Configuration System¶

Overview¶

Frameworm uses a powerful YAML-based configuration system with support for: - Config inheritance - Environment variable interpolation - CLI overrides - Pydantic validation

Basic Usage¶

Loading a Config¶

from frameworm.core import Config

cfg = Config('configs/model.yaml')
print(cfg.model.name)  # Access with dot notation
print(cfg['model']['name'])  # Or dict-style

Config Inheritance¶

Configs can inherit from other configs using the _base_ key:

# configs/base.yaml
training:
  epochs: 100
  batch_size: 32

# configs/gan.yaml
_base_: ./base.yaml
training:
  epochs: 200  # Override
model:
  type: gan  # Add new

Multiple levels of inheritance are supported.

Environment Variables¶

Use ${VAR_NAME} syntax for environment variables:

paths:
  data_dir: ${DATA_ROOT}/images
  output_dir: ./outputs/${EXPERIMENT_NAME}

CLI Overrides¶

Override config values from command line:

cfg = Config.from_cli_args(
    'config.yaml',
    ['training.epochs=500', 'model.dim=256']
)

Validation¶

Define validation schemas with Pydantic:

from pydantic import BaseModel, Field

class ModelConfig(BaseModel):
    name: str
    dim: int = Field(gt=0)

cfg = Config('model.yaml')
validated = cfg.validate(ModelConfig)

Best Practices¶

Use inheritance - Create a base config and extend it
Validate configs - Define schemas for important configs
Environment variables - For paths and secrets
Freeze configs - After loading to prevent accidental modifications
Dump configs - Save merged configs for reproducibility

Examples¶

See examples/config_examples.py for more usage patterns.