-
Notifications
You must be signed in to change notification settings - Fork 5
/
main.py
150 lines (121 loc) · 5.19 KB
/
main.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
import os
import sys
import argparse
import logging
import torch
import torch.nn as nn
import torch.optim as optim
from config import get_cfg_defaults
from utils import valid_model, test_model, get_model, train_clcc
from datasets import get_dataset
from lr_scheduler import LR_Scheduler
from helpers import setup_determinism
from losses import get_loss
import warnings
warnings.filterwarnings("ignore")
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument("--config", type=str, default="",
help="config yaml path")
parser.add_argument("--load", type=str, default="",
help="path to model weight")
parser.add_argument("--valid", action="store_true",
help="enable evaluation mode for validation")
parser.add_argument("--test", action="store_true",
help="enable evaluation mode for testset")
parser.add_argument("-m", "--mode", type=str, default="train",
help="model runing mode (train/valid/test)")
args = parser.parse_args()
if args.valid:
args.mode = "valid"
elif args.test:
args.mode = "test"
return args
def setup_logging(args, cfg):
if not os.path.isdir(cfg.DIRS.LOGS):
os.mkdir(cfg.DIRS.LOGS)
head = '{asctime}:{levelname}: {message}'
handlers = [logging.StreamHandler(sys.stderr), logging.FileHandler(
os.path.join(cfg.DIRS.LOGS, f'{cfg.EXP}_{cfg.MODEL.NAME}_{args.mode}_fold{cfg.TRAIN.FOLD}.log'),
mode='a')]
logging.basicConfig(format=head, style='{', level=logging.DEBUG, handlers=handlers)
logging.info(f'===============================')
logging.info(f'\n\nStart with config {cfg}')
logging.info(f'Command arguments {args}')
def main(args, cfg):
logging.info(f"=========> {cfg.EXP} <=========")
# Declare variables
start_epoch = 0
best_metric = 0.
# Create model
model = get_model(cfg)
if cfg.MODEL.WEIGHT != "":
weight = cfg.MODEL.WEIGHT
model.load_state_dict(torch.load(weight)["state_dict"], strict=False)
# Define Loss and Optimizer
train_criterion = get_loss(cfg)
valid_criterion = get_loss(cfg)
if cfg.SYSTEM.CUDA:
model = model.cuda()
for i in range(len(train_criterion)):
train_criterion[i] = train_criterion[i].cuda()
for i in range(len(valid_criterion)):
valid_criterion[i] = valid_criterion[i].cuda()
# #optimizer
if cfg.OPT.OPTIMIZER == "adamw":
optimizer = optim.AdamW(params=model.parameters(),
lr=cfg.OPT.BASE_LR,
weight_decay=cfg.OPT.WEIGHT_DECAY)
elif cfg.OPT.OPTIMIZER == "adam":
optimizer = optim.Adam(params=model.parameters(),
lr=cfg.OPT.BASE_LR,
weight_decay=cfg.OPT.WEIGHT_DECAY)
elif cfg.OPT.OPTIMIZER == "sgd":
optimizer = optim.SGD(params=model.parameters(),
lr=cfg.OPT.BASE_LR,
weight_decay=cfg.OPT.WEIGHT_DECAY)
else:
raise Exception('OPT.OPTIMIZER should in ["adamw", "adam", "sgd"]')
# Load checkpoint
if args.load != "":
if os.path.isfile(args.load):
print(f"=> loading checkpoint {args.load}")
ckpt = torch.load(args.load, "cpu")
model.load_state_dict(ckpt.pop('state_dict'))
print("resuming optimizer ...")
optimizer.load_state_dict(ckpt.pop('optimizer'))
start_epoch, best_metric = ckpt['epoch'], ckpt['best_metric']
print('start, best_metric =', start_epoch, best_metric)
logging.info(
f"=> loaded checkpoint '{args.load}' (epoch {ckpt['epoch']}, best_metric: {ckpt['best_metric']})")
else:
logging.info(f"=> no checkpoint found at '{args.load}'")
if cfg.SYSTEM.MULTI_GPU:
model = nn.DataParallel(model)
# Load data
train_loader = get_dataset('train', cfg, trainsize=cfg.DATA.SIZE)
valid_loader = get_dataset('valid', cfg, trainsize=cfg.DATA.SIZE)
test_loader = get_dataset('test', cfg, trainsize=cfg.DATA.SIZE)
# Load scheduler
scheduler = LR_Scheduler("cos", cfg.OPT.BASE_LR, cfg.TRAIN.EPOCHS, iters_per_epoch=len(train_loader),
warmup_epochs=cfg.OPT.WARMUP_EPOCHS)
if args.mode == "train":
train_dict = {'CLCC': train_clcc}
train_dict[cfg.TRAIN.METHOD](logging.info, cfg, model, train_loader, valid_loader, train_criterion,
valid_criterion, optimizer, scheduler, start_epoch, best_metric, test_loader)
elif args.mode == "valid":
valid_model(logging.info, cfg, model, valid_criterion, valid_loader)
else:
test_model(logging.info, cfg, model, test_loader, weight=cfg.MODEL.WEIGHT)
if __name__ == "__main__":
args = parse_args()
cfg = get_cfg_defaults()
if args.config != "":
cfg.merge_from_file(args.config)
cfg.freeze()
for _dir in ["WEIGHTS", "OUTPUTS"]:
if not os.path.isdir(cfg.DIRS[_dir]):
os.mkdir(cfg.DIRS[_dir])
setup_logging(args, cfg)
setup_determinism(cfg.SYSTEM.SEED)
main(args, cfg)