Improved Training of Wasserstein GANs
This is a project test Wasserstein GAN objectives on language. The code is built on a fork of the popular project under the same title.
We try to reproduce results from their paper. We clean their code for language generation, try smaller datasets, standard preprocessing and slightly different architectures.
We striped a lot of unused code to better understand the rest.
Datasets
You can download Download Google Billion Word at [http://www.statmt.org/lm-benchmark/] . Other datasets are available at [https://drive.google.com/drive/folders/0B7MLuc1jq3A8eFpVWUZ0eDEwdlE?usp=sharing]
Prerequisites
- Python, NumPy, TensorFlow, SciPy, Matplotlib
- A recent NVIDIA GPU
Important files
Most important is python gan_language.py
: Character-level language model. It has help. Specify directory ith
Improved WGAN.ipynb
: jupyter notebook with my attemt to code it in keras.
Improved WGAN.py
: Improved WGAN.ipynb
exported to pure python script.
directories cuted
, cuted_small
kanye
quora
and romeo
contains graphs for given datasets and some sample generated texts.