-
Notifications
You must be signed in to change notification settings - Fork 1
/
cfg.py
35 lines (31 loc) · 1.06 KB
/
cfg.py
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
# model.py
CHECKPOINT = True # True断点保存
NET = 'DlaNet' # DlaNet
NUM_CLASSES = 20 # @
LAYER = 34
# dataset.py
DATASET_NAME = 'RVOC2007' # @
IMG_EXT = 'jpg'
# HRSC 和 UCAS-AOD的均值和方差
#MEAN = [0.5194416012442385, 0.5378052387430711,0.533462090585746] # @
#STD = [0.3001546018824507, 0.28620901391179554, 0.3014112676161966] # @
# RVOC2007的均值和方差
MEAN = [0.471, 0.448, 0.408]
STD = [0.234, 0.239, 0.242]
# loss.py: smooth
Loss = 'l1' # OR 'gwd'
# train
GPU_ID = '1'
TRAIN_BATCH_SIZE = 2
VAL_BATCH_SIZE = 1
learning_rate = 1.25e-4
NUM_EPOCHS = 800 # RVCO需要训练时间久一点。 150
# predict
RET_IMG = 'img_ret' # 可视化结果保存的位置
CON_SCORE = 0.3
'''
在post_process中,预测值是128*128的[cx,cy,h,w,ang]格式,若采用该格式
直接变回原图尺度,会偶尔出现w出现负值的情况。
因此,需要将[cx,cy,h,w,ang] --> [lx,ly,rx,ry,ang]映射回原图尺度,
在将其变成[cx,cy,h,w,ang]就不会出现该问题了。
'''