-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
264 lines (206 loc) · 10.6 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
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.autograd import Variable
from torch.utils.data import DataLoader
import argparse
import os
from tqdm import tqdm
#from dataset_loader import MyData
from dataset_loader_augment import MyData
#加载网络
from model.model_fusion import FusionNet
#加载损失函数
import pytorch_losses
def Hybrid_Loss(pred, target, reduction='mean'):
#先对输出做归一化处理
pred = torch.sigmoid(pred)
#BCE LOSS
bce_loss = nn.BCELoss()
bce_out = bce_loss(pred, target)
#IOU LOSS
iou_loss = pytorch_losses.IOU(reduction=reduction)
iou_out = iou_loss(pred, target)
#SSIM LOSS
ssim_loss = pytorch_losses.SSIM(window_size=11)
ssim_out = ssim_loss(pred, target)
hybrid_loss = [bce_out, iou_out, ssim_out]
losses = bce_out + iou_out + ssim_out
return hybrid_loss, losses
def cross_entropy2d_edge(input, target, reduction='mean'):
assert (input.size() == target.size())
pos = torch.eq(target, 1).float()
neg = torch.eq(target, 0).float()
num_pos = torch.sum(pos)
num_neg = torch.sum(neg)
num_total = num_pos + num_neg
alpha = num_neg / num_total
beta = 1.1 * num_pos / num_total
# target pixel = 1 -> weight beta
# target pixel = 0 -> weight 1-beta
weights = alpha * pos + beta * neg
return F.binary_cross_entropy_with_logits(input, target, weights, reduction=reduction)
#获取当前学习率
def get_lr(optimizer):
for param_group in optimizer.param_groups:
return param_group['lr']
#训练一个epoch
class Trainer(object):
def __init__(self, cuda, model_fusion, optimizer, scheduler, train_loader, epochs, save_epoch):
self.cuda = cuda
self.model_fusion = model_fusion
self.optimizer = optimizer
self.scheduler = scheduler
self.train_loader = train_loader
self.present_epoch = 0
self.epochs = epochs #总的epoch
self.save_epoch = save_epoch #间隔save_epoch保存一次权重文件
self.train_loss_list = []
self.train_log = './train_log.txt'
with open(self.train_log, 'w') as f:
f.write(('%10s' * 8) % ('Epoch', 'losses', 'loss_1', 'loss_2', 'loss_3', 'loss_4', 'loss_5', 'lr'))
f.close()
def train_epoch(self):
print(('\n' + '%10s' * 8) % ('Epoch', 'losses', 'loss_1', 'loss_2', 'loss_3', 'loss_4', 'loss_5', 'lr'))
# 计算所有的loss
losses_all, loss_1_all, loss_2_all, loss_3_all, loss_4_all, loss_5_all = 0, 0, 0, 0, 0, 0
Hybrid_losses_1,Hybrid_losses_3,Hybrid_losses_4,Hybrid_losses_5 = [0,0,0],[0,0,0],[0,0,0],[0,0,0]
#设置进度条
with tqdm(total=len(self.train_loader)) as pbar:
for batch_idx, (img, mask, depth, edge) in enumerate(self.train_loader):
if self.cuda:
img, mask, depth, edge = img.cuda(), mask.cuda(), depth.cuda(), edge.cuda()
img, mask, depth, edge = Variable(img), Variable(mask), Variable(depth), Variable(edge)
n, c, h, w = img.size() # batch_size, channels, height, weight
#梯度清零
self.optimizer.zero_grad()
depth = depth.view(n, 1, h, w).repeat(1, c, 1, 1) #把深度图变成3个通道
# depth = depth.view(n, 1, h, w)
mask = mask.view(n, 1, h, w)
edge = edge.view(n, 1, h, w)
#前向传播
F1_out, F2_out, F3_out, F4_out, F5_out = self.model_fusion(img, depth)
#计算损失函数
mask = mask.to(torch.float32)
edge = edge.to(torch.float32)
hybrid_loss_1, loss_1 = Hybrid_Loss(F1_out, mask)
loss_2 = cross_entropy2d_edge(F2_out, edge)
hybrid_loss_3, loss_3 = Hybrid_Loss(F3_out, mask)
hybrid_loss_4, loss_4 = Hybrid_Loss(F4_out, mask)
hybrid_loss_5, loss_5 = Hybrid_Loss(F5_out, mask)
#每一个stage的损失函数求和
Hybrid_losses_1 = [Hybrid_losses_1[i] + hybrid_loss_1[i].item() for i in range(len(hybrid_loss_1))]
Hybrid_losses_3 = [Hybrid_losses_3[i] + hybrid_loss_3[i].item() for i in range(len(hybrid_loss_3))]
Hybrid_losses_4 = [Hybrid_losses_4[i] + hybrid_loss_4[i].item() for i in range(len(hybrid_loss_4))]
Hybrid_losses_5 = [Hybrid_losses_5[i] + hybrid_loss_5[i].item() for i in range(len(hybrid_loss_5))]
#计算损失函数用于反向传播
loss_1_all += loss_1.item()
loss_2_all += loss_2.item()
loss_3_all += loss_3.item()
loss_4_all += loss_4.item()
loss_5_all += loss_5.item()
losses = loss_1 + loss_2 + loss_3 + loss_4 + loss_5
losses_all += losses.item()
#实时更新信息
s = ('%10s' * 1 + '%10.4g' * 7) % (self.present_epoch, losses.item(), loss_1.item(), loss_2.item(), loss_3.item(), loss_4.item(), loss_5.item(), get_lr(self.optimizer))
