Implement the inference stage

[inejc]

I.e. training in which only the paragraph matrix is updated (word vectors are pre-trained and frozen), as described in Q. V. Le et al., Distributed Representations of Sentences and Documents.

[ydennisy]

@inejc Do you have a good ref / guide on how this should be implemented?

The best idea I have is to look at gensim source - but maybe there is a better reference?

[inejc]

@ydennisy I believe the original paper mentions the inference step briefly. If I recall correctly, the only difference with training is that during inference all parameters but document vectors are fixed (i.e. word vectors and softmax weights are fixed). Please let me know if that helps.