File size: 8,493 Bytes
c968fc3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2023 Amphion.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn
import numpy as np
import torch.nn.functional as F
from models.tts.naturalspeech2.diffusion import Diffusion
from models.tts.naturalspeech2.diffusion_flow import DiffusionFlow
from models.tts.naturalspeech2.wavenet import WaveNet
from models.tts.naturalspeech2.prior_encoder import PriorEncoder
from modules.naturalpseech2.transformers import TransformerEncoder
from encodec import EncodecModel
from einops import rearrange, repeat

import os
import json


class NaturalSpeech2(nn.Module):
    def __init__(self, cfg):
        super().__init__()
        self.cfg = cfg

        self.latent_dim = cfg.latent_dim
        self.query_emb_num = cfg.query_emb.query_token_num

        self.prior_encoder = PriorEncoder(cfg.prior_encoder)
        if cfg.diffusion.diffusion_type == "diffusion":
            self.diffusion = Diffusion(cfg.diffusion)
        elif cfg.diffusion.diffusion_type == "flow":
            self.diffusion = DiffusionFlow(cfg.diffusion)

        self.prompt_encoder = TransformerEncoder(cfg=cfg.prompt_encoder)
        if self.latent_dim != cfg.prompt_encoder.encoder_hidden:
            self.prompt_lin = nn.Linear(
                self.latent_dim, cfg.prompt_encoder.encoder_hidden
            )
            self.prompt_lin.weight.data.normal_(0.0, 0.02)
        else:
            self.prompt_lin = None

        self.query_emb = nn.Embedding(self.query_emb_num, cfg.query_emb.hidden_size)
        self.query_attn = nn.MultiheadAttention(
            cfg.query_emb.hidden_size, cfg.query_emb.head_num, batch_first=True
        )

        codec_model = EncodecModel.encodec_model_24khz()
        codec_model.set_target_bandwidth(12.0)
        codec_model.requires_grad_(False)
        self.quantizer = codec_model.quantizer

    @torch.no_grad()
    def code_to_latent(self, code):
        latent = self.quantizer.decode(code.transpose(0, 1))
        return latent

    def latent_to_code(self, latent, nq=16):
        residual = latent
        all_indices = []
        all_dist = []
        for i in range(nq):
            layer = self.quantizer.vq.layers[i]
            x = rearrange(residual, "b d n -> b n d")
            x = layer.project_in(x)
            shape = x.shape
            x = layer._codebook.preprocess(x)
            embed = layer._codebook.embed.t()
            dist = -(
                x.pow(2).sum(1, keepdim=True)
                - 2 * x @ embed
                + embed.pow(2).sum(0, keepdim=True)
            )
            indices = dist.max(dim=-1).indices
            indices = layer._codebook.postprocess_emb(indices, shape)
            dist = dist.reshape(*shape[:-1], dist.shape[-1])
            quantized = layer.decode(indices)
            residual = residual - quantized
            all_indices.append(indices)
            all_dist.append(dist)

        out_indices = torch.stack(all_indices)
        out_dist = torch.stack(all_dist)

        return out_indices, out_dist  # (nq, B, T); (nq, B, T, 1024)

    @torch.no_grad()
    def latent_to_latent(self, latent, nq=16):
        codes, _ = self.latent_to_code(latent, nq)
        latent = self.quantizer.vq.decode(codes)
        return latent

    def forward(
        self,
        code=None,
        pitch=None,
        duration=None,
        phone_id=None,
        phone_id_frame=None,
        frame_nums=None,
        ref_code=None,
        ref_frame_nums=None,
        phone_mask=None,
        mask=None,
        ref_mask=None,
    ):
        ref_latent = self.code_to_latent(ref_code)
        latent = self.code_to_latent(code)

        if self.latent_dim is not None:
            ref_latent = self.prompt_lin(ref_latent.transpose(1, 2))

        ref_latent = self.prompt_encoder(ref_latent, ref_mask, condition=None)
        spk_emb = ref_latent.transpose(1, 2)  # (B, d, T')

        spk_query_emb = self.query_emb(
            torch.arange(self.query_emb_num).to(latent.device)
        ).repeat(
            latent.shape[0], 1, 1
        )  # (B, query_emb_num, d)
        spk_query_emb, _ = self.query_attn(
            spk_query_emb,
            spk_emb.transpose(1, 2),
            spk_emb.transpose(1, 2),
            key_padding_mask=~(ref_mask.bool()),
        )  # (B, query_emb_num, d)

