File size: 8,393 Bytes
7e7a152 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 | """Four-task data. Test identity metadata is exclusion-only; no test targets enter training."""
import json,hashlib,collections,random,re
from pathlib import Path
ROOT=Path(__file__).resolve().parents[2]
TASKS=['HOS','sound_class','duration','order']
def read(p):return [json.loads(l) for l in Path(p).open() if l.strip()]
def sha(p):return hashlib.sha256(Path(p).read_bytes()).hexdigest()
def build(cfg):
out=Path(cfg['data_output']);out.mkdir(parents=True,exist_ok=True)
held={s:{k:set() for k in ['scene','clip_id']} for s in ['validation','test']}
for name,meta in [('omini_360_benchmark_v0_1','private/episode_map.jsonl'),('omini_sound_search_v1','private/labels.jsonl')]:
ds=ROOT/'datasets'/name;mp={x['id']:x for x in read(ds/meta)}
for split in held:
for e in read(ds/'public'/f'{split}.jsonl'):
if name.startswith('omini_360') and e['task_type']!='HOS':continue
for k in held[split]:held[split][k].add(mp[e['id']][k])
audit=dict(tasks=TASKS,test_targets_used=False,source_hashes={},splits={},limitations=['Pretrained HVS and audio encoder exposure is unknown.','Object and audio asset disjointness not established.','Existing trajectories use geometric expert supervision; partially observed oracle ambiguity remains.'])
identities={}
for split in ['train','validation']:
rows=[];reject=collections.Counter()
for category in ['static','sound']:
p=Path(cfg['sources'][category])/f'{split}.jsonl';audit['source_hashes'][str(p)]=sha(p)
rs=read(p)
if category=='static':rs=[r for r in rs if r['task']=='HOS' and r.get('source')=='custom']
groups=collections.defaultdict(list)
for r in rs:
assert r['task'] in TASKS and r['split']==split
groups[r['episode_id']].append(r)
for eid,seq in groups.items():
seq.sort(key=lambda r:r['step'])
if any(r[k] in held[h][k] for r in seq for h in (['validation','test'] if split=='train' else ['test']) for k in ['scene','clip_id']):
reject[category]+=len(seq);continue
for i,r in enumerate(seq):
hist=seq[max(0,i-2):i+1]
assert all(hist[j+1]['step']==hist[j]['step']+1 for j in range(len(hist)-1))
# HOS is explicitly visual only. Sound tasks use the full recording before frozen-frame search.
rows.append(dict(id=category+':'+r['id'],task=r['task'],split=split,episode_id=eid,scene=r['scene'],clip_id=r['clip_id'],instruction=r['instruction'],pose=r['pose'],observations=[dict(panorama=x['panorama'],pose=x['pose']) for x in hist],history_actions=[x['answer'] for x in hist[:-1]],audio_path=r.get('audio_path') if category=='sound' else None,audio_end_s=10. if category=='sound' else None,reference=None,dynamic=False,answers=[r['answer']]))
assert set(r['task'] for r in rows)==set(TASKS)
assert len(rows)==len({r['id'] for r in rows})
identities[split]={k:{r[k] for r in rows} for k in ['scene','clip_id']}
p=out/f'{split}.jsonl';content=''.join(json.dumps(r)+'\n' for r in rows)
if p.exists():assert p.read_text()==content,'Frozen dataset drift'
else:p.write_text(content)
audit['splits'][split]=dict(rows=len(rows),tasks=dict(collections.Counter(r['task'] for r in rows)),scenes=len(identities[split]['scene']),clips=len(identities[split]['clip_id']),excluded=dict(reject),sha256=sha(p),first_submit=dict(collections.Counter(r['task'] for r in rows if 'submit(' in r['answers'][0])))
for k in identities['train']:assert not identities['train'][k]&identities['validation'][k]
audit['train_validation_test_scene_clip_disjoint']=True
(out/'audit.json').write_text(json.dumps(audit,indent=2));return audit
def load_image(desc):
import cv2
from PIL import Image
from hvs_core import perspective
if 'image' in desc:
with Image.open(desc['image']) as im:return im.convert('RGB')
if 'video' in desc:
cap=cv2.VideoCapture(desc['video']);assert cap.isOpened();cap.set(cv2.CAP_PROP_POS_FRAMES,desc['frame']);ok,pano=cap.read();cap.release();assert ok,desc
else:pano=cv2.imread(desc['panorama']);assert pano is not None
return Image.fromarray(cv2.cvtColor(perspective(pano,*desc['pose']),cv2.COLOR_BGR2RGB))
def policy_messages(row):
c=[]
if row['reference']:c.extend([{'type':'text','text':'First-frame target identity reference:'},{'type':'image'}])
for i,_ in enumerate(row['observations']):
c.extend([{'type':'text','text':'Current observation:' if i==len(row['observations'])-1 else 'Historical observation:'},{'type':'image'}])
if i<len(row['history_actions']):c.append({'type':'text','text':'Previous action: '+row['history_actions'][i]})
if row.get('audio_path'):c.append({'type':'text','text':'<|audio_start|>'+'<|audio_pad|>'*41+'<|audio_end|>'+f"Audio ends at {row['audio_end_s']:.1f}s; microphone orientation is fixed, independent of camera rotation."})
task='Each rotation advances 0.2 seconds. Track the reference. Only rotate: dyaw in [-30,30], dpitch in [-15,15].' if row['dynamic'] else 'Rotate to find the requested object, then submit the CURRENT pose. dyaw in [-180,180], dpitch in [-90,90].'
