File size: 14,317 Bytes
2a8871c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
import math
import random
import torch
from basicsr.archs.arch_util import default_init_weights
from basicsr.utils.registry import ARCH_REGISTRY
from torch import nn
from torch.nn import functional as F


class NormStyleCode(nn.Module):

    def forward(self, x):
        """Normalize the style codes.

        Args:
            x (Tensor): Style codes with shape (b, c).

        Returns:
            Tensor: Normalized tensor.
        """
        return x * torch.rsqrt(torch.mean(x**2, dim=1, keepdim=True) + 1e-8)


class ModulatedConv2d(nn.Module):
    """Modulated Conv2d used in StyleGAN2.

    There is no bias in ModulatedConv2d.

    Args:
        in_channels (int): Channel number of the input.
        out_channels (int): Channel number of the output.
        kernel_size (int): Size of the convolving kernel.
        num_style_feat (int): Channel number of style features.
        demodulate (bool): Whether to demodulate in the conv layer. Default: True.
        sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None.
        eps (float): A value added to the denominator for numerical stability. Default: 1e-8.
    """

    def __init__(self,
                 in_channels,
                 out_channels,
                 kernel_size,
                 num_style_feat,
                 demodulate=True,
                 sample_mode=None,
                 eps=1e-8):
        super(ModulatedConv2d, self).__init__()
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.kernel_size = kernel_size
        self.demodulate = demodulate
        self.sample_mode = sample_mode
        self.eps = eps

        # modulation inside each modulated conv
        self.modulation = nn.Linear(num_style_feat, in_channels, bias=True)
        # initialization
        default_init_weights(self.modulation, scale=1, bias_fill=1, a=0, mode='fan_in', nonlinearity='linear')

        self.weight = nn.Parameter(
            torch.randn(1, out_channels, in_channels, kernel_size, kernel_size) /
            math.sqrt(in_channels * kernel_size**2))
        self.padding = kernel_size // 2

    def forward(self, x, style):
        """Forward function.

        Args:
            x (Tensor): Tensor with shape (b, c, h, w).
            style (Tensor): Tensor with shape (b, num_style_feat).

        Returns:
            Tensor: Modulated tensor after convolution.
        """
        b, c, h, w = x.shape  # c = c_in
        # weight modulation
        style = self.modulation(style).view(b, 1, c, 1, 1)
        # self.weight: (1, c_out, c_in, k, k); style: (b, 1, c, 1, 1)
        weight = self.weight * style  # (b, c_out, c_in, k, k)

        if self.demodulate:
            demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps)
            weight = weight * demod.view(b, self.out_channels, 1, 1, 1)

        weight = weight.view(b * self.out_channels, c, self.kernel_size, self.kernel_size)

        # upsample or downsample if necessary
        if self.sample_mode == 'upsample':
            x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
        elif self.sample_mode == 'downsample':
            x = F.interpolate(x, scale_factor=0.5, mode='bilinear', align_corners=False)

        b, c, h, w = x.shape
        x = x.view(1, b * c, h, w)
        # weight: (b*c_out, c_in, k, k), groups=b
        out = F.conv2d(x, weight, padding=self.padding, groups=b)
        out = out.view(b, self.out_channels, *out.shape[2:4])

        return out

    def __repr__(self):
        return (f'{self.__class__.__name__}(in_channels={self.in_channels}, out_channels={self.out_channels}, '
                f'kernel_size={self.kernel_size}, demodulate={self.demodulate}, sample_mode={self.sample_mode})')


class StyleConv(nn.Module):
    """Style conv used in StyleGAN2.

