aselimc Config Management
Python ML Projects
Use Hydra + OmegaConf:
@hydra.main(config_path="configs", config_name="train")
def main(cfg: DictConfig):
model = build_model(cfg.model)
trainer = Trainer(**cfg.trainer)
Or dataclasses + env overrides:
@dataclass
class TrainConfig:
lr: float = 1e-4
batch_size: int = 32
epochs: int = 100
Services (12-Factor)
- Config via environment variables
- Use
pydantic-settingsfor typed env parsing .envfiles for local dev (always in.gitignore)- Secrets via vault or cloud secret managers, never in config files
Rules
- No magic numbers in code - all tunables in config
- Hierarchical: defaults < config file < env vars < CLI args
- Config must be serializable and reproducible
- Log the full config at startup for debugging