c.append({'type':'text','text':row['instruction']+f"\nCurrent yaw,pitch={row['pose']}. "+task+' Return exactly one next action in <answer> tags.'})
return [{'role':'system','content':'Control a camera using only causal observations. Positive yaw right, positive pitch up. Return one action without explanation.'},{'role':'user','content':c}]
class ActionDataset:
def __init__(self,path,proc,max_length):self.rows=read(path);self.proc=proc;self.max_length=max_length
def __len__(self):return len(self.rows)
def __getitem__(self,i):
import torch,numpy as np,soundfile as sf
from PIL import Image
r=self.rows[i];images=[]
if r['reference']:
with Image.open(r['reference']) as im:images.append(im.convert('RGB'))
images.extend(load_image(x) for x in r['observations'])
messages=policy_messages(r)
messages[-1]['content'][-1]['text']=messages[-1]['content'][-1]['text'].replace('Return exactly one next action in <answer> tags.','Return exactly one next action in <answer> tags.')
messages[0]['content']='Control a camera using only causal observations. Positive yaw right, positive pitch up. Return one action without explanation.'
text=self.proc.apply_chat_template(messages,tokenize=False,add_generation_prompt=True)
b=dict(self.proc(text=[text],images=images,return_tensors='pt'))
target=torch.tensor([self.proc.tokenizer.encode(r['answers'][0]+'<|im_end|>',add_special_tokens=False)],dtype=torch.long)
n=b['input_ids'].shape[1]
b['input_ids']=torch.cat([b['input_ids'],target],dim=1)
b['attention_mask']=torch.ones_like(b['input_ids'])
b['labels']=b['input_ids'].clone();b['labels'][:,:n]=-100
assert b['input_ids'].shape[1]<=self.max_length
if r.get('audio_path'):
from omini.data.staqa_preprocessing import prepare_staqa_waveform
with sf.SoundFile(r['audio_path']) as f:sr=f.samplerate;wave=f.read(frames=int(sr*r['audio_end_s']),dtype='float32',always_2d=True)
assert wave.shape[1]==2 and np.isfinite(wave).all()
b['audio']=prepare_staqa_waveform(wave,sr).unsqueeze(0)
return b
def inverse_messages():
return [dict(role='system',content='Infer relative camera rotation from overlapping views of a static scene.'),dict(role='user',content=[dict(type='text',text='Before rotation:'),dict(type='image'),dict(type='text',text='After rotation:'),dict(type='image'),dict(type='text',text='Both views have 100-degree horizontal field of view. Positive yaw is right and positive pitch is up. Return exactly <answer>rotate(dyaw,dpitch)</answer> in integer degrees.')])]
def inverse_batch(row,proc,max_length,seed):
import torch
rng=random.Random(seed);before=row['pose'];dy=rng.choice([-45,-30,-15,15,30,45]);dp=rng.choice([-15,0,15])
after=[(before[0]+dy)%360,max(-75,min(75,before[1]+dp))];dp=after[1]-before[1]
assert abs(dp)<=90
desc=row['observations'][-1]
pics=[load_image(dict(panorama=desc['panorama'],pose=p)) for p in [before,after]]
text=proc.apply_chat_template(inverse_messages(),tokenize=False,add_generation_prompt=True)
b=dict(proc(text=[text],images=pics,return_tensors='pt'));n=b['input_ids'].shape[1]
target=torch.tensor([proc.tokenizer.encode(f'<answer>rotate({dy},{dp})</answer><|im_end|>',add_special_tokens=False)])
b['input_ids']=torch.cat([b['input_ids'],target],1);b['attention_mask']=torch.ones_like(b['input_ids']);b['labels']=b['input_ids'].clone();b['labels'][:,:n]=-100
assert b['input_ids'].shape[1]<=max_length
return b
|