neural_lam.models.step_predictors.persistence

neural_lam.models.step_predictors.persistence#

Persistence baseline step predictor.

Classes#

PersistencePredictor

Trivial baseline predictor that returns the previous state unchanged.

Module Contents#

class neural_lam.models.step_predictors.persistence.PersistencePredictor(datastore: neural_lam.datastore.BaseDatastore, output_std: bool = False, output_clamping_lower: dict[str, float] | None = None, output_clamping_upper: dict[str, float] | None = None, **_kwargs: object)[source]#

Bases: neural_lam.models.step_predictors.base.StepPredictor

Trivial baseline predictor that returns the previous state unchanged.

At each AR step the predicted next state is simply the current state, i.e. X_{t+1} = X_t. This provides a persistence (no-change) baseline that can be evaluated through the standard ForecasterModule / ARForecaster pipeline.

Initialize the PersistencePredictor.

Parameters:

  • datastore (BaseDatastore) – The datastore providing grid metadata and data access.

  • output_std (bool, default False) – Ignored, a warning is logged if True. Persistence never predicts uncertainty.

  • output_clamping_lower (dict, optional) – Ignored, persistence returns the raw previous state.

  • output_clamping_upper (dict, optional) – Ignored, persistence returns the raw previous state.

  • **_kwargs – Absorbed so that the standard CLI kwargs (graph_name, hidden_dim, etc.) do not cause errors.

forward(prev_state: torch.Tensor, prev_prev_state: torch.Tensor, forcing: torch.Tensor) → tuple[torch.Tensor, None][source]#

Return the previous state as the prediction.

Parameters:

  • prev_state (torch.Tensor) – Shape (B, num_grid_nodes, num_state_vars). The current state X_t.

  • prev_prev_state (torch.Tensor) – Shape (B, num_grid_nodes, num_state_vars). The previous state X_{t-1} (unused).

  • forcing (torch.Tensor) – Shape (B, num_grid_nodes, num_forcing_vars). External forcings (unused).

Returns:

  • pred_state (torch.Tensor) – Shape (B, num_grid_nodes, num_state_vars). Equal to prev_state.

  • pred_std (None) – Always None, persistence does not predict uncertainty.