Doc2vec: model.docvecs is only of length 10

问题:

I am trying doc2vec for 600000 rows of sentences and my code is below:

model = gensim.models.doc2vec.Doc2Vec(size= 100, min_count = 5,window=4, iter = 50, workers=cores)
model.build_vocab(res) 
model.train(res, total_examples=model.corpus_count, epochs=model.iter)

#len(res) = 663406

#length of unique words 15581
print(len(model.wv.vocab))

#length of doc vectors is 10
len(model.docvecs)

# each of length 100
len(model.docvecs[1])

How do I interpret this result? why is the length of vector only 10 with each of size 100? when the length of 'res' is 663406, it does not make sense. I know something is wrong here.

In Understanding the output of Doc2Vec from Gensim package, they mention that the length of docvec is determined by 'size' which is not clear.

回答1:

The tags of a TaggedDocument should be a list-of-tags. If you instead provided strings, like tags='73215', that would be seen as if the same as the list-of-characters:

tags=['7', '3', '2', '1', '5']

At the end, you'd only have 10 tags in your whole training set, just the 10 digits in various combinations.

That your len(model.docvec[1]) is 100 means you didn't make exactly this error, but perhaps something similar, in constructing your TaggedDocument training data.

Look at the first item in res, to see if its tags property makes sense, and each of the model.docvecs, to see what's being used instead of what you intended.