-
Notifications
You must be signed in to change notification settings - Fork 187
/
config.yml
executable file
·40 lines (33 loc) · 1.26 KB
/
config.yml
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
SLICE_INTV: 2 # slice interval in mm after interpolation. if 0, only use one slice thus no 3d context
NUM_SLICES: 3 # multi-slice input to incorporate 3d context
NORM_SPACING: 1 # normalize image size to norm_spacing mm/pixel (so the smaller this value, the larger the image).
MAX_IM_SIZE: 800
MAX_PATCH_SIZE: 200
ARCH: 'VGG16bn'
EMBEDDING_DIM: 256
SCORE_PROPAGATION: True
SAMPLES_PER_BATCH: 128
BOX_PAD: 60
ROI_METHOD: 'FIXED_CONTEXT'
PAD_BORDER: True
TRAIN:
USE_PRETRAINED_MODEL: False
TEXT_MINED_LABEL: 'RU'
CE_LOSS_WT_1: 1
CE_LOSS_WT_2: 1
CE_LOSS_WEIGHTED: True
CE_POS_WT_CLAMP: 300
RHEM_LOSS_WT: 1
RHEM_POWER: 2
RHEM_BATCH_SIZE: 10000
TRIPLET_LOSS_WT: 5
TRIPLET_LOSS_MARGIN: .1
NUM_TRIPLET: 5000
SIMILAR_LABEL_THRESHOLD: 1
DISSIMILAR_LABEL_THRESHOLD: 0
TEST:
CRITERION: 'mean_perclass_f1'
SCORE_PARAM: 5 # the threshold or top-K for label decision. Only useful if USE_CALIBRATED_TH=False
FILTER_EXCLUSIVE_LABELS: False # if label A's score is higher than B while A and B are exclusive, then remove B
# not always good, because B may be the correct label if A and B are hard to distinguish
USE_CALIBRATED_TH: True # calibrate a threshold for each label in the val set, then use it in the test set