DiffMask

Overview

This library contains a Pytorch implementation of Differentiable Masking Explainer (Diffmask), as presented in [1](https://arxiv.org/abs/2004.14992).

Dependencies

python>=3.6
pytorch>=1.5: https://pytorch.org
pytorch-lightning==0.7.3: https://pytorch-lightning.readthedocs.io
transformers>=2.9.0: https://github.com/huggingface/transformers
torch-optimizer>=0.0.1a9: https://github.com/jettify/pytorch-optimizer
matplotlib>=3.1.1: https://matplotlib.org

Notice that older or newer version could work but they were not tested.

Installation

To install, run

$ python setup.py install

To donwload datasets run

$ ./scripts/download_datasets.sh

To download models, use the following link. Note that these are not the exaclt same models used for the paper.

Structure

diffmask: Contains the source code for DiffMask.

We have 5 jupyter notebbok with the code for reproducing some of the results from our work [1]. Note that since i) the code was refactored and ii) we were not able to realise the exact same models used for the paper, re-generated plots and tables might differ from the ones in our work.

QuestionAnsweringSquadAnalysis Plots and analysis using other methods of a BERT large model for question answering.
QuestionAnsweringSquadDiffMaskAnalysis Plots and analysis using DiffMask of a BERT large model for question answering.
SentimentClassificationSSTAnalysis Plots and analysis using other methods of a BERT base model for sentiment analisys.
SentimentClassificationSSTDiffMaskAnalysis Plots and analysis using DiffMask of a BERT base model for sentiment analisys.
ToyTask Training, analysis and plots for the artificial toy task using both other methods and DiffMask.

Please cite [1] in your work when using this library in your experiments.

Feedback

For questions and comments, feel free to contact Nicola De Cao.

License

MIT

Citation

[1] De Cao, N., Schlichtkrull, M., Aziz, W., Titov, I. (2020).
How do Decisions Emerge across Layers in Neural Models? 
Interpretation with Differentiable Masking
In Proceedings of the 2020 Conference on Empirical
Methods in Natural Language Processing.

BibTeX format:

@article{decao2020decisions,
  title={How do Decisions Emerge across Layers in Neural Models?},
  author={
    De Cao, Nicola and
    Aziz, Wilker and
    Titov, Ivan},
  journal={Proceedings of the 2020 Conference on Empirical
           Methods in Natural Language Processing},
  year={2020}
}

xiye17/diffmask