| """Read an EEF companion and construct explicitly selected OpenWAM-style targets. |
| |
| No temporal shift is applied to command actions. next_state is an explicitly |
| derived, one-retained-frame-ahead achieved-state target, with terminal masking. |
| """ |
| import io |
| import json |
| from pathlib import Path |
| import tarfile |
|
|
| import numpy as np |
|
|
|
|
| def load_index(path): |
| with open(path, encoding='utf-8') as stream: |
| return {row['episode_id']:row for row in map(json.loads,stream)} |
|
|
|
|
| def load_indexed_episode(dataset_root, record, *, action_mode, unify_action=True, allow_copied_commands=False): |
| with (Path(dataset_root)/record['shard']).open('rb') as stream: |
| stream.seek(record['json_offset']);metadata=json.loads(stream.read(record['json_bytes'])) |
| stream.seek(record['npz_offset']);data=stream.read(record['npz_bytes']) |
| if metadata['episode_id']!=record['episode_id']: |
| raise ValueError('EEF episode index does not match shard contents') |
| with np.load(io.BytesIO(data),allow_pickle=False) as archive: |
| arrays={key:archive[key] for key in archive.files} |
| result=build_arrays(arrays,metadata,action_mode=action_mode,unify_action=unify_action, |
| allow_copied_commands=allow_copied_commands) |
| result['metadata']=metadata |
| return result |
|
|
|
|
| def _pose20(pose, grip, pose_valid, grip_valid): |
| q = pose[..., 3:7] |
| x, y, z, w = np.moveaxis(q, -1, 0) |
| col0 = np.stack([1-2*(y*y+z*z), 2*(x*y+z*w), 2*(x*z-y*w)], axis=-1) |
| col1 = np.stack([2*(x*y-z*w), 1-2*(x*x+z*z), 2*(y*z+x*w)], axis=-1) |
| value = np.concatenate([pose[..., :3], col0, col1, grip[..., None]], axis=-1).reshape(len(pose), 20) |
| mask = np.concatenate([np.repeat(pose_valid[..., None], 9, axis=-1), grip_valid[..., None]], axis=-1).reshape(len(pose), 20) |
| return np.where(mask, value, 0).astype(np.float32), mask |
|
|
|
|
| def build_arrays(arrays, metadata, *, action_mode, unify_action=True, allow_copied_commands=False): |
| if action_mode not in {'command', 'next_state'}: |
| raise ValueError('Explicitly choose action_mode=command or next_state') |
| state, state_mask = _pose20(arrays['state_pose'], arrays['state_gripper'], |
| arrays['state_pose_valid'], arrays['state_gripper_valid']) |
| if action_mode == 'next_state': |
| action = np.zeros_like(state); action_mask = np.zeros_like(state_mask) |
| action[:-1] = state[1:]; action_mask[:-1] = state_mask[1:] |
| else: |
| pose = arrays['command_pose'] if 'command_pose' in arrays else arrays['state_pose'] |
| pose_mask = arrays['command_pose_valid'].copy() |
| if not allow_copied_commands: |
| pose_mask[:, np.asarray(metadata.get('command_pose_is_state_copy', [False, False]), dtype=bool)] = False |
| action, action_mask = _pose20(pose, arrays['command_gripper'], pose_mask, arrays['command_gripper_valid']) |
| if metadata.get('command_quality_review_required', False): |
| action_mask[:]=False;action[:]=0 |
| if unify_action: |
| def scatter(value): |
| out = np.zeros((len(value), 80), dtype=value.dtype) |
| out[:, :10] = value[:, :10]; out[:, 34:44] = value[:, 10:20] |
| return out |
| state, action, state_mask, action_mask = map(scatter, [state, action, state_mask, action_mask]) |
| return {'state': state, 'action': action, 'state_mask': state_mask, 'action_mask': action_mask, |
| 'episode_id': metadata['episode_id'], 'fps': metadata['fps'], 'action_mode': action_mode} |
|
|
|
|
| def load_episode(shard_path, episode_id, *, action_mode, unify_action=True, allow_copied_commands=False): |
| with tarfile.open(shard_path, 'r') as shard: |
| metadata = json.load(shard.extractfile(episode_id+'.json')) |
| data = shard.extractfile(episode_id+'.npz').read() |
| with np.load(io.BytesIO(data), allow_pickle=False) as archive: |
| arrays = {key: archive[key] for key in archive.files} |
| result = build_arrays(arrays, metadata, action_mode=action_mode, unify_action=unify_action, |
| allow_copied_commands=allow_copied_commands) |
| result['metadata'] = metadata |
| return result |
|
|
|
|
| def iter_episodes(shard_path, *, action_mode, unify_action=True, allow_copied_commands=False): |
| """Stream companions without repeatedly scanning the TAR directory.""" |
| with tarfile.open(shard_path, 'r|') as shard: |
| pending = {} |
| for member in shard: |
| if not member.isfile(): |
| continue |
| key, suffix = member.name.rsplit('.', 1) |
| data = shard.extractfile(member).read() |
| if suffix == 'npz': |
| pending[key] = data |
| elif suffix == 'json': |
| metadata = json.loads(data) |
| with np.load(io.BytesIO(pending.pop(key)), allow_pickle=False) as archive: |
| arrays = {name: archive[name] for name in archive.files} |
| result = build_arrays(arrays, metadata, action_mode=action_mode, unify_action=unify_action, |
| allow_copied_commands=allow_copied_commands) |
| result['metadata'] = metadata |
| yield result |
|
|
