-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathupdate.txt
150 lines (54 loc) · 2.56 KB
/
update.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
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
1 line1 > line0> line2
2 scale 0.5> scale1 > 128x128
new split 基本没有提升
radam快一丢丢
仅仅rotate基本没有提升
仅仅cutmix训练中
只用rotate = 用多种augment（在adam+ mixup/cutmix的实验中 0.5 scale）
只用cutmix效果好，和radam一起用能更快收敛
①max+avg > gem > max
②just cutmix + rotate > cutmix/mixup +rotate
③0.5 scale line1 better
④alpha=0.4 >alpha=1
⑤rgb gray没有太大区别
⑥patience 和cos差别不是太大，但是收敛略微快一点点
liner1 scale0.5 avg+max rotate cutmix radam coslr cv 0.9904
liner1 scale0.5 rgb avg+max rotate cutmix radam coslr cv 0.9885
liner0 scale0.5 GEM rotate no dropout cutmix radam coslr cv 0.9840
liner1 scale0.5 avg+max rotate cutmix radam reducedlr patience=4 cv 0.9881 alpha=0.4
liner1 scale0.5 avg+max rotate cutmix radam reducedlr patience=2 cv 0.9860 alpha=0.4
liner0 scale0.5 avg+max rotate cutmix radam reducedlr patience=2 cv 0.9869 alpha=0.4
liner1 scale 1 avg+max rotate cutmix radam reducedlr patience=2 cv 0.9796 alpha=0.4
liner1 scale0.5 avg+max rotate cutmix radam reducedlr patience=2 cv 0.9832 alpha=1
liner0 scale0.5 MAXp rotate cutmix adam coslr cv 0.9832 alpha=0.4 cv 0.9797
liner1 scale0.5 avg+max rotate augcv cutmix radam reducedlr patience=4 1e-3 weight_decay focallossroot cv 0.9890 alpha=0.5
最高优先级：
weight_decay 1e-3
cutout
focal loss
ohem loss
224x224 input 直接resize
TODO
1 scale 0.5 with 3channel 基本太大效果
2 focal loss scale 0.5
3 ohem loss scale 0.5
4 去掉mixup, 去掉augment，只用rotate或者只用水平翻转 只用cutmix+rotate有效果
1 加入gridmask
2 regularition add weight decay
3 cutmix alpha to 1
4 gem 使用 效果变差
5 gap后先去掉dropout 使用 效果变差
1 line3 增加tail的特征表示
1 train_resize_128_bs64 scale resize 128 batchsize 64
2 kitti scale 0.5 batchsize 64 RADAM patience 4 factor 0.8
3 kfolder scale 1 batchsize 64
python train.py --model senet50 --outdir 0226/seresnext50_resize224_splitstrict_rotate_liner1_avg+max_radam_cutmix_reducedlr4_lr1e-3 --gpu_ids 2,5 --width 224 --height 224 --feather_data_path BengaliData/feather_resize224/ --mixup 1 --image_mode gray --patience 4 --LR_SCHEDULER REDUCED --optimizer RADAM --lr 1e-3 --lr_ratio 0.65 --batch_size 128
random split strictly
resize 128x128
patience 5 factor 0.65
实验1 augmix + cutmix gray
实验2 augmix + cutmix rgb
no aug no cutmix 9727 115
cutmix 9862 122
fmix 9856 78
cutmix + rotate + patience 3 0.75 radam 9912 9904 120epoch