Advanced Usage

Inheritance Patterns

class BaseConfig(ConfigBase):
    version: str = "1.0.0"
    debug: bool = False

class ModelConfig(BaseConfig):
    dim: int = 512
    dropout: float = 0.1

class TransformerConfig(ModelConfig):
    n_heads: int = 8
    n_layers: int = 12

# All configs inherit version and debug from BaseConfig
config = TransformerConfig()
assert config.version == "1.0.0"
assert config.dim == 512
assert config.n_heads == 8

Type Validation

class AdvancedConfig(ConfigBase):
    # Union types with automatic selection
    layers: Union[int, List[int]] = 12

    # Optional fields with None handling
    checkpoint_path: Optional[Path] = None

    # Complex nested structures
    hyperparams: dict = {"lr": 0.001, "weight_decay": 0.01}

# Usage examples
config = AdvancedConfig(
    layers="[12, 24, 48]",  # String to list conversion
    checkpoint_path="/path/to/checkpoint.pt",  # String to Path
    hyperparams={"lr": "0.0001"}  # Partial dict update
)

Serialization

You can export your configs to different formats.

class ExportableConfig(ConfigBase):
    model: ModelConfig = TransformerConfig()
    optimizer: OptimizerConfig = AdamW()
    training: dict = {"epochs": 100, "batch_size": 32}

config = ExportableConfig()

# Export to different formats
flat_dict = config.to_dict(flatten=True)
nested_dict = config.to_dict(flatten=False)

# Import from dictionary
new_config = make_config_from_flat_dict(ExportableConfig, flat_dict)
new_config_nested = make_config_from_nested_dict(ExportableConfig, nested_dict)

Note that this could be easily extended to other formats.