For example, ReduceLROnPlateau(val_loss) vs ReduceLROnPlateau(val_acc). Its confusing since in some epochs loss increases and accuracy increases. They’re not 100% correlated on every epochs.
This might be the case and you could adapt the patience
value for it.
E.g. if you current training sees a lot of these stats, you might want to use a higher patience to avoid decreasing the learning rate too early.