Create infer_pack/modules.py
Browse files- infer_pack/modules.py +522 -0
infer_pack/modules.py
ADDED
@@ -0,0 +1,522 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import copy
|
2 |
+
import math
|
3 |
+
import numpy as np
|
4 |
+
import scipy
|
5 |
+
import torch
|
6 |
+
from torch import nn
|
7 |
+
from torch.nn import functional as F
|
8 |
+
|
9 |
+
from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
|
10 |
+
from torch.nn.utils import weight_norm, remove_weight_norm
|
11 |
+
|
12 |
+
from infer_pack import commons
|
13 |
+
from infer_pack.commons import init_weights, get_padding
|
14 |
+
from infer_pack.transforms import piecewise_rational_quadratic_transform
|
15 |
+
|
16 |
+
|
17 |
+
LRELU_SLOPE = 0.1
|
18 |
+
|
19 |
+
|
20 |
+
class LayerNorm(nn.Module):
|
21 |
+
def __init__(self, channels, eps=1e-5):
|
22 |
+
super().__init__()
|
23 |
+
self.channels = channels
|
24 |
+
self.eps = eps
|
25 |
+
|
26 |
+
self.gamma = nn.Parameter(torch.ones(channels))
|
27 |
+
self.beta = nn.Parameter(torch.zeros(channels))
|
28 |
+
|
29 |
+
def forward(self, x):
|
30 |
+
x = x.transpose(1, -1)
|
31 |
+
x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
|
32 |
+
return x.transpose(1, -1)
|
33 |
+
|
34 |
+
|
35 |
+
class ConvReluNorm(nn.Module):
|
36 |
+
def __init__(
|
37 |
+
self,
|
38 |
+
in_channels,
|
39 |
+
hidden_channels,
|
40 |
+
out_channels,
|
41 |
+
kernel_size,
|
42 |
+
n_layers,
|
43 |
+
p_dropout,
|
44 |
+
):
|
45 |
+
super().__init__()
|
46 |
+
self.in_channels = in_channels
|
47 |
+
self.hidden_channels = hidden_channels
|
48 |
+
self.out_channels = out_channels
|
49 |
+
self.kernel_size = kernel_size
|
50 |
+
self.n_layers = n_layers
|
51 |
+
self.p_dropout = p_dropout
|
52 |
+
assert n_layers > 1, "Number of layers should be larger than 0."
|
53 |
+
|
54 |
+
self.conv_layers = nn.ModuleList()
|
55 |
+
self.norm_layers = nn.ModuleList()
|
56 |
+
self.conv_layers.append(
|
57 |
+
nn.Conv1d(
|
58 |
+
in_channels, hidden_channels, kernel_size, padding=kernel_size // 2
|
59 |
+
)
|
60 |
+
)
|
61 |
+
self.norm_layers.append(LayerNorm(hidden_channels))
|
62 |
+
self.relu_drop = nn.Sequential(nn.ReLU(), nn.Dropout(p_dropout))
|
63 |
+
for _ in range(n_layers - 1):
|
64 |
+
self.conv_layers.append(
|
65 |
+
nn.Conv1d(
|
66 |
+
hidden_channels,
|
67 |
+
hidden_channels,
|
68 |
+
kernel_size,
|
69 |
+
padding=kernel_size // 2,
|
70 |
+
)
|
71 |
+
)
|
72 |
+
self.norm_layers.append(LayerNorm(hidden_channels))
|
73 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
|
74 |
+
self.proj.weight.data.zero_()
|
75 |
+
self.proj.bias.data.zero_()
|
76 |
+
|
77 |
+
def forward(self, x, x_mask):
|
78 |
+
x_org = x
|
79 |
+
for i in range(self.n_layers):
|
80 |
+
x = self.conv_layers[i](x * x_mask)
