Defined in tensorflow/contrib/timeseries/python/timeseries/

Perform filtering using an exported saved model.

Filtering refers to updating model state based on new observations. Predictions based on the returned model state will be conditioned on these observations.

Starts from the model's default/uninformed state.


  • 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.
  • features: A dictionary mapping keys to Numpy arrays, with several possible shapes (requires keys FilteringFeatures.TIMES and FilteringFeatures.VALUES): Single example; TIMES is a scalar and VALUES is either a scalar or a vector of length [number of features]. Sequence; TIMES is a vector of shape [series length], VALUES either has shape series length or series length x number of features. Batch of sequences; TIMES is a vector of shape [batch size x series length], VALUES has shape [batch size x series length] or [batch size x series length x number of features]. In any case, VALUES and any exogenous features must have their shapes prefixed by the shape of the value corresponding to the TIMES key.


A dictionary containing model state updated to account for the observations in features.