-
Notifications
You must be signed in to change notification settings - Fork 2
/
run.py
171 lines (152 loc) · 6.41 KB
/
run.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
from dataset import *
from res2net import *
from modules import *
import os
import torch
import numpy as np
import torch.utils.data as data
import torch.nn as nn
import torch.optim as optim
from tqdm import tqdm
import time
from evaluator import Eval_Fmeasure,Eval_mae
os.environ['CUDA_VISIBLE_DEVICES'] = '0, 1'
gpu_devices = list(np.arange(torch.cuda.device_count()))
multi_gpu = len(gpu_devices) > 1
output_folder = r'./Outputs/pred/AGNet/EORSSD/Test'
ckpt_folder = r'./Checkpoints'
dataset_root = r'../Dataset/EORSSD'
batch_size = 8
class FLoss(torch.nn.Module):
def __init__(self, beta=0.3, log_like=False):
super(FLoss, self).__init__()
self.beta = beta
self.log_like = log_like
def forward(self, prediction, target):
EPS = 1e-10
N = prediction.size(0)
TP = (prediction * target).view(N, -1).sum(dim=1)
H = self.beta * target.view(N, -1).sum(dim=1) + prediction.view(N, -1).sum(dim=1)
fmeasure = (1 + self.beta) * TP / (H + EPS)
if self.log_like:
floss = -torch.log(fmeasure)
else:
floss = (1 - fmeasure)
return floss.mean()
def iou(pred, mask):
inter = (pred * mask) .sum(dim=(2, 3))
union = (pred + mask) .sum(dim=(2, 3))
wiou = 1 - (inter + 1) / (union - inter + 1)
return wiou.mean()
class BCEloss(nn.Module):
def __init__(self):
super(BCEloss, self).__init__()
self.bce = nn.BCELoss()
self.floss=FLoss()
def forward(self, sm,se,label,edge):
mask_loss = self.bce(sm,label) + 0.6*iou(sm,label) +self.floss(sm,label)
edge_loss = self.bce(se, edge) + 0.6*iou(se,edge) +self.floss(se,edge)
total_loss = mask_loss+0.5*edge_loss
return [total_loss, mask_loss, mask_loss]
def count_param(model):
param_count = 0
for param in model.parameters():
param_count += param.view(-1).size()[0]
return param_count
class Run:
def __init__(self):
self.train_set = EORSSD(dataset_root, 'train', aug=True)
self.train_loader = data.DataLoader(self.train_set, shuffle=True, batch_size=batch_size, num_workers=4, drop_last=True) #true
self.test_set = EORSSD(dataset_root, 'test', aug=False)
self.test_loader = data.DataLoader(self.test_set, shuffle=False, batch_size=1, num_workers=4, drop_last=False)
self.init_lr = 1e-4
self.min_lr = 1e-5
self.train_epoch = 60
self.net = AGNet()
self.net.load_state_dict(torch.load(os.path.join(ckpt_folder, 'trained', 'eorssd.pth')))
self.loss=BCEloss()
def train(self):
self.net.train().cuda()
max_F,mF=0.80,0.85
base, head = [], []
for name, param in self.net.named_parameters():
if 'bkbone' in name:
base.append(param)
else:
head.append(param)
optimizer = optim.Adam([{'params': base}, {'params': head}], lr=self.init_lr, weight_decay=5e-4)
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=self.train_epoch,eta_min=self.min_lr)
for epc in range(1, self.train_epoch + 1):
records = [0] * 3
N = 0
optimizer.param_groups[0]['lr'] = 0.1 * optimizer.param_groups[1]['lr'] # for backbone
optimizer.param_groups[1]['lr'] = optimizer.param_groups[1]['lr']
for image, label, edge in tqdm(self.train_loader):
# prepare input data\n",
image, label, edge = image.cuda(), label.cuda(), edge.cuda()
B = image.size(0)
# forward\n",
optimizer.zero_grad()
sm,se= self.net(image)
losses_list = self.loss(sm,se,label,edge)
# compute loss\n",
total_loss = losses_list[0].mean()
# record loss\n",
N += B
for i in range(len(records)):
records[i] += losses_list[i].mean().item() * B
# backward\n",
total_loss.backward()
optimizer.step()
# update learning rate\n",
scheduler.step()
F,mf=self.test(epc)
if F>max_F:
cache_model(self.net, os.path.join(ckpt_folder, 'trained', 'trained.pth'), multi_gpu)
max_F=F
if mf>mF:
cache_model(self.net, os.path.join(ckpt_folder, 'trained', 'trained_m.pth'), multi_gpu)
mF=mf
if epc==self.train_epoch:
cache_model(self.net, os.path.join(ckpt_folder, 'trained', 'last.pth'), multi_gpu)
# print training information\n",
records = proc_loss(records, N, 4)
print('epoch: {} || total loss: {} || mask loss: {} || cur_mean_F: {:.4f} || cur_max_F:{:.4f}'
.format(epc, records[0], records[1], mF, max_F))
print('finish training.'+'maxF:',max_F)
def test(self,ep):
self.net.eval().cuda()
print("params:",(count_param(self.net)/1e6))
avg_f, mae,img_num = 0.0,0.0, 0
score = torch.zeros(255)
for image, label, prefix in self.test_loader:
with torch.no_grad():
image, label = image.cuda(), label.cuda()
B=image.size(0)
smap,_= self.net(image)
if ep%4==0:
for b in range(B):
path = os.path.join(output_folder, prefix[b] + '.png')
save_smap(smap[b, ...], path)
if torch.mean(label) == 0.0:
continue
if smap.shape != label.shape:
smap = F.interpolate(smap, size=(label.shape[2],label.shape[3]), mode='bilinear')
img_num+=1
mae += Eval_mae(smap, label)
f_score= Eval_Fmeasure(smap, label)
avg_f+=f_score
score = avg_f / img_num
"""if ep%4==0:
for b in range(B):
path = os.path.join(output_folder, prefix[b]+'_'+str(round(f_score.mean().item(),4))[2:] + '.png')
save_smap(smap[b, ...], path)"""
maxF=score.max().item()
mean=score.mean().item()
mae=(mae/img_num)
print('finish testing.', 'F—value : {:.4f}\t'.format(maxF),'mF—value : {:.4f}\t'.format(mean),'MAE : {:.4f}\t'.format(mae))
return maxF,mean
if __name__=='__main__':
run=Run()
#run.train()
run.test(8)