"""Read-only representative endpoint audit; optional bounded HTTP video streaming.""" from __future__ import annotations import argparse import json import sys from pathlib import Path import numpy as np import torch from PIL import Image, ImageDraw sys.path.insert(0,str(Path(__file__).resolve().parents[1])) from resolution_adapt.data import DATASET,DATASET_REVISION,POOLS,pool_of,ResolutionBottleneck,resize_rgb from resolution_adapt.model import LIT_REVISION from training_support import DatasetGripperToImageWAM def audit(root: Path, output: Path, *, stream=False, stats_path=None, processed=None): import av import pyarrow.parquet as pq from huggingface_hub import hf_hub_download,hf_hub_url,get_token output.mkdir(parents=True,exist_ok=True) def file(name): return Path(hf_hub_download(DATASET,name,repo_type='dataset',revision=DATASET_REVISION, local_dir=str(root))) if stream else root/name provenance=[json.loads(line) for line in file('meta/source_episodes.jsonl').read_text().splitlines()] info=json.loads(file('meta/info.json').read_text()) ep_rows=pq.read_table(file('meta/episodes/chunk-000/file-000.parquet')).to_pylist() ep_by_id={e['episode_index']:e for e in ep_rows} if stats_path is None: stats_path=hf_hub_download('linjianman/LIT','imagewam/lit_stage2/dataset_stats.json',revision=LIT_REVISION) stats=json.loads(Path(stats_path).read_text()) def normalize(value,category): st=stats[category]['default']; low=torch.tensor(st['global_min']); high=torch.tensor(st['global_max']) span=high-low; constant=span<1e-4; span[constant]=2 scale=2/span; offset=-1-scale*low; offset[constant]=-low[constant] return (value*scale+offset).clamp(-5,5) report=[]; selected=[next(e for e in provenance if pool_of(e)==p and e['length']>32) for p in POOLS] for prov in selected: pool=pool_of(prov); ep=ep_by_id[prov['episode_index']] shard=info['data_path'].format(chunk_index=ep['data/chunk_index'],file_index=ep['data/file_index']) rows=pq.read_table(file(shard),filters=[('episode_index','=',ep['episode_index'])]).to_pylist() rows.sort(key=lambda r:r['frame_index']) t=0 states=torch.tensor([r['observation.state'] for r in rows[:17]],dtype=torch.float32) actions=torch.tensor([r['action'] for r in rows[:16]],dtype=torch.float32) assert states.shape==(17,8) and actions.shape==(16,7) assert all(r['episode_index']==prov['episode_index'] for r in rows[:17]) original=actions.clone() converted=DatasetGripperToImageWAM().forward({'action':{'default':actions}})['action']['default'] assert torch.equal(converted[:,:6],original[:,:6]) assert torch.equal(converted[:,-1],(1-original[:,-1])/2) cameras=[]; panels=[] for camera in ('front','wrist'): key=f'observation.images.{camera}' video=info['video_path'].format(video_key=key,chunk_index=ep[f'videos/{key}/chunk_index'], file_index=ep[f'videos/{key}/file_index']) timestamps=[ep[f'videos/{key}/from_timestamp']+rows[i]['timestamp'] for i in (0,16)] options={'rw_timeout':'30000000'} location=str(root/video) if stream: location=hf_hub_url(DATASET,video,repo_type='dataset',revision=DATASET_REVISION) token=get_token() if token: options['headers']='Authorization: Bearer '+token+'\r\n' with av.open(location,options=options) as container: video_stream=container.streams.video[0] container.seek(int(timestamps[0]/video_stream.time_base),stream=video_stream,backward=True) decoded={} for frame in container.decode(video_stream): timestamp=float(frame.pts*video_stream.time_base) for i,target in enumerate(timestamps): if abs(timestamp-target)<1e-4: decoded[i]=frame.to_ndarray(format='rgb24') if len(decoded)==2: break if timestamp > timestamps[-1]+.1: raise ValueError('Missing endpoint video frame') raw=torch.from_numpy(np.stack([decoded[i] for i in (0,1)])).permute(0,3,1,2) assert raw.shape==(2,3,256,256) and raw.dtype==torch.uint8 source=raw.float()/255; degraded=resize_rgb(source,128); final=ResolutionBottleneck()(source) assert degraded.shape==(2,3,128,128) and final.shape==(2,3,224,224) cameras.append(final) panels.append((camera,source[0],degraded[0],final[0])) combined=torch.cat(cameras,dim=-1)*2-1 assert combined.shape==(2,3,224,448) and -1<=combined.min()<=combined.max()<=1 normalized_action=normalize(converted,'action'); pose=normalize(states[16],'state') if processed is not None: augment=processed.video_augmentation; processed.video_augmentation=None try: sample=processed[prov['dataset_from_index']+t] finally: processed.video_augmentation=augment torch.testing.assert_close(sample['goal_pose'],pose) torch.testing.assert_close(sample['action'],normalized_action) torch.testing.assert_close(sample['video'],combined.permute(1,0,2,3),atol=.02,rtol=0) assert not sample['goal_pose_is_pad'].any() canvas=Image.new('RGB',(720,560),'white'); draw=ImageDraw.Draw(canvas) for row,(camera,*images) in enumerate(panels): x=0 for label,img in zip(('source 256','bottleneck 128','model 224'),images,strict=True): array=(img.permute(1,2,0).numpy()*255).round().clip(0,255).astype(np.uint8) canvas.paste(Image.fromarray(array),(x,row*280+22)) draw.text((x,row*280),f'{pool} {camera}: {label}',fill='black'); x+=img.shape[-1]+20 canvas.save(output/f'{pool}.png') report.append({'pool':pool,'episode':prov['episode_index'],'anchor':t,'pose_frame':t+16, 'raw':[3,256,256],'degraded':[3,128,128],'model':[3,224,224],'combined':[3,224,448], 'raw_rgb_range':[0,255],'float_range':[0,1],'normalized_image_range':[float(combined.min()),float(combined.max())], 'state_shape':[8],'action_shape':list(converted.shape),'pose_shape':list(pose.shape), 'action_normalized_range':[float(normalized_action.min()),float(normalized_action.max())], 'gripper_exactly_once':True,'stored_rotation':False,'processed_loader_compared':processed is not None}) (output/'dataset_preflight.json').write_text(json.dumps(report,indent=2)+'\n') return report if __name__=='__main__': p=argparse.ArgumentParser();p.add_argument('--dataset-root',type=Path,required=True) p.add_argument('--output',type=Path,required=True);p.add_argument('--stream-from-hub',action='store_true') p.add_argument('--stats',type=Path) a=p.parse_args(); print(json.dumps(audit(a.dataset_root,a.output,stream=a.stream_from_hub,stats_path=a.stats),indent=2))