Quickstart

Run each step from the repo root after uv sync --no-cache.

1. Prepare the three datasets

uv run srn prepare-data

Produces swissrivernetwork/benchmark/dump/<graph>_{train,test}.csv.

2. Tune one method on one graph

uv run srn tune -m lstm -g swiss-2010 -n 50 -wl 90

• -m — method name (see methods).
• -g — graph name (one of swiss-1990, swiss-2010, zurich).
• -n — Ray Tune trial count.
• -wl — training window length.

Checkpoints land under swissrivernetwork/benchmark/outputs/ray_results/.

3. Evaluate

uv run srn evaluate

Writes per-station and per-method tables under swissrivernetwork/benchmark/visualize_results/outputs/stations/.

4. Sweep window lengths

uv run srn sweep

Two phases (ISOLATED → GRAPHLET) write dump/predictions/<path_extra_keys>-evalwl{W}/ files and the CSVs under visualize_results/outputs/win_lens/.

Eval-window leakage

Earlier sweep runs suffered from two bugs that are fixed on main since commit 4daeff3. See Explainers before trusting numbers from pre-fix runs.

5. Visualise

uv run jupyter lab swissrivernetwork/benchmark/visualize_results/

Every notebook under that directory reproduces one paper figure from the CSVs you just wrote.