weilinie/RelGAN

Implementation of RelGAN: Relational Generative Adversarial Networks for Text Generation

RelGAN

This repository contains the code to reproduce the core results from the paper RelGAN: Relational Generative Adversarial Networks for Text Generation.

Dependencies

This project uses Python 3.5.2, with the following lib dependencies:

• Tensorflow 1.4
• Numpy 1.14.1
• Matplotlib 2.2.0
• Scipy 1.0.0
• NLTK 3.2.3
• tqdm 4.19.6

Instructions

The experiments folders contain scripts for starting the different experiments. For example, to reproduce the synthetic data experiments, you can try:

cd oracle/experiments
python3 oracle_relgan.py [job_id] [gpu_id]

or COCO Image Captions:

cd real/experiments
python3 coco_relgan.py [job_id] [gpu_id]

or EMNLP2017 WMT News:

cd real/experiments
python3 emnlp_relgan.py [job_id] [gpu_id]

Note to replace [job_id] and [gpu_id] with appropriate numerical values.

Reference

To cite this work, please use

@INPROCEEDINGS{Nie2019ICLR,
  author = {Nie, Weili and Narodytska, Nina and Patel, Ankit},
  title = {RelGAN: Relational Generative Adversarial Networks for Text Generation},
  booktitle = {International conference on learning representations (ICLR)},
  year = {2019}
}

Acknowledgement

This code is based on the previous benchmarking platform Texygen.