Upload seamless_communication/models/pretssel/ecapa_tdnn.py with huggingface_hub
Browse files
seamless_communication/models/pretssel/ecapa_tdnn.py
ADDED
|
@@ -0,0 +1,477 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates
|
| 2 |
+
# All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# This source code is licensed under the license found in the
|
| 5 |
+
# MIT_LICENSE file in the root directory of this source tree.
|
| 6 |
+
|
| 7 |
+
from typing import List, Optional, Tuple
|
| 8 |
+
|
| 9 |
+
import torch
|
| 10 |
+
import torch.nn.functional as F
|
| 11 |
+
from fairseq2.nn.padding import PaddingMask, to_padding_mask
|
| 12 |
+
from torch import Tensor
|
| 13 |
+
from torch.nn import Conv1d, LayerNorm, Module, ModuleList, ReLU, Sigmoid, Tanh, init
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
class ECAPA_TDNN(Module):
|
| 17 |
+
"""
|
| 18 |
+
Represents the ECAPA-TDNN model described in paper:
|
| 19 |
+
:cite:t`https://doi.org/10.48550/arxiv.2005.07143`.
|
| 20 |
+
|
| 21 |
+
Arguments
|
| 22 |
+
---------
|
| 23 |
+
:param channels:
|
| 24 |
+
Output channels for TDNN/SERes2Net layer.
|
| 25 |
+
:param kernel_sizes:
|
| 26 |
+
List of kernel sizes for each layer.
|
| 27 |
+
:param dilations:
|
| 28 |
+
List of dilations for kernels in each layer.
|
| 29 |
+
:param groups:
|
| 30 |
+
List of groups for kernels in each layer.
|
| 31 |
+
"""
|
| 32 |
+
|
| 33 |
+
def __init__(
|
| 34 |
+
self,
|
| 35 |
+
channels: List[int],
|
| 36 |
+
kernel_sizes: List[int],
|
| 37 |
+
dilations: List[int],
|
| 38 |
+
attention_channels: int,
|
| 39 |
+
res2net_scale: int,
|
| 40 |
+
se_channels: int,
|
| 41 |
+
global_context: bool,
|
| 42 |
+
groups: List[int],
|
| 43 |
+
embed_dim: int,
|
| 44 |
+
input_dim: int,
|
| 45 |
+
):
|
| 46 |
+
super().__init__()
|
| 47 |
+
assert len(channels) == len(kernel_sizes) == len(dilations)
|
| 48 |
+
self.channels = channels
|
| 49 |
+
self.embed_dim = embed_dim
|
| 50 |
+
self.blocks = ModuleList()
|
| 51 |
+
|
| 52 |
+
self.blocks.append(
|
| 53 |
+
TDNNBlock(
|
| 54 |
+
input_dim,
|
| 55 |
+
channels[0],
|
| 56 |
+
kernel_sizes[0],
|
| 57 |
+
dilations[0],
|
| 58 |
+
groups[0],
|
| 59 |
+
)
|
| 60 |
+
)
|
| 61 |
+
|
| 62 |
+
# SE-Res2Net layers
|
| 63 |
+
for i in range(1, len(channels) - 1):
|
| 64 |
+
self.blocks.append(
|
| 65 |
+
SERes2NetBlock(
|
| 66 |
+
channels[i - 1],
|
| 67 |
+
channels[i],
|
| 68 |
+
res2net_scale=res2net_scale,
|
| 69 |
+
se_channels=se_channels,
|
| 70 |
+
kernel_size=kernel_sizes[i],
|
| 71 |
+
dilation=dilations[i],
|
| 72 |
+
groups=groups[i],
|
| 73 |
+
)
|
| 74 |
+
)
|
| 75 |
+
|
| 76 |
+
# Multi-layer feature aggregation
|
| 77 |
+
self.mfa = TDNNBlock(
|
| 78 |
+
channels[-1],
|
| 79 |
+
channels[-1],
|
| 80 |
+
kernel_sizes[-1],
|
| 81 |
+
dilations[-1],
|
| 82 |
+
groups=groups[-1],
|
| 83 |
+
)
|
| 84 |
+
|
| 85 |
+
# Attentive Statistical Pooling
|
| 86 |
+
self.asp = AttentiveStatisticsPooling(
|
| 87 |
+
channels[-1],
|
| 88 |
+
attention_channels=attention_channels,
|
| 89 |
+
global_context=global_context,
|
| 90 |
+
)
|
| 91 |
+
self.asp_norm = LayerNorm(channels[-1] * 2, eps=1e-12)
|
| 92 |
+
|
| 93 |
+
# Final linear transformation
|
| 94 |
+
self.fc = Conv1d(
|
| 95 |
+
in_channels=channels[-1] * 2,
|
| 96 |
+
out_channels=embed_dim,
|
| 97 |
+
kernel_size=1,
|
| 98 |
+
)
|
| 99 |
+
|
| 100 |
+
self.reset_parameters()
|
| 101 |
+
|
| 102 |
+
def reset_parameters(self) -> None:
|
| 103 |
+
"""Reset the parameters and buffers of the module."""
