-
Notifications
You must be signed in to change notification settings - Fork 51
/
models.py
164 lines (142 loc) · 7.43 KB
/
models.py
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
import torch
from torch import nn
from model.encoders import TextEncoder, PosteriorEncoder, AudioEncoder
from model.normalizing_flows import ResidualCouplingBlock
from model.duration_predictors import DurationPredictor, StochasticDurationPredictor
from model.decoder import Generator
from utils.monotonic_align import search_path, generate_path
from utils.model import sequence_mask, rand_slice_segments
class SynthesizerTrn(nn.Module):
"""
Synthesizer for Training
"""
def __init__(
self,
n_vocab,
spec_channels,
segment_size,
inter_channels,
hidden_channels,
filter_channels,
n_heads,
n_layers,
n_layers_q,
n_flows,
kernel_size,
p_dropout,
speaker_cond_layer,
resblock,
resblock_kernel_sizes,
resblock_dilation_sizes,
upsample_rates,
upsample_initial_channel,
upsample_kernel_sizes,
mas_noise_scale,
mas_noise_scale_decay,
use_sdp=True,
use_transformer_flow=True,
n_speakers=0,
gin_channels=0,
**kwargs
):
super().__init__()
self.segment_size = segment_size
self.n_speakers = n_speakers
self.use_sdp = use_sdp
self.mas_noise_scale = mas_noise_scale
self.mas_noise_scale_decay = mas_noise_scale_decay
self.enc_p = TextEncoder(n_vocab, inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, gin_channels=gin_channels, speaker_cond_layer=speaker_cond_layer)
self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, n_layers_q, gin_channels=gin_channels)
# self.enc_q = AudioEncoder(spec_channels, inter_channels, 32, 768, n_heads, 2, kernel_size, p_dropout, gin_channels=gin_channels)
# self.enc_q = AudioEncoder(spec_channels, inter_channels, 32, 32, n_heads, 3, kernel_size, p_dropout, gin_channels=gin_channels)
self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, n_flows=n_flows, gin_channels=gin_channels, mean_only=False, use_transformer_flow=use_transformer_flow)
if use_sdp:
self.dp = StochasticDurationPredictor(hidden_channels, hidden_channels, 3, 0.5, 4, gin_channels=gin_channels)
else:
self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)
if n_speakers > 1:
self.emb_g = nn.Embedding(n_speakers, gin_channels)
def forward(self, x, x_lengths, y, y_lengths, sid=None):
if self.n_speakers > 0:
g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
else:
g = None
z_p_text, m_p_text, logs_p_text, h_text, x_mask = self.enc_p(x, x_lengths, g=g)
z_q_audio, m_q_audio, logs_q_audio, y_mask = self.enc_q(y, y_lengths, g=g)
z_q_dur, m_q_dur, logs_q_dur = self.flow(z_q_audio, m_q_audio, logs_q_audio, y_mask, g=g)
attn = search_path(z_q_dur, m_p_text, logs_p_text, x_mask, y_mask, mas_noise_scale=self.mas_noise_scale)
self.mas_noise_scale = max(self.mas_noise_scale - self.mas_noise_scale_decay, 0.0)
w = attn.sum(2) # [b, 1, t_s]
# * reduce posterior
# TODO Test gain constant
if False:
attn_inv = attn.squeeze(1) * (1 / (w + 1e-9))
m_q_text = torch.matmul(attn_inv.mT, m_q_dur.mT).mT
logs_q_text = torch.matmul(attn_inv.mT, logs_q_dur.mT).mT
# * expand prior
if self.use_sdp:
l_length = self.dp(h_text, x_mask, w, g=g)
l_length = l_length / torch.sum(x_mask)
