def testMarkovChainLogprobMatchesOriginal(self):
model = lorenz_system.ConvectionLorenzBridgeUnknownScales(
use_markov_chain=False)
markov_chain_model = lorenz_system.ConvectionLorenzBridgeUnknownScales(
use_markov_chain=True)
x = self.evaluate(model.prior_distribution().sample(
20, seed=tfp_test_util.test_seed()))
self.assertAllClose(model.unnormalized_log_prob(x),
markov_chain_model.unnormalized_log_prob(
type(markov_chain_model.dtype)(x[0], x[1],
tf.stack(
x[2:],
axis=-2))),
atol=1e-2)
def testConvectionLorenzBridgeHMC(self):
"""Checks approximate samples from the model against the ground truth."""
self.skipTest('b/171518508')
model = lorenz_system.ConvectionLorenzBridgeUnknownScales()
self.validate_ground_truth_using_hmc(
model,
num_chains=2,
num_steps=2000,
num_leapfrog_steps=240,
step_size=0.03,
)
def testConvectionLorenzBridge(self):
"""Checks that unconstrained parameters yield finite joint densities."""
model = lorenz_system.ConvectionLorenzBridgeUnknownScales()
self.validate_log_prob_and_transforms(
model,
sample_transformation_shapes=dict(
identity={
'innovation_scale': [],
'observation_scale': [],
'latents': [30, 3]
}),
check_ground_truth_mean_standard_error=True,
check_ground_truth_mean=True,
check_ground_truth_standard_deviation=True)
