Authors: Juncen Li, Robin Jia, He He, Percy Liang
NOTES:
- The instructions below are still a work in progress.
- If you encounter any problems, please open a GitHub issue, or submit a pull request if you know the fix!
- This is research code meant to serve as a reference implementation. We do not recommend heavily extending or modifying this codebase for other purposes.
- The dataset is distributed under CC BY-SA 4.0 license.
If you have questions, please email Juncen Li at mlijuncenm at gmail.com.
For those who are interested, Reid Pryzant has also released a pytorch implementation of this paper.
Please name the corpora of two styles by "x.0" and "x.1" respectively, and use "x" to refer to them in options. Each file should consist of one sentence per line with tokens separated by a space.
- The
data/yelp/directory contains Yelp review dataset ("x.0" is negative data and "x.1" is positive data). - The
data/amazon/directory contains Amazon review dataset ("x.0" is negative data and "x.1" is positive data). - The
data/imagecaption/directory contains dataset of image captions ("x.0" is humorous data and "x.1" is romantic data).
To train a model, run the following command:
sh run.sh train model_name data_nameTo test a model, run the following command:
sh run.sh test model_name data_name
Where model_name can be "DeleteOnly" or "DeleteAndRetrieve", data_name can be "yelp", "amazon" or "imagecaption". For example you want to reproduce the training process of DeleteOnly on Yelp review dataset, you can run the following command:
```bash
sh run.sh train DeleteOnly yelp
To configure GPU usage, set the [`THEANO_FLAGS`](http://deeplearning.net/software/theano/library/config.html) environment variable, e.g.
```bash
THEANO_FLAGS='device=cuda0,floatX=float32' sh run.sh train model_name data_name
Python == 2.7, Theano >=0.8
Python requirement: numpy==1.15, scipy, nltk, whoosh