File size: 9,604 Bytes
635f007
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from torch import nn
import torch.nn.functional as F
import torchaudio
from transformers import AutoModel


class SpectralConvergengeLoss(torch.nn.Module):
    """Spectral convergence loss module."""

    def __init__(self):
        """Initilize spectral convergence loss module."""
        super(SpectralConvergengeLoss, self).__init__()

    def forward(self, x_mag, y_mag):
        """Calculate forward propagation.
        Args:
            x_mag (Tensor): Magnitude spectrogram of predicted signal (B, #frames, #freq_bins).
            y_mag (Tensor): Magnitude spectrogram of groundtruth signal (B, #frames, #freq_bins).
        Returns:
            Tensor: Spectral convergence loss value.
        """
        return torch.norm(y_mag - x_mag, p=1) / torch.norm(y_mag, p=1)


class STFTLoss(torch.nn.Module):
    """STFT loss module."""

    def __init__(
        self, fft_size=1024, shift_size=120, win_length=600, window=torch.hann_window
    ):
        """Initialize STFT loss module."""
        super(STFTLoss, self).__init__()
        self.fft_size = fft_size
        self.shift_size = shift_size
        self.win_length = win_length
        self.to_mel = torchaudio.transforms.MelSpectrogram(
            sample_rate=24000,
            n_fft=fft_size,
            win_length=win_length,
            hop_length=shift_size,
            window_fn=window,
        )

        self.spectral_convergenge_loss = SpectralConvergengeLoss()

    def forward(self, x, y):
        """Calculate forward propagation.
        Args:
            x (Tensor): Predicted signal (B, T).
            y (Tensor): Groundtruth signal (B, T).
        Returns:
            Tensor: Spectral convergence loss value.
            Tensor: Log STFT magnitude loss value.
        """
        x_mag = self.to_mel(x)
        mean, std = -4, 4
        x_mag = (torch.log(1e-5 + x_mag) - mean) / std

        y_mag = self.to_mel(y)
        mean, std = -4, 4
        y_mag = (torch.log(1e-5 + y_mag) - mean) / std

        sc_loss = self.spectral_convergenge_loss(x_mag, y_mag)
        return sc_loss


class MultiResolutionSTFTLoss(torch.nn.Module):
    """Multi resolution STFT loss module."""

    def __init__(
        self,
        fft_sizes=[1024, 2048, 512],
        hop_sizes=[120, 240, 50],
        win_lengths=[600, 1200, 240],
        window=torch.hann_window,
    ):
        """Initialize Multi resolution STFT loss module.
        Args:
            fft_sizes (list): List of FFT sizes.
            hop_sizes (list): List of hop sizes.
            win_lengths (list): List of window lengths.
            window (str): Window function type.
        """
        super(MultiResolutionSTFTLoss, self).__init__()
        assert len(fft_sizes) == len(hop_sizes) == len(win_lengths)
        self.stft_losses = torch.nn.ModuleList()
        for fs, ss, wl in zip(fft_sizes, hop_sizes, win_lengths):
            self.stft_losses += [STFTLoss(fs, ss, wl, window)]

    def forward(self, x, y):
        """Calculate forward propagation.
        Args:
            x (Tensor): Predicted signal (B, T).
            y (Tensor): Groundtruth signal (B, T).
        Returns:
            Tensor: Multi resolution spectral convergence loss value.
            Tensor: Multi resolution log STFT magnitude loss value.
        """
        sc_loss = 0.0
        for f in self.stft_losses:
            sc_l = f(x, y)
            sc_loss += sc_l
        sc_loss /= len(self.stft_losses)

        return sc_loss


def feature_loss(fmap_r, fmap_g):
    loss = 0
    for dr, dg in zip(fmap_r, fmap_g):
        for rl, gl in zip(dr, dg):
            loss += torch.mean(torch.abs(rl - gl))

    return loss * 2


def discriminator_loss(disc_real_outputs, disc_generated_outputs):
    loss = 0
    r_losses = []
    g_losses = []
    for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
        r_loss = torch.mean((1 - dr) ** 2)
        g_loss = torch.mean(dg**2)
        loss += r_loss + g_loss
        r_losses.append(r_loss.item())
        g_losses.append(g_loss.item())

    return loss, r_losses, g_losses


def generator_loss(disc_outputs):
    loss = 0
    gen_losses = []
    for dg in disc_outputs:
        l = torch.mean((1 - dg) ** 2)
        gen_losses.append(l)
        loss += l

    return loss, gen_losses


""" https://dl.acm.org/doi/abs/10.1145/3573834.3574506 """


def discriminator_TPRLS_loss(disc_real_outputs, disc_generated_outputs):
    loss = 0
    for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
        tau = 0.04
        m_DG = torch.median((dr - dg))
        L_rel = torch.mean((((dr - dg) - m_DG) ** 2)[dr < dg + m_DG])
        loss += tau - F.relu(tau - L_rel)
    return loss


def generator_TPRLS_loss(disc_real_outputs, disc_generated_outputs):
    loss = 0
    for dg, dr in zip(disc_real_outputs, disc_generated_outputs):
        tau = 0.04
        m_DG = torch.median((dr - dg))
        L_rel = torch.mean((((dr - dg) - m_DG) ** 2)[dr < dg + m_DG])
        loss += tau - F.relu(tau - L_rel)
    return loss


