File size: 14,621 Bytes
167f5d0 | 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 | # Copyright 2019 Shigeki Karita
# Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)
"""Transformer encoder definition."""
from typing import List
from typing import Optional
from typing import Tuple
import torch
from torch import nn
import logging
from funasr.models.transformer.attention import MultiHeadedAttention
from funasr.models.lcbnet.attention import MultiHeadedAttentionReturnWeight
from funasr.models.transformer.embedding import PositionalEncoding
from funasr.models.transformer.layer_norm import LayerNorm
from funasr.models.transformer.utils.nets_utils import make_pad_mask
from funasr.models.transformer.positionwise_feed_forward import PositionwiseFeedForward
from funasr.models.transformer.utils.repeat import repeat
from funasr.register import tables
class EncoderLayer(nn.Module):
"""Encoder layer module.
Args:
size (int): Input dimension.
self_attn (torch.nn.Module): Self-attention module instance.
`MultiHeadedAttention` or `RelPositionMultiHeadedAttention` instance
can be used as the argument.
feed_forward (torch.nn.Module): Feed-forward module instance.
`PositionwiseFeedForward`, `MultiLayeredConv1d`, or `Conv1dLinear` instance
can be used as the argument.
dropout_rate (float): Dropout rate.
normalize_before (bool): Whether to use layer_norm before the first block.
concat_after (bool): Whether to concat attention layer's input and output.
if True, additional linear will be applied.
i.e. x -> x + linear(concat(x, att(x)))
if False, no additional linear will be applied. i.e. x -> x + att(x)
stochastic_depth_rate (float): Proability to skip this layer.
During training, the layer may skip residual computation and return input
as-is with given probability.
"""
def __init__(
self,
size,
self_attn,
feed_forward,
dropout_rate,
normalize_before=True,
concat_after=False,
stochastic_depth_rate=0.0,
):
"""Construct an EncoderLayer object."""
super(EncoderLayer, self).__init__()
self.self_attn = self_attn
self.feed_forward = feed_forward
self.norm1 = LayerNorm(size)
self.norm2 = LayerNorm(size)
self.dropout = nn.Dropout(dropout_rate)
self.size = size
self.normalize_before = normalize_before
self.concat_after = concat_after
if self.concat_after:
self.concat_linear = nn.Linear(size + size, size)
self.stochastic_depth_rate = stochastic_depth_rate
def forward(self, x, mask, cache=None):
"""Compute encoded features.
Args:
x_input (torch.Tensor): Input tensor (#batch, time, size).
mask (torch.Tensor): Mask tensor for the input (#batch, time).
cache (torch.Tensor): Cache tensor of the input (#batch, time - 1, size).
Returns:
torch.Tensor: Output tensor (#batch, time, size).
torch.Tensor: Mask tensor (#batch, time).
"""
skip_layer = False
# with stochastic depth, residual connection `x + f(x)` becomes
# `x <- x + 1 / (1 - p) * f(x)` at training time.
stoch_layer_coeff = 1.0
if self.training and self.stochastic_depth_rate > 0:
skip_layer = torch.rand(1).item() < self.stochastic_depth_rate
stoch_layer_coeff = 1.0 / (1 - self.stochastic_depth_rate)
if skip_layer:
if cache is not None:
x = torch.cat([cache, x], dim=1)
return x, mask
residual = x
if self.normalize_before:
x = self.norm1(x)
if cache is None:
x_q = x
else:
assert cache.shape == (x.shape[0], x.shape[1] - 1, self.size)
x_q = x[:, -1:, :]
residual = residual[:, -1:, :]
mask = None if mask is None else mask[:, -1:, :]
if self.concat_after:
x_concat = torch.cat((x, self.self_attn(x_q, x, x, mask)), dim=-1)
x = residual + stoch_layer_coeff * self.concat_linear(x_concat)
else:
x = residual + stoch_layer_coeff * self.dropout(self.self_attn(x_q, x, x, mask))
if not self.normalize_before:
x = self.norm1(x)
residual = x
if self.normalize_before:
x = self.norm2(x)
x = residual + stoch_layer_coeff * self.dropout(self.feed_forward(x))
if not self.normalize_before:
x = self.norm2(x)
if cache is not None:
x = torch.cat([cache, x], dim=1)
return x, mask
@tables.register("encoder_classes", "TransformerTextEncoder")
class TransformerTextEncoder(nn.Module):
"""Transformer text encoder module.
