| |
| |
| import json, os |
| from pathlib import Path |
| from huggingface_hub import HfApi, CommitOperationAdd |
| ROOT=Path('/marimo/disef_cars') |
| def read(path):return json.loads(Path(path).read_text()) |
| def table(rows,heads): |
| return '| '+' | '.join(heads)+' |\n| '+' | '.join(['---']*len(heads))+' |\n'+''.join('| '+' | '.join(map(str,row))+' |\n' for row in rows) |
| def save(name,body): |
| out=ROOT/'reports'/name;out.write_text(body.rstrip()+'\n');return out |
|
|
| def main(): |
| run=ROOT/'runs/cars_disef_spc_seed1_20261025' |
| cfg=read(run/'config/resolved_config.json');gen=read(ROOT/'reports/generation_report.json') |
| test=read(run/'results/final_test.json');remote_base=f"{cfg['hf_run_prefix']}/{cfg['run_id']}" |
| rows=[json.loads(x) for x in (run/'logs/metrics.jsonl').read_text().splitlines() if x.strip()] |
| if len(rows)!=cfg['epochs'] or test['test_total']!=8041 or gen['valid_images']!=12544:raise RuntimeError('final report inputs incomplete') |
| epoch_seconds=rows[-1]['elapsed_seconds'];throughput=cfg['shot']*cfg['class_count']*len(rows)/epoch_seconds if epoch_seconds else None |
| worst=sorted(test['per_class_accuracy'].items(),key=lambda x:x[1])[:10] |
| confusion=test['most_confused_classes'][:10] |
| generation=f'''# Stanford Cars synthetic generation report |
| |
| Variant: **{gen['variant']}**, offline generation completed before training. This is separate from original DisEF-SAP, DisEF-DataDream, and Qwen image generation. |
| |
| - Accepted: {gen['valid_images']} / 12,544, exactly {gen['per_class_min']} to {gen['per_class_max']} per class across 196 classes. |
| - Attempts: {gen['generated_attempts']}; rejected: {gen['rejected']}; retried: {gen['retried']}; historical failed tasks: {gen['failed']}; reference fallback accepted: {gen.get('reference_fallback_accepted',0)}. |
| - Generator: Stable Diffusion 2.1 image-to-image, BF16, DPM-Solver++, 20 nominal steps, starting step 15, CFG 8. LLaVA 1.5 captions use seed-1 permitted support images; Qwen3-VL-Embedding-2B filters by target class. See `configs/generation_config.json` for exact revisions, prompts, seeds, and settings. |
| - Archive: `DanielNguyen121/DisEF-datasets/{gen['archive_filename']}` at revision `{gen['remote_revision']}`; SHA256 `{gen['archive_sha256']}`; bytes {gen['archive_bytes']}. Remote SHA and size verified. |
| - Provenance and per-image acceptance: `cars/_progress/{gen['generation_config']['run_id']}` in the synthetic repository; each source image and caption source is from the leak-free seed-1 support manifest. |
| ''' |
| train=f'''# Stanford Cars fine-tuning report |
| |
| - Run: `{cfg['run_id']}`; variant **{gen['variant']}**. |
| - Backbone: Qwen3-VL-Embedding-2B. LoRA r16, alpha32, dropout0.1 on `qkv`, `proj`, `q_proj`, `v_proj`; both vision and text adapters are trainable. |
| - Optimizer: AdamW, lr 1e-4, weight decay 0.01; no scheduler. Batch size 64; 40 epochs; real/synthetic CE weights 0.8/0.2; RGB conversion only; FP16 autocast with GradScaler. |
| - Support seed: {cfg['split_seed']}; training seed: {cfg['seed']}; {cfg['shot']} real and {cfg['synthetic_per_class']} synthetic images per class. |
| - Model selection: published 1,635-image validation partition. Best epoch {test['best_epoch']}, best validation accuracy {test['best_val_accuracy']:.4%}. |
| - Training elapsed {epoch_seconds:.1f} seconds across {len(rows)} measured epochs. Approximate real-support samples per second {throughput:.2f}; this excludes synthetic samples and validation passes as a separate denominator. |
| - Latest resumable and best adapter checkpoints, epoch logs, and figures are under `{remote_base}` in the public experiment repository. |
| ''' |
| final=f'''# Stanford Cars final test report |
| |
| - Selected checkpoint: epoch {test['best_epoch']} by validation accuracy {test['best_val_accuracy']:.4%}. |
| - Held-out test: {test['test_correct']}/{test['test_total']} correct; Top-1 accuracy **{test['test_accuracy']:.4%}**; Top-5 accuracy **{test.get('top5_accuracy', float('nan')):.4%}**; macro-F1 **{test['macro_f1']:.4f}**. Evaluation runtime: {test.get('evaluation_seconds', float('nan')):.1f} seconds. |
