SQuAD

A pytorch implementation of custom model for Stanford Question and Answering Dataset based on two papers Reading Wikipedia to Answer Open-Domain Questions and QANet: Combining Local Convolution with Global Self-Attention for Reading Comprehension.

Model architecture

Perfomance

Model was trained for 25 epochs which took 7 hours on one Nvidia Tesla P100

As can be noticed, current model's score is less than in original papers, however experiments have shown that more accurate selection of hyperparameters and adding EncoderBlocks from QANet can lead to a higher perfomance.

Quick Start

Requirements

python >=3.5
pytorch 0.4
numpy
msgpack
spacy <=1.9

Setup

Make sure that you installed python 3, pip, wget and unzip, pytorch.

git clone https://github.com/MovElb/SQuAD
cd SQuAD
pip3 install -r requirements.txt
bash download.sh
python3 prepro.py

Training

To run this custom architecture run

# Run training for 25 epochs with batch size 32
python3 train.py -e 25 -bs 32

As this implementation contains code from DrQA, so you can try out original DrQA model. It can be trained by adding flag --qanet_tail=False:

python3 train.py -e 25 -bs 32 --qanet_tail=False

Evalutation

To check the score of the model run

python3 train.py --eval --model_dir=path/to/your/model 

Path of provided by default pretrained weights is SQuAD/models/best_model.pt.

Also, you can run model in the interactive mode:

python3 demo.py --model-file=path/to/your/model

Telegram bot

Install telebot pack for python3

python3 pip install pyTelegramBotAPI

To run telegram bot you need to train your model first, then edit telegram_qanet_bot/config.py to change telegram bot token id. Then run it using command

python3 bot.py --model_file=your_model_path