|
| 104 |
+
|
| 105 |
+
def encoder_init(m: Module) -> None:
|
| 106 |
+
if isinstance(m, Conv1d):
|
| 107 |
+
init.xavier_uniform_(m.weight, init.calculate_gain("relu"))
|
| 108 |
+
|
| 109 |
+
self.apply(encoder_init)
|
| 110 |
+
|
| 111 |
+
def forward(
|
| 112 |
+
self,
|
| 113 |
+
x: Tensor,
|
| 114 |
+
padding_mask: Optional[PaddingMask] = None,
|
| 115 |
+
) -> Tensor:
|
| 116 |
+
"""Returns the embedding vector.
|
| 117 |
+
|
| 118 |
+
Arguments
|
| 119 |
+
---------
|
| 120 |
+
x : torch.Tensor
|
| 121 |
+
Tensor of shape (batch, time, channel).
|
| 122 |
+
"""
|
| 123 |
+
# Minimize transpose for efficiency
|
| 124 |
+
x = x.transpose(1, 2)
|
| 125 |
+
|
| 126 |
+
xl = []
|
| 127 |
+
for layer in self.blocks:
|
| 128 |
+
x = layer(x, padding_mask=padding_mask)
|
| 129 |
+
xl.append(x)
|
| 130 |
+
|
| 131 |
+
# Multi-layer feature aggregation
|
| 132 |
+
x = torch.cat(xl[1:], dim=1)
|
| 133 |
+
x = self.mfa(x)
|
| 134 |
+
|
| 135 |
+
# Attentive Statistical Pooling
|
| 136 |
+
x = self.asp(x, padding_mask=padding_mask)
|
| 137 |
+
x = self.asp_norm(x.transpose(1, 2)).transpose(1, 2)
|
| 138 |
+
|
| 139 |
+
# Final linear transformation
|
| 140 |
+
x = self.fc(x)
|
| 141 |
+
|
| 142 |
+
x = x.transpose(1, 2).squeeze(1) # B x C
|
| 143 |
+
return F.normalize(x, dim=-1)
|
| 144 |
+
|
| 145 |
+
|
| 146 |
+
class TDNNBlock(Module):
|
| 147 |
+
"""An implementation of TDNN.
|
| 148 |
+
|
| 149 |
+
Arguments
|
| 150 |
+
----------
|
| 151 |
+
:param in_channels : int
|
| 152 |
+
Number of input channels.
|
| 153 |
+
:param out_channels : int
|
| 154 |
+
The number of output channels.
|
| 155 |
+
:param kernel_size : int
|
| 156 |
+
The kernel size of the TDNN blocks.
|
| 157 |
+
:param dilation : int
|
| 158 |
+
The dilation of the TDNN block.
|
| 159 |
+
:param groups: int
|
| 160 |
+
The groups size of the TDNN blocks.
