Spaces:
Running
Running
Create model.py
Browse files- GPT_SoVITS/module/model.py +1030 -0
GPT_SoVITS/module/model.py
ADDED
@@ -0,0 +1,1030 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import warnings
|
2 |
+
warnings.filterwarnings("ignore")
|
3 |
+
import copy
|
4 |
+
import math
|
5 |
+
import os
|
6 |
+
import pdb
|
7 |
+
|
8 |
+
import torch
|
9 |
+
from torch import nn
|
10 |
+
from torch.nn import functional as F
|
11 |
+
|
12 |
+
from module import commons
|
13 |
+
from module import modules
|
14 |
+
from module import attentions
|
15 |
+
|
16 |
+
from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
|
17 |
+
from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
|
18 |
+
from module.commons import init_weights, get_padding
|
19 |
+
from module.mrte_model import MRTE
|
20 |
+
from module.quantize import ResidualVectorQuantizer
|
21 |
+
# from text import symbols
|
22 |
+
from text import symbols as symbols_v1
|
23 |
+
from text import symbols2 as symbols_v2
|
24 |
+
from torch.cuda.amp import autocast
|
25 |
+
import contextlib
|
26 |
+
|
27 |
+
|
28 |
+
class StochasticDurationPredictor(nn.Module):
|
29 |
+
def __init__(
|
30 |
+
self,
|
31 |
+
in_channels,
|
32 |
+
filter_channels,
|
33 |
+
kernel_size,
|
34 |
+
p_dropout,
|
35 |
+
n_flows=4,
|
36 |
+
gin_channels=0,
|
37 |
+
):
|
38 |
+
super().__init__()
|
39 |
+
filter_channels = in_channels # it needs to be removed from future version.
|
40 |
+
self.in_channels = in_channels
|
41 |
+
self.filter_channels = filter_channels
|
42 |
+
self.kernel_size = kernel_size
|
43 |
+
self.p_dropout = p_dropout
|
44 |
+
self.n_flows = n_flows
|
45 |
+
self.gin_channels = gin_channels
|
46 |
+
|
47 |
+
self.log_flow = modules.Log()
|
48 |
+
self.flows = nn.ModuleList()
|
49 |
+
self.flows.append(modules.ElementwiseAffine(2))
|
50 |
+
for i in range(n_flows):
|
51 |
+
self.flows.append(
|
52 |
+
modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3)
|
53 |
+
)
|
54 |
+
self.flows.append(modules.Flip())
|
55 |
+
|
56 |
+
self.post_pre = nn.Conv1d(1, filter_channels, 1)
|
57 |
+
self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
|
58 |
+
self.post_convs = modules.DDSConv(
|
59 |
+
filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout
|
60 |
+
)
|
61 |
+
self.post_flows = nn.ModuleList()
|
62 |
+
self.post_flows.append(modules.ElementwiseAffine(2))
|
63 |
+
for i in range(4):
|
64 |
+
self.post_flows.append(
|
65 |
+
modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3)
|
66 |
+
)
|
67 |
+
self.post_flows.append(modules.Flip())
|
68 |
+
|
69 |
+
self.pre = nn.Conv1d(in_channels, filter_channels, 1)
|
70 |
+
self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
|
71 |
+
self.convs = modules.DDSConv(
|
72 |
+
filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout
|
73 |
+
)
|
74 |
+
if gin_channels != 0:
|
75 |
+
self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
|
76 |
+
|
77 |
+
def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
|
78 |
+
x = torch.detach(x)
|
79 |
+
x = self.pre(x)
|
80 |
+
if g is not None:
|
81 |
+
g = torch.detach(g)
|
82 |
+
x = x + self.cond(g)
|
83 |
+
x = self.convs(x, x_mask)
|
84 |
+
x = self.proj(x) * x_mask
|
85 |
+
|
86 |
+
if not reverse:
|
87 |
+
flows = self.flows
|
88 |
+
assert w is not None
|
89 |
+
|
90 |
+
logdet_tot_q = 0
|
91 |
+
h_w = self.post_pre(w)
|
92 |
+
h_w = self.post_convs(h_w, x_mask)
|
93 |
+
h_w = self.post_proj(h_w) * x_mask
|
94 |
+
e_q = (
|
95 |
+
torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype)
|
96 |
+
* x_mask
|
97 |
+
)
|
98 |
+
z_q = e_q
|
99 |
+
for flow in self.post_flows:
|
100 |
+
z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
|
101 |
+
logdet_tot_q += logdet_q
|
102 |
+
z_u, z1 = torch.split(z_q, [1, 1], 1)
|
103 |
+
u = torch.sigmoid(z_u) * x_mask
|
104 |
+
z0 = (w - u) * x_mask
|
105 |
+
logdet_tot_q += torch.sum(
|
106 |
+
(F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1, 2]
|
107 |
+
)
|
108 |
+
logq = (
|
109 |
+
torch.sum(-0.5 * (math.log(2 * math.pi) + (e_q**2)) * x_mask, [1, 2])
|
110 |
+
- logdet_tot_q
|
111 |
+
)
|
112 |
+
|
113 |
+
logdet_tot = 0
|
114 |
+
z0, logdet = self.log_flow(z0, x_mask)
|
115 |
+
logdet_tot += logdet
|
116 |
+
z = torch.cat([z0, z1], 1)
|
117 |
+
for flow in flows:
|
118 |
+
z, logdet = flow(z, x_mask, g=x, reverse=reverse)
|
119 |
+
