Added support for probability calibration and threshold tuning, shifted to new preferred approach where custom thresholds and calibrated probability distributions are used rather than reweighting for imbalanced classes. Other changes: removed reweight_classes constructor parameter (but just always calculate weights in case people want to use them); used random_state for tensorflow seed; tweaked verbose outputs; stopped collecting and reporting average_precision; reverted to current (vs. legacy) Adam optimizer for tensorflow, as they seem to have resolved the problem with cross-validation (and a problem with the legacy Adam optimizer emerged).