Kit-Lemonfoot/vtuber_rvc_models
53
1import copy
2import math
3import numpy as np
4import scipy
5import torch
6from torch import nn
7from torch.nn import functional as F
8
9from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
10from torch.nn.utils import remove_weight_norm
11from torch.nn.utils.parametrizations import weight_norm
12
13
14from lib.infer_pack import commons
15from lib.infer_pack.commons import init_weights, get_padding
16from lib.infer_pack.transforms import piecewise_rational_quadratic_transform
17
18
19LRELU_SLOPE = 0.1
20
21
22class LayerNorm(nn.Module):
23 def __init__(self, channels, eps=1e-5):
24 super().__init__()
25 self.channels = channels
26 self.eps = eps
27
28 self.gamma = nn.Parameter(torch.ones(channels))
29 self.beta = nn.Parameter(torch.zeros(channels))
30
31 def forward(self, x):
32 x = x.transpose(1, -1)
33 x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
34 return x.transpose(1, -1)
35
36
37class ConvReluNorm(nn.Module):
38 def __init__(
39 self,
40 in_channels,
41 hidden_channels,
42 out_channels,
43 kernel_size,
44 n_layers,
45 p_dropout,
46 ):
47 super().__init__()
48 self.in_channels = in_channels
49 self.hidden_channels = hidden_channels
50 self.out_channels = out_channels
51 self.kernel_size = kernel_size
52 self.n_layers = n_layers
53 self.p_dropout = p_dropout
54 assert n_layers > 1, "Number of layers should be larger than 0."
55
56 self.conv_layers = nn.ModuleList()
57 self.norm_layers = nn.ModuleList()
58 self.conv_layers.append(
59 nn.Conv1d(
60 in_channels, hidden_channels, kernel_size, padding=kernel_size // 2
61 )
62 )
63 self.norm_layers.append(LayerNorm(hidden_channels))
64 self.relu_drop = nn.Sequential(nn.ReLU(), nn.Dropout(p_dropout))
65 for _ in range(n_layers - 1):
66 self.conv_layers.append(
67 nn.Conv1d(
68 hidden_channels,
69 hidden_channels,
70 kernel_size,
71 padding=kernel_size // 2,
72 )
73 )
74 self.norm_layers.append(LayerNorm(hidden_channels))
75 self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
76 self.proj.weight.data.zero_()
77 self.proj.bias.data.zero_()
78
79 def forward(self, x, x_mask):
80 x_org = x
81 for i in range(self.n_layers):
82 x = self.conv_layers[i](x * x_mask)
83 x = self.norm_layers[i](x)
84 x = self.relu_drop(x)
85 x = x_org + self.proj(x)
86 return x * x_mask
87
88
89class DDSConv(nn.Module):
90 """
91 Dialted and Depth-Separable Convolution
92 """
93
94 def __init__(self, channels, kernel_size, n_layers, p_dropout=0.0):
95 super().__init__()
96 self.channels = channels
97 self.kernel_size = kernel_size
98 self.n_layers = n_layers
99 self.p_dropout = p_dropout
100
101 self.drop = nn.Dropout(p_dropout)
102 self.convs_sep = nn.ModuleList()
103 self.convs_1x1 = nn.ModuleList()
104 self.norms_1 = nn.ModuleList()
105 self.norms_2 = nn.ModuleList()
106 for i in range(n_layers):
107 dilation = kernel_size**i
108 padding = (kernel_size * dilation - dilation) // 2
109 self.convs_sep.append(
110 nn.Conv1d(
111 channels,
112 channels,
113 kernel_size,
114 groups=channels,
115 dilation=dilation,
116 padding=padding,
117 )
118 )
119 self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
120 self.norms_1.append(LayerNorm(channels))
121 self.norms_2.append(LayerNorm(channels))
122
123 def forward(self, x, x_mask, g=None):
124 if g is not None:
125 x = x + g
126 for i in range(self.n_layers):
