Shape problem in GP priors when `reparameterize=False`

[michaelosthege]

Description of your problem

When passing reparameterize=False to a pm.Latent prior, the resulting variable changes its shape.

with pm.Model() as pmodel:

    ls = pm.LogNormal("ls")
    scaling = pm.LogNormal("scaling")

    mean_func = pm.gp.mean.Zero()
    cov_func = scaling**2 * pm.gp.cov.ExpQuad(input_dim=1, ls=ls)
    gp = pm.gp.Latent(mean_func=mean_func, cov_func=cov_func)
    
    x = numpy.array([1,2,3])
    predA = gp.prior(
        "predA",
        x[:, None],
        size=3,
        reparameterize=True,
    )
    print(predA.shape.eval())
    # [3]

    predB = gp.prior(
        "predB",
        x[:, None],
        size=3,
        reparameterize=False,
    )
    print(predB.shape.eval())
    # [3 3]

Versions and main components

• PyMC/PyMC3 Version: main