| |
| """Export authentic cached evidence for illustrative cases, with hash checks. |
| |
| The input is a small JSON case plan. Extraction runs in the official decoding |
| environment on the data host. Images are never regenerated by an image model. |
| """ |
| import argparse |
| import hashlib |
| import html |
| import json |
| from pathlib import Path |
| import shutil |
| import sys |
| ROOT=Path(__file__).resolve().parents[1] |
| sys.path.insert(0,str(ROOT/'src')) |
| from vbench_prompts_compile import annotation_jobs as J |
| from cache_backend_outputs import sha |
| from cache_matrix_transforms import frames_for_backend,frame_hash |
|
|
|
|
| def main(argv=None): |
| p=argparse.ArgumentParser(description=__doc__) |
| p.add_argument('--plan',required=True);p.add_argument('--out',required=True) |
| p.add_argument('--vbench-root',default='/root/wenbiao_zhao/VBench') |
| args=p.parse_args(argv);sys.path.insert(0,args.vbench_root) |
| import numpy as np |
| from PIL import Image |
| from vbench.utils import load_video |
| cases=json.loads(Path(args.plan).read_text())['cases'] |
| out=Path(args.out);out.mkdir(parents=True,exist_ok=True);result=[] |
| for case in cases: |
| images=[];index=case.get('frame_index',8) |
| if case.get('annotation_image'): |
| source=Path(case['annotation_image']);path=out/(case['id']+'-annotation.png') |
| shutil.copyfile(source,path);images.append(path.name) |
| else: |
| base=case['cache'];source=Path(base['video_path']) |
| if sha(source)!=base['video_sha256']:raise ValueError('video changed: '+str(source)) |
| if case['task']=='scene': |
| frames=load_video(str(source),num_frames=16,return_tensor=False,width=384,height=384) |
| elif case['task']=='action': |
| from vbench.human_action import Compose,Resize,CenterCrop,ClipToTensor,Normalize |
| transform=Compose([Resize(256,interpolation='bilinear'),CenterCrop(size=(224,224)),ClipToTensor(), |
| Normalize(mean=[.485,.456,.406],std=[.229,.224,.225])]) |
| tensor=load_video(str(source),transform,num_frames=16).unsqueeze(0) |
| if frame_hash(tensor.numpy())!=base['model_input_sha256']:raise ValueError('UMT input differs') |
| values=tensor[0].permute(1,2,3,0).numpy() |
| frames=np.rint(np.clip(values*np.asarray([.229,.224,.225])+np.asarray([.485,.456,.406]),0,1)*255).astype('uint8') |
| else:frames=frames_for_backend(source,load_video) |
| if case['task']!='action' and [frame_hash(f) for f in frames]!=base['frame_sha256']: |
| raise ValueError('sampled frames changed: '+str(source)) |
| frame=frames[index];path=out/(case['id']+'-before.png') |
| Image.fromarray(frame.astype('uint8')).save(path);images.append(path.name) |
| changed=case.get('transformed') |
| if changed: |
| kind=changed['evidence_kind'] |
| if kind in {'evidence_mirror','video_mirror'}: |
| after=np.flip(frame,axis=1 if changed['axis']=='hflip' else 0).copy() |
| if kind=='video_mirror': |
| source=Path(changed['transformed_video']) |
| if sha(source)!=changed['transformed_video_sha256']:raise ValueError('mirror file changed') |
| actual=frames_for_backend(source,load_video)[index] |
| if frame_hash(actual)!=changed['frame_sha256'][index] or not np.array_equal(actual,after): |
| raise ValueError('physical mirror pixels differ') |
| elif kind=='occlusion': |
| after=frame.copy();detail=changed['frame_transform'][index] |
| if detail['status']!='ok':raise ValueError('selected occlusion frame incomplete') |
| for x0,y0,x1,y1 in detail['rectangles']:after[y0:y1,x0:x1]=128 |
| if frame_hash(after)!=changed['frame_sha256'][index]:raise ValueError('occlusion pixels differ') |
| else:raise ValueError(kind) |
| path=out/(case['id']+'-after.png');Image.fromarray(after.astype('uint8')).save(path);images.append(path.name) |
| small={k:v for k,v in case.items() if k not in {'cache','transformed'}} |
| small.update(images=images,image_sha256={name:sha(out/name) for name in images}, |
| cache_record_sha256=hashlib.sha256(json.dumps(case.get('cache'),sort_keys=True).encode()).hexdigest()) |
| result.append(small) |
| J.atomic_json(out/'cases.json',{'plan_sha256':sha(args.plan),'exporter_sha256':sha(__file__),'cases':result, |
| 'note':'Illustrative selections, not a frequency estimate. Vision labels are model-derived. Evidence-only mirror pictures illustrate the exact coordinate reflection.'}) |
| document=['<!doctype html><html lang="en"><meta charset="utf-8"><title>VBench evidence cases</title>', |
| '<style>body{font:16px system-ui;margin:2rem;max-width:1200px}figure{display:flex;gap:1rem;margin:1rem 0}img{max-width:47%;object-fit:contain}pre{white-space:pre-wrap;background:#f4f4f4;padding:1rem}section{border-top:1px solid #bbb;margin-top:2rem}</style>', |
| '<h1>Illustrative evidence and interface failures</h1><p>These examples explain failure mechanisms; aggregate frequencies belong to the full matrix tables. Model annotations are not human gold.</p>'] |
| for row in result: |
| document.append('<section><h2>'+html.escape(row['title'])+'</h2><p>'+html.escape(row['note'])+'</p><figure>') |
| document.extend('<img src="'+html.escape(name)+'" alt="'+html.escape(row['title'])+'">' for name in row['images']) |
| document.append('</figure><pre>'+html.escape(json.dumps({k:v for k,v in row.items() if k not in {'title','note','images','image_sha256','cache_record_sha256'}},ensure_ascii=False,indent=2))+'</pre></section>') |
| document.append('</html>');(out/'index.html').write_text('\n'.join(document)+'\n') |
| print(json.dumps({'cases':len(result),'out':str(out)}));return 0 |
|
|
|
|
| if __name__=='__main__':raise SystemExit(main()) |
|
|
