forked from bubbliiiing/unet-pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
241 lines (199 loc) · 10 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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
import os
import time
import numpy as np
import torch
import torch.backends.cudnn as cudnn
import torch.optim as optim
from torch.utils.data import DataLoader
from tqdm import tqdm
from nets.unet import Unet
from nets.unet_training import CE_Loss, Dice_loss, LossHistory
from utils.dataloader import DeeplabDataset, deeplab_dataset_collate
from utils.metrics import f_score
def get_lr(optimizer):
for param_group in optimizer.param_groups:
return param_group['lr']
def fit_one_epoch(net,epoch,epoch_size,epoch_size_val,gen,genval,Epoch,cuda):
net = net.train()
total_loss = 0
total_f_score = 0
val_toal_loss = 0
val_total_f_score = 0
start_time = time.time()
with tqdm(total=epoch_size,desc=f'Epoch {epoch + 1}/{Epoch}',postfix=dict,mininterval=0.3) as pbar:
for iteration, batch in enumerate(gen):
if iteration >= epoch_size:
break
imgs, pngs, labels = batch
with torch.no_grad():
imgs = torch.from_numpy(imgs).type(torch.FloatTensor)
pngs = torch.from_numpy(pngs).type(torch.FloatTensor).long()
labels = torch.from_numpy(labels).type(torch.FloatTensor)
if cuda:
imgs = imgs.cuda()
pngs = pngs.cuda()
labels = labels.cuda()
optimizer.zero_grad()
outputs = net(imgs)
loss = CE_Loss(outputs, pngs, num_classes = NUM_CLASSES)
if dice_loss:
main_dice = Dice_loss(outputs, labels)
loss = loss + main_dice
with torch.no_grad():
#-------------------------------#
# 计算f_score
#-------------------------------#
_f_score = f_score(outputs, labels)
loss.backward()
optimizer.step()
total_loss += loss.item()
total_f_score += _f_score.item()
waste_time = time.time() - start_time
pbar.set_postfix(**{'total_loss': total_loss / (iteration + 1),
'f_score' : total_f_score / (iteration + 1),
's/step' : waste_time,
'lr' : get_lr(optimizer)})
pbar.update(1)
start_time = time.time()
print('Start Validation')
with tqdm(total=epoch_size_val, desc=f'Epoch {epoch + 1}/{Epoch}',postfix=dict,mininterval=0.3) as pbar:
for iteration, batch in enumerate(genval):
if iteration >= epoch_size_val:
break
imgs, pngs, labels = batch
with torch.no_grad():
imgs = torch.from_numpy(imgs).type(torch.FloatTensor)
pngs = torch.from_numpy(pngs).type(torch.FloatTensor).long()
labels = torch.from_numpy(labels).type(torch.FloatTensor)
if cuda:
imgs = imgs.cuda()
pngs = pngs.cuda()
labels = labels.cuda()
outputs = net(imgs)
val_loss = CE_Loss(outputs, pngs, num_classes = NUM_CLASSES)
if dice_loss:
main_dice = Dice_loss(outputs, labels)
val_loss = val_loss + main_dice
#-------------------------------#
# 计算f_score
#-------------------------------#
_f_score = f_score(outputs, labels)
val_toal_loss += val_loss.item()
val_total_f_score += _f_score.item()
pbar.set_postfix(**{'total_loss': val_toal_loss / (iteration + 1),
'f_score' : val_total_f_score / (iteration + 1),
'lr' : get_lr(optimizer)})
pbar.update(1)
loss_history.append_loss(total_loss/(epoch_size+1), val_toal_loss/(epoch_size_val+1))
print('Finish Validation')
print('Epoch:'+ str(epoch+1) + '/' + str(Epoch))
print('Total Loss: %.4f || Val Loss: %.4f ' % (total_loss/(epoch_size+1),val_toal_loss/(epoch_size_val+1)))
print('Saving state, iter:', str(epoch+1))
torch.save(model.state_dict(), 'logs/Epoch%d-Total_Loss%.4f-Val_Loss%.4f.pth'%((epoch+1),total_loss/(epoch_size+1),val_toal_loss/(epoch_size_val+1)))
if __name__ == "__main__":
log_dir = "logs/"
#------------------------------#
# 输入图片的大小
#------------------------------#
inputs_size = [512,512,3]
#---------------------#
# 分类个数+1
# 2+1
#---------------------#
NUM_CLASSES = 21
#--------------------------------------------------------------------#
# 建议选项:
# 种类少(几类)时,设置为True
# 种类多(十几类)时,如果batch_size比较大(10以上),那么设置为True
# 种类多(十几类)时,如果batch_size比较小(10以下),那么设置为False
#---------------------------------------------------------------------#
dice_loss = False
#-------------------------------#
# 主干网络预训练权重的使用
#-------------------------------#
pretrained = False
#-------------------------------#
# Cuda的使用
#-------------------------------#
Cuda = True
#------------------------------#
# 数据集路径
#------------------------------#
dataset_path = "VOCdevkit/VOC2007/"
model = Unet(num_classes=NUM_CLASSES, in_channels=inputs_size[-1], pretrained=pretrained).train()
loss_history = LossHistory("logs/")
#-------------------------------------------#
# 权值文件的下载请看README
# 权值和主干特征提取网络一定要对应
#-------------------------------------------#
model_path = r"model_data/unet_voc.pth"
print('Loading weights into state dict...')
