I am trying out pymc v4 and I can’t seem to use the .predict method anymore with a trained gaussian process sample trace. I get an error saying the values for my hyperparameters must be included in ‘replacements’. .conditional seems to work though.
I am calling the function as so:
with model:
mu, std = gp.predict(xnew, point = trace)
That’s definitely a bug, we should fix that before v4 is all the way out. Thanks for finding this! If you’d like, the best place for bug reports is the github issue tracker: Issues · pymc-devs/pymc · GitHub
Ok, thanks bwengals. FYI, for the same model in v3 i get very good results if I use find_MAP. In v4 I get results but they are completely off, as in useless. I’m guessing this is a bug too so I’ll do a bug report for both separately when I get a chance.