|
81 |
+
x = self.norm_layers[i](x)
|
82 |
+
x = self.relu_drop(x)
|
83 |
+
x = x_org + self.proj(x)
|
84 |
+
return x * x_mask
|
85 |
+
|
86 |
+
|
87 |
+
class DDSConv(nn.Module):
|
88 |
+
"""
|
89 |
+
Dialted and Depth-Separable Convolution
|
90 |
+
"""
|
91 |
+
|
92 |
+
def __init__(self, channels, kernel_size, n_layers, p_dropout=0.0):
|
93 |
+
super().__init__()
|
94 |
+
self.channels = channels
|
95 |
+
self.kernel_size = kernel_size
|
96 |
+
self.n_layers = n_layers
|
97 |
+
self.p_dropout = p_dropout
|
98 |
+
|
99 |
+
self.drop = nn.Dropout(p_dropout)
|
100 |
+
self.convs_sep = nn.ModuleList()
|
101 |
+
self.convs_1x1 = nn.ModuleList()
|
102 |
+
self.norms_1 = nn.ModuleList()
|
103 |
+
self.norms_2 = nn.ModuleList()
|
104 |
+
for i in range(n_layers):
|
105 |
+
dilation = kernel_size**i
|
106 |
+
padding = (kernel_size * dilation - dilation) // 2
|
107 |
+
self.convs_sep.append(
|
108 |
+
nn.Conv1d(
|
109 |
+
channels,
|
110 |
+
channels,
|
111 |
+
kernel_size,
|
112 |
+
groups=channels,
|
113 |
+
dilation=dilation,
|
114 |
+
padding=padding,
|
115 |
+
)
|
116 |
+
)
|
117 |
+
self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
|
118 |
+
self.norms_1.append(LayerNorm(channels))
|
119 |
+
self.norms_2.append(LayerNorm(channels))
|
120 |
+
|
121 |
+
def forward(self, x, x_mask, g=None):
|
122 |
+
if g is not None:
|
123 |
+
x = x + g
|
124 |
+
for i in range(self.n_layers):
|
125 |
+
y = self.convs_sep[i](x * x_mask)
|
126 |
+
y = self.norms_1[i](y)
|
127 |
+
y = F.gelu(y)
|
128 |
+
y = self.convs_1x1[i](y)
|
129 |
+
y = self.norms_2[i](y)
|
130 |
+
y = F.gelu(y)
|
131 |
+
y = self.drop(y)
|
132 |
+
x = x + y
|
133 |
+
return x * x_mask
|
134 |
+
|
135 |
+
|
136 |
+
class WN(torch.nn.Module):
|
137 |
+
def __init__(
|
138 |
+
self,
|
139 |
+
hidden_channels,
|
140 |
+
kernel_size,
|
141 |
+
dilation_rate,
|
142 |
+
n_layers,
|
143 |
+
gin_channels=0,
|
144 |
+
p_dropout=0,
|
145 |
+
):
|
146 |
+
super(WN, self).__init__()
|
147 |
+
assert kernel_size % 2 == 1
|
148 |
+
self.hidden_channels = hidden_channels
|
149 |
+
self.kernel_size = (kernel_size,)
|
150 |
+
self.dilation_rate = dilation_rate
|
151 |
+
self.n_layers = n_layers
|
152 |
+
self.gin_channels = gin_channels
|
153 |
+
self.p_dropout = p_dropout
|
154 |
+
|
155 |
+
self.in_layers = torch.nn.ModuleList()
|
156 |
+
self.res_skip_layers = torch.nn.ModuleList()
|
157 |
+
self.drop = nn.Dropout(p_dropout)
|
158 |
+
|
159 |
+
if gin_channels != 0:
|
160 |
+
cond_layer = torch.nn.Conv1d(
|
161 |
+
gin_channels, 2 * hidden_channels * n_layers, 1
|
162 |
+
)
|
163 |
+
self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name="weight")
|
164 |
+
|
165 |
+
for i in range(n_layers):
|
166 |
+
dilation = dilation_rate**i
|
167 |
+
