50 lines
2.6 KiB
Python
50 lines
2.6 KiB
Python
|
"""老师轨迹回放自检（2026-10-18 首次正式预训练后新增；train 阶段开训前自动跑，约 1 分钟）：
|
|||
|
|
把 gen 写出的老师 50 Hz 轨迹（keep=true）直接放进闭环仿真用的 PyTorch 运动学 / 碰撞 / 到达判据里检查。
|
|||
|
|
老师轨迹在 JS 侧已认证无碰、已到达；如果这里判出碰撞或未到达，说明仿真口径（kin.json、间隙、到达门槛）和数据对不上，
|
|||
|
|
那么闭环指标本身不可信，训练多少步都没用 —— 直接停下（rc 5）。
|
|||
|
|
python -m ml.eval.replay_teacher --n 512 --out ml/runs/replay_teacher.json
|
|||
|
|
判据：到达率（cat≠D）≥ 0.95，碰撞率 ≤ 0.01。"""
|
|||
|
|
import argparse, glob, json, os, sys
|
|||
|
|
import torch
|
|||
|
|
|
|||
|
|
from ml.common.kin_torch import Kin
|
|||
|
|
from ml.sim.closed_loop import goal_err
|
|||
|
|
|
|||
|
|
|
|||
|
|
def main():
|
|||
|
|
ap = argparse.ArgumentParser()
|
|||
|
|
ap.add_argument('--shards', default='ml/data/shard_*.jsonl'); ap.add_argument('--n', type=int, default=512)
|
|||
|
|
ap.add_argument('--out', default='ml/runs/replay_teacher.json')
|
|||
|
|
ap.add_argument('--device', default='cuda' if torch.cuda.is_available() else 'cpu')
|
|||
|
|
a = ap.parse_args()
|
|||
|
|
kin = Kin(device=a.device)
|
|||
|
|
recs = []
|
|||
|
|
for f in sorted(glob.glob(a.shards)):
|
|||
|
|
for line in open(f, encoding='utf-8'):
|
|||
|
|
r = json.loads(line)
|
|||
|
|
if r.get('keep') and r.get('traj') and r.get('cat') != 'D': recs.append(r)
|
|||
|
|
if len(recs) >= a.n: break
|
|||
|
|
if len(recs) >= a.n: break
|
|||
|
|
if not recs: sys.exit('没有可回放的老师轨迹（先跑 gen）')
|
|||
|
|
reach = col = 0; worst_gap = []
|
|||
|
|
for r in recs:
|
|||
|
|
q = torch.tensor(r['traj'], dtype=torch.float32, device=a.device)
|
|||
|
|
mid = 0.5 * (q[1:] + q[:-1])
|
|||
|
|
g = kin.gaps(torch.cat([q, mid], 0), 0.5)['gm'].min(-1).values
|
|||
|
|
hit = bool((g <= 0).any()); col += hit; worst_gap.append(g.min().item())
|
|||
|
|
T = torch.tensor([r['T']], dtype=torch.float32, device=a.device)
|
|||
|
|
R = torch.tensor([r['R'] if r.get('R') else [0.] * 9], dtype=torch.float32, device=a.device)
|
|||
|
|
e, oe = goal_err(kin, q[-1:], T, R)
|
|||
|
|
reach += bool((e < 1).item() and (oe < 0.5).item())
|
|||
|
|
n = len(recs)
|
|||
|
|
res = {'n': n, 'reach': reach / n, 'collision': col / n, 'min_gap_mm_p10': sorted(worst_gap)[n // 10],
|
|||
|
|
'pass': reach / n >= 0.95 and col / n <= 0.01, 'fp': kin.fp}
|
|||
|
|
os.makedirs(os.path.dirname(a.out) or '.', exist_ok=True)
|
|||
|
|
json.dump(res, open(a.out, 'w', encoding='utf-8'), indent=1)
|
|||
|
|
print('老师轨迹回放：到达 %.1f%%、碰撞 %.1f%%（%d 条）' % (100 * res['reach'], 100 * res['collision'], n), flush=True)
|
|||
|
|
sys.exit(0 if res['pass'] else 5)
|
|||
|
|
|
|||
|
|
|
|||
|
|
if __name__ == '__main__':
|
|||
|
|
main()
|