/BiDAF

Bidirectional Attention Flow for Machine Comprehension, Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, Hannaneh Hajishirzi https://arxiv.org/abs/1611.01603

Primary LanguagePythonApache License 2.0Apache-2.0

Bidirectional Attention Flow for Machine Comprehension (Unofficial)

Bidirectional Attention Flow for Machine Comprehension, Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, Hannaneh Hajishirzi https://arxiv.org/abs/1611.01603

This repo contains the implementation of Bi-Directional Attention Flow (BIDAF) network in PyTorch.

Requirements

  • PyTorch: 0.3
  • tqdm (progress bar)

Setup

$ ./download.sh  // download squad dataset
$ python -m squad.prepro    // build documents

And you also need to download glove 6B dataset under dataset directory.

Training

$ python main.py --help
$ python main.py

Test

$ python main.py --resume ./checkpoints/Epoch-12.model --test 1

TODO

  • confirm this model's performance, currently performance does not increasing. (see #1 )
  • Support multi labels
  • Test. The answer span (k, l) where k ≤ l with the maximum value of (p1k, p2l) should be chosen.
  • Imple EM and F1 score
  • clean main.py