TemplateDataset

class neuralhydrology.datasetzoo.template.TemplateDataset(cfg: Config, is_train: bool, period: str, basin: str = None, additional_features: List[Dict[str, pandas.DataFrame]] = [], id_to_int: Dict[str, int] = {}, scaler: Dict[str, pandas.Series | xarray.DataArray] = {})

Bases: BaseDataset

Template class for adding a new data set.

Each dataset class has to derive from BaseDataset, which implements most of the logic for preprocessing data and preparing data for model training. Only two methods have to be implemented for each specific dataset class: _load_basin_data(), which loads the time series data for a single basin, and _load_attributes(), which loads the static attributes for the specific data set.

Usually, we outsource the functions to load the time series and attribute data into separate functions (in the same file), which we then call from the corresponding class methods. This way, we can also use specific basin data or dataset attributes without these classes.

To make this dataset available for model training, don’t forget to add it to the get_dataset() function in ‘neuralhydrology.datasetzoo.__init__.py’

Parameters:

  • cfg (Config) – The run configuration.

  • is_train (bool) – Defines if the dataset is used for training or evaluating. If True (training), means/stds for each feature are computed and stored to the run directory. If one-hot encoding is used, the mapping for the one-hot encoding is created and also stored to disk. If False, a scaler input is expected and similarly the id_to_int input if one-hot encoding is used.

  • period ({'train', 'validation', 'test'}) – Defines the period for which the data will be loaded

  • basin (str, optional) – If passed, the data for only this basin will be loaded. Otherwise the basin(s) are read from the appropriate basin file, corresponding to the period.

  • additional_features (List[Dict[str, pd.DataFrame]], optional) – List of dictionaries, mapping from a basin id to a pandas DataFrame. This DataFrame will be added to the data loaded from the dataset, and all columns are available as ‘dynamic_inputs’, ‘evolving_attributes’ and ‘target_variables’

  • id_to_int (Dict[str, int], optional) – If the config argument ‘use_basin_id_encoding’ is True in the config and period is either ‘validation’ or ‘test’, this input is required. It is a dictionary, mapping from basin id to an integer (the one-hot encoding).

  • scaler (Dict[str, Union[pd.Series, xarray.DataArray]], optional) – If period is either ‘validation’ or ‘test’, this input is required. It contains the centering and scaling for each feature and is stored to the run directory during training (train_data/train_data_scaler.yml).