-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
229 lines (189 loc) · 10.2 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
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
#Implementation taken from https://github.com/easezyc/deep-transfer-learning/blob/master/MUDA/MFSAN/MFSAN_2src/mfsan.py
#Zhu Y, Zhuang F, Wang D. Aligning domain-specific distribution and classifier for cross-domain classification from multiple sources[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2019, 33: 5989-5996.
from __future__ import print_function
import torch
import os
import math
import resnet as models
import copy
import torch.nn.functional as F
import torchvision.transforms as transforms
from torch.autograd import Variable
from torch.utils.data import DataLoader
from dataset_train2D import dataset_train, RandomGenerator
from dataset_test3D import dataset3D
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
# Training settings
batch_size = 32
iteration = 10000
lr = [0.001, 0.01]
momentum = 0.9
cuda = True
seed = 8
log_interval = 20
l2_decay = 5e-4
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Using device:', device)
source1_name = "GE"
source2_name = 'Philips'
target_name = "Siemens"
dataset = "ADNI1"
IMG_PATH = './Dataset/ADNI1'
results_dir = './Results'
img_size = 224
torch.manual_seed(seed)
if cuda:
torch.cuda.manual_seed(seed)
source1_train = dataset_train(base_dir=IMG_PATH, list_dir='./Dataset', split="train_ADNI1_GE_94to125",
transform=transforms.Compose([RandomGenerator(output_size=[img_size, img_size])]))
print("The length of source1 train set is: {}".format(len(source1_train)))
source1_loader = DataLoader(source1_train, batch_size=batch_size, shuffle=True, num_workers=4, pin_memory=True)
source2_train = dataset_train(base_dir=IMG_PATH, list_dir='./Dataset', split="train_ADNI1_Philips_94to125",
transform=transforms.Compose([RandomGenerator(output_size=[img_size, img_size])]))
print("The length of source2 train set is: {}".format(len(source2_train)))
source2_loader = DataLoader(source2_train, batch_size=batch_size, shuffle=True, num_workers=4, pin_memory=True)
db_train = dataset_train(base_dir=IMG_PATH, list_dir='./Dataset', split="train_ADNI1_Siemens_94to125",
transform=transforms.Compose([RandomGenerator(output_size=[img_size, img_size])]))
print("The length of target train set is: {}".format(len(db_train)))
target_train_loader = DataLoader(db_train, batch_size=batch_size, shuffle=True, num_workers=4, pin_memory=True)
db_test = dataset3D(base_dir=IMG_PATH, list_dir='./Dataset', split="test_ADNI1_Siemens_MNI")
target_test_loader = DataLoader(db_test, batch_size=1, shuffle=False, num_workers=1)
print("The length of target test set is: {}".format(len(db_test)))
db_val = dataset3D(base_dir=IMG_PATH, list_dir='./Dataset', split="test_ADNI1_Siemens_MNI")
target_valid_loader = DataLoader(db_val, batch_size=1, shuffle=False, num_workers=1)
print("The length of target validation set is: {}".format(len(db_val)))
def train(model):
source1_iter = iter(source1_loader)
source2_iter = iter(source2_loader)
target_iter = iter(target_train_loader)
correct = 0
optimizer = torch.optim.SGD([
{'params': model.sharedNet.parameters()},
{'params': model.cls_fc_son1.parameters(), 'lr': lr[1]},
{'params': model.cls_fc_son2.parameters(), 'lr': lr[1]},
{'params': model.sonnet1.parameters(), 'lr': lr[1]},
{'params': model.sonnet2.parameters(), 'lr': lr[1]},
], lr=lr[0], momentum=momentum, weight_decay=l2_decay)
for i in range(1, iteration + 1):
model.train()
optimizer.param_groups[0]['lr'] = lr[0] / math.pow((1 + 10 * (i - 1) / (iteration)), 0.75)
optimizer.param_groups[1]['lr'] = lr[1] / math.pow((1 + 10 * (i - 1) / (iteration)), 0.75)
optimizer.param_groups[2]['lr'] = lr[1] / math.pow((1 + 10 * (i - 1) / (iteration)), 0.75)
optimizer.param_groups[3]['lr'] = lr[1] / math.pow((1 + 10 * (i - 1) / (iteration)), 0.75)
optimizer.param_groups[4]['lr'] = lr[1] / math.pow((1 + 10 * (i - 1) / (iteration)), 0.75)
try:
sampled_batch = next(source1_iter)
source_data, source_label = sampled_batch['image'], sampled_batch['label']
except Exception as err:
source1_iter = iter(source1_loader)
sampled_batch = next(source1_iter)
source_data, source_label = sampled_batch['image'], sampled_batch['label']
try:
sampled_batch = next(target_iter)
target_data, __ = sampled_batch['image'], sampled_batch['label']
except Exception as err:
target_iter = iter(target_train_loader)
sampled_batch = next(target_iter)
