Native COCO preparation for RF-DETR and RTMDet

prepare_coco.py

@@ -0,0 +1,48 @@
+"""Materialize the verified public V2.5 dataset as COCO without resizing images."""
+import argparse,io,json,os,hashlib,collections
+from pathlib import Path
+from concurrent.futures import ThreadPoolExecutor,as_completed
+import requests,pyarrow.parquet as pq
+REPO='fireviewer/v2.5-ready-to-train'
+def prepare(output,revision,token=None,progress=None):
+ root=Path(output);root.mkdir(parents=True,exist_ok=True)
+ headers={'Authorization':'Bearer '+token} if token else {}
+ def fetch(path):
+  r=requests.get('https://huggingface.co/datasets/'+REPO+'/resolve/'+revision+'/'+path,headers=headers,timeout=(20,180));r.raise_for_status();return r.content
+ def digest(b):return hashlib.sha256(b).hexdigest()
+ def write(path,data):
+  path.parent.mkdir(parents=True,exist_ok=True);tmp=Path(str(path)+'.pending');tmp.write_bytes(data);os.replace(tmp,path)
+ m=json.loads(fetch('manifest.json'));assert m['ready'] and m['all_native_image_hashes_verified']
+ casebytes=fetch('case-manifest.jsonl');assert digest(casebytes)==m['case_manifest_sha256']
+ rows=[json.loads(l) for l in casebytes.splitlines() if l];mapping={'train':'train','validation':'valid','test':'test'}
+ expected={r['sha256']:r for r in rows};seen=set();counts=collections.Counter()
+ def one(s):
+  raw=fetch(s['path']);assert digest(raw)==s['sha256'] and len(raw)==s['bytes']
+  table=pq.read_table(io.BytesIO(raw));assert table.num_rows==s['rows'];ids=[]
+  for batch in table.to_batches(max_chunksize=32):
+   for d in batch.to_pylist():
+    r=expected[d['sha256']];b=d['image']['bytes'];assert digest(b)==r['sha256'] and d['objects']==r['objects']
+    assert d['split']==r['split'];filename=Path(r['image_file']).name;assert filename==r['image_file']
+    p=root/mapping[r['split']]/filename
+    if not p.exists() or digest(p.read_bytes())!=r['sha256']:write(p,b)
+    ids.append(r['sha256'])
+  return s['split'],ids
+ with ThreadPoolExecutor(max_workers=3) as pool:
+  for i,f in enumerate(as_completed([pool.submit(one,s) for s in m['shards']]),1):
+   split,ids=f.result();assert not seen.intersection(ids);seen.update(ids);counts[split]+=len(ids)
+   if progress:progress(i,len(m['shards']),len(seen))
+ assert len(seen)==m['images'] and dict(counts)==m['splits']
+ for split,folder in mapping.items():
+  raw=fetch('annotations/'+split+'.json');c=json.loads(raw);assert len(c['images'])==m['splits'][split]
+  case_by_id={r['coco_id']:r for r in rows if r['split']==split}
+  for im in c['images']:
+   r=case_by_id[im['id']];assert im['sha256']==r['sha256'] and im['file_name']==r['image_file']
+  write(root/folder/'_annotations.coco.json',raw)
+ for r in rows:r['image_path']=str(root/mapping[r['split']]/r['image_file'])
+ write(root/'rows.jsonl',b''.join(json.dumps(r,ensure_ascii=False).encode()+b'\n' for r in rows))
+ write(root/'ready.json',json.dumps({'repo':REPO,'revision':revision,'images':m['images'],'splits':m['splits'],'native_hashes_verified':True,'metadata_sha256':m['case_manifest_sha256']}).encode())
+ return rows,m
+if __name__=='__main__':
+ p=argparse.ArgumentParser();p.add_argument('--revision',required=True);p.add_argument('--output',required=True);a=p.parse_args()
+ rows,m=prepare(a.output,a.revision,os.environ.get('HF_TOKEN'),lambda i,n,c:print(f'Shards {i}/{n}; native images {c}',flush=True))
+ print(json.dumps({'ready':True,'images':m['images'],'splits':m['splits']}))