def testSignAdaptation(self):
if not tf.executing_eagerly():
return
new_control = fun_mcmc.sign_adaptation(
control=1., output=0.5, set_point=1., adaptation_rate=0.1)
self.assertAllClose(new_control, 1. / 1.1)
new_control = fun_mcmc.sign_adaptation(
control=1., output=0.5, set_point=0., adaptation_rate=0.1)
self.assertAllClose(new_control, 1. * 1.1)
def testSignAdaptation(self):
new_control = fun_mcmc.sign_adaptation(control=1.,
output=0.5,
set_point=1.,
adaptation_rate=0.1)
self.assertAllClose(new_control, 1. / 1.1)
new_control = fun_mcmc.sign_adaptation(control=1.,
output=0.5,
set_point=0.,
adaptation_rate=0.1)
self.assertAllClose(new_control, 1. * 1.1)
def kernel(hmc_state, step_size, step):
hmc_state, hmc_extra = fun_mcmc.hamiltonian_monte_carlo(
hmc_state,
step_size=step_size,
num_integrator_steps=num_leapfrog_steps,
target_log_prob_fn=target_log_prob_fn)
rate = tf.compat.v1.train.polynomial_decay(
0.01,
global_step=step,
power=0.5,
decay_steps=num_adapt_steps,
end_learning_rate=0.)
mean_p_accept = tf.reduce_mean(input_tensor=tf.exp(
tf.minimum(0., hmc_extra.log_accept_ratio)))
step_size = fun_mcmc.sign_adaptation(step_size,
output=mean_p_accept,
set_point=0.9,
adaptation_rate=rate)
return (hmc_state, step_size, step + 1), hmc_extra