[WIP] Implement save_word2vec_format can for Doc2Vec

[cedias]

Hi,
I recently had to export my Doc2Vec model to W2V base format and the function save_word2vec_format wasn't implemented for D2V class and was simply calling W2V one.

Therefore I quickly made this implementation.

If you believe it's worthwhile I'll go ahead and implement the load function to properly test the save/load pipeline, let me know.

[tmylk]

@cedias Apologies for late reply. What will be the advantage of saving in this format? What is your use case?

[cedias]

Well, since Doc2Vec extends Word2Vec, the save_word2vec_format function can already be used to output the word vectors. I believe it's logical to be able to output the document vector as well (At least that is what I was expecting before noticing it was only writing word vectors).

For the save_word2vec_format function the main (and probably only) advantage of this format is legacy. Personnally, I had some code which took classic w2v binary output as input and I wanted to try and input document vectors.

As for the load function i'm not really sure about the use case, besides API consistency, which is why I wanted an opinion about it :)

[gojomo]

I can see this being useful for others as well. Some thoughts:

• should remain possible to also save word-vecs – perhaps even to the same file? – so superclass functionality shouldn't be completely hidden by the override
• should also support case where user's doctags are just plain int indexes rather than strings (and thus model.docvecs.doctags is empty), or a mixture of ints and strings
• load... makes sense as well, subject to the same above points

Achieving those may require some new conventions in the method API and on-disk format to indicate the word-vec/doc-vec distinction... if at all possible those conventions should put a minimal burden on people using older files, other tools, or just one set (word-vecs or doc-vecs) of vectors.

[cedias]

Ok, i'll work on it later this month then. Thanks for the tips.

[tmylk]

Hi @cedias Would you have time to work on this for our release this month?

[cedias]

I believe this feature will be removed/realized in #1107

[tmylk]

@cedias Thanks for following the github development.
The feature that you are proposing is a new feature and is not included in #1107.
That PR doesn't touch on model.docvecs and existing behaviour of simply calling Word2Vec.save_word2vec_format is preserved.

[gojomo]

If Doc2Vec's DocvecsArray is fully adapted to use KeyedVectors, it might then inherit a useful save_word2vec_format() implementation.

Plausibly as per my earlier comment, there could be key-munging conventions for mixing word & doc vectors into the same flattened file on save, or even disentangling them on load. (Some downstream applications might like them mixed-together.)

[parulsethi]

Another use case - I wanted to visualize docvecs in Tensorboard which require vectors to be in text file format, and this functionality would be useful for that.

[menshikh-iv]

Ping @cedias, what status of this PR? Will you finish it soon?

[parulsethi]

@menshikh-iv This feature was added in #1256.

[menshikh-iv]

Then I close this PR, I hope the author agrees with @parulsethi (If not, you can reopen it)