File size: 7,900 Bytes
8d6cd57
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
from typing import Optional, Dict
from functools import partial
import math

import torch
import torch.nn as nn


def get_activation(activation: str = "lrelu"):
    actv_layers = {
        "relu":  nn.ReLU,
        "lrelu": partial(nn.LeakyReLU, 0.2),
    }
    assert activation in actv_layers, f"activation [{activation}] not implemented"
    return actv_layers[activation]


def get_normalization(normalization: str = "batch_norm"):
    norm_layers = {
        "instance_norm": nn.InstanceNorm2d,
        "batch_norm":    nn.BatchNorm2d,
        "group_norm":    partial(nn.GroupNorm, num_groups=8),
        "layer_norm":    partial(nn.GroupNorm, num_groups=1),
    }
    assert normalization in norm_layers, f"normalization [{normalization}] not implemented"
    return norm_layers[normalization]


class ConvLayer(nn.Sequential):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: int = 3,
        stride: int = 1,
        padding: Optional[int] = 1,
        padding_mode: str = "zeros",
        groups: int = 1,
        bias: bool = True,
        transposed: bool = False,
        normalization: Optional[str] = None,
        activation: Optional[str] = "lrelu",
        pre_activate: bool = False,
    ):
        if transposed:
            conv = partial(nn.ConvTranspose2d, output_padding=stride-1)
            padding_mode = "zeros"
        else:
            conv = nn.Conv2d
        layers = [
            conv(
                in_channels,
                out_channels,
                kernel_size=kernel_size,
                stride=stride,
                padding=padding,
                padding_mode=padding_mode,
                groups=groups,
                bias=bias,
            )
        ]

        norm_actv = []
        if normalization is not None:
            norm_actv.append(
                get_normalization(normalization)(
                    num_channels=in_channels if pre_activate else out_channels
                )
            )
        if activation is not None:
            norm_actv.append(
                get_activation(activation)(inplace=True)
            )

        if pre_activate:
            layers = norm_actv + layers
        else:
            layers = layers + norm_actv

        super().__init__(
            *layers
        )


class SubspaceLayer(nn.Module):
    def __init__(
        self,
        dim: int,
        n_basis: int,
    ):
        super().__init__()

        self.U = nn.Parameter(torch.empty(n_basis, dim))
        nn.init.orthogonal_(self.U)
        self.L = nn.Parameter(torch.FloatTensor([3 * i for i in range(n_basis, 0, -1)]))
        self.mu = nn.Parameter(torch.zeros(dim))

    def forward(self, z):
        return (self.L * z) @ self.U + self.mu


class EigenBlock(nn.Module):
    def __init__(
        self,
        width: int,
        height: int,
        in_channels: int,
        out_channels: int,
        n_basis: int,
    ):
        super().__init__()

        self.projection = SubspaceLayer(dim=width*height*in_channels, n_basis=n_basis)
        self.subspace_conv1 = ConvLayer(
            in_channels,
            in_channels,
            kernel_size=1,
            stride=1,
            padding=0,
            transposed=True,
            activation=None,
            normalization=None,
        )
        self.subspace_conv2 = ConvLayer(
            in_channels,
            out_channels,
            kernel_size=3,
            stride=2,
            padding=1,
            transposed=True,
            activation=None,
            normalization=None,
        )

        self.feature_conv1 = ConvLayer(
            in_channels,
            out_channels,
            kernel_size=3,
            stride=2,
            transposed=True,
            pre_activate=True,
        )
        self.feature_conv2 = ConvLayer(
            out_channels,
            out_channels,
            kernel_size=3,
            stride=1,
            transposed=True,
            pre_activate=True,
        )

    def forward(self, z, h):
        phi = self.projection(z).view(h.shape)
        h = self.feature_conv1(h + self.subspace_conv1(phi))
        h = self.feature_conv2(h + self.subspace_conv2(phi))
        return h


class ConditionalGenerator(nn.Module):

    """Conditional generator
    It generates images from one hot label + noise sampled from N(0, 1) with explorable z injection space
    Based on EigenGAN
    """

    def __init__(self,
                 size: int,
                 y_size: int,
                 z_size: int,
                 out_channels: int = 3,
                 n_basis: int = 6,
                 noise_dim: int = 512,
                 base_channels: int = 16,
                 max_channels: int = 512,
                 y_type: str = 'one_hot'):

        if y_type not in ['one_hot', 'multi_label', 'mixed', 'real']:
            raise ValueError('Unsupported `y_type`')

        super(ConditionalGenerator, self).__init__()

        assert (size & (size - 1) == 0) and size != 0, "img size should be a power of 2"

        self.y_type = y_type
        self.y_size = y_size
        self.eps_size = z_size

        self.noise_dim = noise_dim
        self.n_basis = n_basis
        self.n_blocks = int(math.log(size, 2)) - 2

        def get_channels(i_block):
            return min(max_channels, base_channels * (2 ** (self.n_blocks - i_block)))

        self.y_fc = nn.Linear(self.y_size, self.y_size)
        self.concat_fc = nn.Linear(self.y_size + self.eps_size, self.noise_dim)

        self.fc = nn.Linear(self.noise_dim, 4 * 4 * get_channels(0))

        self.blocks = nn.ModuleList()
        for i in range(self.n_blocks):
            self.blocks.append(
                EigenBlock(
                    width=4 * (2 ** i),
                    height=4 * (2 ** i),
                    in_channels=get_channels(i),
                    out_channels=get_channels(i + 1),
                    n_basis=self.n_basis,
                )
            )

        self.out = nn.Sequential(
            ConvLayer(base_channels, out_channels, kernel_size=7, stride=1, padding=3, pre_activate=True),
            nn.Tanh(),
        )

    def forward(self,
                y: torch.Tensor,
                eps: Optional[torch.Tensor] = None,
                zs: Optional[torch.Tensor] = None,
                return_eps: bool = False):

        bs = y.size(0)

        if eps is None:
            eps = self.sample_eps(bs)

        if zs is None:
            zs = self.sample_zs(bs)

        y_out = self.y_fc(y)
        concat = torch.cat((y_out, eps), dim=1)
        concat = self.concat_fc(concat)

        out = self.fc(concat).view(len(eps), -1, 4, 4)
        for block, z in zip(self.blocks, zs.permute(1, 0, 2)):
            out = block(z, out)
        out = self.out(out)

        if return_eps:
            return out, concat

        return out

    def sample_zs(self, batch: int, truncation: float = 1.):
        device = self.get_device()
        zs = torch.randn(batch, self.n_blocks, self.n_basis, device=device)

        if truncation < 1.:
            zs = torch.zeros_like(zs) * (1 - truncation) + zs * truncation
        return zs

    def sample_eps(self, batch: int, truncation: float = 1.):
        device = self.get_device()
        eps = torch.randn(batch, self.eps_size, device=device)

        if truncation < 1.:
            eps = torch.zeros_like(eps) * (1 - truncation) + eps * truncation
        return eps

    def get_device(self):
        return self.fc.weight.device

    def orthogonal_regularizer(self):
        reg = []
        for layer in self.modules():
            if isinstance(layer, SubspaceLayer):
                UUT = layer.U @ layer.U.t()
                reg.append(
                    ((UUT - torch.eye(UUT.shape[0], device=UUT.device)) ** 2).mean()
                )
        return sum(reg) / len(reg)