sentence-VAE

This is a PyTorch implementation of variational auto-encoder (VAE) for natural languages.

Usage instructions

Train model

python main.py train \
--train_file <file_path> \
--valid_file <file_path> \
--vocabulary_file <file_path>

• train_file one-sentence-per-line raw corpus file for training.
• valid_file one-sentence-per-line raw corpus file for validation.
• vocabulary_file vocabulary file.

Sample sentences from prior distribution

python sample.py \
--vocabulary_file <file_path>
--checkpoint_file <file_path> \
--sample_size 10

• vocabulary_file vocabulary file.
• checkpoint_file PyTorch model parameter file.
• sample_size number of samples to generate.

Example

Penn Tree Bank

Download data from here.

Sentence samples.

- a spokesman for the <unk> said he is n't recommending
- to make the comparable market directly comparable each index is based on the close of N equaling N
- you do n't have a <unk>
- but they did n't have to get a <unk> plan
- but pfizer said its third-quarter earnings were hurt by the year earlier
- he said the board had n't yet been scheduled to meet with the situation
- yesterday 's edition of the company 's stock exchange composite trading yesterday
- campeau 's chairman stephen m. wolf had been approached by the end of the year
- these include the tax rate of N N of the common stock and the other N N of the common stock
- but he said the <unk> had been <unk> in the past two years