-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathsearch_config_avmnist.yml
149 lines (102 loc) · 2.51 KB
/
search_config_avmnist.yml
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
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
small_dataset: False
pretrained_path1 : "./saved_supernets/AV-MNIST/supernet_avmnist_image.pth.tar"
pretrained_path2 : "./saved_supernets/AV-MNIST/supernet_avmnist_sound.pth.tar"
hw_lut_path: "./evaluate/backbone_eval/efficiency/lut_ofa_mnist_tx2.json"
fusion_lut: "./evaluate/fusion_eval/lut_fusion_tx2.json"
# Search hyperparameters
evo_search_outer:
parent_popu_size: 8
survival_ratio: .5
mutate_size: 64
crossover_size: 64
mutate_prob: 0.4
crossover_prob: 0.8
evo_iter: 30
evo_search_inner:
survival_ratio: 0.25
# Fusion search parameters
fusion_epochs: 25
C : 192
L : 16
multiplier : 2
node_multiplier : 1
num_outputs : 10
drpt : 0.1
# num_input_nodes : 8
num_keep_edges : 2
arch_learning_rate : 0.0003
arch_weight_decay : 0.001
weight_decay : 0.0001
eta_max : 0.0001
eta_min : 0.000001
Ti : 1
Tm : 2
parallel : False
save : ''
# supernet_arch: 'supernet_dynamic_model'
supernet_arch: 'ofa'
arch: 'supernet_dynamic_model'
exp_name: "unimodal_supernet_avmnist_image"
resume: ""
batch_size_per_gpu: 256
batch_size : 256
sandwich_rule: True
alpha_min: -1.0
alpha_max: 1.0
iw_clip: 5.0
grad_clip_value: 1.0
augment: "auto_augment_tf"
n_gpu_per_node: 8
num_nodes: 8
n_cpu_per_node: 32
memory_per_node: '128g'
epochs: 50
warmup_epochs: 5
start_epoch: 0
label_smoothing: 0.1
inplace_distill: False
#sync-batchnormalization, suggested to use in bignas
bn_calibration: False
sync_bn: False
bn_momentum: 0
bn_eps: 1e-5
post_bn_calibration_batch_num: 64
num_arch_training: 4
models_save_dir: "./saved_models"
#### cloud training resources ####
data_loader_workers_per_gpu: 4
########### regularization ################
# supernet training regularization (the largest network)
dropout: 0.2
drop_connect: 0.2
drop_connect_only_last_two_stages: True
weight_decay_weight: 0.0001
weight_decay_bn_bias: 0.
## =================== optimizer and scheduler======================== #
optimizer:
method: adam
momentum: 0.9
nesterov: True
lr_scheduler:
method: "warmup_cosine_lr"
base_lr: 0.00001
clamp_lr_percent: 0.0
### distributed training settings ###
distributed: False
distributed_val: False
eval_only: True
multiprocessing_distributed: True
# dist_backend: 'nccl'
### multimodal dataset ###
dataset: 'avmnist'
# Path to the 'soundmnist_data' folder
dataset_dir: "/home/imed/Desktop/datasets/AV-MNIST/"
n_classes: 10
drop_last: True
# task : ['multilabel', 'classification']
task: classification
image_resolution: 28
sound_resolution: 20
in_channels: 1
print_freq: 50
seed: 42