In output patch-based rebalancing, we should explicitly handle divide-by-zero when computing class-specific weights. It's possible that an output patch inclues one of the classes exclusively, and the probability of this increases as the output patch size decreases.
https://github.com/seung-lab/znn-release/blob/v4python/python/ZNN_Dataset.py#L612
In output patch-based rebalancing, we should explicitly handle divide-by-zero when computing class-specific weights. It's possible that an output patch inclues one of the classes exclusively, and the probability of this increases as the output patch size decreases.