experiments

Paper Experiments

Scripts and configs to reproduce every table in the Spatial Adapter paper.

1. Download data

See data/README.md for details. Short version:

bash data/download_data.sh
# KAUST: auto-download
# Weather2K: copy weather2k.npy to data/weather2k/
# GWHD: requires a Kaggle API key

2. Run

conda activate spatial-adapter

# Single experiment
python -m examples.experiments.runner \
    --config examples/experiments/configs/production/synthetic_1d.yaml

# All paper experiments
for cfg in examples/experiments/configs/production/*.yaml; do
    python -m examples.experiments.runner --config "$cfg"
done

3. Config → paper table

Config	Paper Table	Description
synthetic_1d.yaml	Table 3	Synthetic 1D reconstruction (10 seeds)
kaust_rmse.yaml	Table 4	KAUST ablation — RMSE (100 seeds)
kaust_covfrob.yaml	Table 4	KAUST ablation — CovFrob (100 seeds)
kaust_svscore.yaml	Table 4	KAUST ablation — SV_score (100 seeds)
weather2k_80_10_10.yaml	Table 5	Weather2K reconstruction — 80/10/10 (100 seeds)
weather2k_10_10_80.yaml	Table 5	Weather2K reconstruction — 10/10/80 (100 seeds)
weather2k_holdout.yaml	Table 6	Weather2K held-out stations (10 seeds)
wheat_head.yaml	Tables 7 & 8	GWHD classification + UQ (3 seeds × 4 backbones)

4. Output

Results land in results/production/<experiment>/:

• results.csv — long-format metrics
• wall_times.csv — per-stage wall times

Layout

examples/experiments/
├── runner.py                # Unified entry point
├── base.py                  # BaseExperiment ABC
├── configs/production/      # Paper experiment configs (YAML)
├── synthetic_1d/experiment.py
├── kaust_2b8/experiment.py
├── weather2k/experiment.py
└── wheat_head/experiment.py

examples/baselines/stdk/     # STDK baseline model