Datasets:
Formats:
csv
Languages:
English
Size:
< 1K
Tags:
continual-learning
catastrophic-forgetting
functional-memory
orthogonal-projection
statistical-certification
pytorch
| #!/usr/bin/env python3 | |
| """Rebuild the two CSV tables from JSON inside the unchanged scientific ZIP.""" | |
| from __future__ import annotations | |
| import argparse | |
| import csv | |
| import hashlib | |
| import json | |
| from pathlib import Path | |
| import statistics | |
| import zipfile | |
| def export(archive, output): | |
| output.mkdir(parents=True, exist_ok=False) | |
| prefix = 'Eidrion_v1.0.0/' | |
| with zipfile.ZipFile(archive) as z: | |
| aggregate = json.loads(z.read(prefix + 'experiments/final_standard/aggregate_results.json')) | |
| summary = [] | |
| for dataset, methods in aggregate['groups'].items(): | |
| for method, g in methods.items(): | |
| m = g['metrics']; a = m['final_micro_accuracy'] | |
| summary.append({'dataset': dataset, 'method': method, 'n_seeds': a['n'], | |
| 'accuracy_mean': a['mean'], 'accuracy_std': a['std'], | |
| 'accuracy_t95_low': a['t95_low'], 'accuracy_t95_high': a['t95_high'], | |
| 'peak_persistent_payload_mean_bytes': m['peak_persistent_payload_bytes']['mean'], | |
| 'protected_class_count_mean': m['protected_class_count']['mean'], | |
| 'fit_mean_seconds': g['timing_seconds']['fit']['mean'], | |
| 'calibration_mean_seconds': g['timing_seconds']['calibration']['mean'], | |
| 'inference_batch1_mean_seconds': g['serial_inference_batch1_seconds']['mean']}) | |
| per_run = [] | |
| for name in sorted(n for n in z.namelist() if n.startswith(prefix + 'experiments/final_standard/runs/') and n.endswith('/result.json')): | |
| raw = z.read(name); d = json.loads(raw) | |
| keys = ['dataset', 'method', 'seed', 'status', 'final_micro_accuracy', 'final_macro_accuracy', | |
| 'final_ece15_full10way_softmax', 'average_acquisition_task_accuracy', | |
| 'average_incremental_seen_micro_accuracy', 'backward_transfer', 'forgetting', | |
| 'forward_transfer_relative_to_initialization', 'protected_class_count', | |
| 'peak_persistent_payload_bytes', 'process_peak_rss_kib', 'source_lock_sha256'] | |
| row = {k: d[k] for k in keys} | |
| row.update(fit_seconds=d['timing_seconds']['fit'], calibration_seconds=d['timing_seconds']['calibration'], | |
| source_result_path=name, source_result_sha256=hashlib.sha256(raw).hexdigest()) | |
| per_run.append(row) | |
| if len(summary) != 26 or len(per_run) != 260: | |
| raise ValueError('Unexpected captured result coverage') | |
| for row in summary: | |
| scores = [r['final_micro_accuracy'] for r in per_run if r['dataset'] == row['dataset'] and r['method'] == row['method']] | |
| if len(scores) != row['n_seeds'] or abs(statistics.mean(scores) - row['accuracy_mean']) > 1e-12: | |
| raise ValueError('Per-run accuracy does not agree with captured aggregate') | |
| for name, rows in [('benchmark_summary.csv', summary), ('per_run_metrics.csv', per_run)]: | |
| with (output / name).open('w', newline='', encoding='utf-8') as f: | |
| writer = csv.DictWriter(f, fieldnames=list(rows[0])) | |
| writer.writeheader(); writer.writerows(rows) | |
| print('Exported 26 summary rows and 260 per-run rows; group means agree with captured JSON.') | |
| def main(): | |
| root = Path(__file__).resolve().parents[1] | |
| parser = argparse.ArgumentParser(description=__doc__) | |
| parser.add_argument('--archive', type=Path, default=root / 'archives/Eidrion_v1.0.0_Research_Package.zip') | |
| parser.add_argument('--output', type=Path, required=True, help='Fresh output directory') | |
| args = parser.parse_args() | |
| export(args.archive, args.output) | |
| if __name__ == '__main__': | |
| main() | |