def cosine(vector1,vector2):
cosV12 = np.dot(vector1, vector2) / (linalg.norm(vector1) * linalg.norm(vector2))
return cosV12
model=gensim.models.doc2vec.Doc2Vec.load('Model_D2V_Game')
string='民生 为了 父亲 我 要 坚强 地 ...'
list=string.split(' ')
vector1=model.infer_vector(doc_words=list,alpha=0.1, min_alpha=0.0001,steps=5)
vector2=model.docvecs.doctag_syn0[0]
print cosine(vector2,vector1)
-0.0232586
I use a train data to train a doc2vec
model. Then, I use infer_vector()
to generate a vector given a document which is in trained data. But they are different. The value of cosine was so small (-0.0232586
) distance between the vector2
which was saved in doc2vec
model and the vector1
which was generated by infer_vector()
. But this is not reasonable ah ...
I find where i have error in. I should use 'string=u'民生 为了 父亲 我 要 坚强 地 ...'' instead 'string='民生 为了 父亲 我 要 坚强 地 ...''. When I correct this way, the cosine distance is up to 0.889342.
Set-up Doc2Vec Training & Evaluation Models In the word2vec architecture, the two algorithm names are “continuous bag of words” (CBOW) and “skip-gram” (SG); in the doc2vec architecture, the corresponding algorithms are “distributed memory” (DM) and “distributed bag of words” (DBOW).
A size of 100 means the vector representing each document will contain 100 elements - 100 values. The vector maps the document to a point in 100 dimensional space. A size of 200 would map a document to a point in 200 dimensional space. The more dimensions, the more differentiation between documents.
As you've noticed, infer_vector()
requires its doc_words
argument to be a list of tokens – matching the same kind of tokenization that was used in training the model. (Passing it a string causes it to just see each individual character as an item in a tokenized list, and even if a few of the tokens are known vocabulary tokens – as with 'a' and 'I' in English – you're unlikely to get good results.)
Additionally, the default parameters of infer_vector()
may be far from optimal for many models. In particular, a larger steps
(at least as large as the number of model training iterations, but perhaps even many times larger) is often helpful. Also, a smaller starting alpha
, perhaps just the common default for bulk training of 0.025, may give better results.
Your test of whether inference gets a vector close to the same vector from bulk-training is a reasonable sanity-check, on both your inference parameters and the earlier training – is the model as a whole learning generalizable patterns in the data? But because most modes of Doc2Vec inherently use randomness, or (during bulk training) can be affected by the randomness introduced by multiple-thread scheduling jitter, you shouldn't expect identical results. They'll just get generally closer, the more training iterations/steps you do.
Finally, note that the most_similar()
method on Doc2Vec
's docvecs
component can also take a raw vector, to give back a list of most-similar already-known vectors. So you can try the following...
ivec = model.infer_vector(doc_words=tokens_list, steps=20, alpha=0.025)
print(model.most_similar(positive=[ivec], topn=10))
...and get a ranked list of the top-10 most-similar (doctag, similarity_score)
pairs.
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With