configs

name	configs
description	Hydra configuration file structure and conventions for models, datasets, callbacks, and training

Configuration Files

All configuration files live in configs/, organized using the Hydra framework.

Directory Structure

• accelerate/ — Distributed training acceleration
• callbacks/ — Lightning callbacks (e.g., model summary, checkpointing)
• data/ — Dataset and dataloader configurations
• datamodule/ — PyTorch Lightning datamodule configurations
• experiment/ — Experiment-specific overrides
• extras/ — Additional utility configurations
• hydra/ — Hydra framework configurations
• logger/ — Logging configurations (e.g., TensorBoard, WandB)
• loss/ — Loss function configurations
• model/ — Model configurations
• opt/ — Optimizer configurations
• paths/ — Path configurations
• plmodule/ — PyTorch Lightning module configurations
• trainer/ — Trainer configurations

Top-level training config files (e.g., train_nop.yaml, train_nop_rollout.yaml) specify which configs to use for a given project.

Callback Configuration

There are two types of callback config files in configs/callbacks/:

• single_callback_name.yaml — Configuration for a single callback.
• many_callbacks_project_name.yaml — Lists all callback configs to be used in a project.

For a new project, create a many_callbacks_project_name.yaml listing the individual callback configs. If you add a new callback, also create a corresponding single_callback_name.yaml.