-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
190 lines (173 loc) · 7.22 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 datetime
import os
from math import sqrt
import cv2
from tqdm import *
import PIL.Image as Image
import scipy.io as sio
import numpy as np
import torchvision.transforms as transforms
import torchvision.utils as vutils
from tensorboardX import SummaryWriter
from torch import optim
from torch.autograd import Variable
import torch
from torch.utils.data import DataLoader
from torch.utils import data
from utils import check_mkdir, evaluate, AverageMeter, CrossEntropyLoss2d
from utils.transforms import RandomGaussianBlur
from psp_net import PSPNet
from gt_colormap import gt_cmap
import torch.nn as nn
import ipdb
import datetime
from eval import eval_main
num_classes = 21
args = {
'train_batch_size': 16,
'lr': 1e-2 ,#/ sqrt(16 / 4),
'lr_decay': 0.9,
'max_iter': 3e4,
'longer_size': 512,
'crop_size': 473,
'stride_rate': 2 / 3.,
'weight_decay': 1e-4,
'momentum': 0.9,
'snapshot': '',
'print_freq': 10,
'val_save_to_img_file': True,
'val_img_sample_rate': 0.01, # randomly sample some validation results to display,
'val_img_display_size': 384,
}
def make_dataset(mode):
assert mode in ['train', 'val', 'test']
items = []
root = 'VOC'
if mode == 'train':
img_path = os.path.join('VOC', 'benchmark_RELEASE', 'dataset', 'img')
mask_path = os.path.join('VOC', 'benchmark_RELEASE', 'dataset', 'cls')
data_list = [l.strip('\n') for l in open(os.path.join(
root, 'benchmark_RELEASE', 'dataset', 'train.txt')).readlines()]
for it in data_list:
item = (os.path.join(img_path, it + '.jpg'), os.path.join(mask_path, it + '.mat'))
items.append(item)
elif mode == 'val':
img_path = os.path.join('VOC', 'benchmark_RELEASE', 'dataset', 'img')
mask_path = os.path.join('VOC', 'benchmark_RELEASE', 'dataset', 'cls')
data_list = [l.strip('\n') for l in open(os.path.join(
root, 'benchmark_RELEASE', 'dataset', 'val.txt')).readlines()]
for it in data_list:
item = (os.path.join(img_path, it + '.jpg'), os.path.join(mask_path, it + '.mat'))
items.append(item)
return items
class VOC(data.Dataset):
def __init__(self, mode):
self.imgs = make_dataset(mode)
if len(self.imgs) == 0:
raise RuntimeError('Found 0 images, please check the data set')
self.mode = mode
def __getitem__(self, index):
if self.mode == 'test':
img_path, mask_path = self.imgs[index]
img = cv2.imread(img_path).astype(float)
mask = sio.loadmat(mask_path)['GTcls']['Segmentation'][0][0]
return torch.from_numpy(img).float(), torch.from_numpy(np.array(mask))
elif self.mode == 'train':
img_path, mask_path = self.imgs[index]
img = Image.open(img_path)
mask = Image.fromarray(sio.loadmat(mask_path)['GTcls']['Segmentation'][0][0])
scale_size = (img.size*(1.5*np.random.rand(1) + 0.5)).astype(np.long)
random_degree = 20*(np.random.rand(1)-0.5)
# Data Augmentation
trans = transforms.Compose([
transforms.Resize(scale_size, Image.NEAREST),
transforms.CenterCrop(img.size),
transforms.RandomRotation((random_degree,random_degree), expand=False),
transforms.Resize((473,473), Image.NEAREST),
])
img = trans(img)
gt = trans(mask)
if np.random.rand(1)>0.5:
img = transforms.RandomHorizontalFlip(1)(img)
gt = transforms.RandomHorizontalFlip(1)(gt)
if np.random.rand(1)>0.5:
img = transforms.RandomVerticalFlip(1)(img)
gt = transforms.RandomVerticalFlip(1)(gt)
img_org = np.copy(img)
img = RandomGaussianBlur()(img)
img = transforms.ToTensor()(img)
img_orig = transforms.ToTensor()(img_org)
gt = np.array(gt)
#img = self.normalize(img)
mean = [torch.mean(img[0]), torch.mean(img[1]), torch.mean(img[2])]
std = [torch.std(img[0]), torch.std(img[1]), torch.std(img[2])]
img = transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))(img)
#img = transforms.Normalize((123.68, 116.779, 103.939), (1, 1, 1))(img)
return img, torch.from_numpy(gt), img_orig
def __len__(self):
return len(self.imgs)
if __name__=='__main__':
# Network define
net = PSPNet(num_classes=num_classes)#.cuda()
'''optimizer = optim.SGD([
{'params': [param for name, param in net.named_parameters() if name[-4:] == 'bias'],
'lr': 2 * args['lr']},
{'params': [param for name, param in net.named_parameters() if name[-4:] != 'bias'],
'lr': args['lr'], 'weight_decay': args['weight_decay']}
], momentum=args['momentum'], nesterov=True)'''
# Summary writer
net.train()
time_current = datetime.datetime.now().strftime("%Y-%m-%d %H:%M")
writer = SummaryWriter('logs/'+time_current)
#dummy_input = Variable(torch.rand(2, 3, 473, 473))
#writer.add_graph(net, (dummy_input, ))
# Net train
net.cuda()
net = torch.nn.DataParallel(net, device_ids=range(torch.cuda.device_count()))
net.train()
train_set = VOC('train')
train_loader = DataLoader(train_set, batch_size=8, num_workers=8, shuffle=True)
train_loss = []
train_accu = []
i = 0
# Loss function
criterion = CrossEntropyLoss2d(size_average=True, ignore_index=255).cuda()
lr = 0.01
while(1):
for data in train_loader:
lr_tmp = lr*(1 - i/args['max_iter'])**0.9
optimizer = optim.SGD(net.parameters(), lr_tmp, momentum=0.9, nesterov=True, weight_decay=0.0001)
net.train()
img, gt, img_org = data
#img.cuda()
#gt.cuda()
#ipdb.set_trace()
img = Variable(img.cuda())
#ipdb.set_trace()
gt = Variable(gt.cuda().long())
optimizer.zero_grad()
outputs, aux = net(img)
main_loss = criterion(outputs, torch.squeeze(gt))#nn.NLLLoss2d(output, gt)
aux_loss = criterion(aux, torch.squeeze(gt))#nn.NLLLoss2d(aux, gt) # calc gradients
loss = main_loss + 0.4 * aux_loss
train_loss.append(loss.data[0])
loss.backward()
optimizer.step() # update gradients
if i % 100 == 0:
#train_accu.append(accuracy)
print('Train Step: {} Loss: {}'.format(i, loss.data[0]))
writer.add_scalar('lr', lr_tmp, i)
writer.add_scalar('train/loss', loss.data[0], i)
#writer.add_image('train/img', img.data, i)
writer.add_image('train/img_orig', img_org, i)
gts = gt_cmap(gt)
writer.add_image('train/gt', gts,i)
#writer.add_image('train/gt', gt.data, i)
i += 1
if i==args['max_iter']:
break
if i % 10000 == 0:# and i!=0:
print('taking snapshot ...')
torch.save(net.module.state_dict(),'logs/'+time_current+'/'+'VOC12_pyramid_'+str(i)+'.pth')
eval_main(net, writer, i)
eval_main(net,writer,i)