Although our work was based on semi-supervised learning, the fully-supervised version still achieves state-of-the-art results in most metrics.