nptdat / qanet

Implementation for QANet using Keras with Tensorflow backend

Overview

Implementation for QANet using Keras with Tensorflow backend.

Preparation (QANet implementation)

Setting variables are defined in src/squad/config.py

0. Preparation

   • Install Python packages in requirements.txt
   • Install English corpus for Spacy: python -m spacy download en

1. Download glove and extract the file glove.6B.300d.txt to data/glove (Setting variable: EMBEDDING_FILE)

2. Download SQuAD data v1.1 and extract the files train-v1.1.json & dev-v1.1.json to data/SQUAD_Data/v1.1 (Setting variable: TRAIN_JSON & DEV_JSON)

For inference using pre-trained model

• Download the model file qanet_ep20.h5 from https://github.com/nptdat/qanet/releases/download/v1.0/qanet_ep20.h5 and put it into model folder. (Setting variable: INFERENCE_MODEL_PATH)

• If you use the above model, I recommend you to download the following files from https://github.com/nptdat/qanet/releases/download/v1.0 to ensure the data consistence:

  • squad_processed-v1.1.pkl.zip: unzip and move the pickle file to data/SQUAD_Data/v1.1/
  • numpy_files.zip: unzip and move all the .npy files to data/SQUAD_Data/v1.1/numpy/
  • Data from these files will overwrite those generated from build_squad_data.py

• Run

$ FLASK_APP=demo_qanet.py flask run --host=0.0.0.0 --port=8080

Then access http://localhost:8080/qanet via browser.

For training

1. Run build_squad_data.py to load SQuAD data from json files, transform the data and save to .pkl files

$ python build_squad_data.py

2. Run train.py

$ python train.py

• Model files will be saved to model folder, 1 model per epoch
• Tensorboard log data will be saved to log/tensorboard
• Please take a look at config.py for further setting

Unit Test

Please read src/squad/test/README.md