Complete experimental artifacts for "When does fusing hand-crafted spectral
knowledge with learned representations pay? A controlled, cost-normalized
benchmark and its organizing law" (Information Fusion, under review).
Every number in the paper traces to a file in this release.
| asset | size | contents |
|---|---|---|
result-tables.tar.gz |
0.9 MB | aggregated tables: all_results.csv (one row per cell), results_by_portion.csv, the combined workbook, law_audit.md, summary.md, summary.tex, and the segmentation and detection tables |
run-records.tar.gz |
1.1 MB | 10,018 final.json records (configuration as executed, accuracy, parameter and FLOP accounting, environment), 2,676 linear-probe records, 60 dense-probe and 30 detection-probe records, 35 robustness records |
training-curves.tar.gz |
35 MB | 9,663 per-epoch metrics.csv files |
logs.tar.gz |
1.8 MB | campaign and wave logs, including the cluster work-queue logs |
configs-and-subsets.tar.gz |
4.5 MB | every cell configuration and the committed subset indices |
SHA256SUMS is attached; verify with sha256sum -c SHA256SUMS.
Scope of the study. 3,054 experimental cells, 9,402 training runs,
3,854 GPU-hours. 13 datasets across six visual domains, 9 backbone families,
150 to 1,281,167 images, 10 to 1000 classes, 32 to 224 px, 5.7M to 86M
parameters. Interventions compared at declared training-cost multiples: a
fixed spectral auxiliary prior (1.02x), SimCLR / SimSiam / DINO
pre-training (2x), ImageNet transfer, DeiT-strength augmentation, and a
learned FitNets teacher (2x), plus their pairwise combinations.
Regenerating the paper's tables
tar -xzf run-records.tar.gz # restores runs/<cell>/seed<N>/final.json
python analysis/aggregate.py # regenerates every table
python analysis/audit_sign_law.py # re-runs the audit reported in the paperNot included: model checkpoints (223 GB) and the source images. All 14
datasets are public and cited in the paper; the committed subset indices
reproduce the exact image selection without redistributing images.
See docs/ARTIFACTS.md for the per-asset detail,
including the scope flags needed to reproduce the law audit and the one
recorded caveat about the ImageNet cells' training curves.