Args:
input_size: input dim
output_size: dimension of attention
attention_heads: the number of heads of multi head attention
linear_units: the number of units of position-wise feed forward
num_blocks: the number of decoder blocks
dropout_rate: dropout rate
attention_dropout_rate: dropout rate in attention
positional_dropout_rate: dropout rate after adding positional encoding
input_layer: input layer type
pos_enc_class: PositionalEncoding or ScaledPositionalEncoding
normalize_before: whether to use layer_norm before the first block
concat_after: whether to concat attention layer's input and output
if True, additional linear will be applied.
i.e. x -> x + linear(concat(x, att(x)))
if False, no additional linear will be applied.
i.e. x -> x + att(x)
positionwise_layer_type: linear of conv1d
positionwise_conv_kernel_size: kernel size of positionwise conv1d layer
padding_idx: padding_idx for input_layer=embed
"""
def __init__(
self,
input_size: int,
output_size: int = 256,
attention_heads: int = 4,
linear_units: int = 2048,
num_blocks: int = 6,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
attention_dropout_rate: float = 0.0,
pos_enc_class=PositionalEncoding,
normalize_before: bool = True,
concat_after: bool = False,
):
super().__init__()
self._output_size = output_size
self.embed = torch.nn.Sequential(
torch.nn.Embedding(input_size, output_size),
pos_enc_class(output_size, positional_dropout_rate),
)
self.normalize_before = normalize_before
positionwise_layer = PositionwiseFeedForward
positionwise_layer_args = (
output_size,
linear_units,
dropout_rate,
)
self.encoders = repeat(
num_blocks,
lambda lnum: EncoderLayer(
output_size,
MultiHeadedAttention(attention_heads, output_size, attention_dropout_rate),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
),
)
if self.normalize_before:
self.after_norm = LayerNorm(output_size)
def output_size(self) -> int:
return self._output_size
def forward(
self,
xs_pad: torch.Tensor,
ilens: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
"""Embed positions in tensor.
Args:
xs_pad: input tensor (B, L, D)
ilens: input length (B)
Returns:
position embedded tensor and mask
"""
masks = (~make_pad_mask(ilens)[:, None, :]).to(xs_pad.device)
xs_pad = self.embed(xs_pad)
xs_pad, masks = self.encoders(xs_pad, masks)
if self.normalize_before:
xs_pad = self.after_norm(xs_pad)
olens = masks.squeeze(1).sum(1)
return xs_pad, olens, None
@tables.register("encoder_classes", "FusionSANEncoder")
class SelfSrcAttention(nn.Module):
"""Single decoder layer module.
Args:
size (int): Input dimension.
self_attn (torch.nn.Module): Self-attention module instance.
`MultiHeadedAttention` instance can be used as the argument.
src_attn (torch.nn.Module): Self-attention module instance.
`MultiHeadedAttention` instance can be used as the argument.
feed_forward (torch.nn.Module): Feed-forward module instance.
`PositionwiseFeedForward`, `MultiLayeredConv1d`, or `Conv1dLinear` instance
can be used as the argument.
dropout_rate (float): Dropout rate.
normalize_before (bool): Whether to use layer_norm before the first block.
concat_after (bool): Whether to concat attention layer's input and output.
if True, additional linear will be applied.
i.e. x -> x + linear(concat(x, att(x)))
if False, no additional linear will be applied. i.e. x -> x + att(x)
"""
def __init__(
self,
size,
attention_heads,
attention_dim,
linear_units,
self_attention_dropout_rate,
src_attention_dropout_rate,
positional_dropout_rate,
dropout_rate,
normalize_before=True,
concat_after=False,
):
"""Construct an SelfSrcAttention object."""
super(SelfSrcAttention, self).__init__()
self.size = size
self.self_attn = MultiHeadedAttention(
attention_heads, attention_dim, self_attention_dropout_rate
)
self.src_attn = MultiHeadedAttentionReturnWeight(
attention_heads, attention_dim, src_attention_dropout_rate
)
self.feed_forward = PositionwiseFeedForward(
attention_dim, linear_units, positional_dropout_rate
)
self.norm1 = LayerNorm(size)
self.norm2 = LayerNorm(size)
self.norm3 = LayerNorm(size)
self.dropout = nn.Dropout(dropout_rate)
self.normalize_before = normalize_before
self.concat_after = concat_after
if self.concat_after:
self.concat_linear1 = nn.Linear(size + size, size)
self.concat_linear2 = nn.Linear(size + size, size)
def forward(self, tgt, tgt_mask, memory, memory_mask, cache=None):
"""Compute decoded features.
