A chainer implementation of Densely Interactive Inference Network (arXiv:1709.04348)
Requirements:
- python 3.6+ (verified)
- Chainer 3.1.0+ (verified)
- Preprocessed MultiNLI/SNLI data from the original DIIN (github)
To Run,
python train.py
For debug with a small data:
python train.py --debug_mode
TODO:
- Learning rate in adadelta
- Use SGD in the middle of training
- Early stopping
- Confusion matrix
- Use ParallelUpdater? (for data parallel processing)
- Improve performance...