lgw1860 / LDA

Three open source versions of LDA with collapsed Gibbs Sampling, modified by nanjunxiao

LDA: Latent Dirichlet Allocation

This repository includes three open source versions of LDA with collapsed Gibbs Sampling, modified by nanjunxiao.

GibbsLDA++ single thread,written in C++

ompi-lda multi-node/multi-threads, written in C++

online_twitter_lda multi-threads,written in Python

collapsed Gibbs LDA reference : my blog

What's New

1. GibbsLDA++

fixed bugs:

1). memory leakage. 'delete[] p' instead of 'delete p',when p points to an Array.

2). Array out of bound. (double)random() / RAND_MAX in [0,1]

int topic = (int)(((double)random() / RAND_MAX) * K);  -->  int topic = (int)(((double)random() / RAND_MAX + 1) * K);
double u = ((double)random() / RAND_MAX) * p[K - 1];   -->  double u = ((double)random() / RAND_MAX + 1) * p[K - 1];

2. ompi-lda

fixed bug:

1). infer.cc bugs.

2). rm 'sampler.UpdateModel(corpus)' in lda.cc.

add features:

1). add theta twords file output.

2). add partial boost's hpp/cpp in include dir, so can make directly.

3. online_twitter_lda

add features:

1). add theta phi mat file output.

TODO

ompi-lda

1). twordsnum can configure.

2). rewrite cmd_flag without boost, so can remove include dir.

3). rewrite makefile.