Configuration Layout#
Start here: experiment configs#
Each YAML file in ``conf/`` is a self-contained experiment config. Run training or inference by selecting one:
python train.py --config-name=bumper_geotransolver_oneshot
python train.py --config-name=crash_geotransolver_oneshot
python inference.py --config-name=crash_geotransolver_oneshot
To add a new experiment, copy an existing file in conf/ and edit data paths, model, and features.
Component configs (advanced)#
The subfolders (model/, datapipe/, reader/, training/, inference/) contain configs referenced by experiments. You rarely need to edit them unless customizing models, readers, or training defaults.
Path |
Purpose |
|---|---|
|
Model architectures (selected via experiment) |
|
Dataset and feature configs |
|
Data format readers (VTP, Zarr) |
|
Training hyperparameters |
|
Inference options |