127 y = self.convs_sep[i](x * x_mask)
128 y = self.norms_1[i](y)
129 y = F.gelu(y)
130 y = self.convs_1x1[i](y)
131 y = self.norms_2[i](y)
132 y = F.gelu(y)
133 y = self.drop(y)
134 x = x + y
135 return x * x_mask
136
137
138class WN(torch.nn.Module):
139 def __init__(
140 self,
141 hidden_channels,
142 kernel_size,
143 dilation_rate,
144 n_layers,
145 gin_channels=0,
146 p_dropout=0,
147 ):
148 super(WN, self).__init__()
149 assert kernel_size % 2 == 1
150 self.hidden_channels = hidden_channels
151 self.kernel_size = (kernel_size,)
152 self.dilation_rate = dilation_rate
153 self.n_layers = n_layers
154 self.gin_channels = gin_channels
155 self.p_dropout = p_dropout
156
157 self.in_layers = torch.nn.ModuleList()
158 self.res_skip_layers = torch.nn.ModuleList()
159 self.drop = nn.Dropout(p_dropout)
160
161 if gin_channels != 0:
162 cond_layer = torch.nn.Conv1d(
163 gin_channels, 2 * hidden_channels * n_layers, 1
164 )
165 self.cond_layer = torch.nn.utils.parametrizations.weight_norm(cond_layer, name="weight")
166
167 for i in range(n_layers):
168 dilation = dilation_rate**i
169 padding = int((kernel_size * dilation - dilation) / 2)
170 in_layer = torch.nn.Conv1d(
171 hidden_channels,
172 2 * hidden_channels,
173 kernel_size,
174 dilation=dilation,
175 padding=padding,
176 )
177 in_layer = torch.nn.utils.parametrizations.weight_norm(in_layer, name="weight")
178 self.in_layers.append(in_layer)
179
180 # last one is not necessary
181 if i < n_layers - 1:
182 res_skip_channels = 2 * hidden_channels
183 else:
184 res_skip_channels = hidden_channels
185
186 res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
187 res_skip_layer = torch.nn.utils.parametrizations.weight_norm(res_skip_layer, name="weight")
188 self.res_skip_layers.append(res_skip_layer)
189
190 def forward(self, x, x_mask, g=None, **kwargs):
191 output = torch.zeros_like(x)
192 n_channels_tensor = torch.IntTensor([self.hidden_channels])
193
194 if g is not None:
195 g = self.cond_layer(g)
196
197 for i in range(self.n_layers):
198 x_in = self.in_layers[i](x)
199 if g is not None:
200 cond_offset = i * 2 * self.hidden_channels
201 g_l = g[:, cond_offset : cond_offset + 2 * self.hidden_channels, :]
202 else:
203 g_l = torch.zeros_like(x_in)
204
205 acts = commons.fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)
206 acts = self.drop(acts)
207
208 res_skip_acts = self.res_skip_layers[i](acts)
209 if i < self.n_layers - 1:
210 res_acts = res_skip_acts[:, : self.hidden_channels, :]
211 x = (x + res_acts) * x_mask
212 output = output + res_skip_acts[:, self.hidden_channels :, :]
213 else:
214 output = output + res_skip_acts
215 return output * x_mask
216
217 def remove_weight_norm(self):
218 if self.gin_channels != 0:
219 torch.nn.utils.remove_weight_norm(self.cond_layer)
220 for l in self.in_layers:
221 torch.nn.utils.remove_weight_norm(l)
222 for l in self.res_skip_layers:
223 torch.nn.utils.remove_weight_norm(l)
224
225
226class ResBlock1(torch.nn.Module):
227 def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
228 super(ResBlock1, self).__init__()
229 self.convs1 = nn.ModuleList(
230 [
231 weight_norm(
232 Conv1d(
233 channels,
234 channels,
235 kernel_size,
236 1,
237 dilation=dilation[0],
238 padding=get_padding(kernel_size, dilation[0]),
239 )
240 ),
241 weight_norm(
242 Conv1d(
243 channels,