        prior_out = self.prior_encoder(
            phone_id=phone_id,
            duration=duration,
            pitch=pitch,
            phone_mask=phone_mask,
            mask=mask,
            ref_emb=spk_emb,
            ref_mask=ref_mask,
            is_inference=False,
        )
        prior_condition = prior_out["prior_out"]  # (B, T, d)

        diff_out = self.diffusion(latent, mask, prior_condition, spk_query_emb)

        return diff_out, prior_out

    @torch.no_grad()
    def inference(
        self, ref_code=None, phone_id=None, ref_mask=None, inference_steps=1000
    ):
        ref_latent = self.code_to_latent(ref_code)

        if self.latent_dim is not None:
            ref_latent = self.prompt_lin(ref_latent.transpose(1, 2))

        ref_latent = self.prompt_encoder(ref_latent, ref_mask, condition=None)
        spk_emb = ref_latent.transpose(1, 2)  # (B, d, T')

        spk_query_emb = self.query_emb(
            torch.arange(self.query_emb_num).to(ref_latent.device)
        ).repeat(
            ref_latent.shape[0], 1, 1
        )  # (B, query_emb_num, d)
        spk_query_emb, _ = self.query_attn(
            spk_query_emb,
            spk_emb.transpose(1, 2),
            spk_emb.transpose(1, 2),
            key_padding_mask=~(ref_mask.bool()),
        )  # (B, query_emb_num, d)

        prior_out = self.prior_encoder(
            phone_id=phone_id,
            duration=None,
            pitch=None,
            phone_mask=None,
            mask=None,
            ref_emb=spk_emb,
            ref_mask=ref_mask,
            is_inference=True,
        )
        prior_condition = prior_out["prior_out"]  # (B, T, d)

        z = torch.randn(
            prior_condition.shape[0], self.latent_dim, prior_condition.shape[1]
        ).to(ref_latent.device) / (1.20)
        x0 = self.diffusion.reverse_diffusion(
            z, None, prior_condition, inference_steps, spk_query_emb
        )

        return x0, prior_out

    @torch.no_grad()
    def reverse_diffusion_from_t(
        self,
        code=None,
        pitch=None,
        duration=None,
        phone_id=None,
        ref_code=None,
        phone_mask=None,
        mask=None,
        ref_mask=None,
        n_timesteps=None,
        t=None,
    ):
        # o Only for debug

        ref_latent = self.code_to_latent(ref_code)
        latent = self.code_to_latent(code)

        if self.latent_dim is not None:
            ref_latent = self.prompt_lin(ref_latent.transpose(1, 2))

        ref_latent = self.prompt_encoder(ref_latent, ref_mask, condition=None)
        spk_emb = ref_latent.transpose(1, 2)  # (B, d, T')

        spk_query_emb = self.query_emb(
            torch.arange(self.query_emb_num).to(latent.device)
        ).repeat(
            latent.shape[0], 1, 1
        )  # (B, query_emb_num, d)
        spk_query_emb, _ = self.query_attn(
            spk_query_emb,
            spk_emb.transpose(1, 2),
            spk_emb.transpose(1, 2),
            key_padding_mask=~(ref_mask.bool()),
        )  # (B, query_emb_num, d)

        prior_out = self.prior_encoder(
            phone_id=phone_id,
            duration=duration,
            pitch=pitch,
            phone_mask=phone_mask,
            mask=mask,
            ref_emb=spk_emb,
            ref_mask=ref_mask,
            is_inference=False,
        )
        prior_condition = prior_out["prior_out"]  # (B, T, d)

        diffusion_step = (
            torch.ones(
                latent.shape[0],
                dtype=latent.dtype,
                device=latent.device,
                requires_grad=False,
            )
            * t
        )
        diffusion_step = torch.clamp(diffusion_step, 1e-5, 1.0 - 1e-5)
        xt, _ = self.diffusion.forward_diffusion(
            x0=latent, diffusion_step=diffusion_step
        )
        # print(torch.abs(xt-latent).max(), torch.abs(xt-latent).mean(), torch.abs(xt-latent).std())

        x0 = self.diffusion.reverse_diffusion_from_t(
            xt, mask, prior_condition, n_timesteps, spk_query_emb, t_start=t
        )

        return x0, prior_out, xt