def assert_gpr_vs_vgp(
m1: tf.Module,
m2: tf.Module,
gamma: float = 1.0,
maxiter: int = 1,
xi_transform: Optional[gpflow.optimizers.natgrad.XiTransform] = None):
assert maxiter >= 1
m2_ll_before = m2.log_likelihood()
m1_ll_before = m1.log_likelihood()
assert m2_ll_before != m1_ll_before
@tf.function(autograph=False)
def loss_cb() -> tf.Tensor:
return -m2.log_marginal_likelihood()
params = (m2.q_mu, m2.q_sqrt)
if xi_transform is not None:
params += (xi_transform, )
opt = NaturalGradient(gamma)
@tf.function(autograph=False)
def minimize_step():
opt.minimize(loss_cb, var_list=[params])
for _ in range(maxiter):
minimize_step()
m2_ll_after = m2.log_likelihood()
m1_ll_after = m1.log_likelihood()
np.testing.assert_allclose(m1_ll_after, m2_ll_after, atol=1e-4)
def assert_sgpr_vs_svgp(m1: tf.Module, m2: tf.Module):
data = m1.data
m1_ll_before = m1.log_likelihood()
m2_ll_before = m2.log_likelihood(data[0], data[1])
assert m2_ll_before != m1_ll_before
@tf.function(autograph=False)
def loss_cb() -> tf.Tensor:
return -m2.log_marginal_likelihood(data[0], data[1])
params = [(m2.q_mu, m2.q_sqrt)]
opt = NaturalGradient(1.)
opt.minimize(loss_cb, var_list=params)
m1_ll_after = m1.log_likelihood()
m2_ll_after = m2.log_likelihood(data[0], data[1])
np.testing.assert_allclose(m1_ll_after, m2_ll_after, atol=1e-4)