LADS/data_configs/base.yaml at main · lisadunlap/LADS · 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
SEED: 0


DATA:
  BATCH_SIZE: 128
  NUM_WORKERS: 4

  # Size of training images (square)
  SIZE: 224

  # Mean and std deviation to use, found in datasets/__init__.py
  NORMALIZATION: 'imagenet'

  ATTENTION_DIR: "NONE"
  USE_GROUP_WEIGHTS: False
  SEPARATE_CLASSES: True

  BALANCE_AVERAGE_MODE: "mean" # mean, harmonic_mean

  REMOVE_BACKGROUND: False
  MASK_PIXELS: False
  ROOT: "/shared/lisabdunlap/vl-attention/data"

EXP:
  FREEZE_BACKBONE: False
  BIAS_DETECTION: False
  SAL_LAYER: 'layer4.2.relu'
  LOG_GCAMS: True
  ATTN_PER_CLASS: False

  LOSSES:
    CLASSIFICATION:
      WEIGHT: 1
    GRADIENT_OUTSIDE:
      COMPUTE: False
      LOG: False
      WEIGHT: 0.01
      CRITERION: "L1"
      GT: "segmentation"
      COMBINE_ATT_MODE: "average_nonzero"
    GRADIENT_INSIDE:
      COMPUTE: False
      LOG: False
      WEIGHT: 1
      CRITERION: "L1"
      GT: "segmentation"
      COMBINE_ATT_MODE: "average_nonzero"
    GRADCAM:
      COMPUTE: False
      LOG: False
      WEIGHT: 1
      CRITERION: "L1"
      GT: "segmentation"
      MODE: "match" # match, suppress_outside
      COMBINE_ATT_MODE: "average_nonzero"

LOGGING:
  # save model every save_every # of epochs. if 0, don't save model on a regular basis
  SAVE_EVERY: 0

  # save best model (updates throughout training)
  SAVE_BEST: True

  # save last model (updates throughout training)
  SAVE_LAST: False

  # Logging attention.
  # Step is # epochs b/w logging. Would also log attention before training and at end.
  LOG_ATTENTION: False
  LOG_ATTENTION_STEP: 10

  SAVE_CONFUSION_MATRIX: False
  SAVE_CONFUSION_MATRIX_PATH: "NONE"

  # Save stats over multiple runs to a CSV file.
  # Gathering stats over multiple trials can also be done by setting EXP.NUM_TRIALS.
  SAVE_STATS_PATH: "NONE"