fantasy-autoencoder/model.py at main · MyNameIsArko/fantasy-autoencoder · 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
from torch import nn
import torch
from utils import reparameterize


class ResidualBlock(nn.Module):
    def __init__(self, inc=64, outc=64):
        super(ResidualBlock, self).__init__()

        if inc != outc:
            self.conv_expand = nn.Conv2d(in_channels=inc, out_channels=outc, kernel_size=1, stride=1, padding=0,
                                         bias=False)
        else:
            self.conv_expand = None

        self.conv = nn.Sequential(
            nn.Conv2d(in_channels=inc, out_channels=outc, kernel_size=3, stride=1, padding=1,
                      bias=False),
            nn.BatchNorm2d(outc),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Conv2d(in_channels=outc, out_channels=outc, kernel_size=3, stride=1, padding=1,
                      bias=False),
            nn.BatchNorm2d(outc)
        )
        self.relu = nn.LeakyReLU(0.2, inplace=True)

    def forward(self, x):
        if self.conv_expand is not None:
            identity_data = self.conv_expand(x)
        else:
            identity_data = x

        output = self.conv(x)
        output = self.relu(torch.add(output, identity_data))
        return output


class Encoder(nn.Module):
    def __init__(self, cdim=3, zdim=512, channels=(64, 128, 256, 512, 512, 512), image_size=256):
        super(Encoder, self).__init__()
        self.zdim = zdim
        self.cdim = cdim
        self.image_size = image_size
        cc = channels[0]
        layers = [
            nn.Conv2d(cdim, cc, 5, 1, 2, bias=False),
            nn.BatchNorm2d(cc),
            nn.LeakyReLU(0.2),
            nn.AvgPool2d(2),
        ]

        sz = image_size // 2
        for ch in channels[1:]:
            layers.append(ResidualBlock(cc, ch))
            layers.append(nn.AvgPool2d(2))
            cc, sz = ch, sz // 2

        layers.append(ResidualBlock(cc, cc))
        self.main = nn.Sequential(*layers)
        self.conv_output_size = self.calc_conv_output_size()
        num_fc_features = torch.zeros(self.conv_output_size).view(-1).shape[0]
        # print("conv shape: ", self.conv_output_size)
        # print("num fc features: ", num_fc_features)
        self.fc = nn.Linear(num_fc_features, 2 * zdim)

    def calc_conv_output_size(self):
        dummy_input = torch.zeros(1, self.cdim, self.image_size, self.image_size)
        dummy_input = self.main(dummy_input)
        return dummy_input[0].shape

    def forward(self, x):
        y = self.main(x).view(x.size(0), -1)
        y = self.fc(y)
        mu, logvar = y.chunk(2, dim=1)
        return mu, logvar


class Decoder(nn.Module):
    def __init__(self, cdim=3, zdim=512, channels=(64, 128, 256, 512, 512, 512), image_size=256,
                 conv_input_size=None):
        super(Decoder, self).__init__()
        self.cdim = cdim
        self.image_size = image_size
        cc = channels[-1]
        self.conv_input_size = conv_input_size
        if conv_input_size is None:
            num_fc_features = cc * 4 * 4
        else:
            num_fc_features = torch.zeros(self.conv_input_size).view(-1).shape[0]
        self.fc = nn.Sequential(
            nn.Linear(zdim, num_fc_features),
            nn.ReLU(True),
        )

        sz = 4

        self.main = nn.Sequential()
        layers = []
        for ch in channels[::-1]:
            layers.append(ResidualBlock(cc, ch))
            layers.append(nn.Upsample(scale_factor=2, mode='nearest'))
            cc, sz = ch, sz * 2

        layers.append(ResidualBlock(cc, cc))
        layers.append(nn.Conv2d(cc, cdim, 5, 1, 2))

        self.main = nn.Sequential(*layers)

    def forward(self, z):
        z = z.view(z.size(0), -1)
        y = self.fc(z)
        y = y.view(z.size(0), *self.conv_input_size)
        y = self.main(y)
        return y


class SoftIntroVAE(nn.Module):
    def __init__(self, cdim=3, zdim=512, channels=(64, 128, 256, 512, 512, 512), image_size=256,
                 cond_dim=10):
        super(SoftIntroVAE, self).__init__()

        self.zdim = zdim
        self.cond_dim = cond_dim

        self.encoder = Encoder(cdim, zdim, channels, image_size)

        self.decoder = Decoder(cdim, zdim, channels, image_size,
                               conv_input_size=self.encoder.conv_output_size)

    def forward(self, x, deterministic=False):
        mu, logvar = self.encode(x)
        if deterministic:
            z = mu
        else:
            z = reparameterize(mu, logvar)
        y = self.decode(z)
        return mu, logvar, z, y

    def sample(self, z):
        y = self.decode(z)
        return y

    def sample_with_noise(self, num_samples=1, device=torch.device("cpu")):
        z = torch.randn(num_samples, self.zdim).to(device)
        return self.decode(z)

    def encode(self, x):
        mu, logvar = self.encoder(x)
        return mu, logvar

    def decode(self, z):
        y = self.decoder(z)
        return y