DEIMv2/configs/deimv2/deimv2_hgnetv2_atto_coco.yml at main · NikhilSandy/DEIMv2 · 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
__include__: [
  '../dataset/coco_detection.yml',
  '../runtime.yml',
  '../base/dataloader.yml',
  '../base/optimizer.yml',
  '../base/deimv2.yml',
]

output_dir: ./outputs/deimv2_hgnetv2_atto_coco

DEIM:
  encoder: LiteEncoder

HGNetv2:
  name: 'Atto'
  return_idx: [2]
  freeze_at: -1
  freeze_norm: False
  use_lab: True

LiteEncoder:
  in_channels: [256]
  feat_strides: [16]
  # intra
  hidden_dim: 64

  # cross
  expansion: 0.34
  depth_mult: 0.5
  act: 'silu'


DEIMTransformer:
  feat_channels: [64, 64]
  feat_strides: [16, 32]
  hidden_dim: 64
  num_levels: 2
  num_points: [4, 2]

  num_layers: 3
  eval_idx: -1
  num_queries: 100

  # FFN
  dim_feedforward: 160

  # New options for DEIMv2
  share_bbox_head: True
  use_gateway: False

# Increase to search for the optimal ema
epoches: 500 # 468 + 32

## Our LR-Scheduler
warmup_iter: 4000
flat_epoch: 250    # 4 + epoch // 2, e.g., 40 = 4 + 72 / 2
no_aug_epoch: 32
lr_gamma: 0.5

optimizer:
  type: AdamW
  params:
    - params: '^(?=.*backbone)(?!.*norm|bn).*$'
      lr: 0.001
    - params: '^(?=.*backbone)(?=.*norm|bn).*$'
      lr: 0.001
      weight_decay: 0.
    - params: '^(?=.*(?:encoder|decoder))(?=.*(?:norm|bn)).*$'    # except bias
      weight_decay: 0.

  lr: 0.002
  betas: [0.9, 0.999]
  weight_decay: 0.0001

eval_spatial_size: [320, 320]
train_dataloader:
  total_batch_size: 128
  dataset:
    transforms:
      ops:
        - {type: Mosaic, output_size: 160, rotation_range: 10, translation_range: [0.1, 0.1], scaling_range: [0.5, 1.5],
           probability: 1.0, fill_value: 0, use_cache: True, max_cached_images: 50, random_pop: True}
        - {type: RandomPhotometricDistort, p: 0.5}
        - {type: RandomZoomOut, fill: 0}
        - {type: RandomIoUCrop, p: 0.8}
        - {type: SanitizeBoundingBoxes, min_size: 12}
        - {type: RandomHorizontalFlip}
        - {type: Resize, size: [320, 320], }
        - {type: SanitizeBoundingBoxes, min_size: 12}
        - {type: ConvertPILImage, dtype: 'float32', scale: True}
        - {type: ConvertBoxes, fmt: 'cxcywh', normalize: True}
      policy:
        epoch: [4, 250, 400]   # list
      mosaic_prob: 0.3

  collate_fn:
    mixup_prob: 0.0
    mixup_epochs: [40000, 15000]
    copyblend_prob: 0.0
    copyblend_epochs: [40000, 15000]

    stop_epoch: 468 # 468 + 32
    ema_restart_decay: 0.9999
    base_size: 320
    base_size_repeat: ~

val_dataloader:
  total_batch_size: 256
  dataset:
    transforms:
      ops:
        - {type: Resize, size: [320, 320], }
        - {type: ConvertPILImage, dtype: 'float32', scale: True}
  shuffle: False
  num_workers: 16


DEIMCriterion:
  losses: ['mal', 'boxes']    # , 'local'
  use_uni_set: False

  matcher:
    matcher_change_epoch: 450   # FIX This