Initial commit. Very early stages of algorithm development.

olavurmortensen committed Sep 25, 2016

Fixed some errors.

olavurmortensen committed Sep 25, 2016

Added online algorithm, removed batch algorithm.

olavurmortensen committed Sep 27, 2016

Using max change instead of mean change criterion. Computing a differ… …

…ent likelihood measure. OnlineAtVb now extends (inherits) LdaModel. Other minor changes.

olavurmortensen committed Sep 29, 2016

Fixed some things with var_mu. Also was passing the wrong arguments t… …

…o eval_liklihood.

olavurmortensen committed Sep 30, 2016

Added 'offline' algorithm, and notebook for experiments.

olavurmortensen committed Sep 30, 2016

Fixed log normalization. Also changed symmetric initilization of hype… …

…rparameters. Updated notebook accordingly.

olavurmortensen committed Oct 9, 2016

Removed offline algorithm class as it is no longer necessary.

olavurmortensen committed Oct 9, 2016

Changed name of online algorithm class and file.

olavurmortensen committed Oct 9, 2016

Made some changes to how the likelihood is computed.

olavurmortensen committed Oct 10, 2016

Changed the name of the online algorithm again.

olavurmortensen committed Oct 10, 2016

Brought the offline algorithm back.

olavurmortensen committed Oct 10, 2016

Working on bound computation.

olavurmortensen committed Oct 11, 2016

Changed the way the data structure is prepared and how the model acce… …

…pts it. Still work to be done in that area.

olavurmortensen committed Oct 12, 2016

Cleaned the code up a bit. Added a simple method to get author topics.

olavurmortensen committed Oct 12, 2016

Removed some comments, mostly TODOs.

olavurmortensen committed Oct 12, 2016

Ran some very successful experiments on 286 documents. Offline algori… …

…thm works. Updated notebook.

olavurmortensen committed Oct 13, 2016

Changed the online algorithm according to all the changes that have h… …

…appened to the offline lately.

olavurmortensen committed Oct 14, 2016

Fixed mistake with mu variable.

olavurmortensen committed Oct 14, 2016

Fixed lambda update, multiplication by size of corpus was missing. Re… …

…moved author_prior_prob from mu update.

olavurmortensen committed Oct 14, 2016

Added a loop for passing over entire corpus. Discarded use of log_nor… …

…malize. Various other changes.

olavurmortensen committed Oct 14, 2016

Moved bound computation out of corpus-wide loop.

olavurmortensen committed Oct 14, 2016

Updated notebook.

olavurmortensen committed Oct 14, 2016

Computing rho in a different way. Added the possibility to evaluate o… …

…nly occasionally. Updated notebook.

olavurmortensen committed Oct 14, 2016

Implemented hyperparam MLE for eta and alpha in offline algo. Removed… …

… use of log_normalize in offline algo. Update notebook.

olavurmortensen committed Oct 17, 2016

Made it possible to sample a subset of documents in lambda update to … …

…speed up large experiments. Made it possible to initialize the model with LDA topics (lambda).

olavurmortensen committed Oct 17, 2016

Now, if LDA topics are supplied lambda is not estimated at all. Added… …

… a small number to mu and phi normalization term to avoid divide by zero. Made some comments (NOTEs and TODOs) about numerical stability.

olavurmortensen committed Oct 18, 2016

Updating notebook.

olavurmortensen committed Oct 19, 2016

Working on line search for hyperparam MLE.

olavurmortensen committed Oct 20, 2016

Made some structural changes to bound and log probability computation.

olavurmortensen committed Oct 20, 2016

In process of updating online algo w.r.t. changes in offline algo.

olavurmortensen committed Oct 20, 2016

Mostly updated the online algorithm according to changes that have be… …

…en done to the offline.

olavurmortensen committed Oct 21, 2016

Fixed a critical mistake in the online algorithm.

olavurmortensen committed Oct 21, 2016

Removed a redundancy in lambda update. Updated notebook.

olavurmortensen committed Nov 4, 2016

Making sure that the model is evaluated after the last iteration, if … …

…eval_every is different from 0. Various comment changes. Updated notebook.

olavurmortensen committed Nov 7, 2016

Updated notebook.

olavurmortensen committed Nov 8, 2016

Fixed mistake in interpolating gamma. Moved lambda update outside of … …

…'iterations' loop. Updated notebook.

