Implementations of a attention model for entailment from this paper in keras and tensorflow.
Compatible with keras v1.0.6 and tensorflow 0.11.0rc2
I implemented the model to learn the APIs for keras and tensorflow, so I have not really tuned on the performance. The models implemented in keras is a little different, as keras does not expose a method to set a LSTMs state.
To train,
- Download snli dataset.
- Create train, dev, test files with tab separated text, hypothesis and label (example file train10.txt). You can find some snippet in
reader.py
for this, if you are lazy. - Train by either running,
python amodel.py -train <TRAIN> -dev <DEV> -test <TEST>
for using the keras implementation, or
python tf_model.py -train <TRAIN> -dev <DEV> -test <TEST>
for using the tensorflow implementation. Look at the get_params()
method in both scripts to see how to specify different parameters.
Log is written out in *.log file with callback for accuracy.
For comments, improvements, bug-reports and suggestions for tuning, email [email protected]