|
| 161 |
+
|
| 162 |
+
Example
|
| 163 |
+
-------
|
| 164 |
+
>>> inp_tensor = torch.rand([8, 120, 64]).transpose(1, 2)
|
| 165 |
+
>>> layer = TDNNBlock(64, 64, kernel_size=3, dilation=1)
|
| 166 |
+
>>> out_tensor = layer(inp_tensor).transpose(1, 2)
|
| 167 |
+
>>> out_tensor.shape
|
| 168 |
+
torch.Size([8, 120, 64])
|
| 169 |
+
"""
|
| 170 |
+
|
| 171 |
+
def __init__(
|
| 172 |
+
self,
|
| 173 |
+
in_channels: int,
|
| 174 |
+
out_channels: int,
|
| 175 |
+
kernel_size: int,
|
| 176 |
+
dilation: int,
|
| 177 |
+
groups: int = 1,
|
| 178 |
+
):
|
| 179 |
+
super().__init__()
|
| 180 |
+
self.conv = Conv1d(
|
| 181 |
+
in_channels=in_channels,
|
| 182 |
+
out_channels=out_channels,
|
| 183 |
+
kernel_size=kernel_size,
|
| 184 |
+
dilation=dilation,
|
| 185 |
+
padding=dilation * (kernel_size - 1) // 2,
|
| 186 |
+
groups=groups,
|
| 187 |
+
)
|
| 188 |
+
self.activation = ReLU()
|
| 189 |
+
self.norm = LayerNorm(out_channels, eps=1e-12)
|
| 190 |
+
|
| 191 |
+
def forward(self, x: Tensor, padding_mask: Optional[PaddingMask] = None) -> Tensor:
|
| 192 |
+
"""Processes the input tensor x and returns an output tensor."""
|
| 193 |
+
x = self.activation(self.conv(x))
|
| 194 |
+
|
| 195 |
+
return self.norm(x.transpose(1, 2)).transpose(1, 2) # type: ignore[no-any-return]
|
| 196 |
+
|
| 197 |
+
|
| 198 |
+
class Res2NetBlock(Module):
|
| 199 |
+
"""An implementation of Res2NetBlock w/ dilation.
|
| 200 |
+
|
| 201 |
+
Arguments
|
| 202 |
+
---------
|
| 203 |
+
:param in_channels : int
|
| 204 |
+
The number of channels expected in the input.
|
| 205 |
+
:param out_channels : int
|
| 206 |
+
The number of output channels.
|
| 207 |
+
:param scale : int
|
| 208 |
+
The scale of the Res2Net block.
|
| 209 |
+
:param kernel_size: int
|
| 210 |
+
The kernel size of the Res2Net block.
|
| 211 |
+
:param dilation : int
|
| 212 |
+
The dilation of the Res2Net block.
|
| 213 |
+
|
| 214 |
+
Example
|
| 215 |
+
-------
|
| 216 |
+
>>> inp_tensor = torch.rand([8, 120, 64]).transpose(1, 2)
|
| 217 |
+
>>> layer = Res2NetBlock(64, 64, scale=4, dilation=3)
|
| 218 |
+
>>> out_tensor = layer(inp_tensor).transpose(1, 2)
|
| 219 |
+
>>> out_tensor.shape
|
| 220 |
+
torch.Size([8, 120, 64])
|
| 221 |
+
"""
|
| 222 |
+
|
| 223 |
+
def __init__(
|
| 224 |
+
self,
|
| 225 |
+
in_channels: int,
|
| 226 |
+
out_channels: int,
|
| 227 |
+
scale: int = 8,
|
| 228 |
+
kernel_size: int = 3,
|
| 229 |
+
dilation: int = 1,
|
| 230 |
+
):
|
| 231 |
+
super().__init__()
|
| 232 |
+
assert in_channels % scale == 0
|
| 233 |
+
assert out_channels % scale == 0
|
| 234 |
+
|
| 235 |
+
in_channel = in_channels // scale
|
| 236 |
+
hidden_channel = out_channels // scale
|
| 237 |
+
self.blocks = ModuleList(
|
| 238 |
+
[
|
| 239 |
+
TDNNBlock(
|
| 240 |
+
in_channel,
|
| 241 |
+
hidden_channel,
|
| 242 |
+
kernel_size=kernel_size,
|
| 243 |
+
dilation=dilation,
|
| 244 |
+
)
|
| 245 |
+
for i in range(scale - 1)
|
| 246 |
+
]
|
| 247 |
+
)
|
| 248 |
+
self.scale = scale
|
| 249 |
+
|
| 250 |
+
def forward(self, x: Tensor) -> Tensor:
|
| 251 |
+
"""Processes the input tensor x and returns an output tensor."""
