Sample prior predictive for Gaussian process model

Questions
1

Hi there!
Is it possible to get prior predictive samples for Gaussian process models? I’d like something like this (MWE taken from the docs):

x = np.linspace(0,1,50)
y = 0.5 *x + np.random.randn(50) * 0.01

with pm.Model() as gp_model:
    cov_func = pm.gp.cov.ExpQuad(1, ls=0.1)
    gp = pm.gp.Marginal(cov_func=cov_func)
    sigma = pm.HalfCauchy("sigma", beta=5)
    y_ = gp.marginal_likelihood("y", X=x[:,None], y=y, noise=sigma)
    y_star = gp.conditional("y_star", x[:,None])

with gp_model:
    prior = pm.sample_prior_predictive()

to create prior predictive samples for y_star? However atm this always returns an array of NaNs.
Thank you!

2

It seems to be a bug? @bwengals?

3

@junpenglao @bwengals So is the way I specify the prior predictive correct? Happy to attempt to fix this if this is a bug!

4

To sample the prior predictive correctly I think you’d need to use gp.Latent. The gp.Marginal implementation marginalizes out the actual GP, which makes it more efficient, but the actual GP f isn’t explicitly included in the model.

1 Like