-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
208 lines (190 loc) · 12.2 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
import os
from time import time
import time
import numpy as np
import torch
import torchvision
from torchvision import transforms
from torch.utils.data import DataLoader,Dataset
from torch.utils.tensorboard import SummaryWriter
from torch.autograd import Variable
from torch import nn
import Config as config
from torch import optim
import torch.nn.functional as F
from tqdm import tqdm
from skimage import io
from net import SA_UNet
use_gpu = torch.cuda.is_available()
'''
clip2
'''
transformers_train = torchvision.transforms.Compose([transforms.ToTensor(),
transforms.Normalize([55.193066, 51.369373, 62.4114, 45.084743, 162.25713, 119.69682, 73.51226, 50.53122, 124.01215, 125.25725, 129.83824],
[25.581112, 27.1253, 30.857647, 31.909128, 40.36898, 38.61341, 36.058147, 30.903032, 44.86116, 28.221159, 27.749435])])
transformers_2015 = torchvision.transforms.Compose([transforms.ToTensor(),transforms.Normalize([39.98135340082576, 38.7783512826099, 49.25017345571914, 43.56729661934147, 142.88135598188185, 110.54692154376903, 69.04691625041669, 45.49868291769581, 119.466705251268, 167.75940195617804, 176.32788100543948],[24.965231234410364, 25.935760307515842, 28.93568655111308, 32.799374033636695, 43.84462333501905, 42.26782425162127, 39.35267893098884, 30.278631742021666, 44.30422049354514, 32.39992776855984, 31.62088668466183])])
transformers_2017 = torchvision.transforms.Compose([transforms.ToTensor(),transforms.Normalize([79.92107939766025, 70.88585301279241, 74.34279705987088, 55.63881129680133, 146.93338202342548, 111.8844097649664, 73.2173914622164, 62.32098328613063, 129.50170305955766, 168.60956106297513, 172.5310738593263],[35.084373205808674, 35.1590432073348, 37.810388689421934, 40.06051029321953, 45.102994590979776, 44.289808957178295, 42.66778313024135, 37.43969000849209, 47.46019891141157, 35.88760521401966, 35.238889194903514])])
transformers_2021 = torchvision.transforms.Compose([transforms.ToTensor(),transforms.Normalize([49.112005585580384, 45.457424385993, 53.89288586569934, 42.91280790030649, 159.31102207196685, 109.68321806074319, 68.05436290348688, 45.75308641813284, 120.84085085607944, 127.30756595035798, 126.23945520467775],
[33.79766356353502, 34.213071803759576, 35.901864071386015, 38.397053918805724, 40.51438517299972, 38.87191699113998, 40.25925825910798, 36.28005575909439, 46.73744708303059, 43.096934210129795, 44.27079847112406])])
'''
randomcrop_1
'''
# transformers_train = torchvision.transforms.Compose([transforms.ToTensor(),
# transforms.Normalize([85.196976, 75.59963, 78.39488, 59.30308, 158.22722, 118.04645, 74.99491, 65.13752, 117.4489, 122.09071, 125.267746],
# [24.819115, 27.262259, 31.153456, 34.482067, 36.620674, 38.301075, 39.15172, 32.24261, 46.38067, 30.155087, 29.005043])])
class MyData(Dataset):
def __init__(self, root_dir, label_dir, transformers = None):
self.root_dir = root_dir
self.label_dir = label_dir
self.image_path = os.listdir(self.root_dir)
self.label_path = os.listdir(self.label_dir)
self.transformers = transformers
def __getitem__(self, idx): #如果想通过item去获取图片，就要先创建图片地址的一个列表
img_name = self.image_path[idx]
label_name = self.label_path[idx]
img_item_path = os.path.join(self.root_dir, img_name) # 每个图片的位置
label_item_path = os.path.join(self.label_dir, label_name)
image = io.imread(img_item_path)
if self.transformers is not None:
image = self.transformers(image)
else:
image = torch.from_numpy(image)
label = np.load(label_item_path)
#label = io.imread(label_item_path)
label = torch.from_numpy(label)
return image,label
def __len__(self):
return len(self.image_path)
#==================================================================================
def train_model(model,criterion,optimizer,scheduler,batch_size,num_epochs=25):
since = time.time()
writer = SummaryWriter(r"./log")
for epoch in range(num_epochs):
b = 0
print('Epoch {}/{}'.format(epoch,num_epochs-1))
print('-' * 10)
for phase in ['img','val']:
running_loss = 0.0
running_corrects = 0.0
train_acc = 0.0
train_loss = 0.0
if phase == 'img':
model.train()
for step,(inputs, labels) in enumerate(dataloaders[phase]):
b = epoch*len(dataloaders[phase]) + step
if use_gpu:
inputs = inputs.float().cuda()
labels = labels.long().cuda()
inputs, labels = Variable(inputs), Variable(labels)
optimizer.zero_grad()
outputs = model(inputs)
out = F.softmax(outputs,dim=1)
preds = torch.argmax(out.data,dim=1)
loss = criterion(outputs,labels)
loss.backward()
#torch.nn.utils.clip_grad_norm_(model.parameters(),max_norm=0.05,norm_type=2.0)
optimizer.step()
running_loss += loss.item()*inputs.size(0)
running_corrects += torch.sum(preds == labels.data)
train_loss += loss.item()
train_acc += torch.sum(preds == labels.data)/(256*256)/batch_size
if step % 50 == 49: # 迭代次数除以300的余数等于299，,每300轮输出一次 0，1，2，3，....299，.....