pbar.set_description(s)
pbar.update(1)
#反向传播
losses.backward()
#更新权重
self.optimizer.step()
# 保存模型
if self.present_epoch % self.save_epoch == 0:
savename_ladder = ('checkpoint/RGBD-SOD_iter_%d.pth' % (self.present_epoch))
torch.save(self.model_fusion.state_dict(), savename_ladder)
total_batch = len(self.train_loader)
#输出最后的损失函数
epoch_information = ('\n' + '%10s' * 1 + '%10.4g' * 7) % ((self.present_epoch),
losses_all / total_batch, loss_1_all / total_batch, loss_2_all / total_batch,
loss_3_all / total_batch, loss_4_all / total_batch, loss_5_all / total_batch,
get_lr(self.optimizer))
print(epoch_information)
#输出各个损失函数的值
print(('Stage' + '%10s' * 3) % ('bce_out', 'iou_out', 'ssim_out'))
print(('1' + '%10.4g' * 3) % (
Hybrid_losses_1[0] / total_batch, Hybrid_losses_1[1] / total_batch, Hybrid_losses_1[2] / total_batch))
print(('3' + '%10.4g' * 3) % (
Hybrid_losses_3[0] / total_batch, Hybrid_losses_3[1] / total_batch, Hybrid_losses_3[2] / total_batch))
print(('4' + '%10.4g' * 3) % (
Hybrid_losses_4[0] / total_batch, Hybrid_losses_4[1] / total_batch, Hybrid_losses_4[2] / total_batch))
print(('5' + '%10.4g' * 3) % (
Hybrid_losses_5[0] / total_batch, Hybrid_losses_5[1] / total_batch, Hybrid_losses_5[2] / total_batch))
#写入文件
with open(self.train_log, 'a') as f:
f.write(epoch_information)
f.close()
#更新学习率
self.scheduler.step()
def train(self):
for epoch in range(self.epochs):
self.train_epoch()
self.present_epoch += 1
if __name__ == '__main__':
"""
opt参数解析:
pre-trained: 是否加载预训练模型
checkpoint: 预训练模型的路径
train-root: 训练数据集路径
epochs: 训练总轮次
batch-size: 批次大小
workers: dataloader的最大worker数量
save-epoch: 间隔save-epoch保存一次模型
cuda: 是否使用GPU进行训练
GPU-id: 使用单块GPU时设置的编号
"""
parser = argparse.ArgumentParser()
parser.add_argument('--checkpoint', type=str, default='./checkpoint', help='path to pre-trained parameters')
parser.add_argument('--train-root', type=str, default='/mnt/02AA93C51773C62F/dataset/train_2985/', help='path to the train dataset')
#超参数设置
parser.add_argument('--epochs', type=int, default=200)
parser.add_argument('--batch-size', type=int, default=8, help='total batch size for all GPUs')
parser.add_argument('--workers', type=int, default=4, help='maximum number of dataloader workers')
parser.add_argument('--lr', type=float, default=0.0005, help='learning rate')
parser.add_argument('--weight_decay', type=float, default=1e-4, help='decay rate of learning rate')
parser.add_argument('--save-epoch', type=int, default=1)
parser.add_argument('--cuda', type=bool, default=True, help='use cuda')
parser.add_argument('--GPU-id', type=int, default=0)
args = parser.parse_args()
#加载训练集
train_loader = torch.utils.data.DataLoader(MyData(args.train_root),batch_size=args.batch_size, shuffle=True,
num_workers=args.workers, pin_memory=True, drop_last=True)
# 定义网络结构
channels = [16, 24, 40, 112, 160]
model_fusion = FusionNet(channels)
# 使用GPU
if args.cuda:
assert torch.cuda.is_available, 'ERROR: cuda can not use'
torch.cuda.set_device(args.GPU_id) #指定显卡
#torch.cuda.set_device('cuda:' + str(gpu_ids)) # 可指定多卡
#torch.backends.cudnn.benchmark = True # GPU网络加速
model_fusion = model_fusion.cuda()
#model_fusion = torch.nn.DataParallel(model_fusion) #多GPU训练
#定义优化器
optimizer = optim.Adam(model_fusion.parameters(), lr=args.lr, weight_decay=args.weight_decay)
#optimizer = optim.SGD(model_fusion.parameters(), lr=cfg['lr'], momentum=cfg['momentum'], weight_decay=cfg['weight_decay'])
# 等间隔调整学习率,每训练step_size个epoch,lr*gamma
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.99)
# 多间隔调整学习率,每训练至milestones中的epoch,lr*gamma
# scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[10, 30, 80], gamma=0.1)
# 指数学习率衰减,lr*gamma**epoch
#lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.98)
# 余弦退火学习率衰减,T_max表示半个周期,lr的初始值作为余弦函数0处的极大值逐渐开始下降,
# 在epoch=T_max时lr降至最小值,即pi/2处,然后进入后半个周期,lr增大
# scheduler = lr_scheduler.CosineAnnealingLR(optimizer, T_max=100, eta_min=0)
#开始训练
training = Trainer(
cuda=args.cuda,
model_fusion=model_fusion,
optimizer=optimizer,
scheduler=lr_scheduler,
train_loader=train_loader,
epochs=args.epochs,
save_epoch=args.save_epoch
)
training.train()