door_state_v20260511 — MobileNetV3-L 224 (4-class)

cvat2 #11 image-level tag (state ∈ unknown/open/closed/mixed). 4-class CE + WeightedRandomSampler。

注意：cvat manifest 中 unknown 類別在 train/val/test 全為 0 樣本 — 實際資料只有 3 個 active class (open/closed/mixed)。模型架構保留 4 個 output 但 unknown logit 未訓練。

Test 指標

accuracy	macro F1	N
0.8366	0.5815	1640

Per-class metrics

class	P	R	F1
unknown	0.0000	0.0000	0.0000
open	0.6222	0.4878	0.5469
closed	0.8646	0.9123	0.8878
mixed	0.8732	0.9099	0.8912

Confusion matrix (rows=true, cols=pred)

	open	closed	mixed	row total
unknown	0	0	0	0
open	140	40	107	287
closed	16	364	19	399
mixed	69	17	868	954
col total	225	421	994	1640

Sample inference (test, 3 per class with samples)

truth=open → pred=open (61%)

truth=open → pred=open (96%)

truth=open → pred=open (99%)

truth=closed → pred=closed (100%)

truth=mixed → pred=mixed (100%)

Training history

epoch	train_loss	val_acc	val_macro_f1
1	1.5521	0.7049	0.5015
2	0.2231	0.8891	0.6558
3	0.0889	0.8918	0.6567
4	0.0445	0.8744	0.6389
5	0.0222	0.8946	0.6486
6	0.0191	0.9221	0.6701
7	0.0122	0.9248	0.6738
8	0.0090	0.9083	0.6674
9	0.0088	0.9166	0.6688
10	0.0045	0.9203	0.6666
11	0.0062	0.9193	0.6747
12	0.0030	0.9175	0.6649
13	0.0029	0.9120	0.6598
14	0.0040	0.9047	0.6619
15	0.0044	0.9102	0.6634
16	0.0028	0.9313	0.6761
17	0.0014	0.9313	0.6761
18	0.0019	0.9258	0.6741
19	0.0021	0.9322	0.6772
20	0.0031	0.9331	0.6830
21	0.0015	0.9258	0.6736
22	0.0026	0.9322	0.6790
23	0.0030	0.9267	0.6723
24	0.0016	0.9239	0.6731
25	0.0021	0.9313	0.6772
26	0.0027	0.9322	0.6778
27	0.0020	0.9303	0.6761
28	0.0029	0.9322	0.6778
29	0.0054	0.9303	0.6761
30	0.0024	0.9322	0.6784

Config

{
  "version": "v20260511",
  "task": "door_state image classification (4-class)",
  "backbone": "mobilenetv3_large_100",
  "params_M": 4.207156,
  "img_size": 224,
  "hyperparams": {
    "batch": 128,
    "epochs": 30,
    "lr": 0.0003,
    "wd": 0.01
  },
  "labels": [
    "unknown",
    "open",
    "closed",
    "mixed"
  ],
  "best_val_macro_f1": 0.6830110617839775,
  "best_epoch": 20,
  "epochs_run": 30,
  "train_time_s": 532.9745559692383,
  "test_metrics": {
    "acc": 0.8365853658536585,
    "macro_f1": 0.5814625773150699,
    "per_class": {
      "unknown": {
        "precision": 0.0,
        "recall": 0.0,
        "f1": 0.0
      },
      "open": {
        "precision": 0.6222222222222222,
        "recall": 0.4878048780487805,
        "f1": 0.546875
      },
      "closed": {
        "precision": 0.8646080760095012,
        "recall": 0.9122807017543859,
        "f1": 0.8878048780487805
      },
      "mixed": {
        "precision": 0.8732394366197183,
        "recall": 0.909853249475891,
        "f1": 0.891170431211499
      }
    },
    "confusion_matrix": [
      [
        0,
        0,
        0,
        0
      ],
      [
        0,
        140,
        40,
        107
      ],
      [
        0,
        16,
        364,
        19
      ],
      [
        0,
        69,
        17,
        868
      ]
    ],
    "n": 1640
  }
}