File size: 5,824 Bytes
998bb30
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import math
from attacks.CGNC.models.attention import SpatialTransformer

# To control feature map in generator
ngf = 64


def snconv2d(eps=1e-12, **kwargs):
    return nn.utils.spectral_norm(nn.Conv2d(**kwargs), eps=eps)


def snlinear(eps=1e-12, **kwargs):
    return nn.utils.spectral_norm(nn.Linear(**kwargs), eps=eps)


def get_gaussian_kernel(kernel_size=3, pad=2, sigma=1, channels=3):
    x_coord = torch.arange(kernel_size)
    x_grid = x_coord.repeat(kernel_size).view(kernel_size, kernel_size)
    y_grid = x_grid.t()
    xy_grid = torch.stack([x_grid, y_grid], dim=-1).float()

    mean = (kernel_size - 1) / 2.
    variance = sigma ** 2.
    gaussian_kernel = (1. / (2. * math.pi * variance)) * \
                      torch.exp(
                          -torch.sum((xy_grid - mean) ** 2., dim=-1) / \
                          (2 * variance)
                      )
    gaussian_kernel = gaussian_kernel / torch.sum(gaussian_kernel)
    gaussian_kernel = gaussian_kernel.view(1, 1, kernel_size, kernel_size)
    gaussian_kernel = gaussian_kernel.repeat(channels, 1, 1, 1)

    gaussian_filter = torch.nn.Conv2d(in_channels=channels, out_channels=channels,
                                      kernel_size=kernel_size, groups=channels, padding=kernel_size - pad, bias=False)

    gaussian_filter.weight.data = gaussian_kernel
    gaussian_filter.weight.requires_grad = False

    return gaussian_filter


class CrossAttenGenerator(nn.Module):
    def __init__(self, inception=False, device='cuda', num_head=1, nz=16, loc=[1, 1, 1], context_dim=512):
        '''
        :param inception: if True crop layer will be added to go from 3x300x300 to 3x299x299.
        '''
        super(CrossAttenGenerator, self).__init__()
        self.inception = inception
        self.device = device
        self.snlinear = snlinear(in_features=512, out_features=nz, bias=False)
        self.loc = loc

        self.block1 = nn.Sequential(
            nn.ReflectionPad2d(3),
            nn.Conv2d(3 + nz * self.loc[0], ngf, kernel_size=7, padding=0, bias=False),
            nn.BatchNorm2d(ngf),
            nn.ReLU(True)
        )
        self.block2 = nn.Sequential(
            nn.Conv2d(ngf + nz * self.loc[1], ngf * 2, kernel_size=3, stride=2, padding=1, bias=False),
            nn.BatchNorm2d(ngf * 2),
            nn.ReLU(True)

        )
        self.block3 = nn.Sequential(
            nn.Conv2d(ngf * 2 + nz * self.loc[2], ngf * 4, kernel_size=3, stride=2, padding=1, bias=False),
            nn.BatchNorm2d(ngf * 4),
            nn.ReLU(True)
        )
        self.resblock1 = ResidualBlock(ngf * 4)
        self.resblock2 = ResidualBlock(ngf * 4)
        self.cross_att2 = SpatialTransformer(ngf * 4, num_head, ngf * 4 // num_head, depth=1, context_dim=context_dim)
        self.resblock3 = ResidualBlock(ngf * 4)
        self.resblock4 = ResidualBlock(ngf * 4)
        self.cross_att4 = SpatialTransformer(ngf * 4, num_head, ngf * 4 // num_head, depth=1, context_dim=context_dim)
        self.resblock5 = ResidualBlock(ngf * 4)
        self.resblock6 = ResidualBlock(ngf * 4)

        self.upsampl1 = nn.Sequential(
            nn.ConvTranspose2d(ngf * 4, ngf * 2, kernel_size=3, stride=2, padding=1, output_padding=1, bias=False),
            nn.BatchNorm2d(ngf * 2),
            nn.ReLU(True)
        )
        self.upsampl2 = nn.Sequential(
            nn.ConvTranspose2d(ngf * 2, ngf, kernel_size=3, stride=2, padding=1, output_padding=1, bias=False),
            nn.BatchNorm2d(ngf),
            nn.ReLU(True)
        )
        self.blockf = nn.Sequential(
            nn.ReflectionPad2d(3),
            nn.Conv2d(ngf, 3, kernel_size=7, padding=0)
        )
        self.crop = nn.ConstantPad2d((0, -1, -1, 0), 0)
        self.alf_layer = get_gaussian_kernel(kernel_size=3, pad=2, sigma=1)

    def forward(self, input, cond, eps=16):
        text_cond = cond.unsqueeze(1).to(torch.float)
        z_cond = self.snlinear(cond.float())

        # loc 0
        z_img = z_cond[:, :, None, None].expand(z_cond.size(0), z_cond.size(1), input.size(2), input.size(3))
        x = self.block1(torch.cat((input, z_img), dim=1))

        # loc 1
        z_img = z_cond[:, :, None, None].expand(z_cond.size(0), z_cond.size(1), x.size(2), x.size(3))
        x = self.block2(torch.cat((x, z_img), dim=1)) if self.loc[1] else self.block2(x)

        # loc 2
        z_img = z_cond[:, :, None, None].expand(z_cond.size(0), z_cond.size(1), x.size(2), x.size(3))
        x = self.block3(torch.cat((x, z_img), dim=1)) if self.loc[2] else self.block3(x)

        x = self.resblock1(x)
        x = self.resblock2(x)
        x = self.cross_att2(x, text_cond)
        x = self.resblock3(x)
        x = self.resblock4(x)
        x = self.cross_att4(x, text_cond)
        x = self.resblock5(x)
        x = self.resblock6(x)

        x = self.upsampl1(x)
        x = self.upsampl2(x)
        x = self.blockf(x)
        if self.inception:
            x = self.crop(x)
        x = torch.tanh(x)
        x = self.alf_layer(x)

        return x * eps


class ResidualBlock(nn.Module):
    def __init__(self, num_filters):
        super(ResidualBlock, self).__init__()
        self.block = nn.Sequential(
            nn.ReflectionPad2d(1),
            nn.Conv2d(in_channels=num_filters, out_channels=num_filters, kernel_size=3, stride=1, padding=0,
                      bias=False),
            nn.BatchNorm2d(num_filters),
            nn.ReLU(True),

            nn.Dropout(0.5),

            nn.ReflectionPad2d(1),
            nn.Conv2d(in_channels=num_filters, out_channels=num_filters, kernel_size=3, stride=1, padding=0,
                      bias=False),
            nn.BatchNorm2d(num_filters)
        )

    def forward(self, x):
        residual = self.block(x)
        return x + residual