Skip to main content

Prediction error

The prediction error (PE) acts as a special evaluator and thus has its own config. If set, it calculates the n-step-ahead PE for all evaluation timeseries both in the latent space (between the encoded and DSR-generated trajectories) and in all modalities it is applicable to. Currently this is the Gaussian and all related evaluators.

Setting the PE adds one modality-independent evaluator, the PEEvaluator and adds PE metrics to all Gaussian- and related evaluators already set in the config.

The PE can be configured through the PE_Args key. This expects a dictionary with following keys:

  • compute_PE (bool, default=False): Whether to calculate the PE.
  • n_steps_ahead (int, default=1): The n-step-ahead prediction's length.
  • min_context_length (int, default=20): The minimal timeseries length given to the evaluator to initialize the first step. This is not given to the DSR model and is only used for the latent initialization. Useful for convolutional networks.
  • batch_size (int, default=-1): The batch size for processing prediction windows. If set to -1, it processes all windows in a single batch.