Skip to content

hackiey/QAnet-pytorch

Repository files navigation

Introduction

A Pytorch implementation of QANet
This repository is based on NLPLearn/QANet and marquezo/qanet-impl

It can get em: 70.155 and f1: 79.432 peformance after 22 epochs(2730 batches per epoch) with EMA.

Requirements

Usage

Preprocess

$ mkdir data
$ python preprocess.py

Train

$ mkdir log
$ mkdir log/qanet
$ python train.py 'some description'

Evaluate

First set the log folder and epoch number in evaluate.py then execute the script.

$ python evaluate.py

Known issues

  • pickle.dump will get an "OSError: [Errno 22] Invalid argument" error on OS X when saving the "train context char" data, it's ok on Ubuntu 16.04.

About

A pytorch implementation of QAnet

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages