Implementation of the paper "Content preserving text generation with attribute controls", Logeswaran et al., which talks 'bout a backtranslation based style transfer network.
First of all, place all the files/folders present in this repository inside a new directory named "my_CPTG", else you will get an alert mentioning the same.
Next, re-install torch and torchtext versions 1.9.1 and 0.10.1 [with(out) cudnn as nccessary] by this command
$ pip3 install torch==1.9.1+cu111 torchvision==0.10.1+cu111 torchaudio==0.9.1 torchtext==0.10.1 -f https://download.pytorch.org/whl/torch_stable.html
Choose the dataset to train on by manipulating dataset_name
variable in main.py
and eval.py
files by names yelp/amazon/imdb.
To train the model run
$ python3 main.py
To test the model run
$ python3 eval.py
Examples of style transferred text produced by the model trained on yelp data set.
- “horrible service .” --> “great staff .”
- “a worthwhile read for the but it is ” --> “a boring read that the is but that a”
- “ very interesting look into the life of our first great ...” --> “very boring rambling to view of of life “