-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathtrain.py
177 lines (146 loc) · 8.44 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
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
##+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
## Toyota Technological Institute
## Author: Yuki Kondo
## Copyright (c) 2024
## yuki.kondo.ab@gmail.com
##
## This source code is licensed under the Apache License license found in the
## LICENSE file in the root directory of this source tree
##+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
import numpy as np
import argparse
import os
import random
import shutil
import datetime
import socket
import torch
from torch.utils.data import DataLoader
from torch.utils.data.sampler import SequentialSampler, BatchSampler
from torch.optim.lr_scheduler import LambdaLR
from model.utils.sync_batchnorm import convert_model
from model.config import cfg
from model.engine.trainer import do_train, do_pretrain_sr
import torchvision.transforms as transforms
from model.data.transforms.data_preprocess import TrainTransforms, TestTransforms
from model.data.transforms.transforms import FactorResize
from model.modeling.build_model import JointModelWithLoss, JointInvModelWithLoss, SRModelWithLoss
from model.data.crack_dataset import CrackDataSet, SRPretrainDataSet
from model.data.retinal_dataset import RetinalDataSet
from model.utils.misc import str2bool, fix_model_state_dict
from model.data import samplers
from model.utils.lr_scheduler import WarmupMultiStepLR, UpDownScheduler
from torch.multiprocessing import Pool, Process, set_start_method
def train(args, cfg):
device = torch.device(cfg.DEVICE)
print('Loading Datasets...')
train_transforms = TrainTransforms(cfg)
sr_transforms = FactorResize(cfg.MODEL.SCALE_FACTOR, cfg.SOLVER.DOWNSCALE_INTERPOLATION)
if cfg.DATASET.ONLY_IMAGES:
trainval_dataset = SRPretrainDataSet(cfg, cfg.DATASET.TRAIN_IMAGE_DIR, transforms=train_transforms, sr_transforms=sr_transforms)
elif "RetinalSeg" in cfg.DATASET.TRAIN_IMAGE_DIR:
trainval_dataset = RetinalDataSet(cfg, cfg.DATASET.TRAIN_IMAGE_DIR, cfg.DATASET.TRAIN_MASK_DIR, transforms=train_transforms, sr_transforms=sr_transforms)
else:
trainval_dataset = CrackDataSet(cfg, cfg.DATASET.TRAIN_IMAGE_DIR, cfg.DATASET.TRAIN_MASK_DIR, transforms=train_transforms, sr_transforms=sr_transforms)
n_samples = len(trainval_dataset)
train_size = int(len(trainval_dataset) * cfg.SOLVER.TRAIN_DATASET_RATIO)
val_size = n_samples - train_size
if "RetinalSeg" in cfg.DATASET.TRAIN_IMAGE_DIR:
train_size, val_size = 12, 3
print(f"Train dataset size: {train_size}, Validation dataset size: {val_size}")
train_dataset, val_dataset = torch.utils.data.random_split(trainval_dataset, [train_size, val_size])
# print(val_dataset.__dict__)
sampler = torch.utils.data.RandomSampler(train_dataset)
batch_sampler = torch.utils.data.sampler.BatchSampler(sampler=sampler, batch_size=cfg.SOLVER.BATCH_SIZE, drop_last=False) #True)
batch_sampler = samplers.IterationBasedBatchSampler(batch_sampler, num_iterations=cfg.SOLVER.MAX_ITER)
train_loader = DataLoader(train_dataset, num_workers=args.num_workers, batch_sampler=batch_sampler, pin_memory=True)
eval_sampler = SequentialSampler(val_dataset)
eval_batch_sampler = BatchSampler(sampler=eval_sampler, batch_size=cfg.SOLVER.BATCH_SIZE, drop_last=False) # True)
eval_loader = DataLoader(val_dataset, num_workers=1, batch_sampler=eval_batch_sampler, pin_memory=False) # num_workers = 1 for memory retrenchment
print('Building model...')
if cfg.DATASET.ONLY_IMAGES and cfg.MODEL.SR != "DSRL":
# print(SRModelWithLoss.__mro__)
model = SRModelWithLoss(cfg, sr_transforms).to(device)
# print(f'------------Model Architecture-------------\n\n<Network SR>\n{model.sr_model}\n')
scheduler_flag = False # for KBPN paper
else:
scheduler_flag = cfg.SOLVER.SCHEDULER
if cfg.MODEL.SR_SEG_INV:
model = JonitInvModelWithLoss(cfg, num_train_ds=train_size, resume_iter=args.resume_iter, sr_transforms=sr_transforms).to(device)
# print(f'------------Model Architecture-------------\n\n<Network SS>\n{model.segmentation_model}\n\n<Network SR>\n{model.sr_model}')
else:
model = JointModelWithLoss(cfg, num_train_ds=train_size, resume_iter=args.resume_iter, sr_transforms=sr_transforms).to(device)
pass
if cfg.MODEL.SR == "DSRL":
# print(f'------------Model Architecture-------------\n{model}')
pass
else:
# print(f'------------Model Architecture-------------\n\n<Network SR>\n{model.sr_model}\n\n<Network SS>\n{model.segmentation_model}')
pass
if cfg.MODEL.OPTIMIZER == "Adam":
optimizer = torch.optim.Adam(model.parameters(), lr=cfg.SOLVER.LR, betas=(0.9, 0.999), eps=1e-8) # betas and eps is referenced from KBPN.