| - One genuine support/training seed pair was run; no mean or standard deviation is reported. |
| - Full per-class accuracy and 196 x 196 confusion matrix: `{remote_base}/results/final_test.json`. Per-image IDs, true labels, Top-1 and Top-5 predictions: `{remote_base}/results/final_test_predictions.json`. |
| |
| ## Ten lowest-accuracy classes |
| |
| '''+table([(a,f'{v:.2%}') for a,v in worst],['Class','Accuracy'])+'\n## Most frequent confusions\n\n'+table([(x['true_class'],x['predicted_class'],x['count']) for x in confusion],['True class','Predicted class','Count'])+''' |
| SPC has no checked-in Cars-specific DisEF result under the audited revision, so no directly matched numeric SPC reference score is available. The original DisEF CLIP+SAP protocol is scientifically different and is not pooled with this run. |
| ''' |
| summary=f'''# DisEF-Cars scientific report |
| |
| **Variant:** {gen['variant']}. This is an SPC Qwen embedding recipe, not the original CLIP DisEF-SAP. |
| |
| **Data:** Stanford Cars, 196 classes and 16,185 verified original images (8,144 official train; 8,041 official test). The DisEF/CoOp split is 6,509 train, 1,635 validation, 8,041 test. Seed-1 16-shot support has 3,136 real images. Fourteen official support selections were replaced because their exact bytes cross split or conflicting-label boundaries; the full audit and both support manifests are retained. Original images are withheld from public redistribution pending permission. |
| |
| **Generation:** {gen['valid_images']} verified synthetic images, 64/class; {gen['generated_attempts']} attempts, {gen['rejected']} rejections, {gen.get('reference_fallback_accepted',0)} documented fallback acceptances. Complete archive `{gen['archive_filename']}` SHA256 `{gen['archive_sha256']}` at `DanielNguyen121/DisEF-datasets@{gen['remote_revision']}`. Generation finished before fine-tuning. |
| |
| **Training:** Qwen3-VL-Embedding-2B with SPC LoRA and 0.8/0.2 real/synthetic CE, AdamW 1e-4, batch64, 40 epochs. Training seed {cfg['seed']}, support seed {cfg['split_seed']}. Best validation epoch {test['best_epoch']}, accuracy {test['best_val_accuracy']:.4%}. Checkpoint selection uses validation only. |
| |
| **Test:** {test['test_correct']}/{test['test_total']} correct; Top-1 accuracy **{test['test_accuracy']:.4%}**, Top-5 accuracy **{test.get('top5_accuracy', float('nan')):.4%}**, macro-F1 **{test['macro_f1']:.4f}**. Full per-class results, confusion matrix, curves, logs, best checkpoint and latest resumable checkpoint are in `{remote_base}`. One genuine seed pair was executed; no seed aggregate is claimed. |
| |
| **Fidelity:** SPC's generic DisEF entrypoint is Qwen-based, while official DisEF is CLIP ViT-B/16 with an original SAP generator. The checked-in SPC entrypoint is aircraft-specific and lacks a Cars-specific result. This run is clearly labeled as the adapted SPC variant. Split duplicate repair and use of the published validation set are declared deviations; test images were not used for generation or checkpoint selection. See `cars_reproducibility_audit.md` for line-level settings and source revisions. |
| ''' |
| outs=[save('generation_report.md',generation),save('training_report.md',train),save('final_test_report.md',final),save('final_scientific_report.md',summary)] |
| for extra in ['cars_reproducibility_audit.md','dataset_audit.md','runtime_benchmark.md','generation_report.json']: |
| outs.append(ROOT/'reports'/extra) |
| outs += [ROOT/'README.md',ROOT/'manifests/dataset_manifest.json',ROOT/'manifests/support_leakage_report.json'] |
| outs += list((ROOT/'source').glob('*.py'))+[ROOT/'source/download_cars.sh',ROOT/'source/requirements-molab-lock.txt'] |
| ops=[CommitOperationAdd(path_in_repo=str(x.relative_to(ROOT)),path_or_fileobj=str(x)) for x in outs] |
| c=HfApi(token=os.environ.get('HF_TOKEN')).create_commit(repo_id='DanielNguyen121/DisEF-Cars',repo_type='dataset',operations=ops, |
| commit_message='Publish measured Cars scientific, generation, training, and test reports') |
| print('FINAL_REPORT_COMMIT',c.oid,flush=True) |
| if __name__=='__main__':main() |
|
|
