This module offers relatively fast implementations of a trust-region method for stochastic variational inference and related algorithms applied to latent Dirichlet allocation. These are realized in the following models:

• OnlineLDA (Hoffman et al., 2010; 2013; Theis & Hoffman, 2015)
• BatchLDA (Blei et al., 2013)
• CumulativeLDA (Broderick et al., 2013)

Additional features include adaptive learning rates (Ranganath et al., 2013) and automatic tuning of hyperparameters via empirical Bayes.

• Python >= 2.7.3
• NumPy >= 1.6.1

I have tested the code with the versions above, but older versions might also work.

python setup.py build
python setup.py install

To test whether the installation worked, you can run the following.

nosetests code/trlda

from trlda.models import OnlineLDA
from trlda.utils import load_documents

# create model
model = OnlineLDA(
	num_words=7000,
	num_topics=100,
	num_documents=1000000,
	alpha=.1,
	eta=.2)

# train model for 10 epochs with a batch size of 200
for epoch in range(10):
	for documents in load_documents('data_train.dat', 200):
		model.update_parameters(
			docs=documents,
			max_iter_tr=10,
			max_iter_inference=20,
			kappa=.7,
			tau=100.,
			update_alpha=True,
			update_eta=True)

L. Theis and M. D. Hoffman
A trust-region method for stochastic variational inference with applications to streaming data
International Conference on Machine Learning, 2015