else:
logw_ = torch.log(w + 1e-6) * x_mask
logw = self.dp(h_text, x_mask, g=g)
l_length = torch.sum((logw - logw_) ** 2, [1, 2]) / torch.sum(x_mask) # for averaging
m_p_dur = torch.matmul(attn.squeeze(1), m_p_text.mT).mT
logs_p_dur = torch.matmul(attn.squeeze(1), logs_p_text.mT).mT
z_p_dur = m_p_dur + torch.randn_like(m_p_dur) * torch.exp(logs_p_dur) * y_mask
z_p_audio, m_p_audio, logs_p_audio = self.flow(z_p_dur, m_p_dur, logs_p_dur, y_mask, g=g, reverse=True)
z_slice, ids_slice = rand_slice_segments(z_q_audio, y_lengths, self.segment_size)
o = self.dec(z_slice, g=g)
return (
o,
l_length,
attn,
ids_slice,
x_mask,
y_mask,
(m_p_text, logs_p_text),
(m_p_dur, logs_p_dur, z_q_dur, logs_q_dur),
(m_p_audio, logs_p_audio, m_q_audio, logs_q_audio),
)
def infer(self, x, x_lengths, sid=None, noise_scale=1, length_scale=1, noise_scale_w=1.0, max_len=None):
if self.n_speakers > 0:
g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
else:
g = None
z_p_text, m_p_text, logs_p_text, h_text, x_mask = self.enc_p(x, x_lengths, g=g)
if self.use_sdp:
logw = self.dp(h_text, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)
else:
logw = self.dp(h_text, x_mask, g=g)
w = torch.exp(logw) * x_mask * length_scale
w_ceil = torch.ceil(w)
y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
y_mask = torch.unsqueeze(sequence_mask(y_lengths, None), 1).to(x_mask.dtype)
attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
attn = generate_path(w_ceil, attn_mask)
m_p_dur = torch.matmul(attn.squeeze(1), m_p_text.mT).mT # [b, t', t], [b, t, d] -> [b, d, t']
logs_p_dur = torch.matmul(attn.squeeze(1), logs_p_text.mT).mT # [b, t', t], [b, t, d] -> [b, d, t']
z_p_dur = m_p_dur + torch.randn_like(m_p_dur) * torch.exp(logs_p_dur) * noise_scale
z_p_audio, m_p_audio, logs_p_audio = self.flow(z_p_dur, m_p_dur, logs_p_dur, y_mask, g=g, reverse=True)
o = self.dec((z_p_audio * y_mask)[:, :, :max_len], g=g)
return o, attn, y_mask, (z_p_dur, m_p_dur, logs_p_dur), (z_p_audio, m_p_audio, logs_p_audio)
def voice_conversion(self, y, y_lengths, sid_src, sid_tgt):
assert self.n_speakers > 0, "n_speakers have to be larger than 0."
g_src = self.emb_g(sid_src).unsqueeze(-1)
g_tgt = self.emb_g(sid_tgt).unsqueeze(-1)
z_q_audio, m_q_audio, logs_q_audio, y_mask = self.enc_q(y, y_lengths, g=g_src)
z_q_dur, m_q_dur, logs_q_dur = self.flow(z_q_audio, m_q_audio, logs_q_audio, y_mask, g=g_src)
z_p_audio, m_p_audio, logs_p_audio = self.flow(z_q_dur, m_q_dur, logs_q_dur, y_mask, g=g_tgt, reverse=True)
o_hat = self.dec(z_p_audio * y_mask, g=g_tgt)
return o_hat, y_mask, (z_q_dur, m_q_dur, logs_q_dur), (z_p_audio, m_p_audio, logs_p_audio)
def voice_restoration(self, y, y_lengths, sid=None):
if self.n_speakers > 0:
g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
else:
g = None
z_q_audio, m_q_audio, logs_q_audio, y_mask = self.enc_q(y, y_lengths, g=g)
z_q_dur, m_q_dur, logs_q_dur = self.flow(z_q_audio, m_q_audio, logs_q_audio, y_mask, g=g)
z_p_audio, m_p_audio, logs_p_audio = self.flow(z_q_dur, m_q_dur, logs_q_dur, y_mask, g=g, reverse=True)
o_hat = self.dec(z_p_audio * y_mask, g=g)
return o_hat, y_mask, (z_q_dur, m_q_dur, logs_q_dur), (z_p_audio, m_p_audio, logs_p_audio)