-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
178 lines (155 loc) · 8.63 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
176
177
178
import argparse
import time
from torch.autograd import Variable
from torch.utils.data import DataLoader
from net import Net
from dataset import *
import matplotlib.pyplot as plt
from metrics import *
import numpy as np
import os
os.environ['KMP_DUPLICATE_LIB_OK'] = 'TRUE'
parser = argparse.ArgumentParser(description="PyTorch BasicIRSTD train")
parser.add_argument("--model_names", default=['ACM', 'ALCNet'], nargs='+',
help="model_name: 'ACM', 'ALCNet', 'DNANet', 'ISNet', 'UIUNet', 'RDIAN', 'ISTDU-Net', 'U-Net', 'RISTDnet'")
parser.add_argument("--dataset_names", default=['NUAA-SIRST'], nargs='+',
help="dataset_name: 'NUAA-SIRST', 'NUDT-SIRST', 'IRSTD-1K', 'SIRST3', 'NUDT-SIRST-Sea', 'IRDST-real'")
parser.add_argument("--img_norm_cfg", default=None, type=dict,
help="specific a img_norm_cfg, default=None (using img_norm_cfg values of each dataset)")
parser.add_argument("--img_norm_cfg_mean", default=None, type=float,
help="specific a mean value img_norm_cfg, default=None (using img_norm_cfg values of each dataset)")
parser.add_argument("--img_norm_cfg_std", default=None, type=float,
help="specific a std value img_norm_cfg, default=None (using img_norm_cfg values of each dataset)")
parser.add_argument("--dataset_dir", default='./datasets', type=str, help="train_dataset_dir")
parser.add_argument("--batchSize", type=int, default=16, help="Training batch sizse")
parser.add_argument("--patchSize", type=int, default=256, help="Training patch size")
parser.add_argument("--save", default='./log', type=str, help="Save path of checkpoints")
parser.add_argument("--resume", default=None, nargs='+', help="Resume from exisiting checkpoints (default: None)")
parser.add_argument("--pretrained", default=None, nargs='+', help="Load pretrained checkpoints (default: None)")
parser.add_argument("--nEpochs", type=int, default=400, help="Number of epochs")
parser.add_argument("--optimizer_name", default='Adam', type=str, help="optimizer name: Adam, Adagrad, SGD")
parser.add_argument("--optimizer_settings", default={'lr': 5e-4}, type=dict, help="optimizer settings")
parser.add_argument("--scheduler_name", default='MultiStepLR', type=str, help="scheduler name: MultiStepLR")
parser.add_argument("--scheduler_settings", default={'step': [200, 300], 'gamma': 0.5}, type=dict, help="scheduler settings")
parser.add_argument("--threads", type=int, default=1, help="Number of threads for data loader to use")
parser.add_argument("--threshold", type=float, default=0.5, help="Threshold for test")
parser.add_argument("--seed", type=int, default=42, help="Threshold for test")
global opt
opt = parser.parse_args()
## Set img_norm_cfg
if opt.img_norm_cfg_mean != None and opt.img_norm_cfg_std != None:
opt.img_norm_cfg = dict()
opt.img_norm_cfg['mean'] = opt.img_norm_cfg_mean
opt.img_norm_cfg['std'] = opt.img_norm_cfg_std
seed_pytorch(opt.seed)
def train():
train_set = TrainSetLoader(dataset_dir=opt.dataset_dir, dataset_name=opt.dataset_name, patch_size=opt.patchSize, img_norm_cfg=opt.img_norm_cfg)
train_loader = DataLoader(dataset=train_set, num_workers=opt.threads, batch_size=opt.batchSize, shuffle=True)
net = Net(model_name=opt.model_name, mode='train').cuda()
net.train()
epoch_state = 0
total_loss_list = []
total_loss_epoch = []
if opt.resume:
for resume_pth in opt.resume:
if opt.dataset_name in resume_pth and opt.model_name in resume_pth:
ckpt = torch.load(resume_pth)
net.load_state_dict(ckpt['state_dict'])
epoch_state = ckpt['epoch']
total_loss_list = ckpt['total_loss']
for i in range(len(opt.scheduler_settings['step'])):
opt.scheduler_settings['step'][i] = opt.scheduler_settings['step'][i] - ckpt['epoch']