logdet_tot = logdet_tot + logdet
|
120 |
+
nll = (
|
121 |
+
torch.sum(0.5 * (math.log(2 * math.pi) + (z**2)) * x_mask, [1, 2])
|
122 |
+
- logdet_tot
|
123 |
+
)
|
124 |
+
return nll + logq # [b]
|
125 |
+
else:
|
126 |
+
flows = list(reversed(self.flows))
|
127 |
+
flows = flows[:-2] + [flows[-1]] # remove a useless vflow
|
128 |
+
z = (
|
129 |
+
torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype)
|
130 |
+
* noise_scale
|
131 |
+
)
|
132 |
+
for flow in flows:
|
133 |
+
z = flow(z, x_mask, g=x, reverse=reverse)
|
134 |
+
z0, z1 = torch.split(z, [1, 1], 1)
|
135 |
+
logw = z0
|
136 |
+
return logw
|
137 |
+
|
138 |
+
|
139 |
+
class DurationPredictor(nn.Module):
|
140 |
+
def __init__(
|
141 |
+
self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0
|
142 |
+
):
|
143 |
+
super().__init__()
|
144 |
+
|
145 |
+
self.in_channels = in_channels
|
146 |
+
self.filter_channels = filter_channels
|
147 |
+
self.kernel_size = kernel_size
|
148 |
+
self.p_dropout = p_dropout
|
149 |
+
self.gin_channels = gin_channels
|
150 |
+
|
151 |
+
self.drop = nn.Dropout(p_dropout)
|
152 |
+
self.conv_1 = nn.Conv1d(
|
153 |
+
in_channels, filter_channels, kernel_size, padding=kernel_size // 2
|
154 |
+
)
|
155 |
+
self.norm_1 = modules.LayerNorm(filter_channels)
|
156 |
+
self.conv_2 = nn.Conv1d(
|
157 |
+
filter_channels, filter_channels, kernel_size, padding=kernel_size // 2
|
158 |
+
)
|
159 |
+
self.norm_2 = modules.LayerNorm(filter_channels)
|
160 |
+
self.proj = nn.Conv1d(filter_channels, 1, 1)
|
161 |
+
|
162 |
+
if gin_channels != 0:
|
163 |
+
self.cond = nn.Conv1d(gin_channels, in_channels, 1)
|
164 |
+
|
165 |
+
def forward(self, x, x_mask, g=None):
|
166 |
+
x = torch.detach(x)
|
167 |
+
if g is not None:
|
168 |
+
g = torch.detach(g)
|
169 |
+
x = x + self.cond(g)
|
170 |
+
x = self.conv_1(x * x_mask)
|
171 |
+
x = torch.relu(x)
|
172 |
+
x = self.norm_1(x)
|
173 |
+
x = self.drop(x)
|
174 |
+
x = self.conv_2(x * x_mask)
|
175 |
+
x = torch.relu(x)
|
176 |
+
x = self.norm_2(x)
|
177 |
+
x = self.drop(x)
|
178 |
+
x = self.proj(x * x_mask)
|
179 |
+
return x * x_mask
|
180 |
+
|
181 |
+
|
182 |
+
class TextEncoder(nn.Module):
|
183 |
+
def __init__(
|
184 |
+
self,
|
185 |
+
out_channels,
|
186 |
+
hidden_channels,
|
187 |
+
filter_channels,
|
188 |
+
n_heads,
|
189 |
+
n_layers,
|
190 |
+
kernel_size,
|
191 |
+
p_dropout,
|
192 |
+
latent_channels=192,
|
193 |
+
version = "v2",
|
194 |
+
):
|
195 |
+
super().__init__()
|
196 |
+
self.out_channels = out_channels
|
197 |
+
self.hidden_channels = hidden_channels
|
198 |
+
self.filter_channels = filter_channels
|
199 |
+
self.n_heads = n_heads
|
200 |
+
self.n_layers = n_layers
|
201 |
+
self.kernel_size = kernel_size
|
202 |
+
self.p_dropout = p_dropout
|
203 |
+
self.latent_channels = latent_channels
|
204 |
+
self.version = version
|
205 |
+
|
206 |
+
self.ssl_proj = nn.Conv1d(768, hidden_channels, 1)
|
207 |
+
|
208 |
+
self.encoder_ssl = attentions.Encoder(
|
209 |
+
hidden_channels,
|
210 |
+
filter_channels,
|
211 |
+
n_heads,
|
212 |
+
n_layers // 2,
|
213 |
+
kernel_size,
|
214 |
+
p_dropout,
|
215 |
+
)
|
216 |
+
|
217 |
+
self.encoder_text = attentions.Encoder(
|
218 |
+
hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
|
219 |
+
)
|
220 |
+
|
221 |
+
if self.version == "v1":
|
222 |
+
symbols = symbols_v1.symbols
|
223 |
+
else:
|
224 |
+
symbols = symbols_v2.symbols
|
225 |
+
self.text_embedding = nn.Embedding(len(symbols), hidden_channels)
|
226 |
+
|
227 |
+
self.mrte = MRTE()
|
228 |
+
|
229 |
+
self.encoder2 = attentions.Encoder(
|
230 |
+
hidden_channels,
|
231 |
+
filter_channels,
|
232 |
+
n_heads,
|
233 |
+
n_layers // 2,
|
234 |
+
kernel_size,
|
235 |
+
p_dropout,
|
236 |
+
)
|
237 |
+
|
238 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
|
239 |
+
|
240 |
+
def forward(self, y, y_lengths, text, text_lengths, ge, speed=1,test=None):
|
241 |
+
y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, y.size(2)), 1).to(
|
242 |
+
y.dtype
|
243 |
+
)
|
244 |
+
|
245 |
+
y = self.ssl_proj(y * y_mask) * y_mask
|
246 |
+
|
247 |
+
y = self.encoder_ssl(y * y_mask, y_mask)
|
248 |
+
|
249 |
+
text_mask = torch.unsqueeze(
|
250 |
+
commons.sequence_mask(text_lengths, text.size(1)), 1
|
251 |
+
).to(y.dtype)
|
252 |
+
if test == 1:
|
253 |
+
text[:, :] = 0
|
254 |
+
text = self.text_embedding(text).transpose(1, 2)
