I have to add some regularizer term to the likelihood equation containing beta terms. I am unable to figure out from the code which part of the code does the M-step update for the conditional multinomial parameter beta

to make it run example , we sould:
1.use --training_iterations , not ----number_of_iterations
2.express the ap.tar.gz and copy the file to root like:

ubgpu@ubgpu:~/github/PyLDA$ tree
.
├── doc.dat
├── input
│   ├── ap
│   └── ap.tar.gz
├── output
├── raw
│   └── ap.tar.gz
├── README.md
├── src
│   └── lda
│   ├── hybrid.py
│   ├── inferencer.py
│   ├── init.py
│   ├── init.pyc
│   ├── launch_test.py
│   ├── launch_train.py
│   ├── monte_carlo.py
│   └── variational_bayes.py
├── test.dat
├── train.dat
└── voc.dat

6 directories, 15 files
ubgpu@ubgpu:/github/PyLDA$ cd src/
ubgpu@ubgpu:/github/PyLDA/src$ python -m lda.launch_train --input_directory=../ input/ap --output_directory=../output/ --number_of_topics=10 --training_iterations=100 --inference_mode=0
successfully load all training docs from /home/ubgpu/github/PyLDA/train.dat...
successfully load all the words from /home/ubgpu/github/PyLDA/voc.dat...
========== ========== ========== ========== ==========
output_directory=../output/../150730-225540-lda-I100-S10-K10-aa0.100000-ab0.000147-im0/
input_directory=..
corpus_name=..
training_iterations=100

In LDA's slice sampling code:

Line 118: r_log_alpha_alpha = old_log_alpha_alpha + hyper_parameter_step;

Should it be:
r_log_alpha_alpha = l_log_alpha_beta + hyper_parameter_step?

(Similar issue with beta as well)

On http://projecteuclid.org/download/pdf_1/euclid.aos/1056562461's page 715:

L = x0 - w * U
R = L + w

Hi @kzhai ,
it's a nice work.
when i am reading you code , i find that the iteration of phi is a little different from the iteration formular:

where is a "lambda" rather then a "eta" in you code,
log_phi = E_log_eta[:, term_ids].T + numpy.tile(scipy.special.psi(gamma_values[[doc_id], :]), (word_ids[doc_id].shape[0], 1));