|
| 252 |
+
y = []
|
| 253 |
+
for i, x_i in enumerate(torch.chunk(x, self.scale, dim=1)):
|
| 254 |
+
if i == 0:
|
| 255 |
+
y_i = x_i
|
| 256 |
+
elif i == 1:
|
| 257 |
+
y_i = self.blocks[i - 1](x_i)
|
| 258 |
+
else:
|
| 259 |
+
y_i = self.blocks[i - 1](x_i + y_i)
|
| 260 |
+
y.append(y_i)
|
| 261 |
+
|
| 262 |
+
y_tensor = torch.cat(y, dim=1)
|
| 263 |
+
return y_tensor
|
| 264 |
+
|
| 265 |
+
|
| 266 |
+
class SEBlock(Module):
|
| 267 |
+
"""An implementation of squeeze-and-excitation block.
|
| 268 |
+
|
| 269 |
+
Arguments
|
| 270 |
+
---------
|
| 271 |
+
in_channels : int
|
| 272 |
+
The number of input channels.
|
| 273 |
+
se_channels : int
|
| 274 |
+
The number of output channels after squeeze.
|
| 275 |
+
out_channels : int
|
| 276 |
+
The number of output channels.
|
| 277 |
+
"""
|
| 278 |
+
|
| 279 |
+
def __init__(
|
| 280 |
+
self,
|
| 281 |
+
in_channels: int,
|
| 282 |
+
se_channels: int,
|
| 283 |
+
out_channels: int,
|
| 284 |
+
):
|
| 285 |
+
super().__init__()
|
| 286 |
+
|
| 287 |
+
self.conv1 = Conv1d(
|
| 288 |
+
in_channels=in_channels, out_channels=se_channels, kernel_size=1
|
| 289 |
+
)
|
| 290 |
+
self.relu = ReLU(inplace=True)
|
| 291 |
+
self.conv2 = Conv1d(
|
| 292 |
+
in_channels=se_channels, out_channels=out_channels, kernel_size=1
|
| 293 |
+
)
|
| 294 |
+
self.sigmoid = Sigmoid()
|
| 295 |
+
|
| 296 |
+
def forward(self, x: Tensor, padding_mask: Optional[PaddingMask] = None) -> Tensor:
|
| 297 |
+
"""Processes the input tensor x and returns an output tensor."""
|
| 298 |
+
if padding_mask is not None:
|
| 299 |
+
mask = padding_mask.materialize().unsqueeze(1)
|
| 300 |
+
s = (x * mask).sum(dim=2, keepdim=True) / padding_mask.seq_lens[
|
| 301 |
+
:, None, None
|
| 302 |
+
]
|
| 303 |
+
else:
|
| 304 |
+
s = x.mean(dim=2, keepdim=True)
|
| 305 |
+
|
| 306 |
+
s = self.relu(self.conv1(s))
|
| 307 |
+
s = self.sigmoid(self.conv2(s))
|
| 308 |
+
|
| 309 |
+
return s * x
|
| 310 |
+
|
| 311 |
+
|
| 312 |
+
class AttentiveStatisticsPooling(Module):
|
| 313 |
+
"""This class implements an attentive statistic pooling layer for each channel.
|
| 314 |
+
It returns the concatenated mean and std of the input tensor.
|
| 315 |
+
|
| 316 |
+
Arguments
|
| 317 |
+
---------
|
| 318 |
+
channels: int
|
| 319 |
+
The number of input channels.
|
| 320 |
+
attention_channels: int
|
| 321 |
+
The number of attention channels.
