ABCDEFGHIJKLMNOPQRSTUVWXYZ
1
This spreadsheet contains results of hyper-parameter study for zero-shot character identification.
2
3
Experiments are conducted with codes in this repository: https://github.com/kosuke1701/optuna-metric-learning
4
The code has been modified as necessary, so the latest version and the implementation at the time of the experiments may differ. However, there should not be that big difference.
5
6
Notes about experiments:
7
* Performance metric is Precision at R on development dataset
8
* All experiments before "cls" uses my private dataset derived from pixiv. 25% of characters are kept for development data.
9
* Since "cls" experiment, I used ZACI-20 dataset. 3.3% of characters are kept for development data.
10
11
* The number of trials by optuna is set to 60 (due to my mistake, there are more trials in some experiments).
12
* The maximum number of epoch is set to 30, and the patience is set to 1.
13
14
15
* Unless otherwise noted, ResNet-18 is used as trunk model.
16
* After (and including) "augmentation" experiment, I used contrastive loss with its previous optimal hyperparameters.
17
* seresnet-152d, epsilon uses m=8 instead of m=16
18
* After "randaugment" experiment, I used RandAugment with its previous optimal hyperparameters.
19
20
* Unless otherwise noted, batch size is set to 64.
21
* efficientnet-b7 uses batch size = 32, m=4
22
23
* data_per_epoch is set to 100000 for batch size 64. It is set to proportional to batch size.
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