-
Notifications
You must be signed in to change notification settings - Fork 4
/
training_log.txt
92 lines (58 loc) · 2.69 KB
/
training_log.txt
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
1.---------------------------------------------------------
kernal size =3 ,original scale, speed:57 images/s,
Downsample 16 shape 64x512--64x256--64x128--64x64------64x32------64x64--64x128--64x256--64x512
original -- 25000 steps
eval
read: 0.017s detect: 0.009s
Accuracy:
car:
Pixel-seg: P: 0.606, R: 0.971, IoU: 0.595
pedestrian:
Pixel-seg: P: 0.299, R: 0.312, IoU: 0.180
cyclist:
Pixel-seg: P: 0.308, R: 0.583, IoU: 0.253
Wait 60s for new checkpoints to be saved ...
2.---------------------------------------------------------
original + downsample 8
kernal size =3 , speed:43 images/s, 25000
in training process
Accuracy:
car:
Pixel-seg: P: , R: , IoU: 0.69
pedestrian:
Pixel-seg: P: , R: , IoU: 0.420
cyclist:
Pixel-seg: P: , R: , IoU: 0.253
---(2) same structure with depth decrease speed 48 images/s
car:
Pixel-seg: P: 0.409, R: 0.962, IoU: 0.402
pedestrian:
Pixel-seg: P: 0.572, R: 0.207, IoU: 0.127
cyclist:
Pixel-seg: P: 0.252, R: 0.684, IoU: 0.151
3.--------------------------------------------------------
original+downsample 8(in fire 11) + downsample 16(in fire12) speed 44 images/s
========================================================================================================================
1-3
(1) conclusion， change downsample is not good for detection, because in the add process, it will become the noise
if all downsample size is 16, the result is also not good.
(2) because we can not use the pre-trained model parameter, the weight in pedestrain and cyclist are hard to coverage.
30000 substract
car:
Pixel-seg: P: 0.566, R: 0.976, IoU: 0.558
pedestrian:
Pixel-seg: P: 0.277, R: 0.369, IoU: 0.188
cyclist:
Pixel-seg: P: 0.249, R: 0.684, IoU: 0.224
Wait 60s for new checkpoints to be saved ...
My idea
-- 64x512 downsample 16 --64x32 --upsample 16 64x512
input(64x512)-- 64x256 downsample 8 --64x32 --upsample 8 64x256
-- 64x128 downsample 4 --64x32 --upsample 4 64x128
The structure in downsample6
conv1b -- GCN -- BR -- - sum
| GCN+BR ---------------------------------- sum |
| | | |
| | | |
| | --sum | |
input - conv1a -Max pooling- fire2 decon13 BR con14