MultiheadForecastLSTM

class neuralhydrology.modelzoo.multihead_forecast_lstm.MultiHeadForecastLSTM(cfg: Config)

Bases: BaseModel

A forecasting model that does not roll out over the forecast horizon.

This is a forecasting model that runs a sequential (LSTM) model up to the forecast issue time, and then directly predicts a sequence of forecast timesteps without using a recurrent rollout. Prediction is done with a custom FC (fully connected) layer, which can include depth.

Do not use this model with forecast_overlap > 0.

Parameters:

cfg (Config) – The run configuration.

Raises:

  • ValueError if forecast_overlap > 0. –

  • ValueError if a forecast_network is not specified. –

forward(data: Dict[str, Tensor]) → Dict[str, Tensor]

Perform a forward pass on the MultiheadForecastLSTM model.

Parameters:

data (Dict[str, torch.Tensor]) – Dictionary, containing input features as key-value pairs.

Returns:

Model outputs and intermediate states as a dictionary.

  • lstm_output_hindcast: Output sequence from the hindcast LSTM.

  • output_forecast: Predictions (before head layer) from the forecast period.

  • h_n_hindcast: Final hidden state of the hindcast model.

  • c_n_hindcast: Final cell state of the hindcast model.

  • y_hat: Predictions over the sequence from the head layer.

Return type:

Dict[str, torch.Tensor]

module_parts = ['forecast_mebedding_net', 'hindcast_embedding_net', 'hindcast_lstm', 'forecast_network', 'hindcast_head', 'forecast_head']