FlowCLAS/src/configs/flowclas/flowclas_city.yaml at main · TRAILab/FlowCLAS · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
# FlowCLAS experiment: Cityscapes + COCO (512×1024 training crops)
# Usage: --data configs/base/cityscapes_coco.yaml -c configs/base/base.yaml -c configs/flowclas/flowclas_city.yaml

# --- Experiment ---
experiment:
  logger: [wandb, csv]
  name: flowclas
  offline: false
  project_name: flowclas_Cityscapes
  wandb_id: null

task: segmentation

# --- Data ---
data:
  divisor: 14
  train_batch_size: 16
  eval_batch_size: 8

  city_train_transform:
    class_path: albumentations.Compose
    init_args:
      transforms:
        - class_path: HorizontalFlip
          init_args:
            p: 0.5
        - class_path: RandomCrop
          init_args:
            height: 512
            width: 1024
            p: 1.0

  coco_train_mix_transform:
    class_path: albumentations.Compose
    init_args:
      transforms:
        - class_path: RandomScaleV2
          init_args:
            scale_factors: [.1, .125, .25, .5, .5, .75]
            interpolation: 3
            p: 1.0
        - class_path: PadIfNeeded
          init_args:
            min_height: 512
            min_width: 1024
            border_mode: 0
            value: 0.0
            mask_value: 255
            always_apply: true
        - class_path: CenterCrop
          init_args:
            height: 512
            width: 1024
            always_apply: true

  coco_train_ood_transform:
    class_path: albumentations.Compose
    init_args:
      transforms:
        - class_path: PadIfNeeded
          init_args:
            min_height: 512
            min_width: 1024
            border_mode: 0
            value: 0.0
            mask_value: 255
            always_apply: true
        - class_path: CenterCrop
          init_args:
            height: 512
            width: 1024
            always_apply: true

# --- Model ---
model:
  class_path: core.models.FlowCLAS
  init_args:
    backbone: rein_dinov2
    backbone_ckpt: ../weights/rein_dinov2l_city.pth
    layers: [0, 1, 2, 3]
    proj_dim: 128
    num_classes: 20
    flow_steps: 16
    use_gaussian: false
    ood_index: 254
    ignore_index: 255
    alpha: 1e-6
    temperature: 0.13
    use_logp_score: false
    use_sam: true
    no_ood: false
    subnet_type: residual
    load_strict: false
    use_seg: true
    slide_inf_cfg:
      enable: false

# --- Metrics (overrides base; full-resolution thresholds) ---
metrics:
  image: []
  pixel:
    - class_path: BinaryAUROC
      init_args:
        thresholds: null
        ignore_index: 255
    - class_path: BinaryAP
      init_args:
        thresholds: null
        ignore_index: 255
    - class_path: RecallAt05
      init_args:
        thresholds: null
        ignore_index: 255
    - class_path: RecallAt01
      init_args:
        thresholds: null
        ignore_index: 255
    - class_path: FPR95
      init_args:
        thresholds: null
        ignore_index: 255
    - class_path: ECEMetric
      init_args:
        ignore_index: 255
    - class_path: MCEMetric
      init_args:
        ignore_index: 255

# --- Optimization ---
optimizer:
  class_path: torch.optim.AdamW
  init_args:
    lr: 1e-6
    weight_decay: 1e-5

lr_scheduler:
  class_path: LinearWarmupCosineAnnealingLR
  init_args:
    max_lr: 1e-6

# --- Visualization ---
visualization:
  save: true
  total: 10
  features_only: false

# --- Checkpoint monitors (per OOD benchmark) ---
monitor1:
  metric:
    - p_FPR95-Fishy
    - p_FPR95-RoadAnomaly
    # - p_FPR95-AnomalyTrack
    # - p_FPR95-ObstacleTrack
  mode: min

monitor2:
  metric:
    - p_BinaryAP-Fishy
    - p_BinaryAP-RoadAnomaly
    # - p_BinaryAP-AnomalyTrack
    # - p_BinaryAP-ObstacleTrack
  mode: max

# --- Trainer overrides ---
trainer:
  max_epochs: 600
  # strategy: ddp_find_unused_parameters_true