regularization

class neuralhydrology.training.regularization.BaseRegularization(cfg: Config, name: str, weight: float = 1.0)

Bases: Module

Base class for regularization terms.

Regularization terms subclass this class by implementing the forward method.

Parameters:

cfg (Config) – The run configuration.
name (str) – The name of the regularization term.
weight (float, optional.) – The weight of the regularization term. Default: 1.

forward(prediction: Dict[str, Tensor], ground_truth: Dict[str, Tensor], other_model_data: Dict[str, Tensor]) → Tensor

Calculate the regularization term.

Parameters:

prediction (Dict[str, torch.Tensor]) – Dictionary of predicted variables for each frequency. If more than one frequency is predicted, the keys must have suffixes _{frequency}. For the required keys, refer to the documentation of the concrete loss.
ground_truth (Dict[str, torch.Tensor]) – Dictionary of ground truth variables for each frequency. If more than one frequency is predicted, the keys must have suffixes _{frequency}. For the required keys, refer to the documentation of the concrete loss.
other_model_data (Dict[str, torch.Tensor]) – Dictionary of all remaining keys-value pairs in the prediction dictionary that are not directly linked to the model predictions but can be useful for regularization purposes, e.g. network internals, weights etc.

Returns:

The regularization value.

Return type:

torch.Tensor

class neuralhydrology.training.regularization.ForecastOverlapMSERegularization(cfg: Config, name: str, weight: float = 1.0)

Bases: BaseRegularization

Squared error regularization for penalizing differences between hindcast and forecast models.

Parameters:: cfg (Config) – The run configuration.

forward(prediction: Dict[str, Tensor], ground_truth: Dict[str, Tensor], *args) → Tensor

Calculate the squared difference between hindcast and forecast model during overlap.

Does not work with multi-frequency models.

Parameters:

prediction (Dict[str, torch.Tensor]) – Dictionary containing y_hindcast_overlap} and y_forecast_overlap.
ground_truth (Dict[str, torch.Tensor]) – Dictionary continaing y_{frequency} for !one! frequency.

Returns:

The sum of mean squared deviations between overlapping portions of hindcast and forecast models.

Return type:

torch.Tensor

Raises:

ValueError if y_hindcast_overlap or y_forecast_overlap is not present in model output. –

class neuralhydrology.training.regularization.TiedFrequencyMSERegularization(cfg: Config, weight: float = 1.0)

Bases: BaseRegularization

Regularization that penalizes inconsistent predictions across frequencies.

This regularization can only be used if at least two frequencies are predicted. For each pair of adjacent frequencies f and f’, where f is a higher frequency than f’, it aggregates the f-predictions to f’ and calculates the mean squared deviation between f’ and aggregated f.

Parameters:

cfg (Config) – The run configuration.
weight (float, optional.) – Weight of the regularization term. Default: 1.

Raises:

ValueError – If the run configuration only predicts one frequency.

forward(prediction: Dict[str, Tensor], ground_truth: Dict[str, Tensor], *args) → Tensor

Calculate the sum of mean squared deviations between adjacent predicted frequencies.

Parameters:

prediction (Dict[str, torch.Tensor]) – Dictionary containing y_hat_{frequency} for each frequency.
ground_truth (Dict[str, torch.Tensor]) – Dictionary continaing y_{frequency} for each frequency.

Returns:

The sum of mean squared deviations for each pair of adjacent frequencies.

Return type:

torch.Tensor