Skip to content

Latest commit

 

History

History
71 lines (57 loc) · 3.81 KB

README.md

File metadata and controls

71 lines (57 loc) · 3.81 KB
Raw

Embedding Perturbed Adversarial AutoEncoders (EPAAE)

This repo contains the code and data of the following paper:
"Towards Robust and Semantically Organised Latent Representations for Unsupervised Text Style Transfer" Sharan Narasimhan, Suvodip Dey, Maunendra Sankar Desarkar (to be published in the Proceedings of NAACL 2022). For now, the arxiv version is available at: https://arxiv.org/abs/2205.02309

For more samples from qualitative experiments using EPAAE, please refer to the Appendix of the paper.

Dependencies

Python 3.8.5+, PyTorch 1.11+

Download data and outputs

Download the 'data' and 'outputs' folders from: https://drive.google.com/drive/folders/1ZvAPAfd_pGFb37vA9gOyTBUiEOgPmHBJ?usp=sharing Place the 'data' folder in the project directory.

Training

. A sample command to train a simple AAE is:

python3 train.py --train data/yelp-shen/train.txt --valid data/yelp-shen/valid.txt --model_type aae --lambda_adv 10 --save-dir checkpoints/yelp-shen/aae-adv10

To train various models, use the following options:

  • VAE: --model_type vae --lambda_kl 0.1 --save-dir checkpoints/snli/vae_kl0.1
  • AAE: `p3 train.py --train ./data/yelp-shen/train.txt --valid data/yelp-shen/valid.txt --model_type aae --lambda_adv 10 --save-dir ./checkpoints/yelp-shen/test --gpu 0
  • LAAE: --model_type aae --lambda_adv 10 --lambda_p 0.01 --save-dir checkpoints/snli/aae_p0.01
  • DAAE(p=0.3): --model_type aae --lambda_adv 10 --noise 0.3,0,0,0 --save-dir checkpoints/snli/daae, where --noise P,P,P,K specifies word drop probability, word blank probability, word substitute probability, max word shuffle distance, respectively
  • EPAAE(zeta=3.0) : --model_type aae --lambda_adv 10 --save-dir checkpoints/scitail/aae-z3.0 --embd-noise --zeta 3.0 --noise-type centered-gau --gpu 5 --batch-size 32
  • EPAAE(p=0.1, zeta=3.0) : --model_type aae --lambda_adv 10 --save-dir checkpoints/snli/aae-z3.0 --embd-noise --zeta 3.0 --noise-type centered-gau --noise 0.1,0,0,0 --gpu 0

Run python3 train.py -h to see all training options.

Testing

  • To generate sentences from the model, an example command is: python test.py --sample --n 10000 --output sample --checkpoint checkpoints/snli/epaae-z2.0/
  • To perform Text Style Transfer via vector arithmetic, an example command is: python3 test.py --tst-on-test --dataset snli --checkpoint checkpoints/supervised/snli/dir_of_pretrained_model --gpu 7 --model-name model_best.pt Please check test.py for more experiment scripts such as TSNE, interpolation, reconstruction. For any questions or reports of bugs, please add an issue and we will resolve it.

Pretrained models and outputs

Please download and use the models we trained as part of this paper: https://drive.google.com/drive/folders/1fWXr8trmpqzaRCYzgnYzurdm3zwJpedN?usp=sharing Use --checkpoint to specify the path of the model. Please feel free to create an issue or email me regarding doubts or issues.

Citation

@article{Narasimhan2022TowardsRA, title={Towards Robust and Semantically Organised Latent Representations for Unsupervised Text Style Transfer}, author={Sharan Narasimhan and Suvodip Dey and Maunendra Sankar Desarkar}, journal={ArXiv}, year={2022}, volume={abs/2205.02309} }

Other considerations

This repo has been built over elements from the clean and high-quality code present in https://github.com/shentianxiao/text-autoencoders, so we encourage you to also check out their work on Text Style Transfer.

Please refer to tips.txt for more details.