Hydra experiment configs¶

Hydra is the recommended entry point when you want to run repeatable experiments without copying and editing a full YAML file for every small change. The old single-file config workflow still works; Hydra is an additional layer that composes the same Data, Model, Training, Generator, Discriminator, Optimizers, Schedulers, and Logging sections that the SRGAN code already understands.

Mental model¶

The legacy command loads one complete YAML file:

python -m opensr_srgan.train --config opensr_srgan/configs/config_training_example.yaml

The Hydra command starts from a small default config, selects reusable config groups, applies an optional experiment preset, and then applies command-line overrides:

python -m opensr_srgan.train_hydra experiment=10m Training.max_epochs=5 Logging.wandb.enabled=false

After composition, opensr_srgan.train_hydra forwards the resolved OmegaConf object to the same opensr_srgan.train.train() function used by the legacy workflow. This keeps the model, data, checkpoint, and logger code paths shared.

Config layout¶

Hydra configs live in opensr_srgan/configs/hydra/:

Path	Purpose
`train.yaml`	Default Hydra entry config. It selects the bundled example dataset, small CPU-safe training settings, CSV logging, and `experiment: null`.
`data/`	Dataset and dataloader presets such as `example`, `sen2naip`, and `lrhr_folder`.
`model/`	Combined `Model`, `Generator`, and `Discriminator` presets.
`training/`	Combined `Training`, `Optimizers`, and `Schedulers` presets.
`logging/`	CSV and W&B logging presets.
`experiment/`	Full experiment presets that reproduce the shipped legacy YAMLs.
`hydra/default.yaml`	Hydra's own run-directory and working-directory settings.

The important convention is that each group writes into the existing top-level SRGAN schema. For example, model/10m_esrgan.yaml contributes Model, Generator, and Discriminator, not a new lowercase model object.

Choosing a config¶

Use experiment=... when you want a known full setup:

python -m opensr_srgan.train_hydra experiment=example
python -m opensr_srgan.train_hydra experiment=10m
python -m opensr_srgan.train_hydra experiment=20m

Use direct overrides when you only want to change one or two values for a run:

python -m opensr_srgan.train_hydra experiment=10m Training.max_epochs=20
python -m opensr_srgan.train_hydra experiment=10m Generator.n_blocks=8 Optimizers.optim_g_lr=5e-5
python -m opensr_srgan.train_hydra experiment=20m Training.gpus=[0,1] Logging.wandb.enabled=false

Use group overrides when you want to mix reusable pieces:

python -m opensr_srgan.train_hydra data=example model=example_srresnet training=cpu_debug logging=csv
python -m opensr_srgan.train_hydra data=sen2naip model=10m_esrgan training=10m logging=wandb_10m

If a field does not already exist in the composed config, Hydra requires a + prefix. For normal SRGAN fields this is rarely needed because the presets define the expected sections up front.

Output directories¶

Hydra creates a run directory under:

logs/<wandb-project>/<timestamp>

The Hydra entry point copies that directory into Logging.output_dir before calling the existing trainer. The trainer then writes checkpoints and the resolved config.yaml there.

Hydra is configured with:

hydra:
  job:
    chdir: false

This means relative paths keep behaving as if you launched the command from the repository root. Dataset paths such as example_dataset/ therefore do not silently start resolving inside the run directory.

Inspecting configs¶

Use Hydra's built-in config printing when you want to check exactly what will be passed to training:

python -m opensr_srgan.train_hydra --cfg job experiment=10m
python -m opensr_srgan.train_hydra --cfg hydra experiment=10m

The first command prints the resolved SRGAN job config. The second prints Hydra's runtime config, including output-directory settings.

When to use legacy YAML¶

Use the legacy YAML entry point when you already have a complete hand-written config file or when a notebook/script expects a path. For the bundled example dataset, use:

python -m opensr_srgan.train --config opensr_srgan/configs/config_training_example.yaml

Use Hydra when you want quick overrides, reusable experiment pieces, or clean per-run config snapshots without creating many near-duplicate files. Use experiment=10m or experiment=20m after configuring the matching real datasets. Both paths call the same trainer once the config is loaded.

References¶

Hydra composes these configs through a Defaults List. See the Hydra Defaults List documentation for the underlying mechanism: https://hydra.cc/docs/1.1/advanced/defaults_list/. The SRGAN entry point also uses an explicit version_base and config_path, following Hydra's upgrade guidance: https://hydra.cc/docs/1.3/upgrades/version_base/.