This implements a network introduced in the arXiv paper: Hierarchical Attention Networks
You first need to place MNIST dataset in ./data/mnist/
- generate_mdist.py generates the MDIST synthetic dataset.
- pack.py packs the generated images and labels into a .npz file.
- train.py trains a progressive attention network with local context of size 3x3 and tests it at every epoch.
If you're using this code in a publication, please cite our paper.
@article{seo2016hierarchical,
title={Hierarchical Attention Networks},
author={Seo, Paul Hongsuck and Lin, Zhe and Cohen, Scott and Shen, Xiaohui and Han, Bohyung},
journal={arXiv preprint arXiv:1606.02393},
year={2016}
}