PV-DBOW or PV-DM?

[jwijffels]

Is this implementation the distributed bag of words ('PV-DBOW') or the distributed memory ('PV-DM') model

[jwijffels]

@hiyijian would be great to have an answer on this

[rekola]

Both models are implemented.

[jwijffels]

@rekola where in the c++ code can you call dm and pv-dbow

[rekola]

It seems my information was based on your work in https://github.com/bnosac/doc2vec/blob/master/R/paragraph2vec.R, which says:

# cbow = 0 = skip-gram                                             = PV-DBOW
# cbow = 1 = continuous bag of words including vector of paragraph = PV-DM

Is this not true?

I've been working on a fork of doc2vec to remove the word and sentence length limits. This original version also crashes, if you have more than 30 million documents in your dataset.

[jwijffels]

Yes, that is indeed my interpretation

# cbow = 0 = skip-gram                                             = PV-DBOW
# cbow = 1 = continuous bag of words including vector of paragraph = PV-DM

and I would prefer to have a validation from @hiyijian as in the R wrapper I call: https://github.com/bnosac/doc2vec/blob/3e947562a0a69e11eb292283116a4fdc9cf5c0f4/src/rcpp_doc2vec.cpp#L14 which calls the train functionality from this repository https://github.com/hiyijian/doc2vec/blob/master/cpp/Doc2Vec.cpp#L65 and I make the above assumption