1. CTPF goes negative on first iteration
2. change UTF8String in README to String
3. put package back in official listing
4. smooth out fLDA and fCTM
5. figure out way to preserve ordering of documents so showdocs() is comprehensible
6. Use feather or paraquet or some fead read-in binary file format for default corpora in order to speed up their loading times.
7. write run tests for package, see issue #3
8. pre-compute ELBOs, or derive formula for various M and K
9. deal with counts situation
10. why is nu variable used in Newton's method for alpha in LDA but not vsq in CTM?
11. when different lex keys map to the same word, need to merge these in fixcorp functions
12. write detect topic code, see issue #2
13. save/load trained models see issue #13
14. gpuCTM signal (6): Abort trap: 6 when too many topics
15. stream documents into GPU models
16. possibly try to collapse phi in gpuLDA and gpuCTM
17. work on HMTM
18. add print_elbo option
19. need to deal with long stretches of negatives some how, see fLDA with 5000 docs and 9 topics.
21. what variables *must* be finite in check_model, currently sigma and invsigma not required to be finite
22. should sigma for CTM be required to be positive definite or just positive semidefinite? test with 2 topics maybe. could set inv(model.sigma) to pinv(model.sigma)
23. get NaNs with LDA with 1 topic. Basically, gamma and alpha go to Inf with 1 topic because that increasing the likelihood
24. pull request to Distributions.jl for Dirichlet with length 1 parameter, entropy should always be zero, but it's not for maxfloat(1.0).
25. findall for docs in corpus not working, need to understand how to do this correctly
26. should possibly define abstract types, AbstractLDA = Union{LDA, fLDA, gpuLDA}, same with CTM and CTPF
27. need to deal with beta and fbeta in filtered models, predict function, etc.
28. need to deal with mu and lambda in CTM with floatmax, etc.
29. when you initialize beta as ones(K, V) / V, it doesn't train.
30. showtitles function
31. errors in showdocs, showtitles need to be standardized, show at the bad doc index, or check all indices at the beginning.
32. need to set up predict to return GPU models

33. in readme explain how people can access raw data: gamma and beta mostly, but also recommendations for CTPF.
34. change hardcoded K, V, M, etc. in check_model functions to string imputed values for model








5. new to temp
4. camelCase to under_score
11. tab align C code
3. decide on fixmodel! function
13. include perplexity function (possibly, not sure if I want to do this)
15. change comments to new style formatting
16. fix checkcorp! situation
17. make extended integers datatype so that chkelbo can be Z U {Inf}
19. fix DTM
19. make sure elbo is matched to model data when it goes into training
20. isposdef fails for model.sigma in fCTM, not exactly symmetric, isposdef function too stringent


