Charm/config.yaml at main · FBehrad/Charm · 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
data:
  dataset: tad66k  # aadb, tad66k, ava, para, baid, spaq, koniq10k
  augmentation: all  # HF, GS, RR, all(without GS), None
  patch_selection: random  # ['saliency', 'random', 'frequency', 'gradient', 'entropy', 'original']
  max_seq_len_from_original_res: 1024  # ViT input length
  patch_size: 14 # 14 for dinov2-small and dinov2-base, 16 for vit-small, and 32 for vit-base
  patch_stride: 14
  factor: 0.5  # the maximum downscaling of unimportant patches
  scales: 2

path:
  ratings_path: D:\Datasets\TAD66k\labels.csv
  img_folder: D:\Datasets\TAD66k\train_imgs
  mask_folder: D:\Datasets\TAD66k\train_depths
  test_folder: D:\Datasets\TAD66k\test_imgs
  test_mask_folder: D:\Datasets\TAD66k\test_depths
  warm_start_checkpoint: null

training:
  batch_size: 16
  shuffle_data: False
  train_size: 0.95 # the portion of the dataset which is used for training, the rest is used for validation
  num_workers: 8
  warm_start: False
  num_epochs: 30
  optimizer: AdamW # Adam or sgd or AdamW
  test_epochs: 3 # This variable specifies the number of epochs after which the model is evaluated on the test set.
  save_model_steps: 3 # This variable specifies the number of epochs after which the model is saved.
  learning_rate:
    lr: 0.00001
    scheduler:
      type: cosine # cosine or step_lr
      cosine:
        cycle_mult: 1.0
        min_lr: 0.000001
        warm_up: 0  # number of epochs, it is better to use  10%  of the whole epochs
        gamma: 0.8
      step_lr:
        milestones:
          - 5
          - 10
        gamma: 0.5
  weight_decay: 0.02 # put 0 for not applying weight decay
  momentum: 0.9
  max_grad_norm: 1.0
  accumulation_steps: 1 # 1 for not using accumulated gradients
  gradient_clipping: True
  LLDR: # Layer-wise learning rate decay
    enable: False
    max_lr_patch: 0.00001
    cosine_scheduler_patch:
      cycle_mult: 1.0
      min_lr: 0.000001
      warm_up: 3 # number of epochs, it is better to use  10%  of the whole epochs
      gamma: 0.8
    weight_decay_patch: 0.01

    max_lr_encoder: 0.00001
    cosine_scheduler_encoder:
      cycle_mult: 1.0
      min_lr: 0.000001
      warm_up: 3  # number of epochs, it is better to use  10%  of the whole epochs
      gamma: 0.8
    weight_decay_encoder: 0.01

    max_lr_mlphead: 0.0001
    cosine_scheduler_mlp:
      cycle_mult: 1.0
      min_lr: 0.00001
      warm_up: 0  # number of epochs, it is better to use  10%  of the whole epochs
      gamma: 0.8
    weight_decay_mlp: 0.01

model:
  encoder_name: facebook/dinov2-small #flexiViT 'facebook/dinov2-small' 'facebook/dinov2-base' microsoft/swin-small-patch4-window7-224 #'facebook/dino-vits16' # WinKawaks/vit-small-patch16-224  #google/vit-base-patch32-224-in21k  google/vit-base-patch32-384 "google/vit-base-patch16-224-in21k" 'facebook/dinov2-large'
  frozen_vit: False
  masked: False
  mixout: False
  mixout_prob: 0.7
  dropout_prob: 0.1  # put zero not to have dropout
  multi_class: False # Just for AADB