Hydra

Hydra is an open-source framework created by Facebook AI Research to simplify the creation, management, and usage of configurations in Python applications. Hydra significantly simplifies the complexity of handling multiple configuration sets by making the configuration content dynamic and composable, especially for applications with a large number of parameters and that need to run in various environments.

You can continue to use Hydra for configuration management while leveraging the powerful features of SwanLab.

Track Metrics

As usual, use swanlab.init and swanlab.log to track your metrics.
Assume your Hydra configuration file is configs/defaults.yaml, then add a few lines:

swanlab:
  project: "my-project"

In your training script, pass the project from the configuration file:

import swanlab
import hydra

@hydra.main(config_path="configs/", config_name="defaults")
def run_experiment(cfg):
    run = swanlab.init(project=cfg.swanlab.project)
    ...
    swanlab.log({"loss": loss})

Track Hyperparameters

Hydra uses omegaconf as the default way to interact with configuration dictionaries.

You can directly pass the OmegaConf dictionary to swanlab.config:

@hydra.main(config_path="configs/", config_name="defaults")
def run_experiment(cfg):
    run = swanlab.init(project=cfg.swanlab.project,
                       config=cfg,
    )
    ...
    swanlab.log({"loss": loss})
    model = Model(**swanlab.config.model.configs)

If unexpected results occur when passing cfg, you can first convert omegaconf.DictConfig to the original type:

@hydra.main(config_path="configs/", config_name="defaults")
def run_experiment(cfg):
    run = swanlab.init(project=cfg.swanlab.project,
                       config=omegaconf.OmegaConf.to_container(cfg, resolve=True, throw_on_missing=True)
    ...
    swanlab.log({"loss": loss})
    model = Model(**swanlab.config.model.configs)