load() in 1.0.0 not rebuilding necessary down-sampling cum_table

[gojomo]

See: https://groups.google.com/d/msg/gensim/F9SLBpik4Yg/Jz7BhRH_DQAJ

The test in Word2Vec load() is only chcking for index2word in its old location, thus a necessary structure for continued training isn't rebuilt.

Surprised no existing save/load tests catch this as the re-build should even be necessary for a save/load within the same latest gensim version.

[tmylk]

Thanks. are you working on a fix or should I submit a pr?

[piskvorky]

@tmylk what is the explanation for this bug making it past the unit tests?

[gojomo]

Not currently working on this; just wanted to make sure there was an issue that could be triaged/assigned. Not sure the understanding above is fully correct - the code looks broken, but I would have expected any save/loadthat involves a (default) negative-sampling model to have triggered the error, so aspects of the problem as-reported are still unclear.

[gojomo]

To be clearer: any [save, load, train] should trigger the error - perhaps no existing tests try training after load? Adding either a train() or some other set of 'model is healthy' assertions after load() may make for better tests.

[tmylk]

There is no unit test that covers training after load in either doc2vec or word2vec. Adding them now.