Spaces:
Runtime error
Runtime error
File size: 19,796 Bytes
0102e16 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 |
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
class _BatchNorm1d(nn.Module):
def __init__(
self,
input_shape=None,
input_size=None,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True,
combine_batch_time=False,
skip_transpose=False,
):
super().__init__()
self.combine_batch_time = combine_batch_time
self.skip_transpose = skip_transpose
if input_size is None and skip_transpose:
input_size = input_shape[1]
elif input_size is None:
input_size = input_shape[-1]
self.norm = nn.BatchNorm1d(
input_size,
eps=eps,
momentum=momentum,
affine=affine,
track_running_stats=track_running_stats,
)
def forward(self, x):
shape_or = x.shape
if self.combine_batch_time:
if x.ndim == 3:
x = x.reshape(shape_or[0] * shape_or[1], shape_or[2])
else:
x = x.reshape(shape_or[0] * shape_or[1], shape_or[3], shape_or[2])
elif not self.skip_transpose:
x = x.transpose(-1, 1)
x_n = self.norm(x)
if self.combine_batch_time:
x_n = x_n.reshape(shape_or)
elif not self.skip_transpose:
x_n = x_n.transpose(1, -1)
return x_n
class _Conv1d(nn.Module):
def __init__(
self,
out_channels,
kernel_size,
input_shape=None,
in_channels=None,
stride=1,
dilation=1,
padding="same",
groups=1,
bias=True,
padding_mode="reflect",
skip_transpose=False,
):
super().__init__()
self.kernel_size = kernel_size
self.stride = stride
self.dilation = dilation
self.padding = padding
self.padding_mode = padding_mode
self.unsqueeze = False
self.skip_transpose = skip_transpose
if input_shape is None and in_channels is None:
raise ValueError("Must provide one of input_shape or in_channels")
if in_channels is None:
in_channels = self._check_input_shape(input_shape)
self.conv = nn.Conv1d(
in_channels,
out_channels,
self.kernel_size,
stride=self.stride,
dilation=self.dilation,
padding=0,
groups=groups,
bias=bias,
)
def forward(self, x):
if not self.skip_transpose:
x = x.transpose(1, -1)
if self.unsqueeze:
x = x.unsqueeze(1)
if self.padding == "same":
x = self._manage_padding(x, self.kernel_size, self.dilation, self.stride)
elif self.padding == "causal":
num_pad = (self.kernel_size - 1) * self.dilation
x = F.pad(x, (num_pad, 0))
elif self.padding == "valid":
pass
else:
raise ValueError(
"Padding must be 'same', 'valid' or 'causal'. Got " + self.padding
)
wx = self.conv(x)
if self.unsqueeze:
wx = wx.squeeze(1)
if not self.skip_transpose:
wx = wx.transpose(1, -1)
return wx
def _manage_padding(
self,
x,
kernel_size: int,
dilation: int,
stride: int,
):
# Detecting input shape
L_in = x.shape[-1]
# Time padding
padding = get_padding_elem(L_in, stride, kernel_size, dilation)
# Applying padding
x = F.pad(x, padding, mode=self.padding_mode)
return x
def _check_input_shape(self, shape):
"""Checks the input shape and returns the number of input channels."""
if len(shape) == 2:
self.unsqueeze = True
in_channels = 1
elif self.skip_transpose:
in_channels = shape[1]
elif len(shape) == 3:
in_channels = shape[2]
else:
raise ValueError("conv1d expects 2d, 3d inputs. Got " + str(len(shape)))
# Kernel size must be odd
if self.kernel_size % 2 == 0:
raise ValueError(
"The field kernel size must be an odd number. Got %s."