padding = int((kernel_size * dilation - dilation) / 2)
|
168 |
+
in_layer = torch.nn.Conv1d(
|
169 |
+
hidden_channels,
|
170 |
+
2 * hidden_channels,
|
171 |
+
kernel_size,
|
172 |
+
dilation=dilation,
|
173 |
+
padding=padding,
|
174 |
+
)
|
175 |
+
in_layer = torch.nn.utils.weight_norm(in_layer, name="weight")
|
176 |
+
self.in_layers.append(in_layer)
|
177 |
+
|
178 |
+
# last one is not necessary
|
179 |
+
if i < n_layers - 1:
|
180 |
+
res_skip_channels = 2 * hidden_channels
|
181 |
+
else:
|
182 |
+
res_skip_channels = hidden_channels
|
183 |
+
|
184 |
+
res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
|
185 |
+
res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name="weight")
|
186 |
+
self.res_skip_layers.append(res_skip_layer)
|
187 |
+
|
188 |
+
def forward(self, x, x_mask, g=None, **kwargs):
|
189 |
+
output = torch.zeros_like(x)
|
190 |
+
n_channels_tensor = torch.IntTensor([self.hidden_channels])
|
191 |
+
|
192 |
+
if g is not None:
|
193 |
+
g = self.cond_layer(g)
|
194 |
+
|
195 |
+
for i in range(self.n_layers):
|
196 |
+
x_in = self.in_layers[i](x)
|
197 |
+
if g is not None:
|
198 |
+
cond_offset = i * 2 * self.hidden_channels
|
199 |
+
g_l = g[:, cond_offset : cond_offset + 2 * self.hidden_channels, :]
|
200 |
+
else:
|
201 |
+
g_l = torch.zeros_like(x_in)
|
202 |
+
|
203 |
+
acts = commons.fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)
|
204 |
+
acts = self.drop(acts)
|
205 |
+
|
206 |
+
res_skip_acts = self.res_skip_layers[i](acts)
|
207 |
+
if i < self.n_layers - 1:
|
208 |
+
res_acts = res_skip_acts[:, : self.hidden_channels, :]
|
209 |
+
x = (x + res_acts) * x_mask
|
210 |
+
output = output + res_skip_acts[:, self.hidden_channels :, :]
|
211 |
+
else:
|
212 |
+
output = output + res_skip_acts
|
213 |
+
return output * x_mask
|
214 |
+
|
215 |
+
def remove_weight_norm(self):
|
216 |
+
if self.gin_channels != 0:
|
217 |
+
torch.nn.utils.remove_weight_norm(self.cond_layer)
|
218 |
+
for l in self.in_layers:
|
219 |
+
torch.nn.utils.remove_weight_norm(l)
|
220 |
+
for l in self.res_skip_layers:
|
221 |
+
torch.nn.utils.remove_weight_norm(l)
|
222 |
+
|
223 |
+
|
224 |
+
class ResBlock1(torch.nn.Module):
|
225 |
+
def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
|
226 |
+
super(ResBlock1, self).__init__()
|
227 |
+
self.convs1 = nn.ModuleList(
|
228 |
+
[
|
229 |
+
weight_norm(
|
230 |
+
Conv1d(
|
231 |
+
channels,
|
232 |
+
channels,
|
233 |
+
kernel_size,
|
234 |
+
1,
|
235 |
+
dilation=dilation[0],
|
236 |
+
padding=get_padding(kernel_size, dilation[0]),
|
237 |
+
)
|
238 |
+
),
|
239 |
+
weight_norm(
|
240 |
+
Conv1d(
|
241 |
+
channels,
|
242 |
+
channels,
|
243 |
+
kernel_size,
|
244 |
+
1,
|
245 |
+
dilation=dilation[1],
|
246 |
+