target_data, __ = sampled_batch['image'], sampled_batch['label']
if cuda:
source_data, source_label = source_data.cuda(), source_label.cuda()
target_data = target_data.cuda()
source_data, source_label = Variable(source_data), Variable(source_label)
target_data = Variable(target_data)
optimizer.zero_grad()
ce_loss, joint_loss, l1_loss = model(source_data, target_data, source_label, mark=1)
gamma = 2 / (1 + math.exp(-10 * (i) / (iteration) )) - 1
loss = ce_loss + gamma * (joint_loss + l1_loss)
loss.backward()
optimizer.step()
if i % log_interval == 0:
print('Train source1 iter: {} [({:.0f}%)]\tLoss: {:.6f}\tCE_Loss: {:.6f}\tjoint_Loss: {:.6f}\tl1_Loss: {:.6f}'.format(
i, 100. * i / iteration, loss.item(), ce_loss.item(), joint_loss.item(), l1_loss.item()))
try:
sampled_batch = next(source2_iter)
source_data, source_label = sampled_batch['image'], sampled_batch['label']
except Exception as err:
source2_iter = iter(source2_loader)
sampled_batch = next(source2_iter)
source_data, source_label = sampled_batch['image'], sampled_batch['label']
try:
sampled_batch = next(target_iter)
target_data, __ = sampled_batch['image'], sampled_batch['label']
except Exception as err:
target_iter = iter(target_train_loader)
sampled_batch = next(target_iter)
target_data, __ = sampled_batch['image'], sampled_batch['label']
if cuda:
source_data, source_label = source_data.cuda(), source_label.cuda()
target_data = target_data.cuda()
source_data, source_label = Variable(source_data), Variable(source_label)
target_data = Variable(target_data)
optimizer.zero_grad()
ce_loss, joint_loss, l1_loss = model(source_data, target_data, source_label, mark=2)
gamma = 2 / (1 + math.exp(-10 * (i) / (iteration))) - 1
loss = ce_loss + gamma * (joint_loss + l1_loss)
loss.backward()
optimizer.step()
if i % log_interval == 0:
print(
'Train source2 iter: {} [({:.0f}%)]\tLoss: {:.6f}\tCE_Loss: {:.6f}\tjoint_Loss: {:.6f}\tl1_Loss: {:.6f}'.format(
i, 100. * i / iteration, loss.item(), ce_loss.item(), joint_loss.item(), l1_loss.item()))
if i % (log_interval * 10) == 0:
t_correct = valid(model, target_valid_loader)
if t_correct >= correct:
correct = t_correct
torch.save(model.state_dict(), results_dir+'/'+ dataset + '_' + source1_name + '_' + source2_name + '_to_' + target_name + '_max_accuracy.pth')
max_model = copy.deepcopy(model)
print("Best performance: ", dataset, source1_name, source2_name, "to", target_name, "max correct:", correct, 'Accuracy: {}/{} ({:.0f}%)'.format(
correct, len(target_valid_loader.dataset), 100. * correct / len(target_valid_loader.dataset)), "\n")
test_correct = valid(max_model, target_test_loader)
print("Testset performance: ", dataset, source1_name, source2_name, "to", target_name, "Testset correct:", test_correct, 'Accuracy: {}/{} ({:.0f}%)'.format(
test_correct, len(target_test_loader.dataset), 100. * test_correct / len(target_test_loader.dataset)), "\n")
def valid(model, loader):
model.eval()
test_loss = 0
correct = 0
correct1 = 0
correct2 = 0
total = 0
with torch.no_grad():
for data_3D, target in loader:
temp_correct = 0
temp_correct1 = 0
temp_correct2 = 0
if cuda:
data_3D, target = data_3D.cuda(), target.cuda()
for slice_number in range(95, 125): #this slice range in the 2D coronal plane used to train the model
temp_x = data_3D[:, :, :, slice_number, :]
temp_x = temp_x.repeat(1, 3, 1, 1)
data, target = Variable(temp_x), Variable(target)
pred1, pred2 = model(data, mark = 0)
pred1 = torch.nn.functional.softmax(pred1, dim=1)
pred2 = torch.nn.functional.softmax(pred2, dim=1)
pred = (pred1 + pred2) / 2
test_loss += F.nll_loss(F.log_softmax(pred, dim=1), target).item()
pred = pred.data.max(1)[1]
temp_correct += pred.eq(target.data.view_as(pred)).cpu().sum()
pred = pred1.data.max(1)[1]
temp_correct1 += pred.eq(target.data.view_as(pred)).cpu().sum()
pred = pred2.data.max(1)[1]
temp_correct2 += pred.eq(target.data.view_as(pred)).cpu().sum()
if temp_correct >= 15:
correct = correct + 1
if temp_correct1 >= 15:
correct1 = correct1 + 1
if temp_correct2 >= 15:
correct2 = correct2 + 1
total = total +1
#print(total, correct, correct1, correct2)
test_loss /= len(loader.dataset)
print(target_name, ' Test set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)'.format(
test_loss, correct, len(loader.dataset),
100. * correct / len(loader.dataset)))
print('source1 correct: {}, source2 correct: {}, Average correct: {}'.format(correct1, correct2, correct))
return correct
if __name__ == '__main__':
model = models.DAMS(num_classes=2)
if cuda:
model.cuda()
train(model)