doc2vec

A simple and readable implementation of doc2vec [1], using Python 3, Keras and TensorFlow.

Installation

pip install -r requirements.txt
python setup.py install

Assumptions

This implementation assumes that your documents are all in the same directory, and named with monotonically incrementing integer IDs, e.g. 0.txt, 1.txt. Each file should contain an ordinary text document, i.e. without any special preprocessing.

Usage

doc2vec path/to/docs/ \
    --save path/to/save/model.hdf5 \
    --save_vocab path/to/save/vocab.vocab \
    --train

By default, this will use the Distributed Memory (DM) model. You can also use the Distributed Bag-Of-Words (DBOW) model with:

doc2vec path/to/docs/ --model dbow

Previously trained models can be loaded with:

doc2vec path/to/docs/ \
    --load path/to/load/model.hdf5 \
    --load_vocab path/to/load/vocab.vocab

And document embeddings can be written to file as follows:

doc2vec path/to/docs/ \
    --load path/to/load/model.hdf5 \
    --load_vocab path/to/load/vocab.vocab \
    --save_doc_embeddings path/to/save/embeddings.hdf5

Finally, you can see all available options and model parameters with:

doc2vec -h

References

1. Le, Quoc, and Mikolov, Tomas. "Distributed representations of sentences and documents." International Conference on Machine Learning. 2014.
2. Bird, Steven, Loper, Edward and Klein, Ewan. "Natural Language Processing with Python." O’Reilly Media Inc. 2009.