|
255 |
+
text = self.encoder_text(text * text_mask, text_mask)
|
256 |
+
y = self.mrte(y, y_mask, text, text_mask, ge)
|
257 |
+
y = self.encoder2(y * y_mask, y_mask)
|
258 |
+
if(speed!=1):
|
259 |
+
y = F.interpolate(y, size=int(y.shape[-1] / speed)+1, mode="linear")
|
260 |
+
y_mask = F.interpolate(y_mask, size=y.shape[-1], mode="nearest")
|
261 |
+
stats = self.proj(y) * y_mask
|
262 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
263 |
+
return y, m, logs, y_mask
|
264 |
+
|
265 |
+
def extract_latent(self, x):
|
266 |
+
x = self.ssl_proj(x)
|
267 |
+
quantized, codes, commit_loss, quantized_list = self.quantizer(x)
|
268 |
+
return codes.transpose(0, 1)
|
269 |
+
|
270 |
+
def decode_latent(self, codes, y_mask, refer, refer_mask, ge):
|
271 |
+
quantized = self.quantizer.decode(codes)
|
272 |
+
|
273 |
+
y = self.vq_proj(quantized) * y_mask
|
274 |
+
y = self.encoder_ssl(y * y_mask, y_mask)
|
275 |
+
|
276 |
+
y = self.mrte(y, y_mask, refer, refer_mask, ge)
|
277 |
+
|
278 |
+
y = self.encoder2(y * y_mask, y_mask)
|
279 |
+
|
280 |
+
stats = self.proj(y) * y_mask
|
281 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
282 |
+
return y, m, logs, y_mask, quantized
|
283 |
+
|
284 |
+
|
285 |
+
class ResidualCouplingBlock(nn.Module):
|
286 |
+
def __init__(
|
287 |
+
self,
|
288 |
+
channels,
|
289 |
+
hidden_channels,
|
290 |
+
kernel_size,
|
291 |
+
dilation_rate,
|
292 |
+
n_layers,
|
293 |
+
n_flows=4,
|
294 |
+
gin_channels=0,
|
295 |
+
):
|
296 |
+
super().__init__()
|
297 |
+
self.channels = channels
|
298 |
+
self.hidden_channels = hidden_channels
|
299 |
+
self.kernel_size = kernel_size
|
300 |
+
self.dilation_rate = dilation_rate
|
301 |
+
self.n_layers = n_layers
|
302 |
+
self.n_flows = n_flows
|
303 |
+
self.gin_channels = gin_channels
|
304 |
+
|
305 |
+
self.flows = nn.ModuleList()
|
306 |
+
for i in range(n_flows):
|
307 |
+
self.flows.append(
|
308 |
+
modules.ResidualCouplingLayer(
|
309 |
+
channels,
|
310 |
+
hidden_channels,
|
311 |
+
kernel_size,
|
312 |
+
dilation_rate,
|
313 |
+
n_layers,
|
314 |
+
gin_channels=gin_channels,
|
315 |
+
mean_only=True,
|
316 |
+
)
|
317 |
+
)
|
318 |
+
self.flows.append(modules.Flip())
|
319 |
+
|
320 |
+
def forward(self, x, x_mask, g=None, reverse=False):
|
321 |
+
if not reverse:
|
322 |
+
for flow in self.flows:
|
323 |
+
x, _ = flow(x, x_mask, g=g, reverse=reverse)
|
324 |
+
else:
|
325 |
+
for flow in reversed(self.flows):
|
326 |
+
x = flow(x, x_mask, g=g, reverse=reverse)
|
327 |
+
return x
|
328 |
+
|
329 |
+
|
330 |
+
class PosteriorEncoder(nn.Module):
|
331 |
+
def __init__(
|
332 |
+
self,
|
333 |
+
in_channels,
|
334 |
+
out_channels,
|
335 |
+
hidden_channels,
|
336 |
+
kernel_size,
|
337 |
+
dilation_rate,
|
338 |
+
n_layers,
|
339 |
+
gin_channels=0,
|
340 |
+
):
|
341 |
+
super().__init__()
|
342 |
+
self.in_channels = in_channels
|
343 |
+
self.out_channels = out_channels
|
344 |
+
self.hidden_channels = hidden_channels
|
345 |
+
self.kernel_size = kernel_size
|
346 |
+
self.dilation_rate = dilation_rate
|
347 |
+
self.n_layers = n_layers
|
348 |
+
self.gin_channels = gin_channels
|
349 |
+
|
350 |
+
self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
|
351 |
+
self.enc = modules.WN(
|
352 |
+
hidden_channels,
|
353 |
+
kernel_size,
|
354 |
+
dilation_rate,
|
355 |
+
n_layers,
|
356 |
+
gin_channels=gin_channels,
|
357 |
+
)
|
358 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
|
359 |
+
|
360 |
+
def forward(self, x, x_lengths, g=None):
|
361 |
+
if g != None:
|
362 |
+
g = g.detach()
|
363 |
+
x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
|
364 |
+
x.dtype
|
365 |
+
)
|
366 |
+
x = self.pre(x) * x_mask
|
367 |
+
x = self.enc(x, x_mask, g=g)
|
368 |
+
stats = self.proj(x) * x_mask
|
369 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
370 |
+
z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
|
371 |
+
return z, m, logs, x_mask
|
372 |
+
|
373 |
+
|
374 |
+
class WNEncoder(nn.Module):
|
375 |
+
def __init__(
|
376 |
+
self,
|
377 |
+
in_channels,
|
378 |
+
out_channels,
|
379 |
+
hidden_channels,
|
380 |
+
kernel_size,
|
381 |
+
dilation_rate,
|
382 |
+
n_layers,
|
383 |
+
gin_channels=0,
|
384 |
+
):
|
385 |
+
super().__init__()
|
386 |
+
self.in_channels = in_channels
|
387 |
+
self.out_channels = out_channels
|
388 |
+
self.hidden_channels = hidden_channels
