Thanks for the reply. That does seem to be the only option for now. It does make reusing trained models for prediction harder though. Suppose I have data (X1, y1) and wish to train a regression model to predict on data X2 using sample_ppc, then using generate_model(data) will create two different models for which you cannot share parameters / traces. One way to work around this is to pool everything together and draw samples from ([X1, X2], [y1, unk]), but that is not very elegant or flexible.