def predict_continuation(continue_from,
signatures,
session,
steps=None,
times=None,
exogenous_features=None):
"""Perform prediction using an exported saved model.
Analogous to _input_pipeline.predict_continuation_input_fn, but operates on a
saved model rather than feeding into Estimator's predict method.
Args:
continue_from: A dictionary containing the results of either an Estimator's
evaluate method or filter_continuation. Used to determine the model state
to make predictions starting from.
signatures: The `MetaGraphDef` protocol buffer returned from
`tf.compat.v1.saved_model.loader.load`. Used to determine the names of
Tensors to feed and fetch. Must be from the same model as `continue_from`.
session: The session to use. The session's graph must be the one into which
`tf.compat.v1.saved_model.loader.load` loaded the model.
steps: The number of steps to predict (scalar), starting after the
evaluation or filtering. If `times` is specified, `steps` must not be; one
is required.
times: A [batch_size x window_size] array of integers (not a Tensor)
indicating times to make predictions for. These times must be after the
corresponding evaluation or filtering. If `steps` is specified, `times`
must not be; one is required. If the batch dimension is omitted, it is
assumed to be 1.
exogenous_features: Optional dictionary. If specified, indicates exogenous
features for the model to use while making the predictions. Values must
have shape [batch_size x window_size x ...], where `batch_size` matches
the batch dimension used when creating `continue_from`, and `window_size`
is either the `steps` argument or the `window_size` of the `times`
argument (depending on which was specified).
Returns:
A dictionary with model-specific predictions (typically having keys "mean"
and "covariance") and a feature_keys.PredictionResults.TIMES key indicating
the times for which the predictions were computed.
Raises:
ValueError: If `times` or `steps` are misspecified.
"""
if exogenous_features is None:
exogenous_features = {}
predict_times = _model_utils.canonicalize_times_or_steps_from_output(
times=times, steps=steps, previous_model_output=continue_from)
features = {_feature_keys.PredictionFeatures.TIMES: predict_times}
features.update(exogenous_features)
predict_signature = signatures.signature_def[
_feature_keys.SavedModelLabels.PREDICT]
output_tensors_by_name, feed_dict = _colate_features_to_feeds_and_fetches(
continue_from=continue_from,
signature=predict_signature,
features=features,
graph=session.graph)
output = session.run(output_tensors_by_name, feed_dict=feed_dict)
output[_feature_keys.PredictionResults.TIMES] = features[
_feature_keys.PredictionFeatures.TIMES]
return output
def predict_continuation(continue_from,
signatures,
session,
steps=None,
times=None,
exogenous_features=None):
"""Perform prediction using an exported saved model.
Analogous to _input_pipeline.predict_continuation_input_fn, but operates on a
saved model rather than feeding into Estimator's predict method.
Args:
continue_from: A dictionary containing the results of either an Estimator's
evaluate method or filter_continuation. Used to determine the model
state to make predictions starting from.
signatures: The `MetaGraphDef` protocol buffer returned from
`tf.saved_model.loader.load`. Used to determine the names of Tensors to
feed and fetch. Must be from the same model as `continue_from`.
session: The session to use. The session's graph must be the one into which
`tf.saved_model.loader.load` loaded the model.
steps: The number of steps to predict (scalar), starting after the
evaluation or filtering. If `times` is specified, `steps` must not be; one
is required.
times: A [batch_size x window_size] array of integers (not a Tensor)
indicating times to make predictions for. These times must be after the
corresponding evaluation or filtering. If `steps` is specified, `times`
must not be; one is required. If the batch dimension is omitted, it is
assumed to be 1.
exogenous_features: Optional dictionary. If specified, indicates exogenous
features for the model to use while making the predictions. Values must
have shape [batch_size x window_size x ...], where `batch_size` matches
the batch dimension used when creating `continue_from`, and `window_size`
is either the `steps` argument or the `window_size` of the `times`
argument (depending on which was specified).
Returns:
A dictionary with model-specific predictions (typically having keys "mean"
and "covariance") and a feature_keys.PredictionResults.TIMES key indicating
the times for which the predictions were computed.
Raises:
ValueError: If `times` or `steps` are misspecified.
