exp(dagger-deploy): 采集支持 fb_gain+safety（训练分布=部署分布）；新增 base_dag2 配置（同 base）

2026-10-20 03:40:15 +08:00
parent 8e5431c460
commit dc27781a72
3 changed files with 15 additions and 2 deletions
@@ -28,6 +28,8 @@ def main():
    ap.add_argument('--parts', type=int, default=8); ap.add_argument('--out', required=True)
    ap.add_argument('--bs', type=int, default=64)
    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')
    a = ap.parse_args()
    gap_min = float(os.environ.get('MIN_GAP_MM') or 3)
    os.makedirs(a.out, exist_ok=True)
@@ -41,7 +43,7 @@ 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=False, seed=a.seed * 100003 + i, return_traj=True)
+            _, 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)
        tr = per['traj'].to(a.device)                                   # [B,T,6]
        B, Tn, _ = tr.shape
        _, T, R, _, _ = tasks_to_tensors(tb, a.device)