Development

Everything needed to work in this repository: environment setup, the data layout, the commands that run training and inference, and the conventions the codebase follows.

Getting Started Install the environment, satisfy the one external prerequisite, lay out the data, and run training, inference and scoring. Project Structure An annotated tour of the tree, how configuration layers resolve, and the conventions to follow when adding an approach or an experiment.

1. Who this section is for

This section assumes you are comfortable with Python, PyTorch and the command line, and that you have a CUDA-capable GPU available. It does not assume any prior familiarity with the Vesuvius Challenge data, with MONAI, or with nnU-Net.

2. The fastest path to a result

If you want to see the pipeline work before reading anything else, this is the shortest route:

Synchronise the locked environment with uv.
Install the external metrics package (see the prerequisite callout below).
Point the pipeline at converted data and run a short training job.
Run inference from the resulting checkpoint.
Score the predictions and read the component metrics.

All five steps are detailed in Getting Started.

One prerequisite is not in the lockfile The competition metric implementation arrives as an external package rather than as a declared dependency, so a fresh uv sync is not sufficient to run a scoring pass. Plan for that step up front rather than discovering it at the end of a training run.