WarpLDA: Cache Efficient Implementation of Latent Dirichlet Allocation
Introduction
WarpLDA is a cache efficient implementation of Latent Dirichlet Allocation, which samples each token in O(1).
Installation
Prerequisites:
- GCC (>=4.8.5)
- CMake (>=2.8.12)
- git
- libnuma
- CentOS:
yum install libnuma-devel
- Ubuntu:
apt-get install libnuma-dev
- CentOS:
Clone this project
git clone https://github.com/thu-ml/warplda
Install third-party dependency
./get_gflags.sh
Download some data, and split it as training and testing set
cd data
wget https://raw.githubusercontent.com/sudar/Yahoo_LDA/master/test/ydir_1k.txt
head -n 900 ydir_1k.txt > ydir_train.txt
tail -n 100 ydir_1k.txt > ydir_test.txt
cd ..
Compile the project
./build.sh
cd release/src
make -j
Quick-start
Format the data
./format -input ../../data/ydir_train.txt -prefix train
./format -input ../../data/ydir_test.txt -vocab_in train.vocab -test -prefix test
Train the model
./warplda --prefix train --k 100 --niter 300
Check the result. Each line is a topic, its id, number of tokens assigned to it, and ten most frequent words with their probabilities.
vim train.info.full.txt
Infer latent topics of some testing data.
./warplda --prefix test --model train.model --inference -niter 40 --perplexity 10
Data format
The data format is identical to Yahoo! LDA. The input data is a text file with a number of lines, where each line is a document. The format of each line is
id1 id2 word1 word2 word3 ...
id1, id2 are two string document identifiers, and each word is a string, separated by white space.
Output format
WarpLDA generates a number of files:
.vocab
(generated by .format
)
Each line of it is a word in the vocabulary.
.info.full.txt
(generated by warplda -estimate
)
The most frequent words for each topic. Each line is a topic, with its topic it, number of tokens assigned to it, and a number of most frequent words in the format (probability, word)
. The number of most frequent words is controlled by -ntop
. .info.words.txt
is a simpler version which only contains words.
.model
(generated by warplda -estimate
)
The word-topic count matrix. The first line contains four integers
<size of vocabulary> <number of topics> <alpha> <beta>
Each of the remaining lines is a row of the word-topic count matrix, represented in the libsvm sparse vector format,
<number of elements> index:count index:count ...
For example, 0:2
on the first line means that the first word in the vocabulary is assigned to topic 0 for 2 times.
.z.estimate
(generated by warplda -estimate
)
The topic assignments of each token in the libsvm format. Each line is a document,
<number of tokens> <word id>:<topic id> <word id>:<topic id> ...
.z.inference
(generated by warplda -inference
)
The format is the same as .z.estimate
.
Other features
-
Use custom prefix for output
-prefix myprefix
-
Output perplexity every 10 iterations
-perplexity 10
-
Tune Dirichlet hyperparameters
-alpha 10 -beta 0.1
-
Use UCI machine learning repository data
wget https://archive.ics.uci.edu/ml/machine-learning-databases/bag-of-words/vocab.nips.txt wget https://archive.ics.uci.edu/ml/machine-learning-databases/bag-of-words/docword.nips.txt.gz gunzip docword.nips.txt.gz ./uci-to-yahoo docword.nips.txt vocab.nips.txt -o nips.txt head -n 1400 nips.txt > nips_train.txt tail -n 100 nips.txt > nips_test.txt
License
MIT
Reference
Please cite WarpLDA if you find it is useful!
@inproceedings{chen2016warplda,
title={WarpLDA: a Cache Efficient O(1) Algorithm for Latent Dirichlet Allocation},
author={Chen, Jianfei and Li, Kaiwei and Zhu, Jun and Chen, Wenguang},
booktitle={VLDB},
year={2016}
}