DANN_py3/main.py at master · wang975679801/DANN_py3 · GitHub

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
import random
import os
import torch.backends.cudnn as cudnn
import torch.optim as optim
import torch.utils.data
from data_loader import GetLoader
from torchvision import datasets
from torchvision import transforms
from model import CNNModel
import numpy as np
from test import test

source_dataset_name = 'MNIST'
target_dataset_name = 'mnist_m'
source_image_root = os.path.join('dataset', source_dataset_name)
target_image_root = os.path.join('dataset', target_dataset_name)
model_root = 'models'
cuda = True
cudnn.benchmark = True
lr = 1e-3
batch_size = 128
image_size = 28
n_epoch = 100

manual_seed = random.randint(1, 10000)
random.seed(manual_seed)
torch.manual_seed(manual_seed)

# load data

img_transform_source = transforms.Compose([
    transforms.Resize(image_size),
    transforms.ToTensor(),
    transforms.Normalize(mean=(0.1307,), std=(0.3081,))
])

img_transform_target = transforms.Compose([
    transforms.Resize(image_size),
    transforms.ToTensor(),
    transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])

dataset_source = datasets.MNIST(
    root='dataset',
    train=True,
    transform=img_transform_source,
    download=True
)

dataloader_source = torch.utils.data.DataLoader(
    dataset=dataset_source,
    batch_size=batch_size,
    shuffle=True,
    num_workers=8)

train_list = os.path.join(target_image_root, 'mnist_m_train_labels.txt')

dataset_target = GetLoader(
    data_root=os.path.join(target_image_root, 'mnist_m_train'),
    data_list=train_list,
    transform=img_transform_target
)

dataloader_target = torch.utils.data.DataLoader(
    dataset=dataset_target,
    batch_size=batch_size,
    shuffle=True,
    num_workers=8)

# load model

my_net = CNNModel()

# setup optimizer

optimizer = optim.Adam(my_net.parameters(), lr=lr)

loss_class = torch.nn.NLLLoss()
loss_domain = torch.nn.NLLLoss()

if cuda:
    my_net = my_net.cuda()
    loss_class = loss_class.cuda()
    loss_domain = loss_domain.cuda()

for p in my_net.parameters():
    p.requires_grad = True

# training

for epoch in range(n_epoch):

    len_dataloader = min(len(dataloader_source), len(dataloader_target))
    data_source_iter = iter(dataloader_source)
    data_target_iter = iter(dataloader_target)

    i = 0
    while i < len_dataloader:

        p = float(i + epoch * len_dataloader) / n_epoch / len_dataloader
        alpha = 2. / (1. + np.exp(-10 * p)) - 1

        # training model using source data
        data_source = data_source_iter.next()
        s_img, s_label = data_source

        my_net.zero_grad()
        batch_size = len(s_label)

        input_img = torch.FloatTensor(batch_size, 3, image_size, image_size)
        class_label = torch.LongTensor(batch_size)
        domain_label = torch.zeros(batch_size)
        domain_label = domain_label.long()

        if cuda:
            s_img = s_img.cuda()
            s_label = s_label.cuda()
            input_img = input_img.cuda()
            class_label = class_label.cuda()
            domain_label = domain_label.cuda()

        input_img.resize_as_(s_img).copy_(s_img)
        class_label.resize_as_(s_label).copy_(s_label)

        class_output, domain_output = my_net(input_data=input_img, alpha=alpha)
        err_s_label = loss_class(class_output, class_label)
        err_s_domain = loss_domain(domain_output, domain_label)

        # training model using target data
        data_target = data_target_iter.next()
        t_img, _ = data_target

        batch_size = len(t_img)

        input_img = torch.FloatTensor(batch_size, 3, image_size, image_size)
        domain_label = torch.ones(batch_size)
        domain_label = domain_label.long()

        if cuda:
            t_img = t_img.cuda()
            input_img = input_img.cuda()
            domain_label = domain_label.cuda()

        input_img.resize_as_(t_img).copy_(t_img)

        _, domain_output = my_net(input_data=input_img, alpha=alpha)
        err_t_domain = loss_domain(domain_output, domain_label)
        err = err_t_domain + err_s_domain + err_s_label
        err.backward()
        optimizer.step()

        i += 1

        print ('epoch: %d, [iter: %d / all %d], err_s_label: %f, err_s_domain: %f, err_t_domain: %f' \
              % (epoch, i, len_dataloader, err_s_label.data.cpu().numpy(),
                 err_s_domain.data.cpu().numpy(), err_t_domain.data.cpu().item()))

    torch.save(my_net, '{0}/mnist_mnistm_model_epoch_{1}.pth'.format(model_root, epoch))
    test(source_dataset_name, epoch)
    test(target_dataset_name, epoch)

print('done')