model.runtime.configs.site_learning_config

Site-learning runtime configuration.

Module Contents

Classes

TargetVariableConfig	One observed target variable under site_learning.targets.<name>.
TargetConfig	All observed target variables under the site_learning.targets block.
TrainingConfig	Optimizer and training-schedule settings under site_learning.training.
CrossValidationConfig	Cross-validation protocol under site_learning.cross_validation.
SiteLearningInitializationConfig	NN-weight initialization under site_learning.initialization.

API

class model.runtime.configs.site_learning_config.TargetVariableConfig

Bases: model.runtime.configs.base_config.BaseConfig

One observed target variable under site_learning.targets.<name>.

Attributes

mapping : str or list Name of the variable in the target NetCDF, optionally [name, scale]. layer : int, optional 1-based soil layer to select when the model output is layer-wise. sample_loss_weight : float Relative weight of this target’s sample-level loss in the total loss. site_loss_weight : float Weight of this target’s site-level (per-site mean) loss; 0 disables it.

mapping: object

None

layer: object

None

sample_loss_weight: float

1.0

site_loss_weight: float

1.0

classmethod from_raw(name, raw)

validate()

apply_layer_selection(tensor, label)

class model.runtime.configs.site_learning_config.TargetConfig

Bases: model.runtime.configs.base_config.BaseConfig

All observed target variables under the site_learning.targets block.

Each key is a target name mapped to a :class:TargetVariableConfig. Sample-loss weights are normalised across targets so their relative magnitude, not absolute scale, determines the multi-objective balance.

specs: dict[str, model.runtime.configs.site_learning_config.TargetVariableConfig]

‘field(…)’

classmethod from_dict(raw)

property normalized_sample_loss_weights

property variable_names

property variables

property loader_mapping

validate()

class model.runtime.configs.site_learning_config.TrainingConfig

Bases: model.runtime.configs.base_config.BaseConfig

Optimizer and training-schedule settings under site_learning.training.

Attributes

num_epochs, lr, weight_decay, seed : int / float Core optimizer settings for the Adam optimizer. train_chunk_size : int, optional Number of timesteps per backpropagation chunk; None uses the full record. Smaller chunks reduce memory and backprop depth. max_grad_norm : float Gradient-norm clipping threshold. val_within_train_enabled, val_fraction : bool / float Hold out a fraction of the training data for in-training validation. early_stopping_* : Patience-based early stopping on the validation metric. reduce_lr_* : ReduceLROnPlateau schedule parameters. min_sites_for_site_loss, min_samples_per_site_for_site_loss : int Thresholds below which the site-level loss is skipped (too few sites or too few samples per site to form a stable per-site mean).

num_epochs: int

10

lr: float

0.001

seed: int

42

train_chunk_size: object

None

chunk_months: int

3

max_grad_norm: float

1.0

weight_decay: float

0.0001

loss_fn: str

‘kge’

kge_weights: object

(1.0, 1.0, 1.0)

debug: bool

False

val_within_train_enabled: bool

True

val_fraction: float

0.3

early_stopping_enabled: bool

True

early_stopping_patience: int

10

early_stopping_min_delta: float

0.0

reduce_lr_enabled: bool

True

reduce_lr_patience: int

5

reduce_lr_factor: float

0.5

reduce_lr_min_delta: float

0.0

reduce_lr_min_lr: float

1e-06

min_sites_for_site_loss: int

10

min_samples_per_site_for_site_loss: int

365

classmethod from_dict(raw)

validate()

class model.runtime.configs.site_learning_config.CrossValidationConfig

Bases: model.runtime.configs.base_config.BaseConfig

Cross-validation protocol under site_learning.cross_validation.

Attributes

enabled : bool Whether to run cross-validation rather than a single train/val split. scheme : str or None spatial (split by site) or temporal (split by time period). n_folds : int Number of CV folds. cv_seed, shuffle : int / bool Control random fold assignment for the spatial-random mode. spatial_mode : str random for randomly assigned folds, or predefined to read fold membership from spatial_fold_path. spatial_fold_path : str, optional Path to a fold-definition file; required when spatial_mode is predefined.

enabled: bool

False

scheme: object

None

n_folds: int

5

cv_seed: int

42

spatial_mode: str

‘random’

spatial_fold_path: object

None

shuffle: bool

True

classmethod from_dict(raw)

validate()

class model.runtime.configs.site_learning_config.SiteLearningInitializationConfig

Bases: model.runtime.configs.base_config.BaseConfig

NN-weight initialization under site_learning.initialization.

Attributes

init_nn_weights_path : str, optional Checkpoint of pretrained NN parameter weights to warm-start from. frozen_parameters : list[str] Names of learnable parameters to hold fixed during training.

init_nn_weights_path: object

None

frozen_parameters: list[str]

‘field(…)’

classmethod from_dict(raw)

validate()