|
389 |
+
self.kernel_size = kernel_size
|
390 |
+
self.dilation_rate = dilation_rate
|
391 |
+
self.n_layers = n_layers
|
392 |
+
self.gin_channels = gin_channels
|
393 |
+
|
394 |
+
self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
|
395 |
+
self.enc = modules.WN(
|
396 |
+
hidden_channels,
|
397 |
+
kernel_size,
|
398 |
+
dilation_rate,
|
399 |
+
n_layers,
|
400 |
+
gin_channels=gin_channels,
|
401 |
+
)
|
402 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
|
403 |
+
self.norm = modules.LayerNorm(out_channels)
|
404 |
+
|
405 |
+
def forward(self, x, x_lengths, g=None):
|
406 |
+
x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
|
407 |
+
x.dtype
|
408 |
+
)
|
409 |
+
x = self.pre(x) * x_mask
|
410 |
+
x = self.enc(x, x_mask, g=g)
|
411 |
+
out = self.proj(x) * x_mask
|
412 |
+
out = self.norm(out)
|
413 |
+
return out
|
414 |
+
|
415 |
+
|
416 |
+
class Generator(torch.nn.Module):
|
417 |
+
def __init__(
|
418 |
+
self,
|
419 |
+
initial_channel,
|
420 |
+
resblock,
|
421 |
+
resblock_kernel_sizes,
|
422 |
+
resblock_dilation_sizes,
|
423 |
+
upsample_rates,
|
424 |
+
upsample_initial_channel,
|
425 |
+
upsample_kernel_sizes,
|
426 |
+
gin_channels=0,
|
427 |
+
):
|
428 |
+
super(Generator, self).__init__()
|
429 |
+
self.num_kernels = len(resblock_kernel_sizes)
|
430 |
+
self.num_upsamples = len(upsample_rates)
|
431 |
+
self.conv_pre = Conv1d(
|
432 |
+
initial_channel, upsample_initial_channel, 7, 1, padding=3
|
433 |
+
)
|
434 |
+
resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
|
435 |
+
|
436 |
+
self.ups = nn.ModuleList()
|
437 |
+
for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
|
438 |
+
self.ups.append(
|
439 |
+
weight_norm(
|
440 |
+
ConvTranspose1d(
|
441 |
+
upsample_initial_channel // (2**i),
|
442 |
+
upsample_initial_channel // (2 ** (i + 1)),
|
443 |
+
k,
|
444 |
+
u,
|
445 |
+
padding=(k - u) // 2,
|
446 |
+
)
|
447 |
+
)
|
448 |
+
)
|
449 |
+
|
450 |
+
self.resblocks = nn.ModuleList()
|
451 |
+
for i in range(len(self.ups)):
|
452 |
+
ch = upsample_initial_channel // (2 ** (i + 1))
|
453 |
+
for j, (k, d) in enumerate(
|
454 |
+
zip(resblock_kernel_sizes, resblock_dilation_sizes)
|
455 |
+
):
|
456 |
+
self.resblocks.append(resblock(ch, k, d))
|
457 |
+
|
458 |
+
self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
|
459 |
+
self.ups.apply(init_weights)
|
460 |
+
|
461 |
+
if gin_channels != 0:
|
462 |
+
self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
|
463 |
+
|
464 |
+
def forward(self, x, g=None):
|
465 |
+
x = self.conv_pre(x)
|
466 |
+
if g is not None:
|
467 |
+
x = x + self.cond(g)
|
468 |
+
|
469 |
+
for i in range(self.num_upsamples):
|
470 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
471 |
+
x = self.ups[i](x)
|
472 |
+
xs = None
|
473 |
+
for j in range(self.num_kernels):
|
474 |
+
if xs is None:
|
475 |
+
xs = self.resblocks[i * self.num_kernels + j](x)
|
476 |
+
else:
|
477 |
+
xs += self.resblocks[i * self.num_kernels + j](x)
|
478 |
+
x = xs / self.num_kernels
|
479 |
+
x = F.leaky_relu(x)
|
480 |
+
x = self.conv_post(x)
|
481 |
+
x = torch.tanh(x)
|
482 |
+
|
483 |
+
return x
|
484 |
+
|
485 |
+
def remove_weight_norm(self):
|
486 |
+
print("Removing weight norm...")
|
487 |
+
for l in self.ups:
|
488 |
+
remove_weight_norm(l)
|
489 |
+
for l in self.resblocks:
|
490 |
+
l.remove_weight_norm()
|
491 |
+
|
492 |
+
|
493 |
+
class DiscriminatorP(torch.nn.Module):
|
494 |
+
def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
|
495 |
+
super(DiscriminatorP, self).__init__()
|
496 |
+
self.period = period
|
497 |
+
self.use_spectral_norm = use_spectral_norm
|
498 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
499 |
+
self.convs = nn.ModuleList(
|
500 |
+
[
|
501 |
+
norm_f(
|
502 |
+
Conv2d(
|
503 |
+
1,
|
504 |
+
32,
|
505 |
+
(kernel_size, 1),
|
506 |
+
(stride, 1),
|
507 |
+
padding=(get_padding(kernel_size, 1), 0),
|
508 |
+
)
|
509 |
+
),
|
510 |
+
norm_f(
|
511 |
+
Conv2d(
|
512 |
+
32,
|
513 |
+
128,
|
514 |
+
(kernel_size, 1),
|
515 |
+
(stride, 1),
|
516 |
+
padding=(get_padding(kernel_size, 1), 0),
|
517 |
+
)
|
518 |
+
),
|
519 |
+
norm_f(
|
520 |
+
Conv2d(
|
521 |
+
128,
|
522 |
+
512,
|
523 |
+
(kernel_size, 1),
|
524 |
+
(stride, 1),
|
525 |
+
padding=(get_padding(kernel_size, 1), 0),
|
526 |
+
)
|
527 |
+
),
|
528 |
+
norm_f(
|
529 |
+
Conv2d(
|
