Description

The aim is to perform style transfer task on text. Here we use the yelp review dataset which can be found in data folder. The dataset consits of negative and positive reviews. We aim to transfer the style of a review from positive to negative and vice versa

Installation

pip3 install -r requirements.txt

This code uses python 3.

Usage

Training (runs inference on the dev set after each epoch)

For Training Negative to Positive Style Transfer run: python3 train.py --config yelp_config.json --bleu

For Training Positive to Negative Style Transfer run: python3 train.py --config yelp_config2.json --bleu

This will reproduce the model on a dataset of yelp reviews:

Checkpoints, logs, model outputs, and TensorBoard summaries are written in the config's working_dir.

See yelp_config.json for all of the training options.

Inference

For Negative to Positive python inference.py --config yelp_config.json --checkpoint path/to/model.ckpt

For Negative to Positive python inference.py --config yelp_config2.json --checkpoint path/to/model.ckpt

To run inference, you can point the src_test and tgt_test fields in your config to new data.

Data prep

Given two pre-tokenized corpus files, use the scripts in tools/ to generate a vocabulary and attribute vocabulary:

python tools/make_vocab.py [entire corpus file (src + tgt cat'd)] [vocab size] > vocab.txt
python tools/make_attribute_vocab.py vocab.txt [corpus src file] [corpus tgt file] [salience ratio] > attribute_vocab.txt
python tools/make_ngram_attribute_vocab.py vocab.txt [corpus src file] [corpus tgt file] [salience ratio] > attribute_vocab.txt