smellingsalt/nb360

NAS-Bench-360 — DASH-ready bundle

A single, verified archive of the ten NAS-Bench-360 datasets used by DASH (NeurIPS 2022), packaged for one-command setup.

Why this exists: the original distribution is slow to download and scattered — per-dataset tarballs/zips across an S3 bucket, a Google Drive pickle, a manually-run preprocessing step — with no checksums. This is that pain, removed: one tar, one sha256, one command.

Contents

nb360.tar extracts to nb360/ — one folder per dataset. The ten tasks, as defined by NAS-Bench-360 (task descriptions and metrics from the NB360 paper, Table 1; sample counts and splits from the NB360 website):

Task	Folder	Inside	Samples (train/val/test)	Objective	Metric
CIFAR-100	cifar100/	cifar-100-python/	60K (40K/10K/10K)	Classify natural images into 100 classes	0–1 error
Spherical	spherical/	s2_cifar100.gz	60K (40K/10K/10K)	Classify spherically projected images into 100 classes	0–1 error
NinaPro	ninapro/	ninaPro/ (6 .npy)	3,916 (2,638/659/659)	Classify sEMG signals into 18 hand-gesture classes	0–1 error
FSD50K	fsd/	chunks/ manifests + 597k/205k/930 WAVs	51K (37K/4K/10K)	Classify sound events in log-mel spectrograms, 200 labels (multi-label)	1–mAP
Darcy Flow	darcy-flow-5/	piececonst_r421_N1024_smooth{1,2}.mat	1,100 (900/100/100)	Predict the final fluid state from initial conditions (dense)	relative ℓ²
PSICOV	psicov/	protein/ (deepcov/psicov/cameo + 4 .lst splits)	3,606 (3,356/100/150)	Predict pairwise distances between residues from sequence features (dense)	MAE₈
Cosmic	cosmic/	npy_train/ 25,239 · npy_test/ 7,573 imgs + *_dirs.npy	5,250 (4,347/483/420)	Predict probabilistic maps identifying cosmic rays in telescope images (dense)	1 – AUROC
ECG	ecg/	challenge2017.pkl	330K (260K/33K/33K)	Detect atrial cardiac disease from ECG recordings (1D)	1 – F1
Satellite	satellite/	satellite_{train,test}.npy	1M (800K/100K/100K)	Classify satellite image pixel time series into 24 land-cover types (1D)	0–1 error
DeepSEA	deepsea/	deepsea_filtered.npz	250K (71K/2.5K/150K)	Predict chromatin and binding states of DNA sequences (1D, multi-label)	1 – AUROC
(unused by the 10 tasks)	extras/	cifar-10-batches-py/	—	shipped for completeness	—

Point = fixed-size input, one prediction per sample; Dense = per-pixel/residue prediction; 1D = sequence input. ℓ² and MAE₈ are as defined in section 2 of the NAS-Bench-360 paper.

Usage

# anonymous, no token needed
curl -L -O https://huggingface.co/datasets/smellingsalt/nb360/resolve/main/nb360.tar
curl -L -O https://huggingface.co/datasets/smellingsalt/nb360/resolve/main/nb360.tar.sha256

sha256sum -c nb360.tar.sha256
tar -xf nb360.tar

Running DASH? Its loaders expect flat paths under one data root (chunks/tr.csv, cosmic/train_dirs.npy, ...), not this per-dataset layout. Use the included script:

curl -L -O https://huggingface.co/datasets/smellingsalt/nb360/resolve/main/flatten_for_dash.py
python3 flatten_for_dash.py nb360 [target_dir]   # flattens into DASH's layout (default nb360/dash_data)
python3 flatten_for_dash.py nb360 --revert       # moves everything back into the per-dataset folders

Then point DASH at the target dir (experiment.data_dir in src/simplification/configs/dash.yaml, or --data_dir).

License & citation

The datasets keep their original licenses (CIFAR-100 CC BY 4.0, FSD50K CC BY 4.0 for annotations, DeepSEA CC BY 4.0, Cosmic CC BY 4.0, ECG ODC-BY 1.0, Darcy Flow MIT, Satellite GPL 3.0, NinaPro CC BY-ND, Spherical CIFAR-100 CC BY-SA, PSICOV GPL). Full license table, the GPL-3.0 text (required to travel with the data), per-dataset papers, BibTeX, and the attribution of the two scripts hosted here (flatten_for_dash.py, preprocess_cosmic.py from DASH): see LICENSES.md.

If you are a rights holder and object to this redistribution, contact Sawansingh Mahara sawansinghmahara@gmail.com and the dataset will be removed.

If you use this bundle, please also cite NAS-Bench-360:

@inproceedings{tu2022nasbench,
  title={{NAS}-Bench-360: Benchmarking Neural Architecture Search on Diverse Tasks},
  author={Tu, Renbo and Roberts, Nicholas and Khodak, Mikhail and Shen, Junhong and Sala, Frederic and Talwalkar, Ameet},
  booktitle={Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year={2022},
  url={https://openreview.net/forum?id=xUXTbq6gWsB}
}