Word2Vec/Doc2Vec offer model-minimization method

[gojomo]

If you're sure you're done training a model, a bunch of its memory-consumptive parts can be discarded:

• syn0 (non-normalized; be sure to save aside normalized versions 1st)
• syn1, syn1neg, syn0_lockf
• doctag_syn0, doctag_syn0_lockf (in Doc2Vec)

There should be a documented method (such as finished_training) to discard these, and testing to ensure there are no lingering unintended dependencies on those sticking around.

Semantics: As a tradeoff, finished_training discards as many model attributes as possible, while still being able to answer infer_vector and __getitem__ queries on the resulting trimmed model. No more continued training on that word2vec/doc2vec model is possible, and any attempt to do so results in a clear, understandable exception.

(Note though that a Doc2Vec model used for future infer_vector() ops needs to keep the syn0 & syn1* values.)

[gojomo]

Since I initially wrote this, I've seen cases where the non-unit-normalized syn0 is preferable to the unit-normed version. (Sometimes the magnitude of the vectors is relevant, with their magnitude in some sense being an indicator of stron/unambiguous meaning.) Also, some Doc2Vec users only want the model for inference, but others would consider the doctag_syn0 to be what they want to keep around for lookups/similarity-rankings.

So utility functions for this model-slimming need to be very carefully named and documented to set expectations properly - and perhaps factored to separate operations, rather than one big finished_training().

[tmylk]

Fixed in #987