writer.add_scalar('loss/img', train_loss / 50, b)
writer.add_scalar('acc/img',train_acc / 50, b)
print('[%d, %5d] loss: %.3f' % (epoch + 1, step + 1, train_loss / 50))
train_loss = 0.0
print('[%d, %5d] acc: %.3f' % (epoch + 1, step + 1, train_acc / 50))
train_acc = 0.0
epoch_loss = running_loss / dataset_sizes[phase]
epoch_acc = float(running_corrects) / dataset_sizes[phase]/256/256
print('{}Train Loss: {:.4f}Train Acc: {:.4f}'.format(phase, epoch_loss, epoch_acc))
writer.add_scalar('loss/epoch_train', epoch_loss, epoch + 1)
writer.add_scalar('acc/epoch_train', epoch_acc, epoch + 1)
if epoch > num_epochs-10:
#torch.save(model.state_dict(), r'.checpoint/UNet/U_Net_ASPP_SAM_class6_The_%d_epoch_model.pth' % (epoch+1))
torch.save(model.state_dict(), r'./checpoint/SA_UNet_%d_epoch_model.pth' % (epoch + 1))
# 1. 记录这个epoch的loss值和准确率
# info = {'loss': epoch_loss, 'accuracy': epoch_acc}
# for tag, value in info.items():
# train_logger.scalar_summary(tag, value, epoch)
#
# # 2. 记录这个epoch的模型的参数和梯度
# for tag, value in model.named_parameters():
# tag = tag.replace('.', '/')
# train_logger.histo_summary(tag, value.data.cpu().numpy(), epoch)
# train_logger.histo_summary(tag + '/grad', value.grad.data.cpu().numpy(), epoch)
# 3. 记录最后一个epoch的图像
# info = {'images': inputs.cpu().numpy()}
# for tag, images in info.items():
# train_logger.image_summary(tag, images, epoch)
else:
#取消验证阶段的梯度
with torch.no_grad():
model.eval()
val_loss = 0.0
val_acc = 0.0
b = 0
for step, (inputs, labels) in enumerate(dataloaders[phase]):
b = epoch*len(dataloaders[phase])+step
# 获取输入
# 判断是否使用gpu
if use_gpu:
inputs = inputs.float().cuda()
labels = labels.long().cuda()
inputs, labels = Variable(inputs), Variable(labels)
outputs = model(inputs)
loss = criterion(outputs, labels)
out = F.softmax(outputs, dim=1)
preds = torch.argmax(out.data, dim=1)
running_loss += loss.item() * inputs.size(0)
running_corrects += torch.sum(preds == labels.data)
val_loss += loss.item()
val_acc += torch.sum(preds == labels.data) / (256 * 256)/batch_size
if step % 50 == 49: # 迭代次数除以300的余数等于299，,每300轮输出一次 0，1，2，3，....299，.....
writer.add_scalar('loss/val', val_loss/50, b)
writer.add_scalar('acc/val', val_acc/50, b)
print('[%d, %5d] val_loss: %.3f' % (epoch + 1, step + 1, val_loss / 50))
val_loss = 0.0
print('[%d, %5d] acc: %.3f' % (epoch + 1, step + 1, val_acc / 50))
val_acc = 0.0
scheduler.step()
lr_exp = scheduler.get_last_lr()[0]
epoch_val_loss = running_loss / dataset_sizes[phase]
epoch_val_acc = float(running_corrects) / dataset_sizes[phase]/256/256
print('{}Val Loss: {:.4f}Val Acc: {:.4f}'.format(phase, epoch_val_loss, epoch_val_acc))
writer.add_scalar('loss/epoch_val', epoch_val_loss, epoch + 1)
writer.add_scalar('acc/epoch_val', epoch_val_acc, epoch + 1)
writer.add_scalar('lr', lr_exp, epoch + 1)
#info = {'loss': epoch_loss, 'accuracy': epoch_acc}
# for tag, value in info.items():
# val_logger.scalar_summary(tag, value, epoch)
#
# for tag, value in model.named_parameters():
# tag = tag.replace('.', '/')
# val_logger.histo_summary(tag, value.data.cpu().numpy(), epoch)
# val_logger.histo_summary(tag + '/grad', value.grad.data.cpu().numpy(), epoch)
writer.close()
# info = {'images': inputs.cpu().numpy()}
# for tag, images in info.items():
# val_logger.image_summary(tag, images, epoch)
time_elapsed = time.time() - since
print('Training complete in {:.0f}m {:.0f}s'.format(time_elapsed // 60, time_elapsed % 60))
if __name__ == '__main__':
train_Data = MyData(config.train_dir, config.train_lable_dir, transformers=transformers_train)
val_Data = MyData(config.vali_dir, config.vali_lable_dir, transformers=transformers_train)
train_DataLoader = DataLoader(train_Data, batch_size=8, shuffle=True, drop_last=True)
val_DataLoader = DataLoader(val_Data, batch_size=8, shuffle=True, drop_last=True)
dataloaders = {'img': train_DataLoader, 'val': val_DataLoader}
dataset_sizes = {'img': train_Data.__len__(), 'val': val_Data.__len__()}
config_vit = config.get_CTranS_config()
net = SA_UNet.Unet(11,8)
net = net.cuda()
criterion = nn.CrossEntropyLoss(ignore_index=0, weight=None, reduction="mean")
criterion = criterion.cuda()
optimzer = torch.optim.Adam(filter(lambda p: p.requires_grad, net.parameters()), lr=0.001)
#optimzer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9, weight_decay=0.0001)
exp_lr_scheduler = torch.optim.lr_scheduler.StepLR(optimzer, step_size=5, gamma=0.1)
train_model(net, criterion, optimzer, num_epochs=20, scheduler=exp_lr_scheduler, batch_size=8)
'''
tensorboard --logdir=log --port=6006
'''