def test_day_of_week_example(self):
# Test that the Seasonal SSM is equivalent to individually modeling
# a random walk on each season's slice of timesteps.
seed = test_util.test_seed(sampler_type='stateless')
drift_scale = 0.6
observation_noise_scale = 0.1
day_of_week = SeasonalStateSpaceModel(
num_timesteps=28,
num_seasons=7,
drift_scale=self._build_placeholder(drift_scale),
observation_noise_scale=observation_noise_scale,
initial_state_prior=tfd.MultivariateNormalDiag(
scale_diag=self._build_placeholder(np.ones([7]))),
num_steps_per_season=1)
random_walk_model = tfd.LinearGaussianStateSpaceModel(
num_timesteps=4,
transition_matrix=self._build_placeholder([[1.]]),
transition_noise=tfd.MultivariateNormalDiag(
scale_diag=self._build_placeholder([drift_scale])),
observation_matrix=self._build_placeholder([[1.]]),
observation_noise=tfd.MultivariateNormalDiag(
scale_diag=self._build_placeholder([observation_noise_scale])),
initial_state_prior=tfd.MultivariateNormalDiag(
scale_diag=self._build_placeholder([1.])))
sampled_time_series = day_of_week.sample(seed=seed)
(sampled_time_series_, total_lp_, prior_mean_,
prior_variance_) = self.evaluate([
sampled_time_series,
day_of_week.log_prob(sampled_time_series),
day_of_week.mean(),
day_of_week.variance()
])
expected_daily_means_, expected_daily_variances_ = self.evaluate(
[random_walk_model.mean(),
random_walk_model.variance()])
# For the (noncontiguous) indices corresponding to each season, assert
# that the model's mean, variance, and log_prob match a random-walk model.
daily_lps = []
for day_idx in range(7):
self.assertAllClose(prior_mean_[day_idx::7], expected_daily_means_)
self.assertAllClose(prior_variance_[day_idx::7],
expected_daily_variances_)
daily_lps.append(
self.evaluate(
random_walk_model.log_prob(
sampled_time_series_[day_idx::7])))
self.assertAlmostEqual(total_lp_, sum(daily_lps), places=3)
def test_day_of_week_example(self):
# Test that the Seasonal SSM is equivalent to individually modeling
# a random walk on each season's slice of timesteps.
drift_scale = 0.6
observation_noise_scale = 0.1
day_of_week = SeasonalStateSpaceModel(
num_timesteps=28,
num_seasons=7,
drift_scale=self._build_placeholder(drift_scale),
observation_noise_scale=observation_noise_scale,
initial_state_prior=tfd.MultivariateNormalDiag(
scale_diag=self._build_placeholder(np.ones([7]))),
num_steps_per_season=1)
random_walk_model = tfd.LinearGaussianStateSpaceModel(
num_timesteps=4,
transition_matrix=self._build_placeholder([[1.]]),
transition_noise=tfd.MultivariateNormalDiag(
scale_diag=self._build_placeholder([drift_scale])),
observation_matrix=self._build_placeholder([[1.]]),
observation_noise=tfd.MultivariateNormalDiag(
scale_diag=self._build_placeholder([observation_noise_scale])),
initial_state_prior=tfd.MultivariateNormalDiag(
scale_diag=self._build_placeholder([1.])))
sampled_time_series = day_of_week.sample()
(sampled_time_series_, total_lp_,
prior_mean_, prior_variance_) = self.evaluate([
sampled_time_series,
day_of_week.log_prob(sampled_time_series),
day_of_week.mean(),
day_of_week.variance()])
expected_daily_means_, expected_daily_variances_ = self.evaluate([
random_walk_model.mean(),
random_walk_model.variance()])
# For the (noncontiguous) indices corresponding to each season, assert
# that the model's mean, variance, and log_prob match a random-walk model.
daily_lps = []
for day_idx in range(7):
self.assertAllClose(prior_mean_[day_idx::7], expected_daily_means_)
self.assertAllClose(prior_variance_[day_idx::7],
expected_daily_variances_)
daily_lps.append(self.evaluate(random_walk_model.log_prob(
sampled_time_series_[day_idx::7])))
self.assertAlmostEqual(total_lp_, sum(daily_lps), places=3)
def test_month_of_year_example(self):
num_days_per_month = np.array(
[31, 28, 31, 30, 30, 31, 31, 31, 30, 31, 30, 31])
# put wildly different near-deterministic priors on the effect for each
# month, so we can easily distinguish the months and diagnose off-by-one
# errors.
monthly_effect_prior_means = np.linspace(-1000, 1000, 12)
monthly_effect_prior_scales = 0.1 * np.ones(12)
drift_scale = 0.3
observation_noise_scale = 0.1
num_timesteps = 365 * 2 # 2 years.
