Skip to content

Swiss River Network Benchmark

ICPR 2026 submission

Open-source reference code, datasets, and figures for Benchmarking Transformers on Spatio-Temporal River Water Temperature Modeling.

The Swiss River Network Benchmark is a reproducible benchmark for river water-temperature forecasting. It ships three real-world graph datasets, eight reference methods, and the exact training / evaluation / sweep pipeline used in the paper.

  • Install in 30 s

    git clone https://github.com/jajupmochi/swiss-river-network-benchmark.git
cd swiss-river-network-benchmark
uv sync --no-cache
uv run srn --help

    Getting started

  • Reproduce the paper

    Run the full training / evaluation / sweep pipeline with one CLI. Every figure in the paper has a matching notebook.

    Paper reproduction

  • Live demo

    Interactive UIs on Hugging Face Space, local Streamlit, and a double-click desktop installer — all sharing one visualisation layer.

    Desktop app

  • API reference

    Generated from the source code via mkdocstrings. Useful when you want to plug a new method into the benchmark harness.

    API reference

What's in the box

  • 3 datasets — swiss-1990, swiss-2010, zurich.
  • 8 methods — LSTM, Graphlet, LSTM + station embedding, ST-GNN, Transformer, Transformer + Graphlet, Transformer + Embedding, Transformer + ST-GNN.
  • 5 installation paths — uv, pip, Docker, desktop installer, LLM-agent paste-and-run.
  • Window-length sweep — the exact configuration behind paper Fig. 4 and the HLE dimension of Fig. 2.

Why open this benchmark

  • Reproducibility first. Bugs that produced the original paper numbers are fixed on main since 4daeff3 — see explainers for the details.
  • Model-agnostic. New methods plug in through one config dict and one dataset adapter; there is no framework lock-in.
  • Operator-friendly. A hydrologist who never touches Python can still run the Streamlit demo from a release installer.

View on GitHub Hugging Face Space