    Args:
        in_channels (int): Channel number of the input.
        out_channels (int): Channel number of the output.
        kernel_size (int): Size of the convolving kernel.
        num_style_feat (int): Channel number of style features.
        demodulate (bool): Whether demodulate in the conv layer. Default: True.
        sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None.
    """

    def __init__(self, in_channels, out_channels, kernel_size, num_style_feat, demodulate=True, sample_mode=None):
        super(StyleConv, self).__init__()
        self.modulated_conv = ModulatedConv2d(
            in_channels, out_channels, kernel_size, num_style_feat, demodulate=demodulate, sample_mode=sample_mode)
        self.weight = nn.Parameter(torch.zeros(1))  # for noise injection
        self.bias = nn.Parameter(torch.zeros(1, out_channels, 1, 1))
        self.activate = nn.LeakyReLU(negative_slope=0.2, inplace=True)

    def forward(self, x, style, noise=None):
        # modulate
        out = self.modulated_conv(x, style) * 2**0.5  # for conversion
        # noise injection
        if noise is None:
            b, _, h, w = out.shape
            noise = out.new_empty(b, 1, h, w).normal_()
        out = out + self.weight * noise
        # add bias
        out = out + self.bias
        # activation
        out = self.activate(out)
        return out


class ToRGB(nn.Module):
    """To RGB (image space) from features.

    Args:
        in_channels (int): Channel number of input.
        num_style_feat (int): Channel number of style features.
        upsample (bool): Whether to upsample. Default: True.
    """

    def __init__(self, in_channels, num_style_feat, upsample=True):
        super(ToRGB, self).__init__()
        self.upsample = upsample
        self.modulated_conv = ModulatedConv2d(
            in_channels, 3, kernel_size=1, num_style_feat=num_style_feat, demodulate=False, sample_mode=None)
        self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))

    def forward(self, x, style, skip=None):
        """Forward function.

        Args:
            x (Tensor): Feature tensor with shape (b, c, h, w).
            style (Tensor): Tensor with shape (b, num_style_feat).
            skip (Tensor): Base/skip tensor. Default: None.

        Returns:
            Tensor: RGB images.
        """
        out = self.modulated_conv(x, style)
        out = out + self.bias
        if skip is not None:
            if self.upsample:
                skip = F.interpolate(skip, scale_factor=2, mode='bilinear', align_corners=False)
            out = out + skip
        return out


class ConstantInput(nn.Module):
    """Constant input.

    Args:
        num_channel (int): Channel number of constant input.
        size (int): Spatial size of constant input.
    """

    def __init__(self, num_channel, size):
        super(ConstantInput, self).__init__()
        self.weight = nn.Parameter(torch.randn(1, num_channel, size, size))

    def forward(self, batch):
        out = self.weight.repeat(batch, 1, 1, 1)
        return out


@ARCH_REGISTRY.register()
class StyleGAN2GeneratorClean(nn.Module):
    """Clean version of StyleGAN2 Generator.

    Args:
        out_size (int): The spatial size of outputs.
        num_style_feat (int): Channel number of style features. Default: 512.
        num_mlp (int): Layer number of MLP style layers. Default: 8.
        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
        narrow (float): Narrow ratio for channels. Default: 1.0.
    """

    def __init__(self, out_size, num_style_feat=512, num_mlp=8, channel_multiplier=2, narrow=1):
        super(StyleGAN2GeneratorClean, self).__init__()
        # Style MLP layers
        self.num_style_feat = num_style_feat
        style_mlp_layers = [NormStyleCode()]
        for i in range(num_mlp):
            style_mlp_layers.extend(
                [nn.Linear(num_style_feat, num_style_feat, bias=True),
                 nn.LeakyReLU(negative_slope=0.2, inplace=True)])
        self.style_mlp = nn.Sequential(*style_mlp_layers)
        # initialization
        default_init_weights(self.style_mlp, scale=1, bias_fill=0, a=0.2, mode='fan_in', nonlinearity='leaky_relu')

        # channel list
        channels = {
            '4': int(512 * narrow),
            '8': int(512 * narrow),
            '16': int(512 * narrow),
            '32': int(512 * narrow),
            '64': int(256 * channel_multiplier * narrow),
            '128': int(128 * channel_multiplier * narrow),
            '256': int(64 * channel_multiplier * narrow),
            '512': int(32 * channel_multiplier * narrow),
            '1024': int(16 * channel_multiplier * narrow)
        }
        self.channels = channels

        self.constant_input = ConstantInput(channels['4'], size=4)
        self.style_conv1 = StyleConv(
            channels['4'],
            channels['4'],
            kernel_size=3,
            num_style_feat=num_style_feat,
            demodulate=True,
            sample_mode=None)
        self.to_rgb1 = ToRGB(channels['4'], num_style_feat, upsample=False)