olavurmortensen committed Nov 8, 2016

Working on an algorithm that tries to process each 'disjoint' set of … …

…authors independently in a mini-batch sort of way.

olavurmortensen committed Nov 8, 2016

Working on a minibatch algorithm. Updated notebook.

olavurmortensen committed Nov 9, 2016

Only updating the necessary expected log theta. Changed the name of O… …

…nlineAtVb2 to DisjointAtVb. Updated notebook. Other minor changes.

olavurmortensen committed Nov 11, 2016

Implemented a new algorithm. It is 5 times faster, more memory effici… …

…ent, and even gives better results.

olavurmortensen committed Nov 13, 2016

Moved all algorithms except the new online one to a 'temp' folder. Ve… …

…ctorization is an option, so I can test it (speed up etc.). Updated notebook.

olavurmortensen committed Nov 15, 2016

Changed the name of the main algorithm (and file). Made a new noteboo… …

…k for old tests, removed all old tests from main notebook. Removed references to old code in __init__.py.

olavurmortensen committed Nov 15, 2016

Cleaning up code. Removed or changed a lot of comments. Removed optio… …

…n of computing log probabilities, although the method still exists. Added a method of computing all the terms of the bound at once.

olavurmortensen committed Nov 16, 2016

Was computing the norm of phi incorrectly, fixed that, speed-up not a… …

…s large as first thought. Made a method for computing phinorm. Updating lambda in a different way. Implemented a numerically stable softmax (phi is a softmax).

olavurmortensen committed Nov 21, 2016

Working on numerically stable phi update and bound computation. Is no… …

…t converging the same way, so it is an option for now.

olavurmortensen committed Nov 22, 2016

Made a separate file for unvectorized code and stored it in , in case… …

… its needed it later.

olavurmortensen committed Nov 22, 2016

Implemented mini-batch algorithm.

olavurmortensen committed Nov 22, 2016

Some minor changes to old code.

olavurmortensen committed Nov 22, 2016

Updated notebook.

olavurmortensen committed Nov 22, 2016

In mini-batch algo, only the terms seen in the current chunk are upda… …

…ted, otherwise problems occur.

olavurmortensen committed Nov 23, 2016

Updated notebook. Finally getting decent results on the entire NIPS d… …

…ataset.

olavurmortensen committed Nov 23, 2016

Optimized phinorm computation by taking expElogbeta out of the loop.

olavurmortensen committed Nov 28, 2016

Computing the bound more efficiently (much faster now). Now not passi… …

…ng entire expElogtheta and expElogbeta matrices to compute_phinorm.

olavurmortensen committed Nov 28, 2016

Removed unnecessary temp file. Updated notebook.

olavurmortensen committed Nov 28, 2016

Merged upstream develop branch into my feature branch.

olavurmortensen committed Nov 28, 2016

In the process of refactoring (atmodel2.py will become the new atmode… …

…l.py).

olavurmortensen committed Nov 30, 2016

Merge branch 'develop' into author-topic_model

olavurmortensen committed Nov 30, 2016

Refactoring the code. A lot left to do.

olavurmortensen committed Nov 30, 2016

The refactored code now runs, converges almost exactly as the old cod… …

…e, produces some decent results, but somehow the results are slightly different. Made some slight changes to LdaModel to make the author-topic model work.

olavurmortensen committed Dec 1, 2016

Refactoring. Various docstring and commenting. Made methods for const… …

…ructing author2doc and doc2author so that the user may do this at will. Assuming that input to bound is seen data, as the converse may be problematic.

olavurmortensen committed Dec 4, 2016

New refactored code now in atmodel.py. Old code is in atmodelold.py, … …

…until I'm confident I don't need it anymore. Working on a new version in atmodel2.py, where I will be looping over authors rather than documents in the update.

olavurmortensen committed Dec 5, 2016

Implemented 'continued training' (call update multiple times) and __g… …

…etitem__ in refactored code (atmodel.py).

olavurmortensen committed Dec 7, 2016

A lot of changes. Most notably, added docstrings, and made it possibl… …

…e to evaluate test set (held-out data).

olavurmortensen committed Dec 8, 2016

Added unit tests. Basically a retrofit of LDA test; some new tests, s… …

…ome altered tests, some tests removed, others unchanged.

olavurmortensen committed Dec 9, 2016

Updated unit tests. Fixed some mistakes. Added some tests; testing up… …

…date fuctionality.

