File size: 10,690 Bytes
a257816
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from abc import ABC

import torch
import torch.nn.functional as F

from VietTTS.flow.decoder import Decoder


class BASECFM(torch.nn.Module, ABC):
    def __init__(

        self,

        n_feats,

        cfm_params,

        n_spks=1,

        spk_emb_dim=128,

    ):
        super().__init__()
        self.n_feats = n_feats
        self.n_spks = n_spks
        self.spk_emb_dim = spk_emb_dim
        self.solver = cfm_params.solver
        if hasattr(cfm_params, "sigma_min"):
            self.sigma_min = cfm_params.sigma_min
        else:
            self.sigma_min = 1e-4

        self.estimator = None

    @torch.inference_mode()
    def forward(self, mu, mask, n_timesteps, temperature=1.0, spks=None, cond=None):
        """Forward diffusion



        Args:

            mu (torch.Tensor): output of encoder

                shape: (batch_size, n_feats, mel_timesteps)

            mask (torch.Tensor): output_mask

                shape: (batch_size, 1, mel_timesteps)

            n_timesteps (int): number of diffusion steps

            temperature (float, optional): temperature for scaling noise. Defaults to 1.0.

            spks (torch.Tensor, optional): speaker ids. Defaults to None.

                shape: (batch_size, spk_emb_dim)

            cond: Not used but kept for future purposes



        Returns:

            sample: generated mel-spectrogram

                shape: (batch_size, n_feats, mel_timesteps)

        """
        z = torch.randn_like(mu) * temperature
        t_span = torch.linspace(0, 1, n_timesteps + 1, device=mu.device)
        return self.solve_euler(z, t_span=t_span, mu=mu, mask=mask, spks=spks, cond=cond)

    def solve_euler(self, x, t_span, mu, mask, spks, cond):
        """

        Fixed euler solver for ODEs.

        Args:

            x (torch.Tensor): random noise

            t_span (torch.Tensor): n_timesteps interpolated

                shape: (n_timesteps + 1,)

            mu (torch.Tensor): output of encoder

                shape: (batch_size, n_feats, mel_timesteps)

            mask (torch.Tensor): output_mask

                shape: (batch_size, 1, mel_timesteps)

            spks (torch.Tensor, optional): speaker ids. Defaults to None.

                shape: (batch_size, spk_emb_dim)

            cond: Not used but kept for future purposes

        """
        t, _, dt = t_span[0], t_span[-1], t_span[1] - t_span[0]

        # I am storing this because I can later plot it by putting a debugger here and saving it to a file
        # Or in future might add like a return_all_steps flag
        sol = []

        for step in range(1, len(t_span)):
            dphi_dt = self.estimator(x, mask, mu, t, spks, cond)

            x = x + dt * dphi_dt
            t = t + dt
            sol.append(x)
            if step < len(t_span) - 1:
                dt = t_span[step + 1] - t

        return sol[-1]

    def compute_loss(self, x1, mask, mu, spks=None, cond=None):
        """Computes diffusion loss



        Args:

            x1 (torch.Tensor): Target

                shape: (batch_size, n_feats, mel_timesteps)

            mask (torch.Tensor): target mask

                shape: (batch_size, 1, mel_timesteps)

            mu (torch.Tensor): output of encoder

                shape: (batch_size, n_feats, mel_timesteps)

            spks (torch.Tensor, optional): speaker embedding. Defaults to None.

                shape: (batch_size, spk_emb_dim)



        Returns:

            loss: conditional flow matching loss

            y: conditional flow

                shape: (batch_size, n_feats, mel_timesteps)

        """
        b, _, t = mu.shape

        # random timestep
        t = torch.rand([b, 1, 1], device=mu.device, dtype=mu.dtype)
        # sample noise p(x_0)
        z = torch.randn_like(x1)

        y = (1 - (1 - self.sigma_min) * t) * z + t * x1
        u = x1 - (1 - self.sigma_min) * z

        loss = F.mse_loss(self.estimator(y, mask, mu, t.squeeze(), spks), u, reduction="sum") / (
            torch.sum(mask) * u.shape[1]
        )
        return loss, y


class CFM(BASECFM):
    def __init__(self, in_channels, out_channel, cfm_params, decoder_params, n_spks=1, spk_emb_dim=64):
        super().__init__(
            n_feats=in_channels,
            cfm_params=cfm_params,
            n_spks=n_spks,
            spk_emb_dim=spk_emb_dim,
        )

        in_channels = in_channels + (spk_emb_dim if n_spks > 1 else 0)
        # Just change the architecture of the estimator here
        self.estimator = Decoder(in_channels=in_channels, out_channels=out_channel, **decoder_params)


class ConditionalCFM(BASECFM):
    def __init__(self, in_channels, cfm_params, n_spks=1, spk_emb_dim=64, estimator: torch.nn.Module = None):
        super().__init__(
            n_feats=in_channels,
            cfm_params=cfm_params,
            n_spks=n_spks,
            spk_emb_dim=spk_emb_dim,
        )
        self.t_scheduler = cfm_params.t_scheduler
        self.training_cfg_rate = cfm_params.training_cfg_rate
        self.inference_cfg_rate = cfm_params.inference_cfg_rate
        in_channels = in_channels + (spk_emb_dim if n_spks > 0 else 0)
        # Just change the architecture of the estimator here
        self.estimator = estimator

