training_label_type instead of a single label_type.  This way, the label
type used for training can be distinct from the type output by the network.
This change breaks backwards compatibility with the previous API.