• Word2vec in line with original word2vec.c (Andrey Kutuzov, #538)
• Same default values. See diff akutuzov@6456cbc
• Standalone script with command line arguments matching those of original C tool.
  Usage ./word2vec_standalone.py -train data.txt -output trained_vec.txt -size 200 -window 2 -sample 1e-4
• load_word2vec_format() performance (@svenkreiss, #555)
  • Remove init_sims() call for performance improvements when normalized vectors are not needed.
  • Remove norm_only parameter (API change). Call init_sims(replace=True) after the load_word2vec_format() call for the old norm_only=True behavior.
• Better internal handling of job batching in word2vec (#535)
  • up to 300% speed up when training on very short documents (~tweets)
• Word2vec allows non-strict unicode error handling (ignore or replace) (Gordon Mohr, #466)
• Doc2Vec model.docvecs[key] now raises KeyError for unknown keys (Gordon Mohr, #520)
• Fix DocvecsArray.index_to_doctag so most_similar() returns string doctags (Gordon Mohr, #560)
• On-demand loading of the pattern library in utils.lemmatize (Jan Zikes, #461)
  • utils.HAS_PATTERN flag moved to utils.has_pattern()
• Threadsafe Word2Vec/Doc2Vec finish-check to avoid hang/unending Word2Vec/Doc2Vec training (Gordon Mohr, #571)
• Tuned TestWord2VecModel.test_cbow_hs() against random failures (Gordon Mohr, #531)
• Prevent ZeroDivisionError when default_timer() indicate no elapsed time (Gordon Mohr, #518)
• Forwards compatibility for NumPy > 1.10 (Matti Lyra, #494, #513)
  • LdaModel and LdaMulticore produce a large number of DeprecationWarnings from
    .inference() because the term ids in each chunk returned from utils.grouper
    are floats. This behaviour has been changed so that the term IDs are now ints.
  • utils.grouper returns a python list instead of a numpy array in .update() when
    LdaModel is called in non distributed mode
  • in distributed mode .update() will still call utils.grouper with as_numpy=True
    to save memory
  • LdaModel.update and LdaMulticore.update have a new keyword parameter
    chunks_as_numpy=True/False (defaults to False) that allows controlling
    this behaviour