File size: 12,834 Bytes
c968fc3 |
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 |
# This module is from [WeNet](https://github.com/wenet-e2e/wenet).
# ## Citations
# ```bibtex
# @inproceedings{yao2021wenet,
# title={WeNet: Production oriented Streaming and Non-streaming End-to-End Speech Recognition Toolkit},
# author={Yao, Zhuoyuan and Wu, Di and Wang, Xiong and Zhang, Binbin and Yu, Fan and Yang, Chao and Peng, Zhendong and Chen, Xiaoyu and Xie, Lei and Lei, Xin},
# booktitle={Proc. Interspeech},
# year={2021},
# address={Brno, Czech Republic },
# organization={IEEE}
# }
# @article{zhang2022wenet,
# title={WeNet 2.0: More Productive End-to-End Speech Recognition Toolkit},
# author={Zhang, Binbin and Wu, Di and Peng, Zhendong and Song, Xingchen and Yao, Zhuoyuan and Lv, Hang and Xie, Lei and Yang, Chao and Pan, Fuping and Niu, Jianwei},
# journal={arXiv preprint arXiv:2203.15455},
# year={2022}
# }
#
"""Decoder definition."""
from typing import Tuple, List, Optional
import torch
from modules.wenet_extractor.transformer.attention import MultiHeadedAttention
from modules.wenet_extractor.transformer.decoder_layer import DecoderLayer
from modules.wenet_extractor.transformer.embedding import PositionalEncoding
from modules.wenet_extractor.transformer.embedding import NoPositionalEncoding
from modules.wenet_extractor.transformer.positionwise_feed_forward import (
PositionwiseFeedForward,
)
from modules.wenet_extractor.utils.mask import subsequent_mask, make_pad_mask
class TransformerDecoder(torch.nn.Module):
"""Base class of Transfomer decoder module.
Args:
vocab_size: output dim
encoder_output_size: dimension of attention
attention_heads: the number of heads of multi head attention
linear_units: the hidden units number of position-wise feedforward
num_blocks: the number of decoder blocks
dropout_rate: dropout rate
self_attention_dropout_rate: dropout rate for attention
input_layer: input layer type
use_output_layer: whether to use output layer
pos_enc_class: PositionalEncoding or ScaledPositionalEncoding
normalize_before:
True: use layer_norm before each sub-block of a layer.
False: use layer_norm after each sub-block of a layer.
src_attention: if false, encoder-decoder cross attention is not
applied, such as CIF model
"""
def __init__(
self,
vocab_size: int,
encoder_output_size: int,
attention_heads: int = 4,
linear_units: int = 2048,
num_blocks: int = 6,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
self_attention_dropout_rate: float = 0.0,
src_attention_dropout_rate: float = 0.0,
input_layer: str = "embed",
use_output_layer: bool = True,
normalize_before: bool = True,
src_attention: bool = True,
):
super().__init__()
attention_dim = encoder_output_size
if input_layer == "embed":
self.embed = torch.nn.Sequential(
torch.nn.Embedding(vocab_size, attention_dim),
PositionalEncoding(attention_dim, positional_dropout_rate),
)
elif input_layer == "none":
self.embed = NoPositionalEncoding(attention_dim, positional_dropout_rate)
else:
raise ValueError(f"only 'embed' is supported: {input_layer}")
self.normalize_before = normalize_before
self.after_norm = torch.nn.LayerNorm(attention_dim, eps=1e-5)
self.use_output_layer = use_output_layer
self.output_layer = torch.nn.Linear(attention_dim, vocab_size)
self.num_blocks = num_blocks
self.decoders = torch.nn.ModuleList(
[
DecoderLayer(
attention_dim,
MultiHeadedAttention(
attention_heads, attention_dim, self_attention_dropout_rate
),
(
MultiHeadedAttention(
attention_heads, attention_dim, src_attention_dropout_rate
)
if src_attention
else None
),
PositionwiseFeedForward(attention_dim, linear_units, dropout_rate),
dropout_rate,
normalize_before,
)
for _ in range(self.num_blocks)
]
)
def forward(
self,
memory: torch.Tensor,
memory_mask: torch.Tensor,
ys_in_pad: torch.Tensor,
ys_in_lens: torch.Tensor,
r_ys_in_pad: torch.Tensor = torch.empty(0),
reverse_weight: float = 0.0,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""Forward decoder.
Args:
memory: encoded memory, float32 (batch, maxlen_in, feat)
memory_mask: encoder memory mask, (batch, 1, maxlen_in)
ys_in_pad: padded input token ids, int64 (batch, maxlen_out)
ys_in_lens: input lengths of this batch (batch)
r_ys_in_pad: not used in transformer decoder, in order to unify api
with bidirectional decoder
reverse_weight: not used in transformer decoder, in order to unify
api with bidirectional decode
Returns:
(tuple): tuple containing:
x: decoded token score before softmax (batch, maxlen_out,
vocab_size) if use_output_layer is True,
torch.tensor(0.0), in order to unify api with bidirectional decoder
olens: (batch, )
"""
tgt = ys_in_pad
maxlen = tgt.size(1)
# tgt_mask: (B, 1, L)
tgt_mask = ~make_pad_mask(ys_in_lens, maxlen).unsqueeze(1)
tgt_mask = tgt_mask.to(tgt.device)
# m: (1, L, L)
m = subsequent_mask(tgt_mask.size(-1), device=tgt_mask.device).unsqueeze(0)
# tgt_mask: (B, L, L)
tgt_mask = tgt_mask & m
x, _ = self.embed(tgt)
for layer in self.decoders:
x, tgt_mask, memory, memory_mask = layer(x, tgt_mask, memory, memory_mask)
if self.normalize_before:
x = self.after_norm(x)
if self.use_output_layer:
x = self.output_layer(x)
olens = tgt_mask.sum(1)
return x, torch.tensor(0.0), olens
def forward_one_step(
self,
memory: torch.Tensor,
memory_mask: torch.Tensor,
tgt: torch.Tensor,
tgt_mask: torch.Tensor,
cache: Optional[List[torch.Tensor]] = None,
) -> Tuple[torch.Tensor, List[torch.Tensor]]:
"""Forward one step.
