exp(dagger-deploy): 轮次21——DAgger-3（部署配置采集500题+15000步微调）负结论：部署口径 39.5% vs dag2 43.8%，开环退化到19.5mm；下一步试冻结编码器只调解码器

2026-10-20 09:41:44 +08:00
parent bb919d33c9
commit 7a8e5aab63
4 changed files with 33 additions and 2 deletions
@@ -30,6 +30,8 @@ def main():
    ap.add_argument('--device', default='cuda' if torch.cuda.is_available() else 'cpu')
    ap.add_argument('--fb_gain', type=float, default=0.0, help='部署分布采集：目标反馈增益（与 rollout 一致）')
    ap.add_argument('--safety', type=int, default=0, help='部署分布采集：是否开安全层')
+    ap.add_argument('--replan', default='r3'); ap.add_argument('--commit_n', type=int, default=3)
+    ap.add_argument('--safety_partial', type=int, default=0)
    a = ap.parse_args()
    gap_min = float(os.environ.get('MIN_GAP_MM') or 3)
    os.makedirs(a.out, exist_ok=True)
@@ -43,7 +45,8 @@ def main():
    for i in range(0, len(tasks), a.bs):
        tb = tasks[i:i + a.bs]
        with torch.no_grad():
-            _, per = rollout(model, kin, tb, a.device, K=8, safety=bool(a.safety), seed=a.seed * 100003 + i, return_traj=True, fb_gain=a.fb_gain)
+            _, per = rollout(model, kin, tb, a.device, K=8, safety=bool(a.safety), seed=a.seed * 100003 + i, return_traj=True,
+                            fb_gain=a.fb_gain, replan=a.replan, commit_n=a.commit_n, safety_partial=a.safety_partial)
        tr = per['traj'].to(a.device)                                   # [B,T,6]
        B, Tn, _ = tr.shape
        _, T, R, _, _ = tasks_to_tensors(tb, a.device)