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
|
from utils.configuration import Configuration
config: Configuration = {
'system': {
# Disable accelerator
'disable_acc': False,
# GPU(s) used in training or testing if available
'CUDA_VISIBLE_DEVICES': '0',
# Directory used in training or testing for temporary storage
'save_dir': 'runs',
# Recorde disentangled image or not
'image_log_on': False,
# The number of subjects for validating (Part of testing set)
'val_size': 10,
},
# Dataset settings
'dataset': {
# Name of dataset (CASIA-B or FVG)
'name': 'CASIA-B',
# Path to dataset root (required)
'root_dir': 'data/CASIA-B-MRCNN-V2/SEG',
# The number of subjects for training
'train_size': 74,
# Number of sampled frames per sequence (Training only)
'num_sampled_frames': 30,
# Truncate clips longer than `truncate_threshold`
'truncate_threshold': 40,
# Discard clips shorter than `discard_threshold`
'discard_threshold': 15,
# Number of input channels of model
'num_input_channels': 3,
# Resolution after resize, can be divided 16
'frame_size': (64, 48),
# Cache dataset or not
'cache_on': False,
},
# Dataloader settings
'dataloader': {
# Batch size (pr, k)
# `pr` denotes number of persons
# `k` denotes number of sequences per person
'batch_size': (4, 6),
# Number of workers of Dataloader
'num_workers': 4,
# Faster data transfer from RAM to GPU if enabled
'pin_memory': True,
},
# Hyperparameter tuning
'hyperparameter': {
'model': {
# Auto-encoder feature channels coefficient
'ae_feature_channels': 64,
# Appearance, canonical and pose feature dimensions
'f_a_c_p_dims': (192, 192, 96),
# HPM pyramid scales, of which sum is number of parts
'hpm_scales': (1, 2, 4, 8),
# Global pooling method
'hpm_use_avg_pool': True,
'hpm_use_max_pool': True,
# Number of parts after Part Net
'tfa_num_parts': 16,
# Attention squeeze ratio
'tfa_squeeze_ratio': 4,
# Embedding dimensions for each part
'embedding_dims': (256, 256),
# Batch Hard or Batch All
'triplet_is_hard': True,
# Use non-zero mean or sum
'triplet_is_mean': True,
# Triplet loss margins for HPM and PartNet, None for soft margin
'triplet_margins': None,
},
'optimizer': {
# Global parameters
# Initial learning rate of Adam Optimizer
'lr': 1e-4,
# Coefficients used for computing running averages of
# gradient and its square
# 'betas': (0.9, 0.999),
# Term added to the denominator
# 'eps': 1e-8,
# Weight decay (L2 penalty)
'weight_decay': 0.001,
# Use AMSGrad or not
# 'amsgrad': False,
# Local parameters (override global ones)
# 'auto_encoder': {
# 'weight_decay': 0.001
# },
},
'scheduler': {
# Step start to decay
'start_step': 500,
# Multiplicative factor of decay in the end
'final_gamma': 0.01,
# Local parameters (override global ones)
# 'hpm': {
# 'final_gamma': 0.001
# }
}
},
# Model metadata
'model': {
# Model name, used for naming checkpoint
'name': 'RGB-GaitPart',
# Restoration iteration from checkpoint (single model)
# 'restore_iter': 0,
# Total iteration for training (single model)
# 'total_iter': 80000,
# Restoration iteration (multiple models, e.g. nm, bg and cl)
'restore_iters': (0, 0, 0),
# Total iteration for training (multiple models)
'total_iters': (30_000, 40_000, 60_000),
},
}
|
