NLP-with-Python | Scikit-Learn , NLTK , Spacy , Gensim , Textblob | Machine Learning library

A gensim Doc2Vec model may be saved and loaded using the .save(filepath) & .load(filepath) methods. (Using these native-to-gensim methods will work on larger models than plain Python pickling can support, and more-efficiently store some of the larger internal arrays as separate files. (If moving the saved model, be sure to keep this subsidiary files alongside the main file that's at exactly the filepath location.)

A previously-trained Doc2Vec model can generate doc-vectors for new texts via the .infer_vector(list_of_words) method.

Note that the list_of_words provided to this method should have been preprocessed/tokenized exactly the same as the training data – and any words that weren't present (or sufficiently min_count frequent) in the training data will be ignored. (At the extreme, this means if you pass in a list_of_words with no recognized words, all words will be ignored, and you'll get back a randomly-initialized but completely-unimproved-by-inference vector.)

Still, if you're re-evaulating or re-training the downstream predictive models on new data from some new domain, you'd often want to re-train the Doc2Vec stage as well, with all available data, so that it has a chance to learn new words from new usage contexts. (It's mainly when your training data was extensive & representative, and your new data comes in incrementally and without major shifts in vocabulary/usage/domain, that you'd want to rely on .infer_vector().)