This is only used for decoding.
Args:
memory: encoded memory, float32 (batch, maxlen_in, feat)
memory_mask: encoded memory mask, (batch, 1, maxlen_in)
tgt: input token ids, int64 (batch, maxlen_out)
tgt_mask: input token mask, (batch, maxlen_out)
dtype=torch.uint8 in PyTorch 1.2-
dtype=torch.bool in PyTorch 1.2+ (include 1.2)
cache: cached output list of (batch, max_time_out-1, size)
Returns:
y, cache: NN output value and cache per `self.decoders`.
y.shape` is (batch, maxlen_out, token)
"""
x, _ = self.embed(tgt)
new_cache = []
for i, decoder in enumerate(self.decoders):
if cache is None:
c = None
else:
c = cache[i]
x, tgt_mask, memory, memory_mask = decoder(
x, tgt_mask, memory, memory_mask, cache=c
)
new_cache.append(x)
if self.normalize_before:
y = self.after_norm(x[:, -1])
else:
y = x[:, -1]
if self.use_output_layer:
y = torch.log_softmax(self.output_layer(y), dim=-1)
return y, new_cache
class BiTransformerDecoder(torch.nn.Module):
"""Base class of Transfomer decoder module.
Args:
vocab_size: output dim
encoder_output_size: dimension of attention
attention_heads: the number of heads of multi head attention
linear_units: the hidden units number of position-wise feedforward
num_blocks: the number of decoder blocks
r_num_blocks: the number of right to left decoder blocks
dropout_rate: dropout rate
self_attention_dropout_rate: dropout rate for attention
input_layer: input layer type
use_output_layer: whether to use output layer
pos_enc_class: PositionalEncoding or ScaledPositionalEncoding
normalize_before:
True: use layer_norm before each sub-block of a layer.
False: use layer_norm after each sub-block of a layer.
"""
def __init__(
self,
vocab_size: int,
encoder_output_size: int,
attention_heads: int = 4,
linear_units: int = 2048,
num_blocks: int = 6,
r_num_blocks: int = 0,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
self_attention_dropout_rate: float = 0.0,
src_attention_dropout_rate: float = 0.0,
input_layer: str = "embed",
use_output_layer: bool = True,
normalize_before: bool = True,
):
super().__init__()
self.left_decoder = TransformerDecoder(
vocab_size,
encoder_output_size,
attention_heads,
linear_units,
num_blocks,
dropout_rate,
positional_dropout_rate,
self_attention_dropout_rate,
src_attention_dropout_rate,
input_layer,
use_output_layer,
normalize_before,
)
self.right_decoder = TransformerDecoder(
vocab_size,
encoder_output_size,
attention_heads,
linear_units,
r_num_blocks,
dropout_rate,
positional_dropout_rate,
self_attention_dropout_rate,
src_attention_dropout_rate,
input_layer,
use_output_layer,
normalize_before,
)
def forward(
self,
memory: torch.Tensor,
memory_mask: torch.Tensor,
ys_in_pad: torch.Tensor,
ys_in_lens: torch.Tensor,
r_ys_in_pad: torch.Tensor,
reverse_weight: float = 0.0,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""Forward decoder.
Args:
memory: encoded memory, float32 (batch, maxlen_in, feat)
memory_mask: encoder memory mask, (batch, 1, maxlen_in)
ys_in_pad: padded input token ids, int64 (batch, maxlen_out)
ys_in_lens: input lengths of this batch (batch)
r_ys_in_pad: padded input token ids, int64 (batch, maxlen_out),
used for right to left decoder
reverse_weight: used for right to left decoder
Returns:
(tuple): tuple containing:
x: decoded token score before softmax (batch, maxlen_out,
vocab_size) if use_output_layer is True,
r_x: x: decoded token score (right to left decoder)
before softmax (batch, maxlen_out, vocab_size)
if use_output_layer is True,
olens: (batch, )
"""
l_x, _, olens = self.left_decoder(memory, memory_mask, ys_in_pad, ys_in_lens)
r_x = torch.tensor(0.0)
if reverse_weight > 0.0:
r_x, _, olens = self.right_decoder(
memory, memory_mask, r_ys_in_pad, ys_in_lens
)
return l_x, r_x, olens
def forward_one_step(
self,
memory: torch.Tensor,
memory_mask: torch.Tensor,
tgt: torch.Tensor,
tgt_mask: torch.Tensor,
cache: Optional[List[torch.Tensor]] = None,
) -> Tuple[torch.Tensor, List[torch.Tensor]]:
"""Forward one step.
This is only used for decoding.
Args:
memory: encoded memory, float32 (batch, maxlen_in, feat)
memory_mask: encoded memory mask, (batch, 1, maxlen_in)
tgt: input token ids, int64 (batch, maxlen_out)
tgt_mask: input token mask, (batch, maxlen_out)
dtype=torch.uint8 in PyTorch 1.2-
dtype=torch.bool in PyTorch 1.2+ (include 1.2)
cache: cached output list of (batch, max_time_out-1, size)
Returns:
y, cache: NN output value and cache per `self.decoders`.
y.shape` is (batch, maxlen_out, token)
"""
return self.left_decoder.forward_one_step(
memory, memory_mask, tgt, tgt_mask, cache
)
|