elif cfg.MODEL.OPTIMIZER == "SGD":
optimizer = torch.optim.SGD(filter(lambda p:p.requires_grad, model.parameters()), lr=cfg.SOLVER.LR, momentum=0.9, weight_decay=5e-4)
sch_func = UpDownScheduler(cfg.SOLVER.SR_PRETRAIN_ITER[1], args.resume_iter, scheduler_flag)
scheduler = LambdaLR(optimizer, lr_lambda = sch_func)
if args.resume_iter > 0:
print('Resume from {}'.format(os.path.join(cfg.OUTPUT_DIR, 'model', 'iteration_{}.pth'.format(args.resume_iter))))
state_dict = fix_model_state_dict(torch.load(os.path.join(cfg.OUTPUT_DIR, 'model', 'iteration_{}.pth'.format(args.resume_iter))))
model.load_state_dict(state_dict , strict=False)
if cfg.SOLVER.SYNC_BATCHNORM:
model = convert_model(model).to(device)
if args.num_gpus > 1:
device_ids = list(range(args.num_gpus))
# device_ids.insert(0, device_ids.pop(cfg.DEVICE_NUM))
print("device_ids:",device_ids)
model = torch.nn.DataParallel(model, device_ids=device_ids) # primaly gpu is last device.
if cfg.DATASET.ONLY_IMAGES:
do_pretrain_sr(args, cfg, model, optimizer, scheduler, train_loader, eval_loader)
else:
do_train(args, cfg, model, optimizer, scheduler, train_loader, eval_loader)
def main():
parser = argparse.ArgumentParser(description='Crack Segmentation with Blind Super Resolution(CSBSR)')
parser.add_argument('--config_file', type=str, default='./config/configs_train.yaml', metavar='FILE', help='path to config file')
parser.add_argument('--output_dirname', type=str, default='', help='')
parser.add_argument('--num_workers', type=int, default=2, help='')
parser.add_argument('--log_step', type=int, default=50, help='')
parser.add_argument('--save_step', type=int, default=2000)
parser.add_argument('--eval_step', type=int, default=2000)
parser.add_argument('--num_gpus', type=int, default=1)
parser.add_argument('--mixed_precision', type=str2bool, default=False)
parser.add_argument('--wandb_flag', type=str2bool, default=True)
parser.add_argument('--resume_iter', type=int, default=0)
parser.add_argument('--debug', type=bool, default=False)
parser.add_argument('--wandb_prj_name', type=str, default="CSBSR") # CSBSR_main_ECCV2022, KBPN_verkondo_iso
# only local project
parser.add_argument('--local', type=bool, default=False)
args = parser.parse_args()
torch.manual_seed(cfg.SEED)
random.seed(cfg.SEED)
np.random.seed(cfg.SEED)
cuda = torch.cuda.is_available()
if cuda:
torch.backends.cudnn.benchmark = True
torch.backends.cudnn.deterministic = True
torch.cuda.manual_seed(cfg.SEED)
if len(args.config_file) > 0:
print('Configration file is loaded from {}'.format(args.config_file))
cfg.merge_from_file(args.config_file)
if "_ds_" in cfg.DATASET.TRAIN_IMAGE_DIR:
cfg.INPUT.IMAGE_SIZE = int(cfg.INPUT.IMAGE_SIZE / cfg.MODEL.SCALE_FACTOR )
if args.local:
train_img_dir_list = (cfg.DATASET.TRAIN_IMAGE_DIR).split('/')
train_img_dir_list[1] = train_img_dir_list[1] + f'_{socket.gethostname()}'
cfg.DATASET.TRAIN_IMAGE_DIR = '/'.join(train_img_dir_list)
train_msk_dir_list = (cfg.DATASET.TRAIN_MASK_DIR).split('/')
train_msk_dir_list[1] = train_msk_dir_list[1] + f'_{socket.gethostname()}'
cfg.DATASET.TRAIN_MASK_DIR = '/'.join(train_msk_dir_list)
cfg.freeze()
if not args.debug and args.resume_iter == 0:
os.makedirs(cfg.OUTPUT_DIR, exist_ok=True)
shutil.copy2(args.config_file, os.path.join(cfg.OUTPUT_DIR, 'config.yaml'))
train(args, cfg)
if __name__ == '__main__':
# os.environ['WANDB_MODE'] = 'offline'
set_start_method('spawn')
main()