R-Net

• A Tensorflow implementation of R-NET: MACHINE READING COMPREHENSION WITH SELF-MATCHING NETWORKS. This project is specially designed for the SQuAD dataset.
• Should you have any question, please contact Wenxuan Zhou ([email protected]).

Requirements

There have been a lot of known problems caused by using different software versions. Please check your versions before opening issues or emailing me.

General

• Python >= 3.4
• unzip, wget

Python Packages

• tensorflow-gpu >= 1.5.0
• spaCy >= 2.0.0
• tqdm
• ujson

Usage

To download and preprocess the data, run

# download SQuAD and Glove
sh download.sh
# preprocess the data
python config.py --mode prepro

Hyper parameters are stored in config.py. To debug/train/test the model, run

python config.py --mode debug/train/test

To get the official score, run

python evaluate-v1.1.py ~/data/squad/dev-v1.1.json log/answer/answer.json

The default directory for tensorboard log file is log/event

See release for trained model.

Detailed Implementaion

• The original paper uses additive attention, which consumes lots of memory. This project adopts scaled multiplicative attention presented in Attention Is All You Need.
• This project adopts variational dropout presented in A Theoretically Grounded Application of Dropout in Recurrent Neural Networks.
• To solve the degradation problem in stacked RNN, outputs of each layer are concatenated to produce the final output.
• When the loss on dev set increases in a certain period, the learning rate is halved.
• During prediction, the project adopts search method presented in Machine Comprehension Using Match-LSTM and Answer Pointer.
• To address efficiency issue, this implementation uses bucketing method (contributed by xiongyifan) and CudnnGRU. The bucketing method can speedup training, but will lower the F1 score by 0.3%.

Performance

Score

Training Time (s/it)

Extensions

These settings may increase the score but not used in the model by default. You can turn these settings on in config.py.

• Pretrained GloVe character embedding. Contributed by yanghanxy.
• Fasttext Embedding. Contributed by xiongyifan. May increase the F1 by 1% (reported by xiongyifan).