forked from qianyu-dlut/MVANet
-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain.py
189 lines (166 loc) · 6.71 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
179
180
181
182
183
184
185
186
187
188
189
import argparse
import os
from datetime import datetime
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from torch.backends import cudnn
from torch.cuda import amp
from torch.utils.tensorboard import SummaryWriter
from torchvision import transforms
from mvanet.model import MVANet
from mvanet.utils.dataset_strategy_fpn import get_loader
from mvanet.utils.misc import AvgMeter, adjust_lr
writer = SummaryWriter()
cudnn.benchmark = True
parser = argparse.ArgumentParser()
parser.add_argument("--epoch", type=int, default=80, help="epoch number")
parser.add_argument("--lr_gen", type=float, default=1e-5, help="learning rate")
parser.add_argument("--batchsize", type=int, default=1, help="training batch size")
parser.add_argument("--trainsize", type=int, default=1024, help="training dataset size")
parser.add_argument(
"--decay_rate", type=float, default=0.9, help="decay rate of learning rate"
)
parser.add_argument(
"--decay_epoch", type=int, default=80, help="every n epochs decay learning rate"
)
opt = parser.parse_args()
print("Generator Learning Rate: {}".format(opt.lr_gen))
# build models
if hasattr(torch.cuda, "empty_cache"):
torch.cuda.empty_cache()
generator = MVANet()
generator.cuda()
generator_params = generator.parameters()
generator_optimizer = torch.optim.Adam(generator_params, opt.lr_gen)
image_root = "./data/DIS5K/DIS-TR/images/"
gt_root = "./data/DIS5K/DIS-TR/masks/"
train_loader = get_loader(
image_root, gt_root, batchsize=opt.batchsize, trainsize=opt.trainsize
)
total_step = len(train_loader)
to_pil = transforms.ToPILImage()
## define loss
CE = torch.nn.BCELoss()
mse_loss = torch.nn.MSELoss(size_average=True, reduce=True)
size_rates = [1]
criterion = nn.BCEWithLogitsLoss().cuda()
criterion_mae = nn.L1Loss().cuda()
criterion_mse = nn.MSELoss().cuda()
use_fp16 = True
scaler = amp.GradScaler(enabled=use_fp16)
def structure_loss(pred, mask):
weit = 1 + 5 * torch.abs(
F.avg_pool2d(mask, kernel_size=31, stride=1, padding=15) - mask
)
wbce = F.binary_cross_entropy_with_logits(pred, mask, reduction="none")
wbce = (weit * wbce).sum(dim=(2, 3)) / weit.sum(dim=(2, 3))
pred = torch.sigmoid(pred)
inter = ((pred * mask) * weit).sum(dim=(2, 3))
union = ((pred + mask) * weit).sum(dim=(2, 3))
wiou = 1 - (inter + 1) / (union - inter + 1)
return (wbce + wiou).mean()
for epoch in range(1, opt.epoch + 1):
torch.cuda.empty_cache()
generator.train()
loss_record = AvgMeter()
print(
"Generator Learning Rate: {}".format(generator_optimizer.param_groups[0]["lr"])
)
for i, pack in enumerate(train_loader, start=1):
torch.cuda.empty_cache()
for rate in size_rates:
torch.cuda.empty_cache()
generator_optimizer.zero_grad()
images, gts = pack
images = Variable(images)
gts = Variable(gts)
images = images.cuda()
gts = gts.cuda()
trainsize = int(round(opt.trainsize * rate / 32) * 32)
if rate != 1:
images = F.upsample(
images,
size=(trainsize, trainsize),
mode="bilinear",
align_corners=True,
)
gts = F.upsample(
gts,
size=(trainsize, trainsize),
mode="bilinear",
align_corners=True,
)
b, c, h, w = gts.size()
target_1 = F.upsample(gts, size=h // 4, mode="nearest")
target_2 = F.upsample(gts, size=h // 8, mode="nearest").cuda()
target_3 = F.upsample(gts, size=h // 16, mode="nearest").cuda()
target_4 = F.upsample(gts, size=h // 32, mode="nearest").cuda()
target_5 = F.upsample(gts, size=h // 64, mode="nearest").cuda()
with amp.autocast(enabled=use_fp16):
(
sideout5,
sideout4,
sideout3,
sideout2,
sideout1,
final,
glb5,
glb4,
glb3,
glb2,
glb1,
tokenattmap4,
tokenattmap3,
tokenattmap2,
tokenattmap1,
) = generator.forward(images)
loss1 = structure_loss(sideout5, target_4)
loss2 = structure_loss(sideout4, target_3)
loss3 = structure_loss(sideout3, target_2)
loss4 = structure_loss(sideout2, target_1)
loss5 = structure_loss(sideout1, target_1)
loss6 = structure_loss(final, gts)
loss7 = structure_loss(glb5, target_5)
loss8 = structure_loss(glb4, target_4)
loss9 = structure_loss(glb3, target_3)
loss10 = structure_loss(glb2, target_2)
loss11 = structure_loss(glb1, target_2)
loss12 = structure_loss(tokenattmap4, target_3)
loss13 = structure_loss(tokenattmap3, target_2)
loss14 = structure_loss(tokenattmap2, target_1)
loss15 = structure_loss(tokenattmap1, target_1)
loss = (
loss1
+ loss2
+ loss3
+ loss4
+ loss5
+ loss6
+ 0.3 * (loss7 + loss8 + loss9 + loss10 + loss11)
+ 0.3 * (loss12 + loss13 + loss14 + loss15)
)
Loss_loc = loss1 + loss2 + loss3 + loss4 + loss5 + loss6
Loss_glb = loss7 + loss8 + loss9 + loss10 + loss11
Loss_map = loss12 + loss13 + loss14 + loss15
writer.add_scalar("loss", loss.item(), epoch * len(train_loader) + i)
generator_optimizer.zero_grad()
scaler.scale(loss).backward()
scaler.step(generator_optimizer)
scaler.update()
if rate == 1:
loss_record.update(loss.data, opt.batchsize)
if i % 10 == 0 or i == total_step:
print(
"{} Epoch [{:03d}/{:03d}], Step [{:04d}/{:04d}], gen Loss: {:.4f}".format(
datetime.now(), epoch, opt.epoch, i, total_step, loss_record.show()
)
)
adjust_lr(generator_optimizer, opt.lr_gen, epoch, opt.decay_rate, opt.decay_epoch)
# save checkpoints every 20 epochs
if epoch % 20 == 0:
save_path = "./saved_model/MVANet/"
if not os.path.exists(save_path):
os.mkdir(save_path)
torch.save(generator.state_dict(), save_path + "Model" + "_%d" % epoch + ".pth")