def _build_sts(self, observed_time_series=None):
max_timesteps = 100
num_features = 3
return DynamicLinearRegression(
design_matrix=tf.random.normal([max_timesteps, num_features]),
observed_time_series=observed_time_series)
def _build_sts(self, observed_time_series=None):
max_timesteps = 100
num_features = 3
return DynamicLinearRegression(
design_matrix=np.random.randn(max_timesteps,
num_features).astype(np.float32),
observed_time_series=observed_time_series)
def test_matrices_from_component(self):
num_timesteps = 4
num_features = 3
drift_scale = 1.23
design_matrix = self._build_placeholder(
np.random.randn(num_timesteps, num_features))
component = DynamicLinearRegression(design_matrix=design_matrix)
ssm = component.make_state_space_model(num_timesteps, [drift_scale])
for t in range(num_timesteps):
observation_matrix = self.evaluate(
ssm.get_observation_matrix_for_timestep(t).to_dense())
self.assertAllClose(self.evaluate(design_matrix[tf.newaxis, t]),
observation_matrix)
observation_noise_mean = self.evaluate(
ssm.get_observation_noise_for_timestep(t).mean())
observation_noise_covariance = self.evaluate(
ssm.get_observation_noise_for_timestep(t).covariance())
self.assertAllClose([0.0], observation_noise_mean)
self.assertAllClose([[0.0]], observation_noise_covariance)
transition_matrix = self.evaluate(
ssm.get_transition_matrix_for_timestep(t).to_dense())
self.assertAllClose(np.eye(num_features), transition_matrix)
transition_noise_mean = self.evaluate(
ssm.get_transition_noise_for_timestep(t).mean())
transition_noise_covariance = self.evaluate(
ssm.get_transition_noise_for_timestep(t).covariance())
self.assertAllClose(np.zeros([num_features]),
transition_noise_mean)
self.assertAllClose(
np.square(drift_scale) * np.eye(num_features),
transition_noise_covariance)
def _build_sts(self, observed_time_series=None):
max_timesteps = 100
num_features = 3
dtype = np.float32
if observed_time_series is not None:
observed_time_series_tensor, _ = (
sts_util.canonicalize_observed_time_series_with_mask(
observed_time_series))
dtype = dtype_util.as_numpy_dtype(observed_time_series_tensor.dtype)
return DynamicLinearRegression(
design_matrix=np.random.randn(
max_timesteps, num_features).astype(dtype),
observed_time_series=observed_time_series)
