`model.runtime.configs.learning_config`¶

Unified learning configuration.

One learning run = a single coupled forward + a single optimiser, with one or more loss heads attached to it. The shared run settings (time, spinup, weight initialization, optimiser/training, output dir) live directly under learning; each loss head contributes only its own data + how to split its entities:

site_level – flux-tower (vertical) constraint: domain + targets.
gauge_level – gauge-discharge (lateral) constraint: gauges + target.

Per-target loss weights live on the targets; k-fold CV is per head (it folds over that head’s entities). At least one loss head must be present.

Module Contents¶

Classes¶

`GaugeSelectionConfig`	`learning.gauge_level.gauges` – gauge discharge data + selection filters.
`SiteLevelConfig`	`learning.site_level` – flux-tower loss head (data + entity split).
`GaugeLevelConfig`	`learning.gauge_level` – gauge-discharge loss head (data + entity split).
`LearningConfig`	Top-level `learning` block: one forward + one optimiser + loss heads.

Data¶

__all__

API¶

class model.runtime.configs.learning_config.GaugeSelectionConfig¶

Bases: model.runtime.configs.base_config.BaseConfig

learning.gauge_level.gauges – gauge discharge data + selection filters.

The river-network file (gauge allocation + routing geometry) is read from model.schemes.river_routing.parameters_path, not repeated here.

discharge_path: object¶: None

allocation_path: object¶: None

selection: object¶: None

area_tol: float¶: 0.2

min_area_km2: float¶: 1000.0

min_days: int¶: 730

min_obs_runoff: float¶: 0.05

classmethod from_dict(raw)¶

validate()¶

class model.runtime.configs.learning_config.SiteLevelConfig¶

Bases: model.runtime.configs.base_config.BaseConfig

learning.site_level – flux-tower loss head (data + entity split).

domain: model.runtime.configs.site_simulation_config.DomainConfig¶: ‘field(…)’

targets_path: object¶: None

targets: model.runtime.configs.site_learning_config.TargetConfig¶: ‘field(…)’

cross_validation: model.runtime.configs.site_learning_config.CrossValidationConfig¶: ‘field(…)’

classmethod from_dict(raw)¶

is_enabled()¶

validate()¶

class model.runtime.configs.learning_config.GaugeLevelConfig¶

Bases: model.runtime.configs.base_config.BaseConfig

learning.gauge_level – gauge-discharge loss head (data + entity split).

gauges: model.runtime.configs.learning_config.GaugeSelectionConfig¶: ‘field(…)’

targets: model.runtime.configs.site_learning_config.TargetConfig¶: ‘field(…)’

cross_validation: model.runtime.configs.site_learning_config.CrossValidationConfig¶: ‘field(…)’

classmethod from_dict(raw)¶

is_enabled()¶

validate()¶

class model.runtime.configs.learning_config.LearningConfig¶

Bases: model.runtime.configs.base_config.BaseConfig

Top-level learning block: one forward + one optimiser + loss heads.

Shared run settings live here; site_level / gauge_level add the loss heads. The parameters to calibrate and their sources are declared in the standard parameterization block.