This is a PyTorch implementation of Language style transfer using cross-alignment technique. Paper link here. It contains code for encoding the text with two different methods.

1. Bag of words (/code_bow/)
2. Byte pair encoding (/code_bpe/)

Make sure you have Python 3.7 or above and pip installed. Execute the below command to install the required packages pip install -r requirements.txt

Below are sample commands to train and test Style transfer and TextCNN models. Replace code_bow with code_bpe to use BPE encoding. Please note that positive and negative tweets/reviews files have same prefix. They end with .0 and .1 for negative and positive reviews/tweets respectively. Example tweets.train.0 and tweets.train.1 for negative and positive tweets respectively. For this example, you have to use the below command to train the Style transfer model

python code_bow/style_transfer.py --saves_path <saves_dir> --train "data/tweets/tweets.train" --max_epochs 20 --vocab ./tmp/twitter.vocab

1. Training

   python code_bow/style_transfer.py --saves_path <saves_dir> --train <train_files_path> --max_epochs 20 --vocab ./tmp/twitter.vocab
   

2. Prediction

   python code_bow/style_transfer.py --predict "the service was great ." --vocab ./tmp/twitter.vocab --model_path <saved_model_path>
   

3. Testing

   python code_bow/style_transfer.py --test <tweets_test_data_files_path> --model_path <saved_model_path> --vocab ./tmp/twitter.vocab 
   

4. Testing the model outputs with a TextCNN model

   python textcnn/main.py -test=<test_outputs_file_path> -snapshot=<saved_textcnn_model_dir>