Hydra (Meta configuration framework)

Meta's open-source Python configuration framework for composing YAML configs and overriding parameters from the command line.

Hydra is an open-source Python configuration framework from Meta (originally Facebook AI Research), built on OmegaConf for reading and writing YAML underneath. It splits an experiment's configuration into several small, composable files — one each for the model, the dataset, the task — assembled as needed at runtime, and lets any field be overridden directly from the command line; its multirun feature can also sweep a whole set of hyperparameters (hyperparameters being values set by hand before training, such as the learning rate) in one go. Robot-learning code tends to have a lot of parameters and a lot of experiments, and Hydra means “change one parameter and rerun” doesn't require touching code — it also automatically creates an output directory for each run and saves the full configuration used at the time, which makes results reproducible. Codebases such as Diffusion Policy and Isaac Lab both use it to manage training configuration.

ExampleRunning python train.py --config-name=train_diffusion_unet_image_workspace task=pusht_image training.seed=42 in the Diffusion Policy repository swaps the task and random seed without touching any code.

Also called

facebookresearch/hydra

Related

Hyperparameter · Random Seed & Reproducibility · Diffusion Policy · NVIDIA Isaac Lab · Weights & Biases (W&B)

Sources

Hydra 官网 (Chinese)
facebookresearch/hydra (GitHub)