-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathtrain.py
215 lines (195 loc) · 8.27 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
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
# coding=utf-8
import torch
import torchvision
from torchvision import transforms
import torch.nn.functional as F
import torch.nn as nn
from torchvision.utils import make_grid
from torch.autograd import Variable
from tensorboardX import SummaryWriter
from datetime import datetime
import os
import pdb
import numpy as np
from PIL import Image
import argparse
import json
from datasets import PriorFolder, Folder
from datasets.saliency import collate_more
from models import Net, FCN
from evaluate import fm_and_mae
from tqdm import tqdm
import random
# random.seed(1996)
home = os.path.expanduser("~")
parser = argparse.ArgumentParser()
parser.add_argument('--train_dir', default='%s/data/datasets/saliency_Dataset/DUT-train' % home) # training dataset
parser.add_argument('--prior_dir', default='%s/data/datasets/saliency_Dataset/results/ECSSD-Sal' % home) # training dataset
parser.add_argument('--val_dir', default='%s/data/datasets/saliency_Dataset/ECSSD' % home) # training dataset
parser.add_argument('--base', default='vgg16') # training dataset
parser.add_argument('--img_size', type=int, default=256) # batch size
parser.add_argument('--b', type=int, default=12) # batch size
parser.add_argument('--max', type=int, default=100000) # epoches
opt = parser.parse_args()
print(opt)
name = 'Train_{}'.format(opt.base)
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
# tensorboard writer
os.system('rm -rf ./runs_%s/*'%name)
writer = SummaryWriter('./runs_%s/'%name + datetime.now().strftime('%B%d %H:%M:%S'))
if not os.path.exists('./runs_%s'%name):
os.mkdir('./runs_%s'%name)
def make_image_grid(img, mean, std):
img = make_grid(img)
for i in range(3):
img[i] *= std[i]
img[i] += mean[i]
return img
def validate(loader, net, output_dir, gt_dir):
if not os.path.exists(output_dir):
os.mkdir(output_dir)
net.eval()
loader = tqdm(loader, desc='validating')
for ib, (data, lbl, prior, img_name, w, h) in enumerate(loader):
with torch.no_grad():
outputs = net(data.cuda(), prior[:, None].cuda())
outputs = F.sigmoid(outputs)
outputs = outputs.squeeze(1).cpu().numpy()
outputs *= 255
for ii, msk in enumerate(outputs):
msk = Image.fromarray(msk.astype(np.uint8))
msk = msk.resize((w[ii], h[ii]))
msk.save('{}/{}.png'.format(output_dir, img_name[ii]), 'PNG')
fm, mae, _, _ = fm_and_mae(output_dir, gt_dir)
net.train()
return fm, mae
def main():
check_dir = '../LPSfiles/' + name
if not os.path.exists(check_dir):
os.mkdir(check_dir)
# data
val_loader = torch.utils.data.DataLoader(
PriorFolder(opt.val_dir, opt.prior_dir, size=256,
mean=mean, std=std),
batch_size=opt.b*3, shuffle=False, num_workers=4, pin_memory=True)
train_loader = torch.utils.data.DataLoader(
Folder(opt.train_dir, scales=[64]*3+[128, 256],
crop=0.9, flip=True, rotate=None,
mean=mean, std=std), collate_fn=collate_more,
batch_size=opt.b*6, shuffle=True, num_workers=4, pin_memory=True)
# models
p = 5
net = Net(base=opt.base)
fcn = FCN(net)
net = nn.DataParallel(net).cuda()
net.train()
"""
# fcn = nn.DataParallel(fcn).cuda()
# sdict =torch.load('/home/crow/LPSfiles/Train2_vgg16/fcn-iter13800.pth')
# fcn.load_state_dict(sdict)
fcn = nn.DataParallel(fcn).cuda()
fcn.train()
optimizer = torch.optim.Adam([
{'params': fcn.parameters(), 'lr': 1e-4},
])
logs = {'best_it':0, 'best': 0}
sal_data_iter = iter(train_loader)
i_sal_data = 0
for it in tqdm(range(opt.max)):
# for it in tqdm(range(1)):
# if it > 1000 and it % 100 == 0:
# optimizer.param_groups[0]['lr'] *= 0.5
if i_sal_data >= len(train_loader):
