[Question] NHP predicted rmse is instable on taxi data. #56

When training the NHP model with the default taxi data and configuration, I observed that while log-likelihood improves over epochs, the time root-mean-square error (RMSE) does not converge, as shown in the figure below.