Pytorch-PLATO

PLATO dialog model with pre-trained parameters in pytorch version

This repository contains:

1. PLATO source code converted in pytorch version.
2. PLATO pre-trained parameters converted in pytorch version: 12-layers, 768-hidden, 12-heads, 132M parameters (uncased model with latent variables).

For simplicity, you can directly download the pytorch model file PLATO.pt from this link.

Requirements

- python >= 3.6
- pytorch == 1.8.0
- numpy
- nltk
- tqdm
- regex

Data preparation

Download data from the link. The tar file contains three processed datasets: DailyDialog, PersonaChat and DSTC7_AVSD.

mv /path/to/data.tar.gz .
tar xzf data.tar.gz

Citation

PLATO: Pre-trained Dialogue Generation Model with Discrete Latent Variable paper link

Original code link in paddlepaddle version: https://github.com/PaddlePaddle/Research/tree/master/NLP/Dialogue-PLATO

Other details can also refer to the descriptions in paddlepaddle version from the above link, we keep the same with it!

@inproceedings{bao2019plato,
    title={PLATO: Pre-trained Dialogue Generation Model with Discrete Latent Variable},
    author={Bao, Siqi and He, Huang and Wang, Fan and Wu, Hua and Wang, Haifeng},
    booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
    pages={85--96},
    year={2020}
}