sal_data_iter = iter(train_loader)
i_sal_data = 0
data, lbls, _ = sal_data_iter.next()
i_sal_data += 1
data = data.cuda()
lbls = [lbl.unsqueeze(1).cuda() for lbl in lbls]
msks = fcn(data)
loss = sum([F.binary_cross_entropy_with_logits(msk, lbl) for msk, lbl in zip(msks, lbls)])
optimizer.zero_grad()
loss.backward()
optimizer.step()
if it % 10 == 0:
writer.add_scalar('loss', loss.item(), it)
image = make_image_grid(data[:6], mean, std)
writer.add_image('Image', torchvision.utils.make_grid(image), it)
big_msk = F.sigmoid(msks[-1]).expand(-1, 3, -1, -1)
writer.add_image('msk', torchvision.utils.make_grid(big_msk.data[:6]), it)
big_msk = lbls[-1].expand(-1, 3, -1, -1)
writer.add_image('gt', torchvision.utils.make_grid(big_msk.data[:6]), it)
# if it % 100 == 0:
if it != 0 and it % 100 == 0:
fm, mae = validate(val_loader, fcn, os.path.join(check_dir, 'results'),
os.path.join(opt.val_dir, 'masks'))
print(u'损失: %.4f'%(loss.item()))
print(u'最大FM: iteration %d的%.4f, 这次FM: %.4f'%(logs['best_it'], logs['best'], fm))
logs[it] = {'FM': fm}
if fm > logs['best']:
logs['best'] = fm
logs['best_it'] = it
torch.save(fcn.state_dict(), '%s/fcn-best.pth' % (check_dir))
with open(os.path.join(check_dir, 'logs.json'), 'w') as outfile:
json.dump(logs, outfile)
torch.save(fcn.state_dict(), '%s/fcn-iter%d.pth' % (check_dir, it))
"""
###################################################################################################
val_loader = torch.utils.data.DataLoader(
PriorFolder(opt.val_dir, opt.prior_dir, size=256,
mean=mean, std=std),
batch_size=opt.b, shuffle=False, num_workers=4, pin_memory=True)
train_loader = torch.utils.data.DataLoader(
Folder(opt.train_dir, scales=[256],
crop=0.9, flip=True, rotate=None,
mean=mean, std=std), collate_fn=collate_more,
batch_size=opt.b, shuffle=True, num_workers=4, pin_memory=True)
optimizer = torch.optim.Adam([
{'params': net.parameters(), 'lr': 1e-4},
])
logs = {'best_it':0, 'best': 0}
sal_data_iter = iter(train_loader)
i_sal_data = 0
for it in tqdm(range(opt.max)):
# if it > 1000 and it % 100 == 0:
# optimizer.param_groups[0]['lr'] *= 0.5
if i_sal_data >= len(train_loader):
sal_data_iter = iter(train_loader)
i_sal_data = 0
data, lbl, _ = sal_data_iter.next()
i_sal_data += 1
data = data.cuda()
lbl = lbl[0].unsqueeze(1)
noisy_label = (lbl.numpy() + np.random.binomial(1, float(p) / 100.0, (256, 256))) % 2
noisy_label = torch.Tensor(noisy_label).cuda()
lbl = lbl.cuda()
msk = net(data, noisy_label)
loss = F.binary_cross_entropy_with_logits(msk, lbl)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if it % 10 == 0:
writer.add_scalar('loss', loss.item(), it)
image = make_image_grid(data[:6], mean, std)
writer.add_image('Image', torchvision.utils.make_grid(image), it)
big_msk = F.sigmoid(msk).expand(-1, 3, -1, -1)
writer.add_image('msk', torchvision.utils.make_grid(big_msk.data[:6]), it)
big_msk = lbl.expand(-1, 3, -1, -1)
writer.add_image('gt', torchvision.utils.make_grid(big_msk.data[:6]), it)
# if it % 200 == 0:
if it != 0 and it % 100 == 0:
fm, mae = validate(val_loader, net, os.path.join(check_dir, 'results'),
os.path.join(opt.val_dir, 'masks'))
print(u'损失: %.4f'%(loss.item()))
print(u'最大FM: iteration %d的%.4f, 这次FM: %.4f'%(logs['best_it'], logs['best'], fm))
logs[it] = {'FM': fm}
if fm > logs['best']:
logs['best'] = fm
logs['best_it'] = it
torch.save(net.state_dict(), '%s/net-best.pth' % (check_dir))
with open(os.path.join(check_dir, 'logs.json'), 'w') as outfile:
json.dump(logs, outfile)
torch.save(net.state_dict(), '%s/net-iter%d.pth' % (check_dir, it))
if __name__ == "__main__":
main()