    @torch.inference_mode()
    def forward(self, mu, mask, n_timesteps, temperature=1.0, spks=None, cond=None):
        """Forward diffusion



        Args:

            mu (torch.Tensor): output of encoder

                shape: (batch_size, n_feats, mel_timesteps)

            mask (torch.Tensor): output_mask

                shape: (batch_size, 1, mel_timesteps)

            n_timesteps (int): number of diffusion steps

            temperature (float, optional): temperature for scaling noise. Defaults to 1.0.

            spks (torch.Tensor, optional): speaker ids. Defaults to None.

                shape: (batch_size, spk_emb_dim)

            cond: Not used but kept for future purposes



        Returns:

            sample: generated mel-spectrogram

                shape: (batch_size, n_feats, mel_timesteps)

        """
        z = torch.randn_like(mu) * temperature
        t_span = torch.linspace(0, 1, n_timesteps + 1, device=mu.device, dtype=mu.dtype)
        if self.t_scheduler == 'cosine':
            t_span = 1 - torch.cos(t_span * 0.5 * torch.pi)
        return self.solve_euler(z, t_span=t_span, mu=mu, mask=mask, spks=spks, cond=cond)

    def solve_euler(self, x, t_span, mu, mask, spks, cond):
        """

        Fixed euler solver for ODEs.

        Args:

            x (torch.Tensor): random noise

            t_span (torch.Tensor): n_timesteps interpolated

                shape: (n_timesteps + 1,)

            mu (torch.Tensor): output of encoder

                shape: (batch_size, n_feats, mel_timesteps)

            mask (torch.Tensor): output_mask

                shape: (batch_size, 1, mel_timesteps)

            spks (torch.Tensor, optional): speaker ids. Defaults to None.

                shape: (batch_size, spk_emb_dim)

            cond: Not used but kept for future purposes

        """
        t, _, dt = t_span[0], t_span[-1], t_span[1] - t_span[0]
        t = t.unsqueeze(dim=0)

        # I am storing this because I can later plot it by putting a debugger here and saving it to a file
        # Or in future might add like a return_all_steps flag
        sol = []

        for step in range(1, len(t_span)):
            dphi_dt = self.forward_estimator(x, mask, mu, t, spks, cond)
            # Classifier-Free Guidance inference introduced in VoiceBox
            if self.inference_cfg_rate > 0:
                cfg_dphi_dt = self.forward_estimator(
                    x, mask,
                    torch.zeros_like(mu), t,
                    torch.zeros_like(spks) if spks is not None else None,
                    torch.zeros_like(cond)
                )
                dphi_dt = ((1.0 + self.inference_cfg_rate) * dphi_dt -
                           self.inference_cfg_rate * cfg_dphi_dt)
            x = x + dt * dphi_dt
            t = t + dt
            sol.append(x)
            if step < len(t_span) - 1:
                dt = t_span[step + 1] - t

        return sol[-1]

    def forward_estimator(self, x, mask, mu, t, spks, cond):
        if isinstance(self.estimator, torch.nn.Module):
            return self.estimator.forward(x, mask, mu, t, spks, cond)
        else:
            ort_inputs = {
                'x': x.cpu().numpy(),
                'mask': mask.cpu().numpy(),
                'mu': mu.cpu().numpy(),
                't': t.cpu().numpy(),
                'spks': spks.cpu().numpy(),
                'cond': cond.cpu().numpy()
            }
            output = self.estimator.run(None, ort_inputs)[0]
            return torch.tensor(output, dtype=x.dtype, device=x.device)

    def compute_loss(self, x1, mask, mu, spks=None, cond=None):
        """Computes diffusion loss



        Args:

            x1 (torch.Tensor): Target

                shape: (batch_size, n_feats, mel_timesteps)

            mask (torch.Tensor): target mask

                shape: (batch_size, 1, mel_timesteps)

            mu (torch.Tensor): output of encoder

                shape: (batch_size, n_feats, mel_timesteps)

            spks (torch.Tensor, optional): speaker embedding. Defaults to None.

                shape: (batch_size, spk_emb_dim)



        Returns:

            loss: conditional flow matching loss

            y: conditional flow

                shape: (batch_size, n_feats, mel_timesteps)

        """
        b, _, t = mu.shape

        # random timestep
        t = torch.rand([b, 1, 1], device=mu.device, dtype=mu.dtype)
        if self.t_scheduler == 'cosine':
            t = 1 - torch.cos(t * 0.5 * torch.pi)
        # sample noise p(x_0)
        z = torch.randn_like(x1)

        y = (1 - (1 - self.sigma_min) * t) * z + t * x1
        u = x1 - (1 - self.sigma_min) * z

        # during training, we randomly drop condition to trade off mode coverage and sample fidelity
        if self.training_cfg_rate > 0:
            cfg_mask = torch.rand(b, device=x1.device) > self.training_cfg_rate
            mu = mu * cfg_mask.view(-1, 1, 1)
            spks = spks * cfg_mask.view(-1, 1)
            cond = cond * cfg_mask.view(-1, 1, 1)

        pred = self.estimator(y, mask, mu, t.squeeze(), spks, cond)
        loss = F.mse_loss(pred * mask, u * mask, reduction="sum") / (torch.sum(mask) * u.shape[1])
        return loss, y