train_lstm.py 文件源码

def _filtering_step(self, current_times, current_values, state, predictions):
    """Update model state based on observations.
    Note that we don't do much here aside from computing a loss. In this case
    it's easier to update the RNN state in _prediction_step, since that covers
    running the RNN both on observations (from this method) and our own
    predictions. This distinction can be important for probabilistic models,
    where repeatedly predicting without filtering should lead to low-confidence
    predictions.
    Args:
      current_times: A [batch size] integer Tensor.
      current_values: A [batch size, self.num_features] floating point Tensor
        with new observations.
      state: The model's state tuple.
      predictions: The output of the previous `_prediction_step`.
    Returns:
      A tuple of new state and a predictions dictionary updated to include a
      loss (note that we could also return other measures of goodness of fit,
      although only "loss" will be optimized).
    """
    state_from_time, prediction, lstm_state = state
    with tf.control_dependencies(
            [tf.assert_equal(current_times, state_from_time)]):
      transformed_values = self._transform(current_values)
      # Use mean squared error across features for the loss.
      predictions["loss"] = tf.reduce_mean(
          (prediction - transformed_values) ** 2, axis=-1)
      # Keep track of the new observation in model state. It won't be run
      # through the LSTM until the next _imputation_step.
      new_state_tuple = (current_times, transformed_values, lstm_state)
    return (new_state_tuple, predictions)