Online learning and hashing trick

• Online learning and better pre-training in GensimFastTex:
  • Hashing trick for building the vocabulary, similar to the original fastText approach
  • It is possible to pre-fit word embeddings from a dataset with word2vec
  • True online earning with partial_fit, the vocabulary is incrementally updated
• New version of fastText.py: 0.8.2
• New version of Gensim: 0.13.4
• Fixed predict_proba output format

Faster prediction and persistence

• Faster prediction for multiple labels with one predict call
• Better persistence with save and load methods
• Fixed parameter names convention

Quasi-stable models

• FastText classifier based on version 0.8.0 of https://github.com/salestock/fastText.py
• GensimFastText has now:
  • negative sampling
  • softmax as alternative output function
  • almost complete LabeledWord2Vec as subclass of Gensim's Word2Vec
• More tests

Fast Cython code

this is broken, valid commit for this release is 263bfb1

• Cython code for fastText in Gensim
• Script for benchmarks

First working model

• GensimFTClassifier can train and predict