Passing dims=(dim1, dim2) in gp.prior leads to shape inconcistency

[ferrine]

User can't safely pass some broadcast dim to create multioutput GP with same parameters

pymc/pymc/gp/gp.py

Lines 152 to 155 in 261862d

	if reparameterize:
	size = np.shape(X)[0]
	v = pm.Normal(name + "_rotated_", mu=0.0, sigma=1.0, size=size, **kwargs)
	f = pm.Deterministic(name, mu + cholesky(cov).dot(v), dims=kwargs.get("dims", None))

Possible implementation

def _build_prior(self, name, X, reparameterize=True, jitter=JITTER_DEFAULT, **kwargs):
    mu = self.mean_func(X)
    cov = stabilize(self.cov_func(X), jitter)
    if reparameterize:
        if "dims" not in kwargs:
            size = kwargs.get("size", np.shape(X)[0])
        else:
            size = None
        v = pm.Normal(name + "_rotated_", mu=0.0, sigma=1.0, size=size, **kwargs)
        f = pm.Deterministic(name, mu + v.dot(cholesky(cov).T), dims=kwargs.get("dims", None))
    else:
        f = pm.MvNormal(name, mu=mu, cov=cov, **kwargs)
    return f