244 channels,
245 kernel_size,
246 1,
247 dilation=dilation[1],
248 padding=get_padding(kernel_size, dilation[1]),
249 )
250 ),
251 weight_norm(
252 Conv1d(
253 channels,
254 channels,
255 kernel_size,
256 1,
257 dilation=dilation[2],
258 padding=get_padding(kernel_size, dilation[2]),
259 )
260 ),
261 ]
262 )
263 self.convs1.apply(init_weights)
264
265 self.convs2 = nn.ModuleList(
266 [
267 weight_norm(
268 Conv1d(
269 channels,
270 channels,
271 kernel_size,
272 1,
273 dilation=1,
274 padding=get_padding(kernel_size, 1),
275 )
276 ),
277 weight_norm(
278 Conv1d(
279 channels,
280 channels,
281 kernel_size,
282 1,
283 dilation=1,
284 padding=get_padding(kernel_size, 1),
285 )
286 ),
287 weight_norm(
288 Conv1d(
289 channels,
290 channels,
291 kernel_size,
292 1,
293 dilation=1,
294 padding=get_padding(kernel_size, 1),
295 )
296 ),
297 ]
298 )
299 self.convs2.apply(init_weights)
300
301 def forward(self, x, x_mask=None):
302 for c1, c2 in zip(self.convs1, self.convs2):
303 xt = F.leaky_relu(x, LRELU_SLOPE)
304 if x_mask is not None:
305 xt = xt * x_mask
306 xt = c1(xt)
307 xt = F.leaky_relu(xt, LRELU_SLOPE)
308 if x_mask is not None:
309 xt = xt * x_mask
310 xt = c2(xt)
311 x = xt + x
312 if x_mask is not None:
313 x = x * x_mask
314 return x
315
316 def remove_weight_norm(self):
317 for l in self.convs1:
318 remove_weight_norm(l)
319 for l in self.convs2:
320 remove_weight_norm(l)
321
322
323class ResBlock2(torch.nn.Module):
324 def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
325 super(ResBlock2, self).__init__()
326 self.convs = nn.ModuleList(
327 [
328 weight_norm(
329 Conv1d(
330 channels,
331 channels,
332 kernel_size,
333 1,
334 dilation=dilation[0],
335 padding=get_padding(kernel_size, dilation[0]),
336 )
337 ),
338 weight_norm(
339 Conv1d(
340 channels,
341 channels,
342 kernel_size,
343 1,
344 dilation=dilation[1],
345 padding=get_padding(kernel_size, dilation[1]),
346 )
347 ),
348 ]
349 )
350 self.convs.apply(init_weights)
351
352 def forward(self, x, x_mask=None):
353 for c in self.convs:
354 xt = F.leaky_relu(x, LRELU_SLOPE)
355 if x_mask is not None:
356 xt = xt * x_mask
357 xt = c(xt)
358 x = xt + x
359 if x_mask is not None:
360 x = x * x_mask
361 return x
362
363 def remove_weight_norm(self):
364 for l in self.convs:
365 remove_weight_norm(l)
366
367
368class Log(nn.Module):
369 def forward(self, x, x_mask, reverse=False, **kwargs):
370 if not reverse:
371 y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
372 logdet = torch.sum(-y, [1, 2])
373 return y, logdet
374 else:
375 x = torch.exp(x) * x_mask
376 return x
377
378
379class Flip(nn.Module):
380 def forward(self, x, *args, reverse=False, **kwargs):
381 x = torch.flip(x, [1])
382 if not reverse:
383 logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
384 return x, logdet
385 else:
386 return x
387
388
389class ElementwiseAffine(nn.Module):
390 def __init__(self, channels):
391 super().__init__()
392 self.channels = channels
393 self.m = nn.Parameter(torch.zeros(channels, 1))
394 self.logs = nn.Parameter(torch.zeros(channels, 1))
395
396 def forward(self, x, x_mask, reverse=False, **kwargs):
397 if not reverse:
398 y = self.m + torch.exp(self.logs) * x
399 y = y * x_mask
400 logdet = torch.sum(self.logs * x_mask, [1, 2])
401 return y, logdet
402 else:
403 x = (x - self.m) * torch.exp(-self.logs) * x_mask
