Latent Diffusion Paraphraser

This is the codebase for the paper Enforcing Paraphrase Generation via Controllable Latent Diffusion.

Reproduce

Training

You personal dataset should be placed in datasets directory, and split into train, valid, test subsets. Each dataset should be in csv format with src, tgt as headers.

When training, you should use main.py

• --config meaning the path to your yaml config file, which should be placed in conf directory
• --mode meaning the train or resume mode
• --ckpt is required only in resume mode

Inference

When inference, you should use seq2seq.py

• --ckpt_dir meaning the checkpoint directory
• --config please use the same config file as training, you can find it in <SAVE_PATH>/conf.yaml

Controlnet Ensemble

Use controlnet_train.py

• --ckpt refers to the original ldp checkpoint path

Controlnet Inference

• --ldp refers to the original ldp checkpoint path
• --ckpt_dir meaning the checkpoint directory

Citation

If you find the code helpful, please cite

@article{zou2024enforcing,
  title={Enforcing Paraphrase Generation via Controllable Latent Diffusion},
  author={Zou, Wei and Zhuang, Ziyuan and Huang, Shujian and Liu, Jia and Chen, Jiajun},
  journal={arXiv preprint arXiv:2404.08938},
  year={2024}
}