class GeneratorLoss(torch.nn.Module):
    def __init__(self, mpd, msd):
        super(GeneratorLoss, self).__init__()
        self.mpd = mpd
        self.msd = msd

    def forward(self, y, y_hat):
        y_df_hat_r, y_df_hat_g, fmap_f_r, fmap_f_g = self.mpd(y, y_hat)
        y_ds_hat_r, y_ds_hat_g, fmap_s_r, fmap_s_g = self.msd(y, y_hat)
        loss_fm_f = feature_loss(fmap_f_r, fmap_f_g)
        loss_fm_s = feature_loss(fmap_s_r, fmap_s_g)
        loss_gen_f, losses_gen_f = generator_loss(y_df_hat_g)
        loss_gen_s, losses_gen_s = generator_loss(y_ds_hat_g)

        loss_rel = generator_TPRLS_loss(y_df_hat_r, y_df_hat_g) + generator_TPRLS_loss(
            y_ds_hat_r, y_ds_hat_g
        )

        loss_gen_all = loss_gen_s + loss_gen_f + loss_fm_s + loss_fm_f + loss_rel

        return loss_gen_all.mean()


class DiscriminatorLoss(torch.nn.Module):
    def __init__(self, mpd, msd):
        super(DiscriminatorLoss, self).__init__()
        self.mpd = mpd
        self.msd = msd

    def forward(self, y, y_hat):
        # MPD
        y_df_hat_r, y_df_hat_g, _, _ = self.mpd(y, y_hat)
        loss_disc_f, losses_disc_f_r, losses_disc_f_g = discriminator_loss(
            y_df_hat_r, y_df_hat_g
        )
        # MSD
        y_ds_hat_r, y_ds_hat_g, _, _ = self.msd(y, y_hat)
        loss_disc_s, losses_disc_s_r, losses_disc_s_g = discriminator_loss(
            y_ds_hat_r, y_ds_hat_g
        )

        loss_rel = discriminator_TPRLS_loss(
            y_df_hat_r, y_df_hat_g
        ) + discriminator_TPRLS_loss(y_ds_hat_r, y_ds_hat_g)

        d_loss = loss_disc_s + loss_disc_f + loss_rel

        return d_loss.mean()


class WavLMLoss(torch.nn.Module):
    def __init__(self, model, wd, model_sr, slm_sr=16000):
        super(WavLMLoss, self).__init__()
        self.wavlm = AutoModel.from_pretrained(model)
        self.wd = wd
        self.resample = torchaudio.transforms.Resample(model_sr, slm_sr)

    def forward(self, wav, y_rec):
        with torch.no_grad():
            wav_16 = self.resample(wav)
            wav_embeddings = self.wavlm(
                input_values=wav_16, output_hidden_states=True
            ).hidden_states
        y_rec_16 = self.resample(y_rec)
        y_rec_embeddings = self.wavlm(
            input_values=y_rec_16.squeeze(), output_hidden_states=True
        ).hidden_states

        floss = 0
        for er, eg in zip(wav_embeddings, y_rec_embeddings):
            floss += torch.mean(torch.abs(er - eg))

        return floss.mean()

    def generator(self, y_rec):
        y_rec_16 = self.resample(y_rec)
        y_rec_embeddings = self.wavlm(
            input_values=y_rec_16, output_hidden_states=True
        ).hidden_states
        y_rec_embeddings = (
            torch.stack(y_rec_embeddings, dim=1)
            .transpose(-1, -2)
            .flatten(start_dim=1, end_dim=2)
        )
        y_df_hat_g = self.wd(y_rec_embeddings)
        loss_gen = torch.mean((1 - y_df_hat_g) ** 2)

        return loss_gen

    def discriminator(self, wav, y_rec):
        with torch.no_grad():
            wav_16 = self.resample(wav)
            wav_embeddings = self.wavlm(
                input_values=wav_16, output_hidden_states=True
            ).hidden_states
            y_rec_16 = self.resample(y_rec)
            y_rec_embeddings = self.wavlm(
                input_values=y_rec_16, output_hidden_states=True
            ).hidden_states

            y_embeddings = (
                torch.stack(wav_embeddings, dim=1)
                .transpose(-1, -2)
                .flatten(start_dim=1, end_dim=2)
            )
            y_rec_embeddings = (
                torch.stack(y_rec_embeddings, dim=1)
                .transpose(-1, -2)
                .flatten(start_dim=1, end_dim=2)
            )

        y_d_rs = self.wd(y_embeddings)
        y_d_gs = self.wd(y_rec_embeddings)

        y_df_hat_r, y_df_hat_g = y_d_rs, y_d_gs

        r_loss = torch.mean((1 - y_df_hat_r) ** 2)
        g_loss = torch.mean((y_df_hat_g) ** 2)

        loss_disc_f = r_loss + g_loss

        return loss_disc_f.mean()

    def discriminator_forward(self, wav):
        with torch.no_grad():
            wav_16 = self.resample(wav)
            wav_embeddings = self.wavlm(
                input_values=wav_16, output_hidden_states=True
            ).hidden_states
            y_embeddings = (
                torch.stack(wav_embeddings, dim=1)
                .transpose(-1, -2)
                .flatten(start_dim=1, end_dim=2)
            )

        y_d_rs = self.wd(y_embeddings)

        return y_d_rs