padding=get_padding(kernel_size, dilation[1]),
|
247 |
+
)
|
248 |
+
),
|
249 |
+
weight_norm(
|
250 |
+
Conv1d(
|
251 |
+
channels,
|
252 |
+
channels,
|
253 |
+
kernel_size,
|
254 |
+
1,
|
255 |
+
dilation=dilation[2],
|
256 |
+
padding=get_padding(kernel_size, dilation[2]),
|
257 |
+
)
|
258 |
+
),
|
259 |
+
]
|
260 |
+
)
|
261 |
+
self.convs1.apply(init_weights)
|
262 |
+
|
263 |
+
self.convs2 = nn.ModuleList(
|
264 |
+
[
|
265 |
+
weight_norm(
|
266 |
+
Conv1d(
|
267 |
+
channels,
|
268 |
+
channels,
|
269 |
+
kernel_size,
|
270 |
+
1,
|
271 |
+
dilation=1,
|
272 |
+
padding=get_padding(kernel_size, 1),
|
273 |
+
)
|
274 |
+
),
|
275 |
+
weight_norm(
|
276 |
+
Conv1d(
|
277 |
+
channels,
|
278 |
+
channels,
|
279 |
+
kernel_size,
|
280 |
+
1,
|
281 |
+
dilation=1,
|
282 |
+
padding=get_padding(kernel_size, 1),
|
283 |
+
)
|
284 |
+
),
|
285 |
+
weight_norm(
|
286 |
+
Conv1d(
|
287 |
+
channels,
|
288 |
+
channels,
|
289 |
+
kernel_size,
|
290 |
+
1,
|
291 |
+
dilation=1,
|
292 |
+
padding=get_padding(kernel_size, 1),
|
293 |
+
)
|
294 |
+
),
|
295 |
+
]
|
296 |
+
)
|
297 |
+
self.convs2.apply(init_weights)
|
298 |
+
|
299 |
+
def forward(self, x, x_mask=None):
|
300 |
+
for c1, c2 in zip(self.convs1, self.convs2):
|
301 |
+
xt = F.leaky_relu(x, LRELU_SLOPE)
|
302 |
+
if x_mask is not None:
|
303 |
+
xt = xt * x_mask
|
304 |
+
xt = c1(xt)
|
305 |
+
xt = F.leaky_relu(xt, LRELU_SLOPE)
|
306 |
+
if x_mask is not None:
|
307 |
+
xt = xt * x_mask
|
308 |
+
xt = c2(xt)
|
309 |
+
x = xt + x
|
310 |
+
if x_mask is not None:
|
311 |
+
x = x * x_mask
|
312 |
+
return x
|
313 |
+
|
314 |
+
def remove_weight_norm(self):
|
315 |
+
for l in self.convs1:
|
316 |
+
remove_weight_norm(l)
|
317 |
+
for l in self.convs2:
|
318 |
+
remove_weight_norm(l)
|
319 |
+
|
320 |
+
|
321 |
+
class ResBlock2(torch.nn.Module):
|
322 |
+
def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
|
323 |
+
super(ResBlock2, self).__init__()
|
324 |
+
self.convs = nn.ModuleList(
|
325 |
+
[
|
326 |
+
weight_norm(
|
327 |
+
Conv1d(
|
328 |
+
channels,
|
329 |
+
channels,
|
330 |
+
kernel_size,
|
331 |
+
1,
|
332 |
+
dilation=dilation[0],
|
333 |
+
padding=get_padding(kernel_size, dilation[0]),
|
334 |
+
)
|
335 |
+
),
|
336 |
+
weight_norm(
|
337 |
+
Conv1d(
|
338 |
+
channels,
|
339 |
+
channels,
|
340 |
+
kernel_size,
|
341 |
+
1,
|
342 |
+
dilation=dilation[1],
|
343 |
+
padding=get_padding(kernel_size, dilation[1]),
|
344 |
+
)
|
345 |
+
),
|
346 |
+
]
|
347 |
+
)
|
348 |
+
self.convs.apply(init_weights)
|
349 |
+
|
350 |
+
def forward(self, x, x_mask=None):
|
351 |
+
for c in self.convs:
|
352 |
+
xt = F.leaky_relu(x, LRELU_SLOPE)
|
353 |
+
if x_mask is not None:
|
354 |
+
xt = xt * x_mask
|
355 |
+
xt = c(xt)
|
356 |
+
x = xt + x
|
357 |
+