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model_dict = model.state_dict()
pretrained_dict = torch.load(model_path, map_location=device)
pretrained_dict = {k: v for k, v in pretrained_dict.items() if np.shape(model_dict[k]) == np.shape(v)}
model_dict.update(pretrained_dict)
model.load_state_dict(model_dict)
print('Finished!')
if Cuda:
net = torch.nn.DataParallel(model)
cudnn.benchmark = True
net = net.cuda()
# 打开数据集的txt
with open(os.path.join(dataset_path, "ImageSets/Segmentation/train.txt"),"r") as f:
train_lines = f.readlines()
# 打开数据集的txt
with open(os.path.join(dataset_path, "ImageSets/Segmentation/val.txt"),"r") as f:
val_lines = f.readlines()
#------------------------------------------------------#
# 主干特征提取网络特征通用,冻结训练可以加快训练速度
# 也可以在训练初期防止权值被破坏。
# Init_Epoch为起始世代
# Interval_Epoch为冻结训练的世代
# Epoch总训练世代
# 提示OOM或者显存不足请调小Batch_size
#------------------------------------------------------#
if True:
lr = 1e-4
Init_Epoch = 0
Interval_Epoch = 50
Batch_size = 2
optimizer = optim.Adam(model.parameters(),lr)
lr_scheduler = optim.lr_scheduler.StepLR(optimizer,step_size=1,gamma=0.92)
train_dataset = DeeplabDataset(train_lines, inputs_size, NUM_CLASSES, True, dataset_path)
val_dataset = DeeplabDataset(val_lines, inputs_size, NUM_CLASSES, False, dataset_path)
gen = DataLoader(train_dataset, batch_size=Batch_size, num_workers=4, pin_memory=True,
drop_last=True, collate_fn=deeplab_dataset_collate)
gen_val = DataLoader(val_dataset, batch_size=Batch_size, num_workers=4,pin_memory=True,
drop_last=True, collate_fn=deeplab_dataset_collate)
epoch_size = len(train_lines) // Batch_size
epoch_size_val = len(val_lines) // Batch_size
if epoch_size == 0 or epoch_size_val == 0:
raise ValueError("数据集过小,无法进行训练,请扩充数据集。")
for param in model.vgg.parameters():
param.requires_grad = False
for epoch in range(Init_Epoch,Interval_Epoch):
fit_one_epoch(model,epoch,epoch_size,epoch_size_val,gen,gen_val,Interval_Epoch,Cuda)
lr_scheduler.step()
if True:
lr = 1e-5
Interval_Epoch = 50
Epoch = 100
Batch_size = 2
optimizer = optim.Adam(model.parameters(),lr)
lr_scheduler = optim.lr_scheduler.StepLR(optimizer,step_size=1,gamma=0.92)
train_dataset = DeeplabDataset(train_lines, inputs_size, NUM_CLASSES, True, dataset_path)
val_dataset = DeeplabDataset(val_lines, inputs_size, NUM_CLASSES, False, dataset_path)
gen = DataLoader(train_dataset, batch_size=Batch_size, num_workers=4, pin_memory=True,
drop_last=True, collate_fn=deeplab_dataset_collate)
gen_val = DataLoader(val_dataset, batch_size=Batch_size, num_workers=4,pin_memory=True,
drop_last=True, collate_fn=deeplab_dataset_collate)
epoch_size = len(train_lines) // Batch_size
epoch_size_val = len(val_lines) // Batch_size
if epoch_size == 0 or epoch_size_val == 0:
raise ValueError("数据集过小,无法进行训练,请扩充数据集。")
for param in model.vgg.parameters():
param.requires_grad = True
for epoch in range(Interval_Epoch,Epoch):
fit_one_epoch(model,epoch,epoch_size,epoch_size_val,gen,gen_val,Epoch,Cuda)
lr_scheduler.step()