"""Prepare short, exactly training-caption-matched inputs for the original v4 protocol.""" import argparse import csv import hashlib import json from pathlib import Path import sys ROOT = Path(__file__).resolve().parents[2] CASES = ( ('simple_walk_forward', 't0000103', 'g00', '直行(训练原文)'), ('simple_walk_circle', 't0003078', 'g02', '逆时针行走(训练原文)'), ('simple_squats', 't0003075', 'g02', '三次深蹲(训练原文)'), ) def prepare(run, output): import numpy as np runtime = json.loads((ROOT / 'config/runtime.json').read_text()) code = Path(runtime['visualization_resources']).parent sys.path.insert(0, str(code / 'OmniHM_bench')) from video_motion import load_motion36, sample_motion36, motion36_to_qpos, G1_MUJOCO_JOINT_NAMES from preview_scale import save manifest_path = output / 'manifest.json' if manifest_path.exists(): m = json.loads(manifest_path.read_text()) if m['source_run'] != str(run) or [c['id'] for c in m['cases']] != [c[0] for c in CASES]: raise ValueError('Existing simple-text selection does not match this run') return manifest = dict(source_run=str(run), selection_only=True, checkpoint_step=None, representation='root40_v1', normalization='g1_root40_fixed_v1', inference_fps=50, denoising_steps=10, action_chunk=30, execution_horizon=20, selection_seed=20261002, protocol='GT frame 0 initialization; free autoregressive rollout; no future GT poses', selection_version='simple_training_text_v1', cases=[]) hashes = {} for index, (identity, motion_id, group, label) in enumerate(CASES): source_manifest = ROOT / f'training_set/text/index/{group}.csv' row = next(r for r in csv.DictReader(source_manifest.open()) if r['move_name'] == motion_id) source = ROOT / 'training_set/text' / row['move_g1_path'] motion = load_motion36(source) # Keep the entire training sequence and the original gallery's 30 Hz display. times = np.arange(int(np.floor((len(motion)-1)*30/50))+1) / 30 qpos = motion36_to_qpos(sample_motion36(motion, source_fps=50, sample_times=times)).astype(np.float32) directory = output / 'text' / identity directory.mkdir(parents=True, exist_ok=True) record = dict(sample_id=identity, caption=row['content_short_description'], source_motion_path=str(source), training_id=f'text_{group} / {motion_id}', source_manifest=str(source_manifest), data_split='训练集内固定样例(不是独立验证集)', label=label, source_sha256={n:hashlib.sha256((source/n).read_bytes()).hexdigest() for n in ('joint_pos.csv','body_pos.csv','body_quat.csv')}) save(directory / 'selection_record.json', record) np.savez_compressed(directory / 'gt.npz', qpos_36=qpos, valid=np.ones(len(qpos),bool), fps=30., joint_names=np.asarray(G1_MUJOCO_JOINT_NAMES), sample_id=identity, score_start_frame=1) case = dict(id=identity, modality='text', directory=f'text/{identity}', caption=row['content_short_description'], seed=20261002+index, fps=30., frames=len(qpos), source_sequence=motion_id, input_media=None, status='selected', label=label) manifest['cases'].append(case) for name in ('selection_record.json','gt.npz'): hashes[f'text/{identity}/{name}'] = hashlib.sha256((directory/name).read_bytes()).hexdigest() contract = [{k:c[k] for k in ('id','directory','seed','fps','frames','source_sequence')} for c in manifest['cases']] manifest['fixed_input_sha256'] = hashes manifest['fixed_case_contract_sha256'] = hashlib.sha256(json.dumps(contract,sort_keys=True,separators=(',',':')).encode()).hexdigest() save(manifest_path, manifest) if __name__ == '__main__': parser = argparse.ArgumentParser(description=__doc__) parser.add_argument('--run', type=Path, required=True) parser.add_argument('--output', type=Path, required=True) args = parser.parse_args() run, output = args.run.resolve(strict=True), args.output.resolve() if not run.is_relative_to(ROOT/'scale/groups') or not output.is_relative_to(ROOT/'scale/groups'): parser.error('Only scale experiment inputs may be prepared') prepare(run, output)