def testInitialValueOfMaskedTimeSeries(self):
if not self.use_static_shape:
return # Dynamic rank is not currently supported.
series = tf.random.stateless_normal(
[2, 1, 3, 4], seed=test_util.test_seed(sampler_type='stateless'))
mask = np.array([
[False, True, False, True],
[True, False, True, False],
# Ensure no error if a batch element is fully masked.
[True, True, True, True]
])
expected_initial_values = distribution_util.move_dimension(
tf.convert_to_tensor(
[series[..., 0, 0], series[..., 1, 1], series[..., 2, 3]]),
source_idx=0,
dest_idx=-1)
initial_values = missing_values_util.initial_value_of_masked_time_series(
self._build_tensor(self.evaluate(series)),
broadcast_mask=self._build_tensor(self.evaluate(
tf.broadcast_to(mask, series.shape)),
dtype=np.bool_))
self.assertAllClose(self.evaluate(initial_values),
expected_initial_values)
def empirical_statistics(observed_time_series):
"""Compute statistics of a provided time series, as heuristic initialization.
Args:
observed_time_series: `Tensor` representing a time series, or batch of time
series, of shape either `batch_shape + [num_timesteps, 1]` or
`batch_shape + [num_timesteps]` (allowed if `num_timesteps > 1`).
Returns:
observed_mean: `Tensor` of shape `batch_shape`, giving the empirical
mean of each time series in the batch.
observed_stddev: `Tensor` of shape `batch_shape`, giving the empirical
standard deviation of each time series in the batch.
observed_initial_centered: `Tensor of shape `batch_shape`, giving the
initial value of each time series in the batch after centering
(subtracting the mean).
"""
with tf.name_scope('empirical_statistics'):
[
observed_time_series,
mask
] = canonicalize_observed_time_series_with_mask(observed_time_series)
squeezed_series = observed_time_series[..., 0]
if mask is None:
observed_mean, observed_variance = tf.nn.moments(
x=squeezed_series, axes=-1)
observed_initial = squeezed_series[..., 0]
else:
broadcast_mask = tf.broadcast_to(tf.cast(mask, tf.bool),
tf.shape(input=squeezed_series))
observed_mean, observed_variance = (
missing_values_util.moments_of_masked_time_series(
squeezed_series, broadcast_mask=broadcast_mask))
try:
observed_initial = (
missing_values_util.initial_value_of_masked_time_series(
squeezed_series, broadcast_mask=broadcast_mask))
except NotImplementedError:
tf1.logging.warn(
'Cannot compute initial values for a masked time series'
'with dynamic shape; using the mean instead. This will'
'affect heuristic priors and may change the results of'
'inference.')
observed_initial = observed_mean
observed_stddev = tf.sqrt(observed_variance)
observed_initial_centered = observed_initial - observed_mean
return observed_mean, observed_stddev, observed_initial_centered
def testInitialValueOfMaskedTimeSeries(self):
if not self.use_static_shape:
return # Dynamic rank is not currently supported.
series = np.random.randn(2, 4)
mask = np.array([[False, True, False, True],
[True, False, True, False]])
expected_initial_values = [series[0, 0], series[1, 1]]
initial_values = missing_values_util.initial_value_of_masked_time_series(
self._build_tensor(series),
broadcast_mask=self._build_tensor(mask, dtype=np.bool))
self.assertAllClose(self.evaluate(initial_values), expected_initial_values)
def empirical_statistics(observed_time_series):
"""Compute statistics of a provided time series, as heuristic initialization.
If a series is entirely unobserved (all values are masked), default statistics
`mean == 0.`, `stddev == 1.`, and `initial_centered == 0.` are returned.
To avoid degenerate models, a value of `1.` is returned for `stddev` whenever
the input series is entirely constant (when the true `stddev` is `0.`).
Args:
observed_time_series: `Tensor` representing a time series, or batch of time
series, of shape either `batch_shape + [num_timesteps, 1]` or
`batch_shape + [num_timesteps]` (allowed if `num_timesteps > 1`).
Returns:
observed_mean: `Tensor` of shape `batch_shape`, giving the empirical
mean of each time series in the batch.
observed_stddev: `Tensor` of shape `batch_shape`, giving the empirical
standard deviation of each time series in the batch.
observed_initial_centered: `Tensor of shape `batch_shape`, giving the
initial value of each time series in the batch after centering
(subtracting the mean).
"""
with tf.name_scope('empirical_statistics'):
[observed_time_series, mask
] = canonicalize_observed_time_series_with_mask(observed_time_series)
squeezed_series = observed_time_series[..., 0]
if mask is None:
observed_mean, observed_variance = tf.nn.moments(x=squeezed_series,
axes=-1)
observed_initial = squeezed_series[..., 0]
else:
broadcast_mask = tf.broadcast_to(tf.cast(mask, tf.bool),
ps.shape(squeezed_series))
observed_mean, observed_variance = (
missing_values_util.moments_of_masked_time_series(
squeezed_series, broadcast_mask=broadcast_mask))
try:
observed_initial = (
missing_values_util.initial_value_of_masked_time_series(
squeezed_series, broadcast_mask=broadcast_mask))
except NotImplementedError:
tf1.logging.warn(
'Cannot compute initial values for a masked time series'
'with dynamic shape; using the mean instead. This will'
'affect heuristic priors and may change the results of'
'inference.')
observed_initial = observed_mean
observed_stddev = tf.sqrt(observed_variance)
observed_initial_centered = observed_initial - observed_mean
# Dividing by zero will estimate `inf` or `nan` for the mean and stddev
# (and thus initial_centered) if a series is entirely masked. Replace these
# with default values.
replace_nans = (lambda x, v: tf.where(tf.math.is_finite(x), x,
tf.cast(v, x.dtype)))
# Avoid stddev of zero from a constant series.
replace_zeros = (
lambda x, v: tf.where(tf.equal(x, 0.), tf.cast(v, x.dtype), x))
return (replace_nans(observed_mean, 0.),
replace_zeros(replace_nans(observed_stddev, 1.),
1.), replace_nans(observed_initial_centered, 0.))
