| from pathlib import Path |
| import math |
| import threading |
| from .config import MAX_UPLOAD_BYTES, MODELS |
| _REQUEST_LOCK = threading.RLock() |
| _PROFILES = {'human': 'Human', 'yam': 'YAM', 'dex': 'Dex', 'tianji': 'Tianji'} |
|
|
| class DataEnginePipeline: |
|
|
| @staticmethod |
| def _inputs(video, profile, left_wrist, right_wrist, yam32, instruction): |
| from .media import InputError, probe, check_synchronized |
| if not isinstance(profile, str) or profile.lower() not in _PROFILES: |
| raise InputError('profile must be human, yam, dex, or tianji') |
| canonical = _PROFILES[profile.lower()] |
| if bool(left_wrist) != bool(right_wrist): |
| raise InputError('Provide both left and right wrist videos together.') |
| if canonical == 'Human' and left_wrist: |
| raise InputError('The human profile accepts one video.') |
| if not isinstance(yam32, bool): |
| raise InputError('yam32 must be true or false.') |
| if yam32 and (canonical != 'YAM' or not left_wrist): |
| raise InputError('yam32 requires YAM and all three camera views.') |
| if not isinstance(instruction, str) or len(instruction.strip()) > 2000: |
| raise InputError('instruction must be a string of at most 2000 characters.') |
| paths = {'main': video} |
| if left_wrist: |
| paths.update(left_wrist=left_wrist, right_wrist=right_wrist) |
| videos = {} |
| for role, value in paths.items(): |
| if not isinstance(value, (str, Path)) or not str(value).strip(): |
| raise InputError('Each camera path must be a nonempty local path.') |
| path = Path(value).expanduser().resolve(strict=True) |
| if not path.is_file() or not 0 < path.stat().st_size <= MAX_UPLOAD_BYTES: |
| raise InputError('Each input must be a nonempty local video of at most 256 MiB.') |
| videos[role] = probe(path) |
| check_synchronized(videos) |
| return (videos, canonical) |
|
|
| def inspect_inputs(self, video, *, profile='human', left_wrist=None, right_wrist=None, yam32=False, instruction=''): |
| videos, canonical = self._inputs(video, profile, left_wrist, right_wrist, yam32, instruction) |
| return {'status': 'inputs_validated', 'profile': canonical, 'yam32': yam32, 'duration_s': videos['main'].duration_s, 'model_revisions': MODELS, 'new_model_calls': 0, 'cameras': {role: {'sha256': v.sha256, 'source_fps': str(v.rate), 'frames': len(v.pts_us), 'size': list(v.size), 'constant_frame_rate': v.cfr, 'duration_s': v.duration_s} for role, v in videos.items()}} |
|
|
| def annotate(self, video, *, profile='human', left_wrist=None, right_wrist=None, yam32=False, instruction='', progress=None, include_audit=False): |
| import torch |
| if not torch.cuda.is_available(): |
| raise RuntimeError('CUDA is required for inference. Use --validate-only to check inputs on CPU.') |
| from . import pipeline |
| with _REQUEST_LOCK: |
| videos, canonical = self._inputs(video, profile, left_wrist, right_wrist, yam32, instruction) |
| result = pipeline.annotate(videos, canonical, yam32, instruction, progress or (lambda *a, **k: None)) |
| duration = result['duration_s'] |
| for segment in result['segments']: |
| if not (math.isfinite(segment['start_s']) and math.isfinite(segment['end_s']) and (0 <= segment['start_s'] < segment['end_s'] <= duration + 0.001)): |
| raise RuntimeError('The pipeline returned an invalid caption interval.') |
| if not include_audit: |
| result.pop('pipeline', None) |
| return result |
|
|
