QANet

A Pytorch implementation of QANet
The code is mostly based on the two repositories: hengruo/QANet-pytorch NLPLearn/QANet

Performance

*The results of hidden size 128 with 8 heads were run with 12 batches.

Requirements

• python 3.6

• pytorch 0.4.0

• tqdm

• spacy 2.0.11

• tensorboardX

• absl-py

Usage

Download and preprocess the data

# download SQuAD and Glove$ sh download.sh# preprocess$ python3.6 main.py --mode data

Train the model

# model/model.pt will be generated every epoch$ python3.6 main.py --mode train

Tensorboard

# Run tensorboard for visualisation$ tensorboard --logdir ./log/

TODO

•  Add Exponential Moving Average

•  Reach the performance of the paper with hidden size 96, 1 head.

•  Test on hidden size 128, 8 head.