topics.dat vocab.dat topics.dat 10

will produce a topic list with top 10 words selected. For help, run print.topics.R

---

My Notes:

INSTALL (OS X)

brew install gsl
make

TRAIN

cd hdp-faster
./hdp --help

mkdir train_dir
./hdp --train_data /Volumes/HDD2/ren_data/dev_hdd/bluebrain/9_lda/_preproc/20130617_pubmed_preproc/dca/dca.txtbag.noheader --directory train_dir

GSL on viz-cluster

cd ~/public/lib/
wget ftp://ftp.gnu.org/gnu/gsl/gsl-1.7.tar.gz
tar -zxvf gsl-1.7.tar.gz
rm gsl-1.7.tar.gz
cd gsl-1.7

./configure --prefix=/home/richarde/public/lib
make
make install

Now in hdp-faster's Makefile:

GSL_INCLUDE = /home/richarde/public/include
GSL_LIB = /home/richarde/public/lib

Then

make linux linux-d


./hdp --verbose --train_data /home/richarde/public/corpora/pubmed_abstracts_100k.ldac-txtbag --directory train_dir