        self.log_size = int(math.log(out_size, 2))
        self.num_layers = (self.log_size - 2) * 2 + 1
        self.num_latent = self.log_size * 2 - 2

        self.style_convs = nn.ModuleList()
        self.to_rgbs = nn.ModuleList()
        self.noises = nn.Module()

        in_channels = channels['4']
        # noise
        for layer_idx in range(self.num_layers):
            resolution = 2**((layer_idx + 5) // 2)
            shape = [1, 1, resolution, resolution]
            self.noises.register_buffer(f'noise{layer_idx}', torch.randn(*shape))
        # style convs and to_rgbs
        for i in range(3, self.log_size + 1):
            out_channels = channels[f'{2**i}']
            self.style_convs.append(
                StyleConv(
                    in_channels,
                    out_channels,
                    kernel_size=3,
                    num_style_feat=num_style_feat,
                    demodulate=True,
                    sample_mode='upsample'))
            self.style_convs.append(
                StyleConv(
                    out_channels,
                    out_channels,
                    kernel_size=3,
                    num_style_feat=num_style_feat,
                    demodulate=True,
                    sample_mode=None))
            self.to_rgbs.append(ToRGB(out_channels, num_style_feat, upsample=True))
            in_channels = out_channels

    def make_noise(self):
        """Make noise for noise injection."""
        device = self.constant_input.weight.device
        noises = [torch.randn(1, 1, 4, 4, device=device)]

        for i in range(3, self.log_size + 1):
            for _ in range(2):
                noises.append(torch.randn(1, 1, 2**i, 2**i, device=device))

        return noises

    def get_latent(self, x):
        return self.style_mlp(x)

    def mean_latent(self, num_latent):
        latent_in = torch.randn(num_latent, self.num_style_feat, device=self.constant_input.weight.device)
        latent = self.style_mlp(latent_in).mean(0, keepdim=True)
        return latent

    def forward(self,
                styles,
                input_is_latent=False,
                noise=None,
                randomize_noise=True,
                truncation=1,
                truncation_latent=None,
                inject_index=None,
                return_latents=False):
        """Forward function for StyleGAN2GeneratorClean.

        Args:
            styles (list[Tensor]): Sample codes of styles.
            input_is_latent (bool): Whether input is latent style. Default: False.
            noise (Tensor | None): Input noise or None. Default: None.
            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
            truncation (float): The truncation ratio. Default: 1.
            truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
            inject_index (int | None): The injection index for mixing noise. Default: None.
            return_latents (bool): Whether to return style latents. Default: False.
        """
        # style codes -> latents with Style MLP layer
        if not input_is_latent:
            styles = [self.style_mlp(s) for s in styles]
        # noises
        if noise is None:
            if randomize_noise:
                noise = [None] * self.num_layers  # for each style conv layer
            else:  # use the stored noise
                noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
        # style truncation
        if truncation < 1:
            style_truncation = []
            for style in styles:
                style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
            styles = style_truncation
        # get style latents with injection
        if len(styles) == 1:
            inject_index = self.num_latent

            if styles[0].ndim < 3:
                # repeat latent code for all the layers
                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
            else:  # used for encoder with different latent code for each layer
                latent = styles[0]
        elif len(styles) == 2:  # mixing noises
            if inject_index is None:
                inject_index = random.randint(1, self.num_latent - 1)
            latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
            latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
            latent = torch.cat([latent1, latent2], 1)

        # main generation
        out = self.constant_input(latent.shape[0])
        out = self.style_conv1(out, latent[:, 0], noise=noise[0])
        skip = self.to_rgb1(out, latent[:, 1])

        i = 1
        for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
                                                        noise[2::2], self.to_rgbs):
            out = conv1(out, latent[:, i], noise=noise1)
            out = conv2(out, latent[:, i + 1], noise=noise2)
            skip = to_rgb(out, latent[:, i + 2], skip)  # feature back to the rgb space
            i += 2

        image = skip

        if return_latents:
            return image, latent
        else:
            return image, None