Style Transformer: Unpaired Text Style Transfer without Disentangled Latent Representation

This folder contains the code for the paper 《Style Transformer: Unpaired Text Style Transfer without Disentangled Latent Representation》

Requirements

1. pip packages

   • pytorch>=0.4.0
   • torchtext>=0.4.0
   • nltk
   • fasttext==0.8.3
   • pypi-kenlm
   • tqdm
   • wandb

2. LM for evaluator

   • Download https://drive.google.com/open?id=1pklyWxzNPPxnKNy_TmA8h_tmGmiZttPN
   • Put ppl_yelp.binary in the folder evaluator

3. nltk setup

   • Run the following command to download (or double-check) nltk tokenizer dependency:

   python -c "import nltk; nltk.download('punkt')"
   

Usage

The hyperparameters for the Style Transformer can be found in main.py or with python main.py -h.

• To train:

python main.py --do_train

• To inference on test set:

SAVE=./save/Feb15141010/ckpts/2000 # just example
python main.py --do_test \
-test_out ./submission.txt \
-preload_F ${SAVE}_F.pth \
-preload_D ${SAVE}_D.pth

You can modify other parameters to suit your need.

• To run hw5-2

SAVE=./save/Feb15203331/ckpts/ #just example
STEP=1300
python main.py --part2 \
--part2_model_dir $SAVE \
--part2_step $STEP