Configs¶

Hydra configuration files defining models, data modules, trainers, logging, and experiment settings.

All training and evaluation behaviour is controlled via these YAML configs.

How it works¶

The top-level config.yml defines the default configuration by including entries from the subfolders:

defaults:
  - model: simple
  - data: dummy
  - trainer: default
  - logging: mlflow
  - experiment: base

Hydra then loads the corresponding files:

These files are merged together into one final config object, which is passed into the training script.

model/

Defines model architectures and their hyperparameters. Each file here must contain a _target_ field that points to a Python class implementing the model. Example: selecting different neural network variants or LightningModules.
data/

Configures DataModules, dataloaders, dataset paths, and batch sizes. Each file should describe how training/validation/test data is created or loaded.
trainer/

Contains configurations for PyTorch Lightning’s Trainer: number of epochs, devices, precision settings, logging frequency, etc.
logging/

Parameters for MLflow or other logging backends. This keeps experiment tracking separate from model or data settings.
experiment/

Optional presets for organising runs (name, notes, parameter overrides). Useful for keeping track of different experiment configurations.

Hydra is invoked in train.py via:

@hydra.main(config_path="../../configs", config_name="config")
def main(cfg):
    ...

You can start training with the default config:

python -m FRAME_FM.training.train.py

Or override specific config options from the command line:

python -m FRAME_FM.training.train.py model=simple data=cifar10 trainer.max_epochs=50