README.md

metadata

license: mit

DRNO sanity outputs

This repository mirrors outputs/sanity/, the full sanity experiment output directory for Discrete-Residual Neural Operators (DRNO). The artifacts were generated by the code at https://github.com/soundai2016/DRNO with configs/sanity.yaml.

Download

pip install --upgrade huggingface_hub
hf download soundai2016/drno --local-dir outputs/sanity

Contents

outputs/sanity/
  data/              generated PDE datasets: .npz arrays + .json metadata
  checkpoints/       PyTorch checkpoints for each task/model/seed/lambda/selector
  histories/         per-epoch training logs
  results/           CSV metrics, Pareto grids, uncertainty, runtime and audit ledgers
  figures/           analysis figures from the plot stage
  tables/            JSON tables derived from results/*.csv
  pdebench_hdf5/     PDEBench-style HDF5 exports, manifest and audit summary
  publication/       manuscript-ready figures, tables and source data
  resolved_config.json
  run_environment.json
  publication_package.json
  publication_figures_run.json
  pdebench_manifest_summary.json

Scope

The sanity run covers four PDE tasks: darcy, burgers, schrodinger, and ns2d_kolmogorov; train/val/test/OOD/mesh-transfer splits; seeds 0..4; FNO, PINO-style residual, DRNO solver-consistent residual, MC-dropout FNO, DeepONet (Darcy/Burgers) and weak-form residual (Darcy/Burgers) baselines.

Naming convention

data/{task}_{split}.npz
histories/{task}_{model}_seed{seed}_lam{lambda}_{tag}.csv
checkpoints/{task}_{model}_seed{seed}_lam{lambda}_{tag}_{selector}.pt

Common tag values are main and sweep. Common checkpoint selectors are best_rel, best_res, best_qoi, best_pareto, and final. Decimal residual weights are filename-safe, e.g. 0.2 -> 0p2.

Reproduce or extend

git clone https://github.com/soundai2016/DRNO
cd DRNO
python -m pip install -e .

# Recreate the full suite from scratch.
bash scripts/run_sanity.sh

# Or reuse downloaded artifacts under outputs/sanity/ and rerun a stage.
python -m drno.run_all --config configs/sanity.yaml --stage audit
python -m drno.run_all --config configs/sanity.yaml --stage plot --force

The full pipeline is:

generate -> train -> sweep -> eval -> pareto -> uncertainty -> benchmark -> runtime -> pdebench -> plot -> audit

Notes

• resolved_config.json records the exact normalized configuration used for the run.
• run_environment.json records the software and hardware environment.
• .pt checkpoint files are PyTorch serialized artifacts; load them only in a trusted Python environment.
• CSV/JSON/HDF5 files are intended for analysis and reproduction; figures and publication/ are derived display artifacts.