olavurmortensen committed Dec 9, 2016

Forgot to add num_docs to ids of new authors in id2author. Some comme… …

…nts.

olavurmortensen committed Dec 9, 2016

Made it possible to use serialized corpora (MmCorpus), and made unit … …

…tests of that functionality. There are some caveats to the use of serialized corpora in the author-topic model. Updated docstring.

olavurmortensen committed Dec 11, 2016

Removed code in unit tests that silence logging (useful when doing lo… …

…cal testing), as it broke some other tests.

olavurmortensen committed Dec 11, 2016

Just removed a comment.

olavurmortensen committed Dec 12, 2016

Reverted some changes that were made to ldamodel.py that were no long… …

…er necessary.

olavurmortensen committed Dec 12, 2016

get_author_topics now takes author name instead of integer ID; change… …

…d unit tests accordingly. Overwrote get_document_topics to raise a NotImplementedError in the author-topic model.

olavurmortensen committed Dec 12, 2016

Logging silencing again causing unit test failures. Fixed.

olavurmortensen committed Dec 13, 2016

Updated docstring. Changed __getitem__ method.

olavurmortensen committed Dec 28, 2016

Added a new notebook where a stackexchange dataset is used. Started w… …

…riting a tutorial. Updated old notebook as well.

olavurmortensen committed Dec 28, 2016

Updated notebooks (just to trigger rebuild).

olavurmortensen committed Dec 28, 2016

Updated code w.r.t. comments from Lev (@tmylk).

olavurmortensen committed Jan 5, 2017

Updated all notebooks.

olavurmortensen committed Jan 5, 2017

Two algorithms in 'temp' used to test the difference between blocking… …

… and standard variational Bayes AT model.

olavurmortensen committed Jan 5, 2017

Added the deepcopy again. Without it, the program can fail and the sy… …

…stem crashes (looking into it).

olavurmortensen committed Jan 5, 2017

Removed minimum_phi_value test (was already commented out).

olavurmortensen committed Jan 10, 2017

Comments and docstrings. Responding to comments from Lev, and working… …

… a bit with Sphinx documentation.

olavurmortensen committed Jan 10, 2017

Added the author-topic model to the API reference. Also slight change… …

… to author-topic model rst file.

olavurmortensen committed Jan 10, 2017

Added a test for gamma in persistency.

olavurmortensen committed Jan 10, 2017

Removed test for single author in persistency test (test is simplifie… …

…d with this and the previous commit 7ac77b7).

olavurmortensen committed Jan 10, 2017

Removed save and load methods, using LdaModel's methods directly work… …

…s. Also fixed the 'endclass' comment.

olavurmortensen committed Jan 10, 2017

Removed all temporary files.

olavurmortensen committed Jan 12, 2017

Made changes to model and test in preperation for a merge with upstream.

olavurmortensen committed Jan 12, 2017

Merge remote-tracking branch 'upstream/develop' into author-topic_model

olavurmortensen committed Jan 12, 2017

Modified the bound method; it was somewhat confusing, and there were … …

…even some mistakes. Cleaned up the code (atmodel.py and tests) w.r.t. PEP8 (disregarding E501, E731, E12 and W503) and removing vertical indent.

olavurmortensen committed Jan 13, 2017

Simplified sum in phi norm computation.

olavurmortensen committed Jan 13, 2017

Fixed some mistakes introduced in bound method in recent commit.

olavurmortensen committed Jan 13, 2017

Updated algorithm and tests w.r.t. comments from Lev. Other changes a… …

…s well: removed do_mstep method (using LdaModels version directly), using minimum_probability in get_author_topics, removed statement (in log) that said perplexity is evaluated on held-out data.

olavurmortensen committed Jan 14, 2017

Updated tutorial. Removed test notebook.

olavurmortensen committed Jan 14, 2017

Updated notebook.

olavurmortensen committed Jan 14, 2017

Updated tutorial.

olavurmortensen committed Jan 14, 2017

Updated tutorial (introduction).

olavurmortensen committed Jan 14, 2017

Changes w.r.t. change requests from @tmylk, plus some other changes.

olavurmortensen committed Jan 16, 2017

Added the URL to view notebook in HTML to tutorial.

olavurmortensen committed Jan 16, 2017

Telling users to view the notebook in nbviewer instead. Docstring lin… …

…ks to github.

olavurmortensen committed Jan 16, 2017