-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_UG2L.py
134 lines (110 loc) · 4.65 KB
/
train_UG2L.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
#!/usr/bin/python3
#coding=utf-8
from misc import AvgMeter, check_mkdir
import sys
import datetime
sys.path.insert(0, '../')
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
from torch import optim
from torch.optim.lr_scheduler import _LRScheduler
from timm.scheduler import create_scheduler
from timm.optim import create_optimizer
from src.UG2L import net as Net
from ORSI_SOD_dataset import ORSI_SOD_Dataset
from parser import parser_args
from evaluator import *
args = {
'iter_num': 7500,
'epoch': 300,
'train_batch_size': 8,
'last_iter': 0,
'snapshot': ''
}
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
torch.manual_seed(2022)
"""
smaps : BCE + wIOU
edges: BCE
"""
def structure_loss(pred, mask):
#mask = mask.detach()
wbce = F.binary_cross_entropy_with_logits(pred, mask)
pred = torch.sigmoid(pred)
inter = (pred*mask).sum(dim=(2,3))
union = (pred+mask).sum(dim=(2,3))
wiou = 1-(inter+1)/(union-inter+1)
return wbce.mean()+wiou.mean()#
"""
define calculate uncertainty loss
"""
def cal_usl(seg_logits, seg_gts):
assert seg_logits.shape == seg_gts.shape, (seg_logits.shape, seg_gts.shape)
pred = seg_logits.sigmoid()
#np.log(2 - abs((2*i-1.0))) / np.log(2)
ucl = torch.log(2 - (2 * pred - 1).abs())
return ucl.mean() / torch.log(torch.Tensor([2])).cuda()
"""
perform cosine strategy
"""
def get_coef(iter_percentage, method='cos'):
coef_range = (0, 1)
min_coef, max_coef = min(coef_range), max(coef_range)
normalized_coef = 1 + np.cos(iter_percentage * np.pi)
ual_coef = 1/2 * normalized_coef * (max_coef - min_coef) + min_coef
return ual_coef
def main(dataset):
#define dataset
data_path = '/data/iopen/lyf/SaliencyOD_in_RSIs/'+ dataset +' dataset/'
train_dataset = ORSI_SOD_Dataset(root = data_path, mode = 'train', aug = True)
train_loader = DataLoader(train_dataset, batch_size = args['train_batch_size'], shuffle = True, num_workers = 8)
test_set = ORSI_SOD_Dataset(root = data_path, mode = "test", aug = False)
test_loader = DataLoader(test_set, batch_size = 1, shuffle = True, num_workers = 1)
args['iter_num'] = args['epoch'] * len(train_loader)
#############################ResNet pretrained###########################
#res18[2,2,2,2],res34[3,4,6,3],res50[3,4,6,3],res101[3,4,23,3],res152[3,8,36,3]
model = Net()
##############################Optim setting###############################
net = model.cuda().train()
optimizer = create_optimizer(parser_args, net)
lr_scheduler, _ = create_scheduler(parser_args, optimizer)
#########################################################################
curr_iter = args['last_iter']
for epoch in range(0, args['epoch']):
loss_sod_record, loss_ucl_record, total_loss_record = AvgMeter(), AvgMeter(), AvgMeter()
net.train()
for i, data in enumerate(train_loader):
# data\binarizing\Variable
inputs, labels, _, _ = data
batch_size = inputs.size(0)
inputs = inputs.type(torch.FloatTensor).cuda()
labels = labels.type(torch.FloatTensor).cuda()
optimizer.zero_grad()
smap1, smap2, smap3, smap4, smap5 = net(inputs)
##########loss#############
loss1_1 = structure_loss(smap1, labels)
loss1_2 = structure_loss(smap2, labels)
loss1_3 = structure_loss(smap3, labels)
loss1_4 = structure_loss(smap4, labels)
loss1_5 = structure_loss(smap5, labels)
loss_sod = loss1_1 + loss1_2 + (loss1_3 / 2) + (loss1_4 / 4) + (loss1_5 / 8)
loss_ucl = cal_usl(smap1, labels) + cal_usl(smap2, labels) + (cal_usl(smap3, labels) / 2) + (cal_usl(smap4, labels)/ 4) + (cal_usl(smap5, labels) / 8)
loss_ucl = get_coef(iter_percentage= curr_iter/args['iter_num']) * loss_ucl
total_loss = loss_sod + loss_ucl
total_loss.backward()
optimizer.step()
lr_scheduler.step(epoch)
loss_sod_record.update(loss_sod.item(), batch_size)
loss_ucl_record.update(loss_ucl.item(), batch_size)
total_loss_record.update(total_loss.item(), batch_size)
#############log###############
curr_iter += 1
log = '[epoch: %03d] [iter: %05d] [total loss %.5f] [sod loss %.5f] [ucl loss %.5f] [lr %.13f] ' % \
(epoch, curr_iter, total_loss_record.avg, loss_sod_record.avg, loss_ucl_record.avg, optimizer.param_groups[0]['lr'])
print(log)
if __name__ == '__main__':
dataset = "ORS_4199"
main(dataset)