530 |
+
512,
|
531 |
+
1024,
|
532 |
+
(kernel_size, 1),
|
533 |
+
(stride, 1),
|
534 |
+
padding=(get_padding(kernel_size, 1), 0),
|
535 |
+
)
|
536 |
+
),
|
537 |
+
norm_f(
|
538 |
+
Conv2d(
|
539 |
+
1024,
|
540 |
+
1024,
|
541 |
+
(kernel_size, 1),
|
542 |
+
1,
|
543 |
+
padding=(get_padding(kernel_size, 1), 0),
|
544 |
+
)
|
545 |
+
),
|
546 |
+
]
|
547 |
+
)
|
548 |
+
self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
|
549 |
+
|
550 |
+
def forward(self, x):
|
551 |
+
fmap = []
|
552 |
+
|
553 |
+
# 1d to 2d
|
554 |
+
b, c, t = x.shape
|
555 |
+
if t % self.period != 0: # pad first
|
556 |
+
n_pad = self.period - (t % self.period)
|
557 |
+
x = F.pad(x, (0, n_pad), "reflect")
|
558 |
+
t = t + n_pad
|
559 |
+
x = x.view(b, c, t // self.period, self.period)
|
560 |
+
|
561 |
+
for l in self.convs:
|
562 |
+
x = l(x)
|
563 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
564 |
+
fmap.append(x)
|
565 |
+
x = self.conv_post(x)
|
566 |
+
fmap.append(x)
|
567 |
+
x = torch.flatten(x, 1, -1)
|
568 |
+
|
569 |
+
return x, fmap
|
570 |
+
|
571 |
+
|
572 |
+
class DiscriminatorS(torch.nn.Module):
|
573 |
+
def __init__(self, use_spectral_norm=False):
|
574 |
+
super(DiscriminatorS, self).__init__()
|
575 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
576 |
+
self.convs = nn.ModuleList(
|
577 |
+
[
|
578 |
+
norm_f(Conv1d(1, 16, 15, 1, padding=7)),
|
579 |
+
norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
|
580 |
+
norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
|
581 |
+
norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
|
582 |
+
norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
|
583 |
+
norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
|
584 |
+
]
|
585 |
+
)
|
586 |
+
self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
|
587 |
+
|
588 |
+
def forward(self, x):
|
589 |
+
fmap = []
|
590 |
+
|
591 |
+
for l in self.convs:
|
592 |
+
x = l(x)
|
593 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
594 |
+
fmap.append(x)
|
595 |
+
x = self.conv_post(x)
|
596 |
+
fmap.append(x)
|
597 |
+
x = torch.flatten(x, 1, -1)
|
598 |
+
|
599 |
+
return x, fmap
|
600 |
+
|
601 |
+
|
602 |
+
class MultiPeriodDiscriminator(torch.nn.Module):
|
603 |
+
def __init__(self, use_spectral_norm=False):
|
604 |
+
super(MultiPeriodDiscriminator, self).__init__()
|
605 |
+
periods = [2, 3, 5, 7, 11]
|
606 |
+
|
607 |
+
discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
|
608 |
+
discs = discs + [
|
609 |
+
DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
|
610 |
+
]
|
611 |
+
self.discriminators = nn.ModuleList(discs)
|
612 |
+
|
613 |
+
def forward(self, y, y_hat):
|
614 |
+
y_d_rs = []
|
615 |
+
y_d_gs = []
|
616 |
+
fmap_rs = []
|
617 |
+
fmap_gs = []
|
618 |
+
for i, d in enumerate(self.discriminators):
|
619 |
+
y_d_r, fmap_r = d(y)
|
620 |
+
y_d_g, fmap_g = d(y_hat)
|
621 |
+
y_d_rs.append(y_d_r)
|
622 |
+
y_d_gs.append(y_d_g)
|
623 |
+
fmap_rs.append(fmap_r)
|
624 |
+
fmap_gs.append(fmap_g)
|
625 |
+
|
626 |
+
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
627 |
+
|
628 |
+
|
629 |
+
class ReferenceEncoder(nn.Module):
|
630 |
+
"""
|
631 |
+
inputs --- [N, Ty/r, n_mels*r] mels
|
632 |
+
outputs --- [N, ref_enc_gru_size]
|
633 |
+
"""
|
634 |
+
|
635 |
+
def __init__(self, spec_channels, gin_channels=0):
|
636 |
+
super().__init__()
|
637 |
+
self.spec_channels = spec_channels
|
638 |
+
ref_enc_filters = [32, 32, 64, 64, 128, 128]
|
639 |
+
K = len(ref_enc_filters)
|
640 |
+
filters = [1] + ref_enc_filters
|
641 |
+
convs = [
|
642 |
+
weight_norm(
|
643 |
+
nn.Conv2d(
|
644 |
+
in_channels=filters[i],
|
645 |
+
out_channels=filters[i + 1],
|
646 |
+
kernel_size=(3, 3),
|
647 |
+
stride=(2, 2),
|
648 |
+
padding=(1, 1),
|
649 |
+
)
|
650 |
+
)
|
651 |
+
for i in range(K)
|
652 |
+
]
|
653 |
+
self.convs = nn.ModuleList(convs)
|
654 |
+
# self.wns = nn.ModuleList([weight_norm(num_features=ref_enc_filters[i]) for i in range(K)])
|
655 |
+
|
656 |
+
out_channels = self.calculate_channels(spec_channels, 3, 2, 1, K)
|
657 |
+
self.gru = nn.GRU(
|
658 |
+
input_size=ref_enc_filters[-1] * out_channels,
|
659 |
+
hidden_size=256 // 2,
|
660 |
+
batch_first=True,
|
661 |
+
)
|
662 |
+
self.proj = nn.Linear(128, gin_channels)
|
663 |
+
|
664 |
+
def forward(self, inputs):
|
665 |
+
N = inputs.size(0)
|
666 |
+
