-
Notifications
You must be signed in to change notification settings - Fork 9
/
train.py
144 lines (124 loc) · 5.87 KB
/
train.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
import argparse
import os
import os.path as osp
import pprint
import random
import warnings
import numpy as np
import yaml
import torch
from torch.utils import data
from advent.model.deeplabv2 import get_deeplab_v2
from advent.dataset.gta5 import GTA5DataSet
from advent.dataset.cityscapes import CityscapesDataSet
from advent.domain_adaptation.config import cfg, cfg_from_file
from advent.domain_adaptation.train_UDA import train_domain_adaptation
warnings.filterwarnings("ignore", message="numpy.dtype size changed")
warnings.filterwarnings("ignore")
def get_arguments():
"""
Parse input arguments
"""
parser = argparse.ArgumentParser(description="Code for domain adaptation (DA) training")
parser.add_argument('--cfg', type=str, default=None,
help='optional config file', )
parser.add_argument("--random-train", action="store_true",
help="not fixing random seed.")
parser.add_argument("--tensorboard", action="store_true",
help="visualize training loss with tensorboardX.")
parser.add_argument("--viz-every-iter", type=int, default=None,
help="visualize results.")
parser.add_argument("--exp-suffix", type=str, default=None,
help="optional experiment suffix")
return parser.parse_args()
def main():
# LOAD ARGS
args = get_arguments()
print('Called with args:')
print(args)
assert args.cfg is not None, 'Missing cfg file'
cfg_from_file(args.cfg)
# auto-generate exp name if not specified
if cfg.EXP_NAME == '':
cfg.EXP_NAME = f'{cfg.SOURCE}2{cfg.TARGET}_{cfg.TRAIN.MODEL}_{cfg.TRAIN.DA_METHOD}'
if args.exp_suffix:
cfg.EXP_NAME += f'_{args.exp_suffix}'
# auto-generate snapshot path if not specified
if cfg.TRAIN.SNAPSHOT_DIR == '':
cfg.TRAIN.SNAPSHOT_DIR = osp.join(cfg.EXP_ROOT_SNAPSHOT, cfg.EXP_NAME)
os.makedirs(cfg.TRAIN.SNAPSHOT_DIR, exist_ok=True)
# tensorboard
if args.tensorboard:
if cfg.TRAIN.TENSORBOARD_LOGDIR == '':
cfg.TRAIN.TENSORBOARD_LOGDIR = osp.join(cfg.EXP_ROOT_LOGS, 'tensorboard', cfg.EXP_NAME)
os.makedirs(cfg.TRAIN.TENSORBOARD_LOGDIR, exist_ok=True)
if args.viz_every_iter is not None:
cfg.TRAIN.TENSORBOARD_VIZRATE = args.viz_every_iter
else:
cfg.TRAIN.TENSORBOARD_LOGDIR = ''
print('Using config:')
pprint.pprint(cfg)
# INIT
_init_fn = None
if not args.random_train:
torch.manual_seed(cfg.TRAIN.RANDOM_SEED)
torch.cuda.manual_seed(cfg.TRAIN.RANDOM_SEED)
np.random.seed(cfg.TRAIN.RANDOM_SEED)
random.seed(cfg.TRAIN.RANDOM_SEED)
def _init_fn(worker_id):
np.random.seed(cfg.TRAIN.RANDOM_SEED + worker_id)
if os.environ.get('ADVENT_DRY_RUN', '0') == '1':
return
# LOAD SEGMENTATION NET
assert osp.exists(cfg.TRAIN.RESTORE_FROM), f'Missing init model {cfg.TRAIN.RESTORE_FROM}'
if cfg.TRAIN.MODEL == 'DeepLabv2':
model = get_deeplab_v2(num_classes=cfg.NUM_CLASSES, multi_level=cfg.TRAIN.MULTI_LEVEL)
saved_state_dict = torch.load(cfg.TRAIN.RESTORE_FROM)
if 'DeepLab_resnet_pretrained_imagenet' in cfg.TRAIN.RESTORE_FROM:
new_params = model.state_dict().copy()
for i in saved_state_dict:
i_parts = i.split('.')
if not i_parts[1] == 'layer5':
new_params['.'.join(i_parts[1:])] = saved_state_dict[i]
model.load_state_dict(new_params)
else:
model.load_state_dict(saved_state_dict)
else:
raise NotImplementedError(f"Not yet supported {cfg.TRAIN.MODEL}")
print('Model loaded')
# DATALOADERS
source_dataset = GTA5DataSet(root=cfg.DATA_DIRECTORY_SOURCE,
list_path=cfg.DATA_LIST_SOURCE,
set=cfg.TRAIN.SET_SOURCE,
max_iters=cfg.TRAIN.MAX_ITERS * cfg.TRAIN.BATCH_SIZE_SOURCE,
crop_size=cfg.TRAIN.INPUT_SIZE_SOURCE,
mean=cfg.TRAIN.IMG_MEAN,
labels_size=(1280, 720)
)
source_loader = data.DataLoader(source_dataset,
batch_size=cfg.TRAIN.BATCH_SIZE_SOURCE,
num_workers=cfg.NUM_WORKERS,
shuffle=True,
pin_memory=True,
worker_init_fn=_init_fn)
#
target_dataset = CityscapesDataSet(root=cfg.DATA_DIRECTORY_TARGET,
list_path=cfg.DATA_LIST_TARGET,
set=cfg.TRAIN.SET_TARGET,
info_path=cfg.TRAIN.INFO_TARGET,
max_iters=cfg.TRAIN.MAX_ITERS * cfg.TRAIN.BATCH_SIZE_TARGET,
crop_size=cfg.TRAIN.INPUT_SIZE_TARGET,
mean=cfg.TRAIN.IMG_MEAN,
labels_size=(1024, 512))
target_loader = data.DataLoader(target_dataset,
batch_size=cfg.TRAIN.BATCH_SIZE_TARGET,
num_workers=cfg.NUM_WORKERS,
shuffle=False,
pin_memory=True,
worker_init_fn=_init_fn)
with open(osp.join(cfg.TRAIN.SNAPSHOT_DIR, 'train_cfg.yml'), 'w') as yaml_file:
yaml.dump(cfg, yaml_file, default_flow_style=False)
# UDA TRAINING
train_domain_adaptation(model, source_loader, target_loader, cfg, _init_fn)
if __name__ == '__main__':
main()