|
| 322 |
+
"""
|
| 323 |
+
|
| 324 |
+
def __init__(
|
| 325 |
+
self, channels: int, attention_channels: int = 128, global_context: bool = True
|
| 326 |
+
):
|
| 327 |
+
super().__init__()
|
| 328 |
+
|
| 329 |
+
self.eps = 1e-12
|
| 330 |
+
self.global_context = global_context
|
| 331 |
+
if global_context:
|
| 332 |
+
self.tdnn = TDNNBlock(channels * 3, attention_channels, 1, 1)
|
| 333 |
+
else:
|
| 334 |
+
self.tdnn = TDNNBlock(channels, attention_channels, 1, 1)
|
| 335 |
+
|
| 336 |
+
self.tanh = Tanh()
|
| 337 |
+
self.conv = Conv1d(
|
| 338 |
+
in_channels=attention_channels, out_channels=channels, kernel_size=1
|
| 339 |
+
)
|
| 340 |
+
|
| 341 |
+
def forward(self, x: Tensor, padding_mask: Optional[PaddingMask] = None) -> Tensor:
|
| 342 |
+
"""Calculates mean and std for a batch (input tensor).
|
| 343 |
+
|
| 344 |
+
Arguments
|
| 345 |
+
---------
|
| 346 |
+
x : torch.Tensor
|
| 347 |
+
Tensor of shape [N, C, L].
|
| 348 |
+
"""
|
| 349 |
+
L = x.shape[-1]
|
| 350 |
+
|
| 351 |
+
def _compute_statistics(
|
| 352 |
+
x: Tensor, m: Tensor, dim: int = 2, eps: float = self.eps
|
| 353 |
+
) -> Tuple[Tensor, Tensor]:
|
| 354 |
+
mean = (m * x).sum(dim)
|
| 355 |
+
std = torch.sqrt((m * (x - mean.unsqueeze(dim)).pow(2)).sum(dim).clamp(eps))
|
| 356 |
+
return mean, std
|
| 357 |
+
|
| 358 |
+
# if lengths is None:
|
| 359 |
+
# lengths = [x.shape[0]]
|
| 360 |
+
|
| 361 |
+
# Make binary mask of shape [N, 1, L]
|
| 362 |
+
# mask = to_padding_mask(lengths, max(lengths))
|
| 363 |
+
if padding_mask is not None:
|
| 364 |
+
mask = padding_mask.materialize()
|
| 365 |
+
else:
|
| 366 |
+
mask = to_padding_mask(torch.IntTensor([L]), L).repeat(x.shape[0], 1).to(x)
|
| 367 |
+
mask = mask.unsqueeze(1)
|
| 368 |
+
|
| 369 |
+
# Expand the temporal context of the pooling layer by allowing the
|
| 370 |
+
# self-attention to look at global properties of the utterance.
|
| 371 |
+
if self.global_context:
|
| 372 |
+
# torch.std is unstable for backward computation
|
| 373 |
+
# https://github.com/pytorch/pytorch/issues/4320
|
| 374 |
+
total = mask.sum(dim=2, keepdim=True).to(x)
|
| 375 |
+
mean, std = _compute_statistics(x, mask / total)
|
| 376 |
+
mean = mean.unsqueeze(2).repeat(1, 1, L)
|
| 377 |
+
std = std.unsqueeze(2).repeat(1, 1, L)
|
| 378 |
+
attn = torch.cat([x, mean, std], dim=1)
|
| 379 |
+
else:
|
| 380 |
+
attn = x
|
| 381 |
+
|
| 382 |
+
# Apply layers
|
| 383 |
+
attn = self.conv(self.tanh(self.tdnn(attn)))
|
| 384 |
+
|
| 385 |
+
# Filter out zero-paddings
|
| 386 |
+
attn = attn.masked_fill(mask == 0, float("-inf"))
|
| 387 |
+
|
| 388 |
+
attn = F.softmax(attn, dim=2)
|
| 389 |
+
mean, std = _compute_statistics(x, attn)
|
| 390 |
+
# Append mean and std of the batch
|
| 391 |
+
pooled_stats = torch.cat((mean, std), dim=1)
|
| 392 |
+
pooled_stats = pooled_stats.unsqueeze(2)
|
| 393 |
+
|
| 394 |
+
return pooled_stats
|
| 395 |
+
|
| 396 |
+
|
| 397 |
+
class SERes2NetBlock(Module):
|
| 398 |
+
"""An implementation of building block in ECAPA-TDNN, i.e.,
|
| 399 |
+
TDNN-Res2Net-TDNN-SEBlock.
