Implementation of word-by-word attention in "Reasoning about Entailment with Neural Attention" Paper.
Requirement:
Python 2.7
Tensorflow 1.0.0
Recommended OS: Linux
How to run:
Fill all_data directory with snli data, and preferably fill all_data/embed directory with glove pretrained embedding, change the embedding file name to embedding.N (N is the embedding dimension, for example, embedding.100 means this is a embedding file that store the 100dim embedding)
When all the data is in place, you can train the model as follows:
# generate a config file in save01 dir, if there already is a config,
python model.py --weight-path ./savings/save01
# load config file in ./savings/save01, and start training
python model.py --weight-path ./savings/save01 --load-config
# load the config file in ./savings/save01 along with saved parameters and start run test
python model.py --weight-path ./savings/save01 --load-config --train-test test
# The config file Generating command line is not necessary if you already have one inside the target dir
# config file is used to configure the structure of the model
# You can edit this config file (should comply with some rules of course)
# You can always load an existing config file rather than generating one