out = inputs.view(N, 1, -1, self.spec_channels) # [N, 1, Ty, n_freqs]
|
667 |
+
for conv in self.convs:
|
668 |
+
out = conv(out)
|
669 |
+
# out = wn(out)
|
670 |
+
out = F.relu(out) # [N, 128, Ty//2^K, n_mels//2^K]
|
671 |
+
|
672 |
+
out = out.transpose(1, 2) # [N, Ty//2^K, 128, n_mels//2^K]
|
673 |
+
T = out.size(1)
|
674 |
+
N = out.size(0)
|
675 |
+
out = out.contiguous().view(N, T, -1) # [N, Ty//2^K, 128*n_mels//2^K]
|
676 |
+
|
677 |
+
self.gru.flatten_parameters()
|
678 |
+
memory, out = self.gru(out) # out --- [1, N, 128]
|
679 |
+
|
680 |
+
return self.proj(out.squeeze(0)).unsqueeze(-1)
|
681 |
+
|
682 |
+
def calculate_channels(self, L, kernel_size, stride, pad, n_convs):
|
683 |
+
for i in range(n_convs):
|
684 |
+
L = (L - kernel_size + 2 * pad) // stride + 1
|
685 |
+
return L
|
686 |
+
|
687 |
+
|
688 |
+
class Quantizer_module(torch.nn.Module):
|
689 |
+
def __init__(self, n_e, e_dim):
|
690 |
+
super(Quantizer_module, self).__init__()
|
691 |
+
self.embedding = nn.Embedding(n_e, e_dim)
|
692 |
+
self.embedding.weight.data.uniform_(-1.0 / n_e, 1.0 / n_e)
|
693 |
+
|
694 |
+
def forward(self, x):
|
695 |
+
d = (
|
696 |
+
torch.sum(x**2, 1, keepdim=True)
|
697 |
+
+ torch.sum(self.embedding.weight**2, 1)
|
698 |
+
- 2 * torch.matmul(x, self.embedding.weight.T)
|
699 |
+
)
|
700 |
+
min_indicies = torch.argmin(d, 1)
|
701 |
+
z_q = self.embedding(min_indicies)
|
702 |
+
return z_q, min_indicies
|
703 |
+
|
704 |
+
|
705 |
+
class Quantizer(torch.nn.Module):
|
706 |
+
def __init__(self, embed_dim=512, n_code_groups=4, n_codes=160):
|
707 |
+
super(Quantizer, self).__init__()
|
708 |
+
assert embed_dim % n_code_groups == 0
|
709 |
+
self.quantizer_modules = nn.ModuleList(
|
710 |
+
[
|
711 |
+
Quantizer_module(n_codes, embed_dim // n_code_groups)
|
712 |
+
for _ in range(n_code_groups)
|
713 |
+
]
|
714 |
+
)
|
715 |
+
self.n_code_groups = n_code_groups
|
716 |
+
self.embed_dim = embed_dim
|
717 |
+
|
718 |
+
def forward(self, xin):
|
719 |
+
# B, C, T
|
720 |
+
B, C, T = xin.shape
|
721 |
+
xin = xin.transpose(1, 2)
|
722 |
+
x = xin.reshape(-1, self.embed_dim)
|
723 |
+
x = torch.split(x, self.embed_dim // self.n_code_groups, dim=-1)
|
724 |
+
min_indicies = []
|
725 |
+
z_q = []
|
726 |
+
for _x, m in zip(x, self.quantizer_modules):
|
727 |
+
_z_q, _min_indicies = m(_x)
|
728 |
+
z_q.append(_z_q)
|
729 |
+
min_indicies.append(_min_indicies) # B * T,
|
730 |
+
z_q = torch.cat(z_q, -1).reshape(xin.shape)
|
731 |
+
loss = 0.25 * torch.mean((z_q.detach() - xin) ** 2) + torch.mean(
|
732 |
+
(z_q - xin.detach()) ** 2
|
733 |
+
)
|
734 |
+
z_q = xin + (z_q - xin).detach()
|
735 |
+
z_q = z_q.transpose(1, 2)
|
736 |
+
codes = torch.stack(min_indicies, -1).reshape(B, T, self.n_code_groups)
|
737 |
+
return z_q, loss, codes.transpose(1, 2)
|
738 |
+
|
739 |
+
def embed(self, x):
|
740 |
+
# idx: N, 4, T
|
741 |
+
x = x.transpose(1, 2)
|
742 |
+
x = torch.split(x, 1, 2)
|
743 |
+
ret = []
|
744 |
+
for q, embed in zip(x, self.quantizer_modules):
|
745 |
+
q = embed.embedding(q.squeeze(-1))
|
746 |
+
ret.append(q)
|
747 |
+
ret = torch.cat(ret, -1)
|
748 |
+
return ret.transpose(1, 2) # N, C, T
|
749 |
+
|
750 |
+
|
751 |
+
class CodePredictor(nn.Module):
|
752 |
+
def __init__(
|
753 |
+
self,
|
754 |
+
hidden_channels,
|
755 |
+
filter_channels,
|
756 |
+
n_heads,
|
757 |
+
n_layers,
|
758 |
+
kernel_size,
|
759 |
+
p_dropout,
|
760 |
+
n_q=8,
|
761 |
+
dims=1024,
|
762 |
+
ssl_dim=768,
|
763 |
+
):
|
764 |
+
super().__init__()
|
765 |
+
self.hidden_channels = hidden_channels
|
766 |
+
self.filter_channels = filter_channels
|
767 |
+
self.n_heads = n_heads
|
768 |
+
self.n_layers = n_layers
|
769 |
+
self.kernel_size = kernel_size
|
770 |
+
self.p_dropout = p_dropout
|
771 |
+
|
772 |
+
self.vq_proj = nn.Conv1d(ssl_dim, hidden_channels, 1)
|
773 |
+
self.ref_enc = modules.MelStyleEncoder(
|
774 |
+
ssl_dim, style_vector_dim=hidden_channels
|
775 |
+
)
|
776 |
+
|
777 |
+
self.encoder = attentions.Encoder(
|
778 |
+
hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
|
779 |
+
)
|
780 |
+
|
781 |
+
self.out_proj = nn.Conv1d(hidden_channels, (n_q - 1) * dims, 1)
|
782 |
+
self.n_q = n_q
|
783 |
+
self.dims = dims
|
784 |
+
|
785 |
+
def forward(self, x, x_mask, refer, codes, infer=False):
|
786 |
+
x = x.detach()
|
787 |
+
x = self.vq_proj(x * x_mask) * x_mask
|
788 |
+
