File size: 6,723 Bytes
4efe6b5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from torch.nn.utils.parametrizations import spectral_norm, weight_norm

from rvc.lib.algorithm.commons import get_padding
from rvc.lib.algorithm.residuals import LRELU_SLOPE


class MultiPeriodDiscriminator(torch.nn.Module):
    """
    Multi-period discriminator.

    This class implements a multi-period discriminator, which is used to
    discriminate between real and fake audio signals. The discriminator
    is composed of a series of convolutional layers that are applied to
    the input signal at different periods.

    Args:
        use_spectral_norm (bool): Whether to use spectral normalization.
            Defaults to False.
    """

    def __init__(self, use_spectral_norm=False):
        super(MultiPeriodDiscriminator, self).__init__()
        periods = [2, 3, 5, 7, 11, 17]
        self.discriminators = torch.nn.ModuleList(
            [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
            + [DiscriminatorP(p, use_spectral_norm=use_spectral_norm) for p in periods]
        )

    def forward(self, y, y_hat):
        """
        Forward pass of the multi-period discriminator.

        Args:
            y (torch.Tensor): Real audio signal.
            y_hat (torch.Tensor): Fake audio signal.
        """
        y_d_rs, y_d_gs, fmap_rs, fmap_gs = [], [], [], []
        for d in self.discriminators:
            y_d_r, fmap_r = d(y)
            y_d_g, fmap_g = d(y_hat)
            y_d_rs.append(y_d_r)
            y_d_gs.append(y_d_g)
            fmap_rs.append(fmap_r)
            fmap_gs.append(fmap_g)

        return y_d_rs, y_d_gs, fmap_rs, fmap_gs


class MultiPeriodDiscriminatorV2(torch.nn.Module):
    """
    Multi-period discriminator V2.

    This class implements a multi-period discriminator V2, which is used
    to discriminate between real and fake audio signals. The discriminator
    is composed of a series of convolutional layers that are applied to
    the input signal at different periods.

    Args:
        use_spectral_norm (bool): Whether to use spectral normalization.
            Defaults to False.
    """

    def __init__(self, use_spectral_norm=False):
        super(MultiPeriodDiscriminatorV2, self).__init__()
        periods = [2, 3, 5, 7, 11, 17, 23, 37]
        self.discriminators = torch.nn.ModuleList(
            [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
            + [DiscriminatorP(p, use_spectral_norm=use_spectral_norm) for p in periods]
        )

    def forward(self, y, y_hat):
        """
        Forward pass of the multi-period discriminator V2.

        Args:
            y (torch.Tensor): Real audio signal.
            y_hat (torch.Tensor): Fake audio signal.
        """
        y_d_rs, y_d_gs, fmap_rs, fmap_gs = [], [], [], []
        for d in self.discriminators:
            y_d_r, fmap_r = d(y)
            y_d_g, fmap_g = d(y_hat)
            y_d_rs.append(y_d_r)
            y_d_gs.append(y_d_g)
            fmap_rs.append(fmap_r)
            fmap_gs.append(fmap_g)

        return y_d_rs, y_d_gs, fmap_rs, fmap_gs


class DiscriminatorS(torch.nn.Module):
    """
    Discriminator for the short-term component.

    This class implements a discriminator for the short-term component
    of the audio signal. The discriminator is composed of a series of
    convolutional layers that are applied to the input signal.
    """

    def __init__(self, use_spectral_norm=False):
        super(DiscriminatorS, self).__init__()
        norm_f = spectral_norm if use_spectral_norm else weight_norm
        self.convs = torch.nn.ModuleList(
            [
                norm_f(torch.nn.Conv1d(1, 16, 15, 1, padding=7)),
                norm_f(torch.nn.Conv1d(16, 64, 41, 4, groups=4, padding=20)),
                norm_f(torch.nn.Conv1d(64, 256, 41, 4, groups=16, padding=20)),
                norm_f(torch.nn.Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
                norm_f(torch.nn.Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
                norm_f(torch.nn.Conv1d(1024, 1024, 5, 1, padding=2)),
            ]
        )
        self.conv_post = norm_f(torch.nn.Conv1d(1024, 1, 3, 1, padding=1))

    def forward(self, x):
        """
        Forward pass of the discriminator.

        Args:
            x (torch.Tensor): Input audio signal.
        """
        fmap = []
        for conv in self.convs:
            x = torch.nn.functional.leaky_relu(conv(x), LRELU_SLOPE)
            fmap.append(x)
        x = self.conv_post(x)
        fmap.append(x)
        x = torch.flatten(x, 1, -1)
        return x, fmap


class DiscriminatorP(torch.nn.Module):
    """
    Discriminator for the long-term component.

    This class implements a discriminator for the long-term component
    of the audio signal. The discriminator is composed of a series of
    convolutional layers that are applied to the input signal at a given
    period.

    Args:
        period (int): Period of the discriminator.
        kernel_size (int): Kernel size of the convolutional layers.
            Defaults to 5.
        stride (int): Stride of the convolutional layers. Defaults to 3.
        use_spectral_norm (bool): Whether to use spectral normalization.
            Defaults to False.
    """

    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
        super(DiscriminatorP, self).__init__()
        self.period = period
        norm_f = spectral_norm if use_spectral_norm else weight_norm

        in_channels = [1, 32, 128, 512, 1024]
        out_channels = [32, 128, 512, 1024, 1024]

        self.convs = torch.nn.ModuleList(
            [
                norm_f(
                    torch.nn.Conv2d(
                        in_ch,
                        out_ch,
                        (kernel_size, 1),
                        (stride, 1),
                        padding=(get_padding(kernel_size, 1), 0),
                    )
                )
                for in_ch, out_ch in zip(in_channels, out_channels)
            ]
        )

        self.conv_post = norm_f(torch.nn.Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))

    def forward(self, x):
        """
        Forward pass of the discriminator.

        Args:
            x (torch.Tensor): Input audio signal.
        """
        fmap = []
        b, c, t = x.shape
        if t % self.period != 0:
            n_pad = self.period - (t % self.period)
            x = torch.nn.functional.pad(x, (0, n_pad), "reflect")
        x = x.view(b, c, -1, self.period)

        for conv in self.convs:
            x = torch.nn.functional.leaky_relu(conv(x), LRELU_SLOPE)
            fmap.append(x)

        x = self.conv_post(x)
        fmap.append(x)
        x = torch.flatten(x, 1, -1)
        return x, fmap