CSTnet/models.py at master · bing199/CSTnet · 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
import torch
import torch.nn as nn
from transfer_losses import TransferLoss
import backbones

class NewTransferNet(nn.Module):
    def __init__(self, num_class, base_net='rsp_resnet50', transfer_loss='mmd', use_bottleneck=True, bottleneck_width=256, max_iter=1000, input_channels=3,**kwargs):
        super(NewTransferNet, self).__init__()
        self.num_class = num_class
        self.base_network = backbones.get_backbone(base_net, input_channels=input_channels)
        self.use_bottleneck = use_bottleneck
        self.transfer_loss = transfer_loss
        if self.use_bottleneck:
            bottleneck_list = [
                nn.Linear(self.base_network.output_num(), bottleneck_width),
                nn.ReLU()
            ]
            self.bottleneck_layer = nn.Sequential(*bottleneck_list)
            feature_dim = bottleneck_width
        else:
            feature_dim = self.base_network.output_num()

        self.classifier_layer = nn.Linear(feature_dim, num_class)
        transfer_loss_args = {
            "loss_type": self.transfer_loss,
            "max_iter": max_iter,
            "num_class": num_class
        }
        self.adapt_loss = TransferLoss(**transfer_loss_args)
        self.criterion = torch.nn.CrossEntropyLoss()

    def forward(self, source, target, source_label):
        source_outputs, \
        source_x_IN_1, source_x_1, source_x_style_1a, \
        source_x_IN_2, source_x_2, source_x_style_2a, \
        source_x_IN_3, source_x_3, source_x_style_3a = self.base_network(source)

        target_outputs, \
        target_x_IN_1, target_x_1, target_x_style_1a, \
        target_x_IN_2, target_x_2, target_x_style_2a, \
        target_x_IN_3, target_x_3, target_x_style_3a = self.base_network(target)

        if self.use_bottleneck:
            source = self.bottleneck_layer(source_outputs)
            target = self.bottleneck_layer(target_outputs)

        # classification loss
        source_clf = self.classifier_layer(source)
        clf_loss = self.criterion(source_clf, source_label)

        # dis_loss
        dis_criterion = torch.nn.SmoothL1Loss()

        dis_content = dis_criterion(source_x_IN_1, source_x_1) \
                + dis_criterion(source_x_IN_2, source_x_2) \
                + dis_criterion(source_x_IN_3, source_x_3) \
                + dis_criterion(target_x_IN_1, target_x_1) \
                + dis_criterion(target_x_IN_2, target_x_2) \
                + dis_criterion(target_x_IN_3, target_x_3)
        dis_style = dis_criterion(source_x_style_1a, source_x_1) \
                    + dis_criterion(source_x_style_2a, source_x_2) \
                    + dis_criterion(source_x_style_3a, source_x_3) \
                    + dis_criterion(target_x_style_1a, target_x_1) \
                    + dis_criterion(target_x_style_2a, target_x_2) \
                    + dis_criterion(target_x_style_3a, target_x_3)
        """
        dis_content = dis_criterion(target_x_IN_1, target_x_1) \
                + dis_criterion(target_x_IN_2, target_x_2) \
                + dis_criterion(target_x_IN_3, target_x_3)
        dis_style = dis_criterion(target_x_style_1a, target_x_1) \
                    + dis_criterion(target_x_style_2a, target_x_2) \
                    + dis_criterion(target_x_style_3a, target_x_3)
        """
        dis_loss = dis_content - dis_style
        # dis_loss = dis_content
        # dis_loss = - dis_style

        """
        # class-level alignment loss
        target_clf = self.classifier_layer(target)
        source_clf = self.classifier_layer(source)
        source_probs = torch.softmax(source_clf, dim=1)
        class_loss = torch.sum(source_probs * torch.nn.functional.cross_entropy(target_clf, source_label))
        class_loss = class_loss / source_probs.size(0)
        """

        # transfer loss
        kwargs = {}
        if self.transfer_loss == "lmmd":
            kwargs['source_label'] = source_label
            target_clf = self.classifier_layer(target)
            kwargs['target_logits'] = torch.nn.functional.softmax(target_clf, dim=1)
        elif self.transfer_loss == "daan":
            source_clf = self.classifier_layer(source)
            kwargs['source_logits'] = torch.nn.functional.softmax(source_clf, dim=1)
            target_clf = self.classifier_layer(target)
            kwargs['target_logits'] = torch.nn.functional.softmax(target_clf, dim=1)
        elif self.transfer_loss == 'bnm':
            tar_clf = self.classifier_layer(target)
            target = nn.Softmax(dim=1)(tar_clf)

        transfer_loss = self.adapt_loss(source, target, **kwargs)

        return clf_loss, dis_loss, transfer_loss

    def get_parameters(self, initial_lr=1.0):
        params = [
            {'params': self.base_network.parameters(), 'lr': 0.1 * initial_lr},
            {'params': self.classifier_layer.parameters(), 'lr': 1.0 * initial_lr},
        ]
        if self.use_bottleneck:
            params.append(
                {'params': self.bottleneck_layer.parameters(), 'lr': 1.0 * initial_lr}
            )
        # Loss-dependent
        if self.transfer_loss == "adv":
            params.append(
                {'params': self.adapt_loss.loss_func.domain_classifier.parameters(), 'lr': 1.0 * initial_lr}
            )
        elif self.transfer_loss == "daan":
            params.append(
                {'params': self.adapt_loss.loss_func.domain_classifier.parameters(), 'lr': 1.0 * initial_lr}
            )
            params.append(
                {'params': self.adapt_loss.loss_func.local_classifiers.parameters(), 'lr': 1.0 * initial_lr}
            )
        return params

    def predict(self, x):
        x_4, \
        x_IN_1, x_1, x_style_1a, \
        x_IN_2, x_2, x_style_2a, \
        x_IN_3, x_3, x_style_3a = self.base_network(x)
        x = self.bottleneck_layer(x_4)
        clf = self.classifier_layer(x)
        return clf

    def get_features(self,x):
        x_4, \
        x_IN_1, x_1, x_style_1a, \
        x_IN_2, x_2, x_style_2a, \
        x_IN_3, x_3, x_style_3a = self.base_network(x)
        return self.bottleneck_layer(x_4)

    def epoch_based_processing(self, *args, **kwargs):
        if self.transfer_loss == "daan":
            self.adapt_loss.loss_func.update_dynamic_factor(*args, **kwargs)
        else:
            pass