% (self.kernel_size)
)
return in_channels
def get_padding_elem(L_in: int, stride: int, kernel_size: int, dilation: int):
if stride > 1:
n_steps = math.ceil(((L_in - kernel_size * dilation) / stride) + 1)
L_out = stride * (n_steps - 1) + kernel_size * dilation
padding = [kernel_size // 2, kernel_size // 2]
else:
L_out = (L_in - dilation * (kernel_size - 1) - 1) // stride + 1
padding = [(L_in - L_out) // 2, (L_in - L_out) // 2]
return padding
# Skip transpose as much as possible for efficiency
class Conv1d(_Conv1d):
def __init__(self, *args, **kwargs):
super().__init__(skip_transpose=True, *args, **kwargs)
class BatchNorm1d(_BatchNorm1d):
def __init__(self, *args, **kwargs):
super().__init__(skip_transpose=True, *args, **kwargs)
def length_to_mask(length, max_len=None, dtype=None, device=None):
assert len(length.shape) == 1
if max_len is None:
max_len = length.max().long().item() # using arange to generate mask
mask = torch.arange(max_len, device=length.device, dtype=length.dtype).expand(
len(length), max_len
) < length.unsqueeze(1)
if dtype is None:
dtype = length.dtype
if device is None:
device = length.device
mask = torch.as_tensor(mask, dtype=dtype, device=device)
return mask
class TDNNBlock(nn.Module):
def __init__(
self,
in_channels,
out_channels,
kernel_size,
dilation,
activation=nn.ReLU,
groups=1,
):
super(TDNNBlock, self).__init__()
self.conv = Conv1d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
dilation=dilation,
groups=groups,
)
self.activation = activation()
self.norm = BatchNorm1d(input_size=out_channels)
def forward(self, x):
return self.norm(self.activation(self.conv(x)))
class Res2NetBlock(torch.nn.Module):
"""An implementation of Res2NetBlock w/ dilation.
Arguments
---------
in_channels : int
The number of channels expected in the input.
out_channels : int
The number of output channels.
scale : int
The scale of the Res2Net block.
kernel_size: int
The kernel size of the Res2Net block.
dilation : int
The dilation of the Res2Net block.
Example
-------
>>> inp_tensor = torch.rand([8, 120, 64]).transpose(1, 2)
>>> layer = Res2NetBlock(64, 64, scale=4, dilation=3)
>>> out_tensor = layer(inp_tensor).transpose(1, 2)
>>> out_tensor.shape
torch.Size([8, 120, 64])
"""
def __init__(self, in_channels, out_channels, scale=8, kernel_size=3, dilation=1):
super(Res2NetBlock, self).__init__()
assert in_channels % scale == 0
assert out_channels % scale == 0
in_channel = in_channels // scale
hidden_channel = out_channels // scale
self.blocks = nn.ModuleList(
[
TDNNBlock(
in_channel,
hidden_channel,
kernel_size=kernel_size,
dilation=dilation,
)
for i in range(scale - 1)
]
)
self.scale = scale
def forward(self, x):
y = []
for i, x_i in enumerate(torch.chunk(x, self.scale, dim=1)):
if i == 0:
y_i = x_i
elif i == 1:
y_i = self.blocks[i - 1](x_i)
else:
y_i = self.blocks[i - 1](x_i + y_i)
y.append(y_i)
y = torch.cat(y, dim=1)
return y
class SEBlock(nn.Module):
"""An implementation of squeeze-and-excitation block.
Arguments
---------
in_channels : int
The number of input channels.
se_channels : int
The number of output channels after squeeze.
out_channels : int
The number of output channels.
Example
-------
>>> inp_tensor = torch.rand([8, 120, 64]).transpose(1, 2)
>>> se_layer = SEBlock(64, 16, 64)
>>> lengths = torch.rand((8,))
>>> out_tensor = se_layer(inp_tensor, lengths).transpose(1, 2)
>>> out_tensor.shape
torch.Size([8, 120, 64])
"""
def __init__(self, in_channels, se_channels, out_channels):
super(SEBlock, self).__init__()
self.conv1 = Conv1d(
in_channels=in_channels, out_channels=se_channels, kernel_size=1
)
self.relu = torch.nn.ReLU(inplace=True)
self.conv2 = Conv1d(
in_channels=se_channels, out_channels=out_channels, kernel_size=1
)
self.sigmoid = torch.nn.Sigmoid()
def forward(self, x, lengths=None):
L = x.shape[-1]
if lengths is not None:
mask = length_to_mask(lengths * L, max_len=L, device=x.device)
mask = mask.unsqueeze(1)
total = mask.sum(dim=2, keepdim=True)
s = (x * mask).sum(dim=2, keepdim=True) / total
else:
s = x.mean(dim=2, keepdim=True)
s = self.relu(self.conv1(s))
s = self.sigmoid(self.conv2(s))
return s * x
class AttentiveStatisticsPooling(nn.Module):
"""This class implements an attentive statistic pooling layer for each channel.
It returns the concatenated mean and std of the input tensor.
