📦 模型下載：https://pub-478929a98a5c440cb22c2241c0bde314.r2.dev/door_state_v20260515/best.pt

door_state_v20260515 — hatch_swa stack (2 binary heads + SWA-4)

cvat2 #11 image-level tag。Schema 從 4-class softmax 改成 2 binary heads (has_open / has_close)。
unknown 在 manifest build 階段過濾。ppe-demo HatchHandler hot-swap 即可上線。

State → (has_open, has_close)：

mixed → (1, 1)
open → (1, 0)
closed → (0, 1)
unknown → filtered (no rows in cvat 樣本 → no-op)

三版對照

	door_v20260511 (4-class softmax)	door_v20260515 best.pt	door_v20260515 SWA-4
schema	4-class softmax (unknown/open/closed/mixed)	2 binary heads	2 binary heads
backbone	MobileNetV3-L 224	MobileNetV3-L 384	MobileNetV3-L 384
test accuracy	0.8366	—	—
test mAP (active heads)	—	0.9939	0.9806
test macro F1	0.5815*	0.9607	0.9645
ckpt schema	4-class softmax (handler 不相容)	HatchHandler 相容	HatchHandler 相容
ckpt thresholds	none	per-attr val-best F1	per-attr val-best F1

* v20260511 macro_f1 0.5815 包含 unknown=0 support 拖累；active 3 class mean F1 0.775。

Per-attr metrics (test set)

attr	best.pt						SWA-4
attr	AP	F1	P	R	thr	valid	AP	F1	P	R	thr
has_open	0.9967	0.9742	0.9750	0.9734	0.733	1642	0.9984	0.9924	0.9904	0.9944	0.313
has_close	0.9911	0.9472	0.9612	0.9335	0.914	1642	0.9628	0.9366	0.9090	0.9660	0.005

Training history

ep	train_loss	val_mAP	val_macroF1	val_open_AP	val_close_AP
1	0.5041	0.9555	0.9150	0.914	0.996
2	0.0813	0.9938	0.9649	0.988	1.000
3	0.0233	0.9875	0.9570	0.975	1.000
4	0.0164	0.9834	0.9547	0.969	0.998
5	0.0117	0.9920	0.9788	0.984	1.000
6	0.0055	0.9622	0.9606	0.927	0.998
7	0.0081	0.9840	0.9647	0.969	0.999
8	0.0071	0.9644	0.9614	0.929	0.999
9	0.0028	0.9926	0.9770	0.986	1.000
10	0.0056	0.9800	0.9648	0.962	0.998
11	0.0039	0.9737	0.9681	0.948	1.000
12	0.0018	0.9909	0.9742	0.982	0.999
13	0.0013	0.9927	0.9779	0.986	0.999
14	0.0021	0.9937	0.9759	0.988	1.000
15	0.0011	0.9752	0.9690	0.951	1.000
16	0.0011	0.9813	0.9687	0.963	0.999
17	0.0009	0.9863	0.9700	0.973	0.999
18	0.0007	0.9853	0.9702	0.971	0.999
19	0.0018	0.9831	0.9712	0.967	0.999
20	0.0006	0.9755	0.9684	0.952	0.999

Sample inference (test, 4 per state，SWA-4 + val-thr)

truth=open pred=unknown/? open=100% close=0% thr_open=1.00 thr_close=0.18

truth=closed pred=closed open=0% close=100% thr_open=1.00 thr_close=0.18

truth=mixed pred=closed open=100% close=100% thr_open=1.00 thr_close=0.18

truth=mixed pred=closed open=71% close=100% thr_open=1.00 thr_close=0.18

FP audit — top-8 最自信錯判（sampled 300）

truth=mixed → pred=closed open=100% close=100%

Config

{
  "version": "v20260515",
  "backbone_name": "mobilenetv3_large_100.ra_in1k",
  "arch": "MobileNetV3-L 2 binary heads (has_open/has_close), partial-label BCE, hatch_swa stack, SWA-4",
  "params_M": 4.204594,
  "feat_dim": 1280,
  "attrs": [
    "has_open",
    "has_close"
  ],
  "img_size": 384,
  "best_val_mAP": 0.993847131190981,
  "best_epoch": 2,
  "epochs_run": 20,
  "total_train_time_s": 495.99095845222473,
  "test_metrics_best": {
    "mAP": 0.993896870987621,
    "macro_f1": 0.9606927842247622,
    "per_attr": {
      "has_open": {
        "ap": 0.9967186921084682,
        "f1": 0.9742143432715552,
        "p": 0.975,
        "r": 0.9734299516908212,
        "thr": 0.73291015625,
        "n": 1242,
        "valid": 1642,
        "tp": 1209,
        "fp": 31,
        "fn": 33
      },
      "has_close": {
        "ap": 0.991075049866774,
        "f1": 0.9471712251779693,
        "p": 0.9612167300380228,
        "r": 0.9335302806499262,
        "thr": 0.9140625,
        "n": 1354,
        "valid": 1642,
        "tp": 1264,
        "fp": 51,
        "fn": 90
      }
    }
  },
  "test_metrics_swa": {
    "mAP": 0.9806083384001334,
    "macro_f1": 0.9644968473528424,
    "per_attr": {
      "has_open": {
        "ap": 0.9984425119276941,
        "f1": 0.9923664122137404,
        "p": 0.9903769045709703,
        "r": 0.9943639291465378,
        "thr": 0.31298828125,
        "n": 1242,
        "valid": 1642,
        "tp": 1235,
        "fp": 12,
        "fn": 7
      },
      "has_close": {
        "ap": 0.9627741648725727,
        "f1": 0.9366272824919442,
        "p": 0.9089645587213343,
        "r": 0.96602658788774,
        "thr": 0.005489349365234375,
        "n": 1354,
        "valid": 1642,
        "tp": 1308,
        "fp": 131,
        "fn": 46
      }
    }
  },
  "val_metrics_swa": {
    "mAP": 0.9829408931150647,
    "macro_f1": 0.9705861404390816,
    "per_attr": {
      "has_open": {
        "ap": 0.9665191589075289,
        "f1": 0.9520697167755992,
        "p": 0.9700332963374029,
        "r": 0.9347593582887701,
        "thr": 1.0,
        "n": 935,
        "valid": 1972,
        "tp": 874,
        "fp": 27,
        "fn": 61
      },
      "has_close": {
        "ap": 0.9993626273226005,
        "f1": 0.9891025641025641,
        "p": 0.9993523316062176,
        "r": 0.9790609137055838,
        "thr": 0.1759033203125,
        "n": 1576,
        "valid": 1972,
        "tp": 1543,
        "fp": 1,
        "fn": 33
      }
    }
  },
  "swa_n": 4,
  "swa_source_ckpts": [
    "ep17.pt",
    "ep18.pt",
    "ep19.pt",
    "ep20.pt"
  ],
  "hyperparams": {
    "batch": 64,
    "epochs": 20,
    "lr": 0.0003,
    "wd": 0.01,
    "patience": 20,
    "img_size": 384,
    "aug": "cam_erase",
    "sampler": "weighted",
    "loss": "bce",
    "focal_gamma": 2.0
  }
}