def _run_mc_step():
"""Run one mini-batch step of the Monte-Carlo method."""
feed_dict.update({key_placeholder: num_steps})
batch_mean_est, batch_mean_sq_est = tf_session.run(
(mean_est, mean_sq_est), feed_dict=feed_dict)
new_mean_est_eval = util.running_mean_estimate(mean_est_eval,
batch_mean_est,
num_samples, batch_size)
new_mean_sq_est_eval = util.running_mean_estimate(
mean_sq_est_eval, batch_mean_sq_est, num_samples, batch_size)
std_est_eval = util.stddev_est(new_mean_est_eval, new_mean_sq_est_eval)
half_conf_interval = util.half_clt_conf_interval(
confidence, num_samples + batch_size, std_est_eval)
return new_mean_est_eval, new_mean_sq_est_eval, half_conf_interval
def test_running_mean_estimate_is_correct(self):
np.random.seed(0)
num_dims = 8
mean_est = np.random.normal(size=[num_dims])
batch_est = np.random.normal(size=[num_dims])
num_samples = 128
batch_size = 16
updated_mean_est = util.running_mean_estimate(
mean_est, batch_est, num_samples, batch_size)
mean_est_fraction = float(num_samples) / float(num_samples + batch_size)
batch_est_fraction = float(batch_size) / float(num_samples + batch_size)
self.assertAllAlmostEqual(
updated_mean_est,
mean_est_fraction * mean_est + batch_est_fraction * batch_est)