-
Notifications
You must be signed in to change notification settings - Fork 31
/
train_config.py
executable file
·125 lines (97 loc) · 4.98 KB
/
train_config.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
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
124
#-*-coding:utf-8-*-
import os
import numpy as np
from easydict import EasyDict as edict
config = edict()
os.environ["CUDA_VISIBLE_DEVICES"] = "0" ##if u use muti gpu set them visiable there and then set config.TRAIN.num_gpu
config.TRAIN = edict()
#### below are params for dataiter
config.TRAIN.process_num = 3 ### process_num for data provider
config.TRAIN.prefetch_size = 100 ### prefect Q size for data provider
config.TRAIN.num_gpu = 1 ##match with os.environ["CUDA_VISIBLE_DEVICES"]
config.TRAIN.batch_size = 6 ###A big batch size may achieve a better result, but the memory is a problem
config.TRAIN.log_interval = 10
config.TRAIN.epoch = 2000 ###just keep training , evaluation shoule be take care by yourself,
### generally 10,0000 iters is enough
config.TRAIN.train_set_size=13000 ###widerface train size
config.TRAIN.val_set_size=3000 ###widerface val size
config.TRAIN.iter_num_per_epoch = config.TRAIN.train_set_size // config.TRAIN.num_gpu // config.TRAIN.batch_size
config.TRAIN.val_iter=config.TRAIN.val_set_size// config.TRAIN.num_gpu // config.TRAIN.batch_size
config.TRAIN.lr_value_every_step = [0.00001,0.0001,0.0005,0.0001,0.00001,0.000001] ##warm up is used
config.TRAIN.lr_decay_every_step = [500,1000,60000,80000,100000]
config.TRAIN.weight_decay_factor = 5.e-4 ##l2 regular
config.TRAIN.vis=False ##check data flag
config.TRAIN.norm='BN' ##'GN' OR 'BN'
config.TRAIN.lock_basenet_bn=False
config.TRAIN.frozen_stages=-1 ##no freeze
config.TEST = edict()
config.TEST.parallel_iterations=8
config.TEST.score_thres = 0.05
config.TEST.iou_thres = 0.3
config.TEST.max_detect = 1500
config.DATA = edict()
config.DATA.root_path=''
config.DATA.train_txt_path='train.txt'
config.DATA.val_txt_path='val.txt'
config.DATA.num_category=1 ###face 1 voc 20 coco 80
config.DATA.num_class = config.DATA.num_category + 1 # +1 background
config.DATA.PIXEL_MEAN = [123., 116., 103.] ###rgb
config.DATA.PIXEL_STD = [58., 57., 57.]
config.DATA.hin = 640 # input size
config.DATA.win = 640
config.DATA.max_size=[config.DATA.hin,config.DATA.win] ##h,w
config.DATA.cover_small_face=5 ###cover the small faces
config.DATA.mutiscale=False #if muti scale set False then config.DATA.MAX_SIZE will be the inputsize
config.DATA.scales=(320,640)
# anchors -------------------------
config.ANCHOR = edict()
config.ANCHOR.rect=True
config.ANCHOR.rect_longer=True #### make anchor h/w=1.5
config.ANCHOR.ANCHOR_STRIDE = 16
config.ANCHOR.ANCHOR_SIZES = (16,32,64, 128, 256, 512) # sqrtarea of the anchor box
config.ANCHOR.ANCHOR_STRIDES = (4, 8,16, 32, 64, 128) # strides for each FPN level. Must be the same length as ANCHOR_SIZES
config.ANCHOR.ANCHOR_RATIOS = (1., 4.) ###### 1:2 in size,
config.ANCHOR.POSITIVE_ANCHOR_THRESH = 0.35
config.ANCHOR.NEGATIVE_ANCHOR_THRESH = 0.35
config.ANCHOR.AVG_MATCHES=20
config.ANCHOR.super_match=True
from lib.core.anchor.anchor import Anchor
config.ANCHOR.achor=Anchor()
# # basemodel ---------------------- fddb 0.983
# config.MODEL = edict()
# config.MODEL.continue_train=False ### revover from a trained model
# config.MODEL.model_path = './model' # save directory
# config.MODEL.net_structure='resnet_v1_50'#resnet_v1_50,resnet_v1_101,mobilenet
# config.MODEL.pretrained_model='resnet_v1_50.ckpt'
# config.MODEL.fpn_dims=[256,512,1024,2048,512,256]
# config.MODEL.fem_dims=512
###resnet_v1_101 as basemodel
# config.MODEL = edict()
# config.MODEL.continue_train=False ### revover from a trained model
# config.MODEL.model_path = './model/' # save directory
# config.MODEL.net_structure='resnet_v1_101' ######'resnet_v1_50,resnet_v1_101,mobilenet
# config.MODEL.pretrained_model='resnet_v1_101.ckpt'
# config.MODEL.fpn_dims=[256,512,1024,2048,512,256]
# config.MODEL.fem_dims=512
#vgg as basemodel. if vgg, set config.TRAIN.norm ='None', achieves fddb 0.987
config.MODEL = edict()
config.TRAIN.norm='None'
config.MODEL.l2_norm=[10,8,5]
config.MODEL.continue_train=False ### revover from a trained model
config.MODEL.model_path = './model/' # save directory
config.MODEL.net_structure='vgg_16'
config.MODEL.pretrained_model='vgg_16.ckpt'
config.MODEL.fpn_dims=[256,512,512,1024,512,256]
config.MODEL.fem_dims=512
# ##mobilenet as basemodel, mobile net is not that fast
# config.MODEL = edict()
# config.MODEL.continue_train=False ### revover from a trained model
# config.MODEL.model_path = './model/' # save directory
# config.MODEL.net_structure='MobilenetV2' ######'resnet_v1_50,resnet_v1_101,MobilenetV2
# config.MODEL.pretrained_model='mobilenet_v2_0.5_224.ckpt'
# config.MODEL.fpn_dims=[96,96,288,160,256,256]
# config.MODEL.fem_dims=256
config.MODEL.fpn=True
config.MODEL.dual_mode=True
config.MODEL.maxout=True
config.MODEL.max_negatives_per_positive= 3.0