"""
if exogenous_features is None:
exogenous_features = {}
predict_times = _model_utils.canonicalize_times_or_steps_from_output(
times=times, steps=steps, previous_model_output=continue_from)
features = {_feature_keys.PredictionFeatures.TIMES: predict_times}
features.update(exogenous_features)
predict_signature = signatures.signature_def[
_feature_keys.SavedModelLabels.PREDICT]
output_tensors_by_name, feed_dict = _colate_features_to_feeds_and_fetches(
continue_from=continue_from,
signature=predict_signature,
features=features,
graph=session.graph)
output = session.run(output_tensors_by_name, feed_dict=feed_dict)
output[_feature_keys.PredictionResults.TIMES] = features[
_feature_keys.PredictionFeatures.TIMES]
return output
def predict_continuation_input_fn(evaluation,
steps=None,
times=None,
exogenous_features=None):
"""An Estimator input_fn for running predict() after evaluate().
If the call to evaluate() we are making predictions based on had a batch_size
greater than one, predictions will start after each of these windows
(i.e. will have the same batch dimension).
Args:
evaluation: The dictionary returned by `Estimator.evaluate`, with keys
FilteringResults.STATE_TUPLE and FilteringResults.TIMES.
steps: The number of steps to predict (scalar), starting after the
evaluation. If `times` is specified, `steps` must not be; one is required.
times: A [batch_size x window_size] array of integers (not a Tensor)
indicating times to make predictions for. These times must be after the
corresponding evaluation. If `steps` is specified, `times` must not be;
one is required. If the batch dimension is omitted, it is assumed to be 1.
exogenous_features: Optional dictionary. If specified, indicates exogenous
features for the model to use while making the predictions. Values must
have shape [batch_size x window_size x ...], where `batch_size` matches
the batch dimension used when creating `evaluation`, and `window_size` is
either the `steps` argument or the `window_size` of the `times` argument
(depending on which was specified).
Returns:
An `input_fn` suitable for passing to the `predict` function of a time
series `Estimator`.
Raises:
ValueError: If `times` or `steps` are misspecified.
"""
if exogenous_features is None:
exogenous_features = {}
predict_times = model_utils.canonicalize_times_or_steps_from_output(
times=times, steps=steps, previous_model_output=evaluation)
features = {
feature_keys.PredictionFeatures.STATE_TUPLE:
evaluation[feature_keys.FilteringResults.STATE_TUPLE],
feature_keys.PredictionFeatures.TIMES:
predict_times
}
features.update(exogenous_features)
def _predict_input_fn():
"""An input_fn for predict()."""
# Prevents infinite iteration with a constant output in an Estimator's
# predict().
limited_features = {}
for key, values in features.items():
limited_values = nest.map_structure(
lambda value: training.limit_epochs(value, num_epochs=1),
values)
limited_features[key] = limited_values
return (limited_features, None)
return _predict_input_fn
def predict_continuation_input_fn(evaluation,
steps=None,
times=None,
exogenous_features=None):
"""An Estimator input_fn for running predict() after evaluate().
If the call to evaluate() we are making predictions based on had a batch_size
greater than one, predictions will start after each of these windows
(i.e. will have the same batch dimension).
Args:
evaluation: The dictionary returned by `Estimator.evaluate`, with keys
FilteringResults.STATE_TUPLE and FilteringResults.TIMES.
steps: The number of steps to predict (scalar), starting after the
evaluation. If `times` is specified, `steps` must not be; one is required.
times: A [batch_size x window_size] array of integers (not a Tensor)
indicating times to make predictions for. These times must be after the
corresponding evaluation. If `steps` is specified, `times` must not be;
one is required. If the batch dimension is omitted, it is assumed to be 1.
exogenous_features: Optional dictionary. If specified, indicates exogenous
features for the model to use while making the predictions. Values must
have shape [batch_size x window_size x ...], where `batch_size` matches
the batch dimension used when creating `evaluation`, and `window_size` is
either the `steps` argument or the `window_size` of the `times` argument
(depending on which was specified).
Returns:
An `input_fn` suitable for passing to the `predict` function of a time
series `Estimator`.
Raises:
ValueError: If `times` or `steps` are misspecified.
"""
if exogenous_features is None:
exogenous_features = {}
predict_times = model_utils.canonicalize_times_or_steps_from_output(
times=times, steps=steps, previous_model_output=evaluation)
features = {
feature_keys.PredictionFeatures.STATE_TUPLE:
evaluation[feature_keys.FilteringResults.STATE_TUPLE],
feature_keys.PredictionFeatures.TIMES:
predict_times
}
features.update(exogenous_features)
def _predict_input_fn():
"""An input_fn for predict()."""
# Prevents infinite iteration with a constant output in an Estimator's
# predict().
limited_features = {}
for key, values in features.items():
limited_values = nest.map_structure(
lambda value: training.limit_epochs(value, num_epochs=1), values)
limited_features[key] = limited_values
return (limited_features, None)
return _predict_input_fn
