March 19, 2020: The articles in wiki10k and wiki1k are not guaranteed to be disjoint from each other. Is the random seed really enabling replicability?

March 20, 2020: the function that is hardest to convert to Python 3 from Python 2 is wikirandom.py, so we leave as is. We also leave onlinewikipedia.py as Python 2 since we don't use it.

March 21, 2020: Currently representing the variational parameter of per-word topic assignment explicitly in SB-LDA's do-e-step. Correctness is the priority now. Later, to save time and memory, might switch to implicit representation.

March 22, 2020: Skeleton of SB-LDA is completed. It's encouraging that as training progresses, held-out log-likelihood improves but we're missing unit tests. For instance, we should report if the e-step fails to converge.

It takes 30 minutes to train LDA 1/K but 2 hours to train SB-LDA. The LL of SB-LDA is 
significantly worse than SB-LDA; could be issue with batch size (SVI paper Figure 13).

March 25, 2020: To isolate the effect of optimization, should load the topics learned from LDA 1/K to initialize the training of SB-LDA, and vice-versa.