|
| 400 |
+
|
| 401 |
+
Arguments
|
| 402 |
+
----------
|
| 403 |
+
out_channels: int
|
| 404 |
+
The number of output channels.
|
| 405 |
+
res2net_scale: int
|
| 406 |
+
The scale of the Res2Net block.
|
| 407 |
+
kernel_size: int
|
| 408 |
+
The kernel size of the TDNN blocks.
|
| 409 |
+
dilation: int
|
| 410 |
+
The dilation of the Res2Net block.
|
| 411 |
+
groups: int
|
| 412 |
+
Number of blocked connections from input channels to output channels.
|
| 413 |
+
|
| 414 |
+
Example
|
| 415 |
+
-------
|
| 416 |
+
>>> x = torch.rand(8, 120, 64).transpose(1, 2)
|
| 417 |
+
>>> conv = SERes2NetBlock(64, 64, res2net_scale=4)
|
| 418 |
+
>>> out = conv(x).transpose(1, 2)
|
| 419 |
+
>>> out.shape
|
| 420 |
+
torch.Size([8, 120, 64])
|
| 421 |
+
"""
|
| 422 |
+
|
| 423 |
+
def __init__(
|
| 424 |
+
self,
|
| 425 |
+
in_channels: int,
|
| 426 |
+
out_channels: int,
|
| 427 |
+
res2net_scale: int = 8,
|
| 428 |
+
se_channels: int = 128,
|
| 429 |
+
kernel_size: int = 1,
|
| 430 |
+
dilation: int = 1,
|
| 431 |
+
groups: int = 1,
|
| 432 |
+
):
|
| 433 |
+
super().__init__()
|
| 434 |
+
self.out_channels = out_channels
|
| 435 |
+
self.tdnn1 = TDNNBlock(
|
| 436 |
+
in_channels,
|
| 437 |
+
out_channels,
|
| 438 |
+
kernel_size=1,
|
| 439 |
+
dilation=1,
|
| 440 |
+
groups=groups,
|
| 441 |
+
)
|
| 442 |
+
self.res2net_block = Res2NetBlock(
|
| 443 |
+
out_channels,
|
| 444 |
+
out_channels,
|
| 445 |
+
res2net_scale,
|
| 446 |
+
kernel_size,
|
| 447 |
+
dilation,
|
| 448 |
+
)
|
| 449 |
+
self.tdnn2 = TDNNBlock(
|
| 450 |
+
out_channels,
|
| 451 |
+
out_channels,
|
| 452 |
+
kernel_size=1,
|
| 453 |
+
dilation=1,
|
| 454 |
+
groups=groups,
|
| 455 |
+
)
|
| 456 |
+
self.se_block = SEBlock(out_channels, se_channels, out_channels)
|
| 457 |
+
|
| 458 |
+
self.shortcut = None
|
| 459 |
+
if in_channels != out_channels:
|
| 460 |
+
self.shortcut = Conv1d(
|
| 461 |
+
in_channels=in_channels,
|
| 462 |
+
out_channels=out_channels,
|
| 463 |
+
kernel_size=1,
|
| 464 |
+
)
|
| 465 |
+
|
| 466 |
+
def forward(self, x: Tensor, padding_mask: Optional[PaddingMask] = None) -> Tensor:
|
| 467 |
+
"""Processes the input tensor x and returns an output tensor."""
|
| 468 |
+
residual = x
|
| 469 |
+
if self.shortcut:
|
| 470 |
+
residual = self.shortcut(x)
|
| 471 |
+
|
| 472 |
+
x = self.tdnn1(x)
|
| 473 |
+
x = self.res2net_block(x)
|
| 474 |
+
x = self.tdnn2(x)
|
| 475 |
+
x = self.se_block(x, padding_mask=padding_mask)
|
| 476 |
+
|
| 477 |
+
return x + residual
|