Real-time collaboration for Jupyter Notebooks, Linux Terminals, LaTeX, VS Code, R IDE, and more,
all in one place. Commercial Alternative to JupyterHub.

1
import math
2
import torch
3
import torch.nn.functional as F
4
from torch import nn
5
from torch.nn import Parameter
6
from .config import device, num_classes
7

8

9
def create_model(opt):
10
    if opt.model == 'pix2pixHD':
11
        #from .pix2pixHD_model import Pix2PixHDModel, InferenceModel
12
        from .fs_model import fsModel
13
        model = fsModel()
14
    else:
15
        from .ui_model import UIModel
16
        model = UIModel()
17

18
    model.initialize(opt)
19
    if opt.verbose:
20
        print("model [%s] was created" % (model.name()))
21

22
    if opt.isTrain and len(opt.gpu_ids) and not opt.fp16:
23
        model = torch.nn.DataParallel(model, device_ids=opt.gpu_ids)
24

25
    return model
26

27

28

29
class SEBlock(nn.Module):
30
    def __init__(self, channel, reduction=16):
31
        super(SEBlock, self).__init__()
32
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
33
        self.fc = nn.Sequential(
34
            nn.Linear(channel, channel // reduction),
35
            nn.PReLU(),
36
            nn.Linear(channel // reduction, channel),
37
            nn.Sigmoid()
38
        )
39

40
    def forward(self, x):
41
        b, c, _, _ = x.size()
42
        y = self.avg_pool(x).view(b, c)
43
        y = self.fc(y).view(b, c, 1, 1)
44
        return x * y
45

46

47
class IRBlock(nn.Module):
48
    expansion = 1
49

50
    def __init__(self, inplanes, planes, stride=1, downsample=None, use_se=True):
51
        super(IRBlock, self).__init__()
52
        self.bn0 = nn.BatchNorm2d(inplanes)
53
        self.conv1 = conv3x3(inplanes, inplanes)
54
        self.bn1 = nn.BatchNorm2d(inplanes)
55
        self.prelu = nn.PReLU()
56
        self.conv2 = conv3x3(inplanes, planes, stride)
57
        self.bn2 = nn.BatchNorm2d(planes)
58
        self.downsample = downsample
59
        self.stride = stride
60
        self.use_se = use_se
61
        if self.use_se:
62
            self.se = SEBlock(planes)
63

64
    def forward(self, x):
65
        residual = x
66
        out = self.bn0(x)
67
        out = self.conv1(out)
68
        out = self.bn1(out)
69
        out = self.prelu(out)
70

71
        out = self.conv2(out)
72
        out = self.bn2(out)
73
        if self.use_se:
74
            out = self.se(out)
75

76
        if self.downsample is not None:
77
            residual = self.downsample(x)
78

79
        out += residual
80
        out = self.prelu(out)
81

82
        return out
83

84

85
class ResNet(nn.Module):
86

87
    def __init__(self, block, layers, use_se=True):
88
        self.inplanes = 64
89
        self.use_se = use_se
90
        super(ResNet, self).__init__()
91
        self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, bias=False)
92
        self.bn1 = nn.BatchNorm2d(64)
93
        self.prelu = nn.PReLU()
94
        self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2)
95
        self.layer1 = self._make_layer(block, 64, layers[0])
96
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
97
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
98
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
99
        self.bn2 = nn.BatchNorm2d(512)
100
        self.dropout = nn.Dropout()
101
        self.fc = nn.Linear(512 * 7 * 7, 512)
102
        self.bn3 = nn.BatchNorm1d(512)
103

104
        for m in self.modules():
105
            if isinstance(m, nn.Conv2d):
106
                nn.init.xavier_normal_(m.weight)
107
            elif isinstance(m, nn.BatchNorm2d) or isinstance(m, nn.BatchNorm1d):
108
                nn.init.constant_(m.weight, 1)
109
                nn.init.constant_(m.bias, 0)
110
            elif isinstance(m, nn.Linear):
111
                nn.init.xavier_normal_(m.weight)
112
                nn.init.constant_(m.bias, 0)
113

114
    def _make_layer(self, block, planes, blocks, stride=1):
115
        downsample = None
116
        if stride != 1 or self.inplanes != planes * block.expansion:
117
            downsample = nn.Sequential(
118
                nn.Conv2d(self.inplanes, planes * block.expansion,
119
                          kernel_size=1, stride=stride, bias=False),
120
                nn.BatchNorm2d(planes * block.expansion),
121
            )
122

123
        layers = []
124
        layers.append(block(self.inplanes, planes, stride, downsample, use_se=self.use_se))
125
        self.inplanes = planes
126
        for i in range(1, blocks):
127
            layers.append(block(self.inplanes, planes, use_se=self.use_se))
128

129
        return nn.Sequential(*layers)
130

131
    def forward(self, x):
132
        x = self.conv1(x)
133
        x = self.bn1(x)
134
        x = self.prelu(x)
135
        x = self.maxpool(x)
136

137
        x = self.layer1(x)
138
        x = self.layer2(x)
139
        x = self.layer3(x)
140
        x = self.layer4(x)
141

142
        x = self.bn2(x)
143
        x = self.dropout(x)
144
        x = x.view(x.size(0), -1)
145
        x = self.fc(x)
146
        x = self.bn3(x)
147

148
        return x
149

150

151
class ArcMarginModel(nn.Module):
152
    def __init__(self, args):
153
        super(ArcMarginModel, self).__init__()
154

155
        self.weight = Parameter(torch.FloatTensor(num_classes, args.emb_size))
156
        nn.init.xavier_uniform_(self.weight)
157

158
        self.easy_margin = args.easy_margin
159
        self.m = args.margin_m
160
        self.s = args.margin_s
161

162
        self.cos_m = math.cos(self.m)
163
        self.sin_m = math.sin(self.m)
164
        self.th = math.cos(math.pi - self.m)
165
        self.mm = math.sin(math.pi - self.m) * self.m
166

167
    def forward(self, input, label):
168
        x = F.normalize(input)
169
        W = F.normalize(self.weight)
170
        cosine = F.linear(x, W)
171
        sine = torch.sqrt(1.0 - torch.pow(cosine, 2))
172
        phi = cosine * self.cos_m - sine * self.sin_m  # cos(theta + m)
173
        if self.easy_margin:
174
            phi = torch.where(cosine > 0, phi, cosine)
175
        else:
176
            phi = torch.where(cosine > self.th, phi, cosine - self.mm)
177
        one_hot = torch.zeros(cosine.size(), device=device)
178
        one_hot.scatter_(1, label.view(-1, 1).long(), 1)
179
        output = (one_hot * phi) + ((1.0 - one_hot) * cosine)
180
        output *= self.s
181
        return output
182

183