Step-unrolled Denoising Autoencoders for Text Generation

Unofficial PyTorch implementation of "Step-unrolled Denoising Autoencoders for Text Generation"

Work-in-progress, more experiments soon.

Currently implements unconditional character-level text generation on the text8 dataset.

Requires x-transformers and ptpt to be installed.

Usage

Run SUNDAE training using config file cfg.toml:

python main.py --cfg-path cfg.toml

Sample text from trained model saved at ckpt.pt:

python main.py --cfg-path cfg.toml --resume ckpt.pt --sample --nb-samples 16

Other useful flags:

--seed          # set RNG seed 
--no-save       # disable saving of checkpoints [False]
--no-cuda       # disable the use of CUDA device [False]
--no-amp        # disable the use of automatic mixed precision [False]
--nb-workers    # set number of dataloader workers. [4]

TODO:

• [ ] Add argmax-unrolled sampling
• [ ] Encoder-decoder translation experiments
• [ ] Pre-trained checkpoints

Citation

Step-unrolled Denoising Autoencoders for Text Generation

Nikolay Savinov, Junyoung Chung, Mikolaj Binkowski, Erich Elsen, Aaron van den Oord

@misc{savinov2021stepunrolled,
title={Step-unrolled Denoising Autoencoders for Text Generation}, 
author={Nikolay Savinov and Junyoung Chung and Mikolaj Binkowski and Erich Elsen and Aaron van den Oord},
year={2021},
eprint={2112.06749},
archivePrefix={arXiv},
primaryClass={cs.CL}
}