GP conditional prediction when GP was construced inside user defined function

Questions v5
1

I have a model where I end up generating a lot of GPs, so I made a function:

def do_gp(name):
    lengthscale = pm.Weibull(f'{name}_lengthscale', alpha=2, beta=1) 
    cov = pm.gp.cov.ExpQuad(1, ls=lengthscale, active_dims=[0]) 
    gp = pm.gp.Latent(cov_func = cov)
    out = gp.prior(f'{name}_gp_samples', X=time[:,None])
    return(out)

Used as:

with pm.Model() as model:
    time = pm.MutableData('time', time_np)
    a_samples = do_gp('a')
    b_samples = do_gp('b')
    # etc

But now I’d like to use the gp.conditional() function for each such created gp to obtain posterior predictive samples and I can’t quite discern how to go about doing that when each of the gp objects were declared inside a function as I’ve done. Any suggestions?

2

I think you have to return the GPs from the helper functions

3

Is there a way to assign them as named objects perchance? I’ll eventually be doing some nesting and I’d like to use the name argument to set the name of the gp instances

4

You can always do python stuff. Global variables, attach stuff to PyMC models, wrap everything in a fancy class… and so on