This is an attempt at a TSA Kaggle competition

I used minimal preprocessing and treat the problem as multi-class, though I also attempted to build a suite of 16 binary classifiers to process each region. That worked better, but training the many-layer CNNs was very expensive