404 return x
405
406
407class ResidualCouplingLayer(nn.Module):
408 def __init__(
409 self,
410 channels,
411 hidden_channels,
412 kernel_size,
413 dilation_rate,
414 n_layers,
415 p_dropout=0,
416 gin_channels=0,
417 mean_only=False,
418 ):
419 assert channels % 2 == 0, "channels should be divisible by 2"
420 super().__init__()
421 self.channels = channels
422 self.hidden_channels = hidden_channels
423 self.kernel_size = kernel_size
424 self.dilation_rate = dilation_rate
425 self.n_layers = n_layers
426 self.half_channels = channels // 2
427 self.mean_only = mean_only
428
429 self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
430 self.enc = WN(
431 hidden_channels,
432 kernel_size,
433 dilation_rate,
434 n_layers,
435 p_dropout=p_dropout,
436 gin_channels=gin_channels,
437 )
438 self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
439 self.post.weight.data.zero_()
440 self.post.bias.data.zero_()
441
442 def forward(self, x, x_mask, g=None, reverse=False):
443 x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
444 h = self.pre(x0) * x_mask
445 h = self.enc(h, x_mask, g=g)
446 stats = self.post(h) * x_mask
447 if not self.mean_only:
448 m, logs = torch.split(stats, [self.half_channels] * 2, 1)
449 else:
450 m = stats
451 logs = torch.zeros_like(m)
452
453 if not reverse:
454 x1 = m + x1 * torch.exp(logs) * x_mask
455 x = torch.cat([x0, x1], 1)
456 logdet = torch.sum(logs, [1, 2])
457 return x, logdet
458 else:
459 x1 = (x1 - m) * torch.exp(-logs) * x_mask
460 x = torch.cat([x0, x1], 1)
461 return x
462
463 def remove_weight_norm(self):
464 self.enc.remove_weight_norm()
465
466
467class ConvFlow(nn.Module):
468 def __init__(
469 self,
470 in_channels,
471 filter_channels,
472 kernel_size,
473 n_layers,
474 num_bins=10,
475 tail_bound=5.0,
476 ):
477 super().__init__()
478 self.in_channels = in_channels
479 self.filter_channels = filter_channels
480 self.kernel_size = kernel_size
481 self.n_layers = n_layers
482 self.num_bins = num_bins
483 self.tail_bound = tail_bound
484 self.half_channels = in_channels // 2
485
486 self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
487 self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.0)
488 self.proj = nn.Conv1d(
489 filter_channels, self.half_channels * (num_bins * 3 - 1), 1
490 )
491 self.proj.weight.data.zero_()
492 self.proj.bias.data.zero_()
493
494 def forward(self, x, x_mask, g=None, reverse=False):
495 x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
496 h = self.pre(x0)
497 h = self.convs(h, x_mask, g=g)
498 h = self.proj(h) * x_mask
499
500 b, c, t = x0.shape
501 h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2) # [b, cx?, t] -> [b, c, t, ?]
502
503 unnormalized_widths = h[..., : self.num_bins] / math.sqrt(self.filter_channels)
504 unnormalized_heights = h[..., self.num_bins : 2 * self.num_bins] / math.sqrt(
505 self.filter_channels
506 )
507 unnormalized_derivatives = h[..., 2 * self.num_bins :]
508
509 x1, logabsdet = piecewise_rational_quadratic_transform(
510 x1,
511 unnormalized_widths,
512 unnormalized_heights,
513 unnormalized_derivatives,
514 inverse=reverse,
515 tails="linear",
516 tail_bound=self.tail_bound,
517 )
518
519 x = torch.cat([x0, x1], 1) * x_mask
520 logdet = torch.sum(logabsdet * x_mask, [1, 2])
521 if not reverse:
522 return x, logdet
523 else:
524 return x
525