g = self.ref_enc(refer, x_mask)
|
789 |
+
x = x + g
|
790 |
+
x = self.encoder(x * x_mask, x_mask)
|
791 |
+
x = self.out_proj(x * x_mask) * x_mask
|
792 |
+
logits = x.reshape(x.shape[0], self.n_q - 1, self.dims, x.shape[-1]).transpose(
|
793 |
+
2, 3
|
794 |
+
)
|
795 |
+
target = codes[1:].transpose(0, 1)
|
796 |
+
if not infer:
|
797 |
+
logits = logits.reshape(-1, self.dims)
|
798 |
+
target = target.reshape(-1)
|
799 |
+
loss = torch.nn.functional.cross_entropy(logits, target)
|
800 |
+
return loss
|
801 |
+
else:
|
802 |
+
_, top10_preds = torch.topk(logits, 10, dim=-1)
|
803 |
+
correct_top10 = torch.any(top10_preds == target.unsqueeze(-1), dim=-1)
|
804 |
+
top3_acc = 100 * torch.mean(correct_top10.float()).detach().cpu().item()
|
805 |
+
|
806 |
+
print("Top-10 Accuracy:", top3_acc, "%")
|
807 |
+
|
808 |
+
pred_codes = torch.argmax(logits, dim=-1)
|
809 |
+
acc = 100 * torch.mean((pred_codes == target).float()).detach().cpu().item()
|
810 |
+
print("Top-1 Accuracy:", acc, "%")
|
811 |
+
|
812 |
+
return pred_codes.transpose(0, 1)
|
813 |
+
|
814 |
+
|
815 |
+
class SynthesizerTrn(nn.Module):
|
816 |
+
"""
|
817 |
+
Synthesizer for Training
|
818 |
+
"""
|
819 |
+
|
820 |
+
def __init__(
|
821 |
+
self,
|
822 |
+
spec_channels,
|
823 |
+
segment_size,
|
824 |
+
inter_channels,
|
825 |
+
hidden_channels,
|
826 |
+
filter_channels,
|
827 |
+
n_heads,
|
828 |
+
n_layers,
|
829 |
+
kernel_size,
|
830 |
+
p_dropout,
|
831 |
+
resblock,
|
832 |
+
resblock_kernel_sizes,
|
833 |
+
resblock_dilation_sizes,
|
834 |
+
upsample_rates,
|
835 |
+
upsample_initial_channel,
|
836 |
+
upsample_kernel_sizes,
|
837 |
+
n_speakers=0,
|
838 |
+
gin_channels=0,
|
839 |
+
use_sdp=True,
|
840 |
+
semantic_frame_rate=None,
|
841 |
+
freeze_quantizer=None,
|
842 |
+
version = "v2",
|
843 |
+
**kwargs
|
844 |
+
):
|
845 |
+
super().__init__()
|
846 |
+
self.spec_channels = spec_channels
|
847 |
+
self.inter_channels = inter_channels
|
848 |
+
self.hidden_channels = hidden_channels
|
849 |
+
self.filter_channels = filter_channels
|
850 |
+
self.n_heads = n_heads
|
851 |
+
self.n_layers = n_layers
|
852 |
+
self.kernel_size = kernel_size
|
853 |
+
self.p_dropout = p_dropout
|
854 |
+
self.resblock = resblock
|
855 |
+
self.resblock_kernel_sizes = resblock_kernel_sizes
|
856 |
+
self.resblock_dilation_sizes = resblock_dilation_sizes
|
857 |
+
self.upsample_rates = upsample_rates
|
858 |
+
self.upsample_initial_channel = upsample_initial_channel
|
859 |
+
self.upsample_kernel_sizes = upsample_kernel_sizes
|
860 |
+
self.segment_size = segment_size
|
861 |
+
self.n_speakers = n_speakers
|
862 |
+
self.gin_channels = gin_channels
|
863 |
+
self.version = version
|
864 |
+
|
865 |
+
self.use_sdp = use_sdp
|
866 |
+
self.enc_p = TextEncoder(
|
867 |
+
inter_channels,
|
868 |
+
hidden_channels,
|
869 |
+
filter_channels,
|
870 |
+
n_heads,
|
871 |
+
n_layers,
|
872 |
+
kernel_size,
|
873 |
+
p_dropout,
|
874 |
+
version = version,
|
875 |
+
)
|
876 |
+
self.dec = Generator(
|
877 |
+
inter_channels,
|
878 |
+
resblock,
|
879 |
+
resblock_kernel_sizes,
|
880 |
+
resblock_dilation_sizes,
|
881 |
+
upsample_rates,
|
882 |
+
upsample_initial_channel,
|
883 |
+
upsample_kernel_sizes,
|
884 |
+
gin_channels=gin_channels,
|
885 |
+
)
|
886 |
+
self.enc_q = PosteriorEncoder(
|
887 |
+
spec_channels,
|
888 |
+
inter_channels,
|
889 |
+
hidden_channels,
|
890 |
+
5,
|
891 |
+
1,
|
892 |
+
16,
|
893 |
+
gin_channels=gin_channels,
|
894 |
+
)
|
895 |
+
self.flow = ResidualCouplingBlock(
|
896 |
+
inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels
|
897 |
+
)
|
898 |
+
|
899 |
+
# self.version=os.environ.get("version","v1")
|
900 |
+
if(self.version=="v1"):
|
901 |
+
self.ref_enc = modules.MelStyleEncoder(spec_channels, style_vector_dim=gin_channels)
|
902 |
+
else:
|
903 |
+
self.ref_enc = modules.MelStyleEncoder(spec_channels, style_vector_dim=gin_channels)
|
904 |
+
|
905 |
+
ssl_dim = 768
|
906 |
+
assert semantic_frame_rate in ["25hz", "50hz"]
|
907 |
+
self.semantic_frame_rate = semantic_frame_rate
|
908 |
+
if semantic_frame_rate == "25hz":
|
909 |
+
self.ssl_proj = nn.Conv1d(ssl_dim, ssl_dim, 2, stride=2)
|
910 |
+
else:
|
911 |
+
self.ssl_proj = nn.Conv1d(ssl_dim, ssl_dim, 1, stride=1)
|
912 |
+
|
913 |
+
self.quantizer = ResidualVectorQuantizer(dimension=ssl_dim, n_q=1, bins=1024)
|
914 |
+
