This repository contains a pytorch implementation of end-to-end memory networks (MemNNs) for machine reading comprehension. The implementation is based on this repository.
The code was run using Python 3.5, and pytorch v0.4.1.
Train and evaluate a window-based model on the NE part of the CBT dataset:
python3.5 main.py --mode win --train 1 --lr 0.001 --hops 1 --eval 1 --data-dir CBTest/data/ --ent-setup ent --cuda 1 --epochs 20 --log-epochs 1 --dataset cbt --memory-size 105 --embed-size 100 --win-size-kv 2 --dataset-part NE --exclude-unseen-ans 0
To train on CliCR, first request the dataset, then:
python3.5 main.py --mode win --train 1 --lr 0.001 --hops 1 --eval 1 --data-dir clicr/ --ent-setup ent --cuda 1 --epochs 10 --log-epochs 1 --dataset clicr --memory-size 300 --embed-size 100 --win-size-kv 2 --exclude-unseen-ans 0 --anonymize
See main.py for the full list of options.