Arguments
---------
channels: int
The number of input channels.
attention_channels: int
The number of attention channels.
Example
-------
>>> inp_tensor = torch.rand([8, 120, 64]).transpose(1, 2)
>>> asp_layer = AttentiveStatisticsPooling(64)
>>> lengths = torch.rand((8,))
>>> out_tensor = asp_layer(inp_tensor, lengths).transpose(1, 2)
>>> out_tensor.shape
torch.Size([8, 1, 128])
"""
def __init__(self, channels, attention_channels=128, global_context=True):
super().__init__()
self.eps = 1e-12
self.global_context = global_context
if global_context:
self.tdnn = TDNNBlock(channels * 3, attention_channels, 1, 1)
else:
self.tdnn = TDNNBlock(channels, attention_channels, 1, 1)
self.tanh = nn.Tanh()
self.conv = Conv1d(
in_channels=attention_channels, out_channels=channels, kernel_size=1
)
def forward(self, x, lengths=None):
"""Calculates mean and std for a batch (input tensor).
Arguments
---------
x : torch.Tensor
Tensor of shape [N, C, L].
"""
L = x.shape[-1]
def _compute_statistics(x, m, dim=2, eps=self.eps):
mean = (m * x).sum(dim)
std = torch.sqrt((m * (x - mean.unsqueeze(dim)).pow(2)).sum(dim).clamp(eps))
return mean, std
if lengths is None:
lengths = torch.ones(x.shape[0], device=x.device)
# Make binary mask of shape [N, 1, L]
mask = length_to_mask(lengths * L, max_len=L, device=x.device)
mask = mask.unsqueeze(1)
# Expand the temporal context of the pooling layer by allowing the
# self-attention to look at global properties of the utterance.
if self.global_context:
# torch.std is unstable for backward computation
# https://github.com/pytorch/pytorch/issues/4320
total = mask.sum(dim=2, keepdim=True).float()
mean, std = _compute_statistics(x, mask / total)
mean = mean.unsqueeze(2).repeat(1, 1, L)
std = std.unsqueeze(2).repeat(1, 1, L)
attn = torch.cat([x, mean, std], dim=1)
else:
attn = x
# Apply layers
attn = self.conv(self.tanh(self.tdnn(attn)))
# Filter out zero-paddings
attn = attn.masked_fill(mask == 0, float("-inf"))
attn = F.softmax(attn, dim=2)
mean, std = _compute_statistics(x, attn)
# Append mean and std of the batch
pooled_stats = torch.cat((mean, std), dim=1)
pooled_stats = pooled_stats.unsqueeze(2)
return pooled_stats
class SERes2NetBlock(nn.Module):
"""An implementation of building block in ECAPA-TDNN, i.e.,
TDNN-Res2Net-TDNN-SEBlock.
Arguments
----------
out_channels: int
The number of output channels.
res2net_scale: int
The scale of the Res2Net block.
kernel_size: int
The kernel size of the TDNN blocks.
dilation: int
The dilation of the Res2Net block.
activation : torch class
A class for constructing the activation layers.
groups: int
Number of blocked connections from input channels to output channels.
Example
-------
>>> x = torch.rand(8, 120, 64).transpose(1, 2)
>>> conv = SERes2NetBlock(64, 64, res2net_scale=4)
>>> out = conv(x).transpose(1, 2)
>>> out.shape
torch.Size([8, 120, 64])
"""
def __init__(
self,
in_channels,
out_channels,
res2net_scale=8,
se_channels=128,
kernel_size=1,
dilation=1,
activation=torch.nn.ReLU,
groups=1,
):
super().__init__()
self.out_channels = out_channels
self.tdnn1 = TDNNBlock(
in_channels,
out_channels,
kernel_size=1,
dilation=1,
activation=activation,
groups=groups,
)
self.res2net_block = Res2NetBlock(
out_channels, out_channels, res2net_scale, kernel_size, dilation
)
self.tdnn2 = TDNNBlock(
out_channels,
out_channels,
kernel_size=1,
dilation=1,
activation=activation,
groups=groups,
)
self.se_block = SEBlock(out_channels, se_channels, out_channels)
self.shortcut = None
if in_channels != out_channels:
self.shortcut = Conv1d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
)
def forward(self, x, lengths=None):
residual = x
if self.shortcut:
residual = self.shortcut(x)
x = self.tdnn1(x)
x = self.res2net_block(x)
x = self.tdnn2(x)
x = self.se_block(x, lengths)
return x + residual
class ECAPA_TDNN(torch.nn.Module):
"""An implementation of the speaker embedding model in a paper.
"ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in
TDNN Based Speaker Verification" (https://arxiv.org/abs/2005.07143).
Arguments
---------
activation : torch class
A class for constructing the activation layers.
channels : list of ints
Output channels for TDNN/SERes2Net layer.
kernel_sizes : list of ints
List of kernel sizes for each layer.
dilations : list of ints
List of dilations for kernels in each layer.
lin_neurons : int
Number of neurons in linear layers.
groups : list of ints
List of groups for kernels in each layer.
Example
-------
>>> input_feats = torch.rand([5, 120, 80])
>>> compute_embedding = ECAPA_TDNN(80, lin_neurons=192)
>>> outputs = compute_embedding(input_feats)
>>> outputs.shape
torch.Size([5, 1, 192])
"""
def __init__(
self,
input_size,
lin_neurons=192,
activation=torch.nn.ReLU,
channels=[512, 512, 512, 512, 1536],
kernel_sizes=[5, 3, 3, 3, 1],
dilations=[1, 2, 3, 4, 1],
attention_channels=128,
res2net_scale=8,
se_channels=128,
global_context=True,
groups=[1, 1, 1, 1, 1],
window_size=20,
window_shift=1,
):
super().__init__()
assert len(channels) == len(kernel_sizes)
assert len(channels) == len(dilations)
self.channels = channels
self.blocks = nn.ModuleList()
self.window_size = window_size
self.window_shift = window_shift
# The initial TDNN layer
self.blocks.append(
TDNNBlock(
input_size,
channels[0],
kernel_sizes[0],
dilations[0],
activation,
groups[0],
)
)
# SE-Res2Net layers
for i in range(1, len(channels) - 1):
self.blocks.append(
SERes2NetBlock(
channels[i - 1],
channels[i],
res2net_scale=res2net_scale,
se_channels=se_channels,
kernel_size=kernel_sizes[i],
dilation=dilations[i],
activation=activation,
groups=groups[i],
)
)
# Multi-layer feature aggregation
self.mfa = TDNNBlock(
channels[-1],
channels[-1],
kernel_sizes[-1],
dilations[-1],
activation,
groups=groups[-1],
)
# Attentive Statistical Pooling
self.asp = AttentiveStatisticsPooling(
channels[-1],
attention_channels=attention_channels,
global_context=global_context,
)
self.asp_bn = BatchNorm1d(input_size=channels[-1] * 2)
# Final linear transformation
self.fc = Conv1d(
in_channels=channels[-1] * 2,
out_channels=lin_neurons,
kernel_size=1,
)
def windowed_pooling(self, x, lengths=None):
# x: Batch, Channel, Time
tt = x.shape[2]
num_chunk = int(math.ceil(tt / self.window_shift))
pad = self.window_size // 2
x = F.pad(x, (pad, pad, 0, 0), "reflect")
stat_list = []
for i in range(num_chunk):
# B x C
st, ed = i * self.window_shift, i * self.window_shift + self.window_size
x = self.asp(
x[:, :, st:ed],
lengths=(
torch.clamp(lengths - i, 0, self.window_size)
if lengths is not None
else None
),
)
x = self.asp_bn(x)
x = self.fc(x)
stat_list.append(x)
return torch.cat(stat_list, dim=2)
def forward(self, x, lengths=None):
"""Returns the embedding vector.
Arguments
---------
x : torch.Tensor
Tensor of shape (batch, time, channel).
lengths: torch.Tensor
Tensor of shape (batch, )
"""
# Minimize transpose for efficiency
x = x.transpose(1, 2)
xl = []
for layer in self.blocks:
try:
x = layer(x, lengths=lengths)
except TypeError:
x = layer(x)
xl.append(x)
# Multi-layer feature aggregation
x = torch.cat(xl[1:], dim=1)
x = self.mfa(x)
if self.window_size is None:
# Attentive Statistical Pooling
x = self.asp(x, lengths=lengths)
x = self.asp_bn(x)
# Final linear transformation
x = self.fc(x)
# x = x.transpose(1, 2)
x = x.squeeze(2) # -> B, C
else:
x = self.windowed_pooling(x, lengths)
x = x.transpose(1, 2) # -> B, T, C
return x
|