tf.contrib.timeseries.OneShotPredictionHead

View source on GitHub

A time series head which exports a single stateless serving signature.

tf.contrib.timeseries.OneShotPredictionHead(
    model, state_manager, optimizer, input_statistics_generator=None, name=None
)

The serving default signature exported by this head expects times, values, and any exogenous features, but no state. values has shape [batch_size, filter_length, num_features] and times has shape [batch_size, total_length], where total_length > filter_length. Any exogenous features must have their shapes prefixed by the shape of the times feature.

When serving, first performs filtering on the series up to filter_length starting from the default start state for the model, then computes predictions on the remainder of the series, returning them.

Model state is neither accepted nor returned, so filtering must be performed each time predictions are requested when using this head.

Methods

create_estimator_spec

View source

create_estimator_spec(
    features, mode, labels=None
)

Performs basic error checking and returns an EstimatorSpec.

create_loss

View source

create_loss(
    features, mode, logits=None, labels=None
)

See _Head.