-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain_options.py
222 lines (204 loc) · 10.5 KB
/
train_options.py
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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
from __future__ import absolute_import, division, print_function
import argparse
from models.rec_decoders.swiftnet_rec.rec_decoders import SUPPORTED_REC_DECODER
class TrainOptions:
def __init__(self):
self.parser = argparse.ArgumentParser(description="SemSeg Evaluator options")
# SYSTEM options
self.parser.add_argument("--no_cuda",
type=int,
choices=[0, 1],
default=0,
help="if set disables CUDA")
self.parser.add_argument("--num_workers",
type=int,
help="number of dataloader workers",
default=2)
self.parser.add_argument("--verbose",
type=int,
choices=[0, 1],
default=0,
help="If set, script will be more talkative...")
self.parser.add_argument("--deterministic",
type=int,
choices=[0, 1],
default=1,
help="If set, training is performed deterministic")
self.parser.add_argument("--global_seed",
type=int,
help="Set seed for initialization",
default=1234)
self.parser.add_argument("--worker_seed",
type=int,
help="Set seed for initialization",
default=3333)
# DATA options
self.parser.add_argument("--dataset",
type=str,
help="dataset to use",
choices=['kitti_2015', 'cityscapes'],
default='cityscapes')
self.parser.add_argument("--height",
type=int,
help="input image height",
default=1024)
self.parser.add_argument("--width",
type=int,
help="input image width",
default=2048)
self.parser.add_argument("--crop_height",
type=int,
help="input crop image height",
default=768)
self.parser.add_argument("--crop_width",
type=int,
help="input crop image width",
default=768)
self.parser.add_argument("--trainvaltest_split",
type=str,
help="Which split of dataset to load",
choices=['train','validation','test'],
default='train')
self.parser.add_argument("--zeromean",
type=int,
choices=[0, 1],
default=0,
help="Input data is normalized to zero mean/var")
# MODEL options
self.parser.add_argument("--model_name",
type=str,
help="the name of the model architecture this network is based on",
default="SwiftNet")
self.parser.add_argument("--encoder",
type=str,
help="select backbone",
default="resnet18")
self.parser.add_argument("--atrous",
type=int,
help="Enables the atrous convolution",
default=1)
# TRAINING options
self.parser.add_argument("--classweighing",
help="If set, use classweighing for dataset",
type=int,
choices=[0,1],
default=0)
self.parser.add_argument("--LRscheduler",
type=str,
help="specify LR scheduler",
default="CosineAnnealing",
choices=["CosineAnnealing", "Poly", "Polynomial"])
self.parser.add_argument("--warmup",
help="Enable warm up training",
type=int,
default=0)
self.parser.add_argument("--warmup_iters",
help="Num of epochs where warmup lasts.",
type=int,
default=0)
self.parser.add_argument("--warmup_lr_init",
help="Initial learning rate on warming up.",
type=float,
default=0.0)
# OPTIMIZATION options
self.parser.add_argument("--optimizer",
type=str,
help="optimizer for training",
default="Adam",
choices=["Adam", "SGD", "AdamW"])
self.parser.add_argument("--batch_size_train",
type=int,
help="batch size for training",
default=12)
self.parser.add_argument("--learning_rate_fine_tune",
type=float,
help="initial learning rate for fine_tune params",
default=1e-4)
self.parser.add_argument("--learning_rate_random_init",
type=float,
help="initial learning rate for random_init params",
default=4e-4)
self.parser.add_argument("--eta_min_fine_tune",
type=float,
help="eta min for fine_tune params",
default=1e-7)
self.parser.add_argument("--eta_min_random_init",
type=float,
help="eta min for random init params",
default=1e-6)
self.parser.add_argument("--num_epochs",
type=int,
help="number of epochs for training",
default=200)
self.parser.add_argument("--momentum",
type=float,
help="SGD momentum",
default=0.9)
self.parser.add_argument("--wd_fine_tune",
type=float,
help="weighting factor of weight decay",
default=0.25e-4) # 1e-4 if ImageNet pretrained weights are NOT used
self.parser.add_argument("--wd_random_init",
type=float,
help="weighting factor of weight decay",
default=1e-4)
self.parser.add_argument("--group_params",
type=int,
help="Group parameters based on weight decay apply.",
default=0)
# EVALUATION options
self.parser.add_argument("--val_frequency",
type=int,
help="number of epochs between each validation. For standalone, any number > 0 will produce an output",
default=1)
self.parser.add_argument("--batch_size_val",
type=int,
help="batch size for validation",
default=1)
# SAVE & LOGGING options
self.parser.add_argument("--savedir",
type=str,
default='')
self.parser.add_argument("--save_frequency",
type=int,
help="number of epochs between each save",
default=10)
def parse(self):
self.options = self.parser.parse_args()
return self.options
class TrainRecOptions(TrainOptions):
def __init__(self):
super().__init__()
# OPTIMIZATION options
self.parser.add_argument("--rec_decoder",
type=str,
help="Choose the reconstruction decoder for SwiftNetRec",
default="swiftnet")
# choices=SUPPORTED_REC_DECODER)
# Lateral connection between encoder and decoder (ResNet only)
self.parser.add_argument("--lateral",
type=int,
help="Whether to connect encoder to decoder (ResNet only)",
choices=[0, 1],
default=0)
self.parser.add_argument("--load_model_state_name",
type=str,
help="name of model state checkpoint to load",
default=None)
self.parser.add_argument("--decoder_init",
type=str,
help="Initialization method for training the reconstruction decoder.",
default='random')
self.parser.add_argument("--delta_d",
type=int,
help="Handler to freeze certain reconstruction decoder layer. See more on ...",
default=0)
self.parser.add_argument("--route",
type=str,
help="Freezing route",
default="lat_bw")
self.parser.add_argument('--idlc',
type=int,
nargs='+',
default=[0,0,0],
help="Decoder connection gate.")