self.freeze_quantizer = freeze_quantizer
|
915 |
+
|
916 |
+
def forward(self, ssl, y, y_lengths, text, text_lengths):
|
917 |
+
y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, y.size(2)), 1).to(
|
918 |
+
y.dtype
|
919 |
+
)
|
920 |
+
if(self.version=="v1"):
|
921 |
+
ge = self.ref_enc(y * y_mask, y_mask)
|
922 |
+
else:
|
923 |
+
ge = self.ref_enc(y * y_mask, y_mask)
|
924 |
+
with autocast(enabled=False):
|
925 |
+
maybe_no_grad = torch.no_grad() if self.freeze_quantizer else contextlib.nullcontext()
|
926 |
+
with maybe_no_grad:
|
927 |
+
if self.freeze_quantizer:
|
928 |
+
self.ssl_proj.eval()
|
929 |
+
self.quantizer.eval()
|
930 |
+
ssl = self.ssl_proj(ssl)
|
931 |
+
quantized, codes, commit_loss, quantized_list = self.quantizer(
|
932 |
+
ssl, layers=[0]
|
933 |
+
)
|
934 |
+
|
935 |
+
if self.semantic_frame_rate == "25hz":
|
936 |
+
quantized = F.interpolate(
|
937 |
+
quantized, size=int(quantized.shape[-1] * 2), mode="nearest"
|
938 |
+
)
|
939 |
+
|
940 |
+
x, m_p, logs_p, y_mask = self.enc_p(
|
941 |
+
quantized, y_lengths, text, text_lengths, ge
|
942 |
+
)
|
943 |
+
z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=ge)
|
944 |
+
z_p = self.flow(z, y_mask, g=ge)
|
945 |
+
|
946 |
+
z_slice, ids_slice = commons.rand_slice_segments(
|
947 |
+
z, y_lengths, self.segment_size
|
948 |
+
)
|
949 |
+
o = self.dec(z_slice, g=ge)
|
950 |
+
return (
|
951 |
+
o,
|
952 |
+
commit_loss,
|
953 |
+
ids_slice,
|
954 |
+
y_mask,
|
955 |
+
y_mask,
|
956 |
+
(z, z_p, m_p, logs_p, m_q, logs_q),
|
957 |
+
quantized,
|
958 |
+
)
|
959 |
+
|
960 |
+
def infer(self, ssl, y, y_lengths, text, text_lengths, test=None, noise_scale=0.5):
|
961 |
+
y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, y.size(2)), 1).to(
|
962 |
+
y.dtype
|
963 |
+
)
|
964 |
+
if(self.version=="v1"):
|
965 |
+
ge = self.ref_enc(y * y_mask, y_mask)
|
966 |
+
else:
|
967 |
+
ge = self.ref_enc(y * y_mask, y_mask)
|
968 |
+
|
969 |
+
ssl = self.ssl_proj(ssl)
|
970 |
+
quantized, codes, commit_loss, _ = self.quantizer(ssl, layers=[0])
|
971 |
+
if self.semantic_frame_rate == "25hz":
|
972 |
+
quantized = F.interpolate(
|
973 |
+
quantized, size=int(quantized.shape[-1] * 2), mode="nearest"
|
974 |
+
)
|
975 |
+
|
976 |
+
x, m_p, logs_p, y_mask = self.enc_p(
|
977 |
+
quantized, y_lengths, text, text_lengths, ge, test=test
|
978 |
+
)
|
979 |
+
z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
|
980 |
+
|
981 |
+
z = self.flow(z_p, y_mask, g=ge, reverse=True)
|
982 |
+
|
983 |
+
o = self.dec((z * y_mask)[:, :, :], g=ge)
|
984 |
+
return o, y_mask, (z, z_p, m_p, logs_p)
|
985 |
+
|
986 |
+
@torch.no_grad()
|
987 |
+
def decode(self, codes, text, refer, noise_scale=0.5,speed=1):
|
988 |
+
def get_ge(refer):
|
989 |
+
ge = None
|
990 |
+
if refer is not None:
|
991 |
+
refer_lengths = torch.LongTensor([refer.size(2)]).to(refer.device)
|
992 |
+
refer_mask = torch.unsqueeze(
|
993 |
+
commons.sequence_mask(refer_lengths, refer.size(2)), 1
|
994 |
+
).to(refer.dtype)
|
995 |
+
if (self.version == "v1"):
|
996 |
+
ge = self.ref_enc(refer * refer_mask, refer_mask)
|
997 |
+
else:
|
998 |
+
ge = self.ref_enc(refer * refer_mask, refer_mask)
|
999 |
+
return ge
|
1000 |
+
if(type(refer)==list):
|
1001 |
+
ges=[]
|
1002 |
+
for _refer in refer:
|
1003 |
+
ge=get_ge(_refer)
|
1004 |
+
ges.append(ge)
|
1005 |
+
ge=torch.stack(ges,0).mean(0)
|
1006 |
+
else:
|
1007 |
+
ge=get_ge(refer)
|
1008 |
+
|
1009 |
+
y_lengths = torch.LongTensor([codes.size(2) * 2]).to(codes.device)
|
1010 |
+
text_lengths = torch.LongTensor([text.size(-1)]).to(text.device)
|
1011 |
+
|
1012 |
+
quantized = self.quantizer.decode(codes)
|
1013 |
+
if self.semantic_frame_rate == "25hz":
|
1014 |
+
quantized = F.interpolate(
|
1015 |
+
quantized, size=int(quantized.shape[-1] * 2), mode="nearest"
|
1016 |
+
)
|
1017 |
+
x, m_p, logs_p, y_mask = self.enc_p(
|
1018 |
+
quantized, y_lengths, text, text_lengths, ge,speed
|
1019 |
+
)
|
1020 |
+
z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
|
1021 |
+
|
1022 |
+
z = self.flow(z_p, y_mask, g=ge, reverse=True)
|
1023 |
+
|
1024 |
+
o = self.dec((z * y_mask)[:, :, :], g=ge)
|
1025 |
+
return o
|
1026 |
+
|
1027 |
+
def extract_latent(self, x):
|
1028 |
+
ssl = self.ssl_proj(x)
|
1029 |
+
quantized, codes, commit_loss, quantized_list = self.quantizer(ssl)
|
1030 |
+
return codes.transpose(0, 1)
|