Args:
tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size).
tgt_mask (torch.Tensor): Mask for input tensor (#batch, maxlen_out).
memory (torch.Tensor): Encoded memory, float32 (#batch, maxlen_in, size).
memory_mask (torch.Tensor): Encoded memory mask (#batch, maxlen_in).
cache (List[torch.Tensor]): List of cached tensors.
Each tensor shape should be (#batch, maxlen_out - 1, size).
Returns:
torch.Tensor: Output tensor(#batch, maxlen_out, size).
torch.Tensor: Mask for output tensor (#batch, maxlen_out).
torch.Tensor: Encoded memory (#batch, maxlen_in, size).
torch.Tensor: Encoded memory mask (#batch, maxlen_in).
"""
residual = tgt
if self.normalize_before:
tgt = self.norm1(tgt)
if cache is None:
tgt_q = tgt
tgt_q_mask = tgt_mask
else:
# compute only the last frame query keeping dim: max_time_out -> 1
assert cache.shape == (
tgt.shape[0],
tgt.shape[1] - 1,
self.size,
), f"{cache.shape} == {(tgt.shape[0], tgt.shape[1] - 1, self.size)}"
tgt_q = tgt[:, -1:, :]
residual = residual[:, -1:, :]
tgt_q_mask = None
if tgt_mask is not None:
tgt_q_mask = tgt_mask[:, -1:, :]
if self.concat_after:
tgt_concat = torch.cat((tgt_q, self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)), dim=-1)
x = residual + self.concat_linear1(tgt_concat)
else:
x = residual + self.dropout(self.self_attn(tgt_q, tgt, tgt, tgt_q_mask))
if not self.normalize_before:
x = self.norm1(x)
residual = x
if self.normalize_before:
x = self.norm2(x)
if self.concat_after:
x_concat = torch.cat((x, self.src_attn(x, memory, memory, memory_mask)), dim=-1)
x = residual + self.concat_linear2(x_concat)
else:
x, score = self.src_attn(x, memory, memory, memory_mask)
x = residual + self.dropout(x)
if not self.normalize_before:
x = self.norm2(x)
residual = x
if self.normalize_before:
x = self.norm3(x)
x = residual + self.dropout(self.feed_forward(x))
if not self.normalize_before:
x = self.norm3(x)
if cache is not None:
x = torch.cat([cache, x], dim=1)
return x, tgt_mask, memory, memory_mask
@tables.register("encoder_classes", "ConvBiasPredictor")
class ConvPredictor(nn.Module):
def __init__(
self,
size=256,
l_order=3,
r_order=3,
attention_heads=4,
attention_dropout_rate=0.1,
linear_units=2048,
):
super().__init__()
self.atten = MultiHeadedAttention(attention_heads, size, attention_dropout_rate)
self.norm1 = LayerNorm(size)
self.feed_forward = PositionwiseFeedForward(size, linear_units, attention_dropout_rate)
self.norm2 = LayerNorm(size)
self.pad = nn.ConstantPad1d((l_order, r_order), 0)
self.conv1d = nn.Conv1d(size, size, l_order + r_order + 1, groups=size)
self.output_linear = nn.Linear(size, 1)
def forward(self, text_enc, asr_enc):
# stage1 cross-attention
residual = text_enc
text_enc = residual + self.atten(text_enc, asr_enc, asr_enc, None)
# stage2 FFN
residual = text_enc
text_enc = self.norm1(text_enc)
text_enc = residual + self.feed_forward(text_enc)
# stage Conv predictor
text_enc = self.norm2(text_enc)
context = text_enc.transpose(1, 2)
queries = self.pad(context)
memory = self.conv1d(queries)
output = memory + context
output = output.transpose(1, 2)
output = torch.relu(output)
output = self.output_linear(output)
if output.dim() == 3:
output = output.squeeze(2)
return output
|