if opt.pretrained:
for pretrained_pth in opt.pretrained:
if opt.dataset_name in pretrained_pth and opt.model_name in pretrained_pth:
ckpt = torch.load(resume_pth)
net.load_state_dict(ckpt['state_dict'])
### Default settings
if opt.optimizer_name == 'Adam':
opt.optimizer_settings = {'lr': 5e-4}
opt.scheduler_name = 'MultiStepLR'
opt.scheduler_settings = {'epochs':400, 'step': [200, 300], 'gamma': 0.1}
opt.scheduler_settings['epochs'] = opt.nEpochs
### Default settings of DNANet
if opt.optimizer_name == 'Adagrad':
opt.optimizer_settings = {'lr': 0.05}
opt.scheduler_name = 'CosineAnnealingLR'
opt.scheduler_settings = {'epochs':1500, 'min_lr':1e-5}
opt.scheduler_settings['epochs'] = opt.nEpochs
opt.nEpochs = opt.scheduler_settings['epochs']
net = torch.nn.DataParallel(net)
optimizer, scheduler = get_optimizer(net, opt.optimizer_name, opt.scheduler_name, opt.optimizer_settings, opt.scheduler_settings)
for idx_epoch in range(epoch_state, opt.nEpochs):
for idx_iter, (img, gt_mask) in enumerate(train_loader):
img, gt_mask = Variable(img).cuda(), Variable(gt_mask).cuda()
if img.shape[0] == 1:
continue
pred = net.forward(img)
loss = net.module.loss(pred, gt_mask)
total_loss_epoch.append(loss.detach().cpu())
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
if (idx_epoch + 1) % 10 == 0:
total_loss_list.append(float(np.array(total_loss_epoch).mean()))
print(time.ctime()[4:-5] + ' Epoch---%d, total_loss---%f,'
% (idx_epoch + 1, total_loss_list[-1]))
opt.f.write(time.ctime()[4:-5] + ' Epoch---%d, total_loss---%f,\n'
% (idx_epoch + 1, total_loss_list[-1]))
total_loss_epoch = []
if (idx_epoch + 1) % 50 == 0:
save_pth = opt.save + '/' + opt.dataset_name + '/' + opt.model_name + '_' + str(idx_epoch + 1) + '.pth.tar'
save_checkpoint({
'epoch': idx_epoch + 1,
'state_dict': net.module.state_dict(),
'total_loss': total_loss_list,
}, save_pth)
test(save_pth)
if (idx_epoch + 1) == opt.nEpochs and (idx_epoch + 1) % 50 != 0:
save_pth = opt.save + '/' + opt.dataset_name + '/' + opt.model_name + '_' + str(idx_epoch + 1) + '.pth.tar'
save_checkpoint({
'epoch': idx_epoch + 1,
'state_dict': net.module.state_dict(),
'total_loss': total_loss_list,
}, save_pth)
test(save_pth)
def test(save_pth):
test_set = TestSetLoader(opt.dataset_dir, opt.dataset_name, opt.dataset_name, img_norm_cfg=opt.img_norm_cfg)
test_loader = DataLoader(dataset=test_set, num_workers=1, batch_size=1, shuffle=False)
net = Net(model_name=opt.model_name, mode='test').cuda()
ckpt = torch.load(save_pth)
net.load_state_dict(ckpt['state_dict'])
net.eval()
eval_mIoU = mIoU()
eval_PD_FA = PD_FA()
for idx_iter, (img, gt_mask, size, _) in enumerate(test_loader):
img = Variable(img).cuda()
pred = net.forward(img)
pred = pred[:,:,:size[0],:size[1]]
gt_mask = gt_mask[:,:,:size[0],:size[1]]
eval_mIoU.update((pred>opt.threshold).cpu(), gt_mask)
eval_PD_FA.update((pred[0,0,:,:]>opt.threshold).cpu(), gt_mask[0,0,:,:], size)
results1 = eval_mIoU.get()
results2 = eval_PD_FA.get()
print("pixAcc, mIoU:\t" + str(results1))
print("PD, FA:\t" + str(results2))
opt.f.write("pixAcc, mIoU:\t" + str(results1) + '\n')
opt.f.write("PD, FA:\t" + str(results2) + '\n')
def save_checkpoint(state, save_path):
if not os.path.exists(os.path.dirname(save_path)):
os.makedirs(os.path.dirname(save_path))
torch.save(state, save_path)
return save_path
if __name__ == '__main__':
for dataset_name in opt.dataset_names:
opt.dataset_name = dataset_name
for model_name in opt.model_names:
opt.model_name = model_name
if not os.path.exists(opt.save):
os.makedirs(opt.save)
opt.f = open(opt.save + '/' + opt.dataset_name + '_' + opt.model_name + '_' + (time.ctime()).replace(' ', '_').replace(':', '_') + '.txt', 'w')
print(opt.dataset_name + '\t' + opt.model_name)
train()
print('\n')
opt.f.close()