Using Variational Auto-Encoder For Generating Texts

Overview

Tensorflow implementation of Generating Sentences from a Continuous Space.

Usage

Training:

python vae_lstm-lstm.py

Parameters

Parameters can be set directly in Parameters class in vae_lstm-lst.py file. (or specify through command line parameters). You can choose to train on PTB dataset or on Game of Thrones dataset. Parameter debug can be set to true for calling Tensorflow debugger.

• --dataset PTB or GOT
• --beam_search use beam search (default=sample from the multinulli distribution)
• --beam_size beam size, default set to 2

During training generated text will not appear from the first iteration, be patiant and have fun :) I will add pre-trained model later.

Specific requirements

• zhusuan - probabilistic framework https://github.com/thu-ml/zhusuan/
• tensorflow >= 1.0
• gensim (for pretrained w2vec)

Other files

• train_rnnlm.py - RNN Word-level language model. Can be trained on PTB or Game of Thrones dataset.