Code of paper Affective Decoding for Empathetic Response Generation
5 branches for different experiment settings: master(Transfo), adde(AD, AD+DE), adm(AD + multi-task learning), tml(Transfo + multi-task learning), prepend (Transfo + prepending emotion label predicted by fasttext).
Usage
mkdir log
mkdir save
mkdir save/pretrained_lm // download
download the pretrained model params (GPT) from here. Put the files into save/pretrained_lm
Train
Run the command (GPU will be used if available, make sure CUDA is installed):
python train.py --save_path save/model
Interact with model
git checkout adm
python play.py --model_path save/model --turns 2
Requirements
- PyTorch (version >=1.4)
- tqdm
- sklearn
- spacy (version < 3)
- ftfy
- pandas
- tensorboardX