This is a code used in paper "End-To-End Memory Networks" for training MemN2N on bAbI question-answering tasks.
You need Matlab to run this code. We tested it on MATLAB versions R2014a and R2015a. In addition, you have to download bAbI data from fb.ai/babi. Then set base_dir variable in run_babi.m to the path where the data is stored.
Run the following command in Matlab
>> run_babi
This will start training on task 1. Change variable t to train on other tasks. To train on all tasks simultaniously, run
>> run_babi_joint
You can try different model configurations that used in the paper by changing various options in config_babi.m (config_babi_joint.m for joint training). For example, setting
use_bow = true; linear_start = false; randomize_time = 0;
you can train the simplest BOW model in the paper.
You might notice that the performance varies lot in some tasks. In the paper, we repeat training 10 times and picked the one with lowest training error.
- Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, Rob Fergus, "End-To-End Memory Networks", arXiv:1503.08895 [cs.NE].