if x_mask is not None:
|
358 |
+
x = x * x_mask
|
359 |
+
return x
|
360 |
+
|
361 |
+
def remove_weight_norm(self):
|
362 |
+
for l in self.convs:
|
363 |
+
remove_weight_norm(l)
|
364 |
+
|
365 |
+
|
366 |
+
class Log(nn.Module):
|
367 |
+
def forward(self, x, x_mask, reverse=False, **kwargs):
|
368 |
+
if not reverse:
|
369 |
+
y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
|
370 |
+
logdet = torch.sum(-y, [1, 2])
|
371 |
+
return y, logdet
|
372 |
+
else:
|
373 |
+
x = torch.exp(x) * x_mask
|
374 |
+
return x
|
375 |
+
|
376 |
+
|
377 |
+
class Flip(nn.Module):
|
378 |
+
def forward(self, x, *args, reverse=False, **kwargs):
|
379 |
+
x = torch.flip(x, [1])
|
380 |
+
if not reverse:
|
381 |
+
logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
|
382 |
+
return x, logdet
|
383 |
+
else:
|
384 |
+
return x
|
385 |
+
|
386 |
+
|
387 |
+
class ElementwiseAffine(nn.Module):
|
388 |
+
def __init__(self, channels):
|
389 |
+
super().__init__()
|
390 |
+
self.channels = channels
|
391 |
+
self.m = nn.Parameter(torch.zeros(channels, 1))
|
392 |
+
self.logs = nn.Parameter(torch.zeros(channels, 1))
|
393 |
+
|
394 |
+
def forward(self, x, x_mask, reverse=False, **kwargs):
|
395 |
+
if not reverse:
|
396 |
+
y = self.m + torch.exp(self.logs) * x
|
397 |
+
y = y * x_mask
|
398 |
+
logdet = torch.sum(self.logs * x_mask, [1, 2])
|
399 |
+
return y, logdet
|
400 |
+
else:
|
401 |
+
x = (x - self.m) * torch.exp(-self.logs) * x_mask
|
402 |
+
return x
|
403 |
+
|
404 |
+
|
405 |
+
class ResidualCouplingLayer(nn.Module):
|
406 |
+
def __init__(
|
407 |
+
self,
|
408 |
+
channels,
|
409 |
+
hidden_channels,
|
410 |
+
kernel_size,
|
411 |
+
dilation_rate,
|
412 |
+
n_layers,
|
413 |
+
p_dropout=0,
|
414 |
+
gin_channels=0,
|
415 |
+
mean_only=False,
|
416 |
+
):
|
417 |
+
assert channels % 2 == 0, "channels should be divisible by 2"
|
418 |
+
super().__init__()
|
419 |
+
self.channels = channels
|
420 |
+
self.hidden_channels = hidden_channels
|
421 |
+
self.kernel_size = kernel_size
|
422 |
+
self.dilation_rate = dilation_rate
|
423 |
+
self.n_layers = n_layers
|
424 |
+
self.half_channels = channels // 2
|
425 |
+
self.mean_only = mean_only
|
426 |
+
|
427 |
+
self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
|
428 |
+
self.enc = WN(
|
429 |
+
hidden_channels,
|
430 |
+
kernel_size,
|
431 |
+
dilation_rate,
|
432 |
+
n_layers,
|
433 |
+
p_dropout=p_dropout,
|
434 |
+
gin_channels=gin_channels,
|
435 |
+
)
|
436 |
+
self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
|
437 |
+
self.post.weight.data.zero_()
|
438 |
+
self.post.bias.data.zero_()
|
439 |
+
|
440 |
+
def forward(self, x, x_mask, g=None, reverse=False):
|
441 |
+
x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
|
442 |
+