initial_step = 22
month_of_year = SeasonalStateSpaceModel(
num_timesteps=num_timesteps,
num_seasons=12,
num_steps_per_season=num_days_per_month,
drift_scale=self._build_placeholder(drift_scale),
observation_noise_scale=observation_noise_scale,
initial_state_prior=tfd.MultivariateNormalDiag(
loc=self._build_placeholder(monthly_effect_prior_means),
scale_diag=self._build_placeholder(
monthly_effect_prior_scales)),
initial_step=initial_step)
sampled_series_, prior_mean_, prior_variance_ = self.evaluate(
(month_of_year.sample()[..., 0], month_of_year.mean()[..., 0],
month_of_year.variance()[..., 0]))
# For each month, ensure the mean (and samples) of each day matches the
# expected effect for that month, and that the variances all match
# including expected drift from year to year
current_idx = -initial_step # start at beginning of year
current_drift_variance = 0.
for _ in range(
3): # loop over three years to include entire 2-year period
for (num_days_in_month, prior_effect_mean,
prior_effect_scale) in zip(num_days_per_month,
monthly_effect_prior_means,
monthly_effect_prior_scales):
month_indices = range(current_idx,
current_idx + num_days_in_month)
current_idx += num_days_in_month
# Trim any indices outside of the observed window.
month_indices = [
idx for idx in month_indices if 0 <= idx < num_timesteps
]
if month_indices:
ones = np.ones(len(month_indices))
self.assertAllClose(sampled_series_[month_indices],
prior_effect_mean * ones,
atol=20.)
self.assertAllClose(prior_mean_[month_indices],
prior_effect_mean * ones,
atol=1e-4)
self.assertAllClose(
prior_variance_[month_indices],
(prior_effect_scale**2 + current_drift_variance +
observation_noise_scale**2) * ones,
atol=1e-4)
# At the end of each year, allow the effects to drift.
current_drift_variance += drift_scale**2
def test_month_of_year_example(self):
num_days_per_month = np.array(
[31, 28, 31, 30, 30, 31, 31, 31, 30, 31, 30, 31])
# put wildly different near-deterministic priors on the effect for each
# month, so we can easily distinguish the months and diagnose off-by-one
# errors.
monthly_effect_prior_means = np.linspace(-1000, 1000, 12)
monthly_effect_prior_scales = 0.1 * np.ones(12)
drift_scale = 0.3
observation_noise_scale = 0.1
num_timesteps = 365 * 2 # 2 years.
initial_step = 22
month_of_year = SeasonalStateSpaceModel(
num_timesteps=num_timesteps,
num_seasons=12,
num_steps_per_season=num_days_per_month,
drift_scale=self._build_placeholder(drift_scale),
observation_noise_scale=observation_noise_scale,
initial_state_prior=tfd.MultivariateNormalDiag(
loc=self._build_placeholder(monthly_effect_prior_means),
scale_diag=self._build_placeholder(monthly_effect_prior_scales)),
initial_step=initial_step)
sampled_series_, prior_mean_, prior_variance_ = self.evaluate(
(month_of_year.sample()[..., 0],
month_of_year.mean()[..., 0],
month_of_year.variance()[..., 0]))
# For each month, ensure the mean (and samples) of each day matches the
# expected effect for that month, and that the variances all match
# including expected drift from year to year
current_idx = -initial_step # start at beginning of year
current_drift_variance = 0.
for _ in range(3): # loop over three years to include entire 2-year period
for (num_days_in_month, prior_effect_mean, prior_effect_scale) in zip(
num_days_per_month,
monthly_effect_prior_means,
monthly_effect_prior_scales):
month_indices = range(current_idx, current_idx + num_days_in_month)
current_idx += num_days_in_month
# Trim any indices outside of the observed window.
month_indices = [idx for idx in month_indices
if 0 <= idx < num_timesteps]
if month_indices:
ones = np.ones(len(month_indices))
self.assertAllClose(sampled_series_[month_indices],
prior_effect_mean * ones, atol=20.)
self.assertAllClose(prior_mean_[month_indices],
prior_effect_mean * ones, atol=1e-4)
self.assertAllClose(prior_variance_[month_indices],
(prior_effect_scale**2 +
current_drift_variance +
observation_noise_scale**2) * ones,
atol=1e-4)
# At the end of each year, allow the effects to drift.
current_drift_variance += drift_scale**2