nh_results_ensemble module

Utility script to average the predictions of several runs.

neuralhydrology.utils.nh_results_ensemble.create_results_ensemble(run_dirs: List[Path], best_k: int = None, metrics: List[str] = None, period: str = 'test', epoch: int = None) → dict

Average the predictions of several runs for the specified period and calculate new metrics.

If best_k is provided, only the k runs with the best validation NSE will be used in the generated ensemble.

Parameters:

  • run_dirs (List[Path]) – List of directories of the runs to be merged

  • best_k (int, optional) – If provided, will only merge the k best runs based on validation NSE.

  • metrics (List[str], optional) – Use this parameter to override the metrics from the config files in the run directories.

  • period ({'test', 'validation', 'train'}, optional) – One of train, val, test. If best_k is used, only ‘test’ is allowed. The run_directories must contain results files for the specified period.

  • epoch (int, optional) – If provided, will ensemble the model predictions of this epoch otherwise of the last epoch

Returns:

Dictionary of ensemble predictions and metrics per basin and frequency.

Return type:

dict