h = self.pre(x0) * x_mask
|
443 |
+
h = self.enc(h, x_mask, g=g)
|
444 |
+
stats = self.post(h) * x_mask
|
445 |
+
if not self.mean_only:
|
446 |
+
m, logs = torch.split(stats, [self.half_channels] * 2, 1)
|
447 |
+
else:
|
448 |
+
m = stats
|
449 |
+
logs = torch.zeros_like(m)
|
450 |
+
|
451 |
+
if not reverse:
|
452 |
+
x1 = m + x1 * torch.exp(logs) * x_mask
|
453 |
+
x = torch.cat([x0, x1], 1)
|
454 |
+
logdet = torch.sum(logs, [1, 2])
|
455 |
+
return x, logdet
|
456 |
+
else:
|
457 |
+
x1 = (x1 - m) * torch.exp(-logs) * x_mask
|
458 |
+
x = torch.cat([x0, x1], 1)
|
459 |
+
return x
|
460 |
+
|
461 |
+
def remove_weight_norm(self):
|
462 |
+
self.enc.remove_weight_norm()
|
463 |
+
|
464 |
+
|
465 |
+
class ConvFlow(nn.Module):
|
466 |
+
def __init__(
|
467 |
+
self,
|
468 |
+
in_channels,
|
469 |
+
filter_channels,
|
470 |
+
kernel_size,
|
471 |
+
n_layers,
|
472 |
+
num_bins=10,
|
473 |
+
tail_bound=5.0,
|
474 |
+
):
|
475 |
+
super().__init__()
|
476 |
+
self.in_channels = in_channels
|
477 |
+
self.filter_channels = filter_channels
|
478 |
+
self.kernel_size = kernel_size
|
479 |
+
self.n_layers = n_layers
|
480 |
+
self.num_bins = num_bins
|
481 |
+
self.tail_bound = tail_bound
|
482 |
+
self.half_channels = in_channels // 2
|
483 |
+
|
484 |
+
self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
|
485 |
+
self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.0)
|
486 |
+
self.proj = nn.Conv1d(
|
487 |
+
filter_channels, self.half_channels * (num_bins * 3 - 1), 1
|
488 |
+
)
|
489 |
+
self.proj.weight.data.zero_()
|
490 |
+
self.proj.bias.data.zero_()
|
491 |
+
|
492 |
+
def forward(self, x, x_mask, g=None, reverse=False):
|
493 |
+
x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
|
494 |
+
h = self.pre(x0)
|
495 |
+
h = self.convs(h, x_mask, g=g)
|
496 |
+
h = self.proj(h) * x_mask
|
497 |
+
|
498 |
+
b, c, t = x0.shape
|
499 |
+
h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2) # [b, cx?, t] -> [b, c, t, ?]
|
500 |
+
|
501 |
+
unnormalized_widths = h[..., : self.num_bins] / math.sqrt(self.filter_channels)
|
502 |
+
unnormalized_heights = h[..., self.num_bins : 2 * self.num_bins] / math.sqrt(
|
503 |
+
self.filter_channels
|
504 |
+
)
|
505 |
+
unnormalized_derivatives = h[..., 2 * self.num_bins :]
|
506 |
+
|
507 |
+
x1, logabsdet = piecewise_rational_quadratic_transform(
|
508 |
+
x1,
|
509 |
+
unnormalized_widths,
|
510 |
+
unnormalized_heights,
|
511 |
+
unnormalized_derivatives,
|
512 |
+
inverse=reverse,
|
513 |
+
tails="linear",
|
514 |
+
tail_bound=self.tail_bound,
|
515 |
+
)
|
516 |
+
|
517 |
+
x = torch.cat([x0, x1], 1) * x_mask
|
518 |
+
logdet = torch.sum(logabsdet * x_mask, [1, 2])
|
519 |
+
if not reverse:
|
520 |
+
return x, logdet
|
521 |
+
else:
|
522 |
+
return x
|