def testBasic(self, centered, use_fft):
"""Checks that you get finite values given unconstrained samples.
We check `log_prob` as well as the values of the expectations.
Args:
centered: Whether or not to use the centered parameterization.
use_fft: Whether or not to use FFT-based convolution to implement the
centering transformation.
"""
model = vectorized_stochastic_volatility.VectorizedStochasticVolatility(
centered_returns=_test_dataset(), centered=centered, use_fft=use_fft)
self.validate_log_prob_and_transforms(
model,
sample_transformation_shapes=dict(
identity={
'persistence_of_volatility': [],
'mean_log_volatility': [],
'white_noise_shock_scale': [],
'log_volatility': [5]
}))
def testParameterizationsConsistent(self):
"""Run HMC for both parameterizations, and compare posterior means."""
self.skipTest('Broken by omnistaging b/168705919')
centered_returns = _test_dataset()
centered_model = (
vectorized_stochastic_volatility.VectorizedStochasticVolatility(
centered_returns, centered=True))
non_centered_model = (
vectorized_stochastic_volatility.VectorizedStochasticVolatility(
centered_returns, centered=False, use_fft=False))
non_centered_fft_model = (
vectorized_stochastic_volatility.VectorizedStochasticVolatility(
centered_returns, centered=False, use_fft=True))
logging.info('Centered:')
centered_results = self.evaluate(
test_util.run_hmc_on_model(
centered_model,
num_chains=4,
num_steps=1000,
num_leapfrog_steps=10,
step_size=0.1,
# TF XLA is very slow on this problem
use_xla=BACKEND == 'backend_jax',
target_accept_prob=0.7))
logging.info('Acceptance rate: %s', centered_results.accept_rate)
logging.info('ESS: %s', centered_results.ess)
logging.info('r_hat: %s', centered_results.r_hat)
logging.info('Non-centered (without FFT):')
non_centered_results = self.evaluate(
test_util.run_hmc_on_model(
non_centered_model,
num_chains=4,
num_steps=1000,
num_leapfrog_steps=10,
step_size=0.1,
# TF XLA is very slow on this problem
use_xla=BACKEND == 'backend_jax',
target_accept_prob=0.7))
logging.info('Acceptance rate: %s', non_centered_results.accept_rate)
logging.info('ESS: %s', non_centered_results.ess)
logging.info('r_hat: %s', non_centered_results.r_hat)
logging.info('Non-centered (with FFT):')
non_centered_fft_results = self.evaluate(
test_util.run_hmc_on_model(
non_centered_fft_model,
num_chains=4,
num_steps=1000,
num_leapfrog_steps=10,
step_size=0.1,
# TF XLA is very slow on this problem
use_xla=BACKEND == 'backend_jax',
target_accept_prob=0.7))
logging.info('Acceptance rate: %s', non_centered_fft_results.accept_rate)
logging.info('ESS: %s', non_centered_fft_results.ess)
logging.info('r_hat: %s', non_centered_fft_results.r_hat)
centered_params = self.evaluate(
tf.nest.map_structure(
tf.identity, centered_model.sample_transformations['identity'](
centered_results.chain)))
non_centered_params = self.evaluate(
tf.nest.map_structure(
tf.identity, non_centered_model.sample_transformations['identity'](
non_centered_results.chain)))
non_centered_fft_params = self.evaluate(
tf.nest.map_structure(
tf.identity,
non_centered_fft_model.sample_transformations['identity'](
non_centered_fft_results.chain)))
def get_mean_and_var(chain):
def one_part(chain):
mean = chain.mean((0, 1))
var = chain.var((0, 1))
ess = 1. / (1. / self.evaluate(
tfp.mcmc.effective_sample_size(
chain, filter_beyond_positive_pairs=True))).mean(0)
return mean, var / ess
mean_var = tf.nest.map_structure(one_part, chain)
return (nest.map_structure_up_to(chain, lambda x: x[0], mean_var),
nest.map_structure_up_to(chain, lambda x: x[1], mean_var))
centered_mean, centered_var = get_mean_and_var(centered_params)
non_centered_mean, non_centered_var = get_mean_and_var(non_centered_params)
non_centered_fft_mean, non_centered_fft_var = get_mean_and_var(
non_centered_fft_params)
def get_atol(var1, var2):
# TODO(b/144290399): Use the full atol vector.
max_var_per_rv = tf.nest.map_structure(
lambda v1, v2: (3. * np.sqrt(v1 + v2)).max(), var1, var2)
return functools.reduce(max, tf.nest.flatten(max_var_per_rv))
self.assertAllCloseNested(
centered_mean,
non_centered_mean,
atol=get_atol(centered_var, non_centered_var),
)
self.assertAllCloseNested(
centered_mean,
non_centered_fft_mean,
atol=get_atol(centered_var, non_centered_fft_var),
)
