File size: 13,838 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 |
# 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}
# }
#
from typing import Dict, Optional, Tuple
import torch
from modules.wenet_extractor.cif.predictor import MAELoss
from modules.wenet_extractor.paraformer.search.beam_search import Hypothesis
from modules.wenet_extractor.transformer.asr_model import ASRModel
from modules.wenet_extractor.transformer.ctc import CTC
from modules.wenet_extractor.transformer.decoder import TransformerDecoder
from modules.wenet_extractor.transformer.encoder import TransformerEncoder
from modules.wenet_extractor.utils.common import IGNORE_ID, add_sos_eos, th_accuracy
from modules.wenet_extractor.utils.mask import make_pad_mask
class Paraformer(ASRModel):
"""Paraformer: Fast and Accurate Parallel Transformer for
Non-autoregressive End-to-End Speech Recognition
see https://arxiv.org/pdf/2206.08317.pdf
"""
def __init__(
self,
vocab_size: int,
encoder: TransformerEncoder,
decoder: TransformerDecoder,
ctc: CTC,
predictor,
ctc_weight: float = 0.5,
predictor_weight: float = 1.0,
predictor_bias: int = 0,
ignore_id: int = IGNORE_ID,
reverse_weight: float = 0.0,
lsm_weight: float = 0.0,
length_normalized_loss: bool = False,
):
assert 0.0 <= ctc_weight <= 1.0, ctc_weight
assert 0.0 <= predictor_weight <= 1.0, predictor_weight
super().__init__(
vocab_size,
encoder,
decoder,
ctc,
ctc_weight,
ignore_id,
reverse_weight,
lsm_weight,
length_normalized_loss,
)
self.predictor = predictor
self.predictor_weight = predictor_weight
self.predictor_bias = predictor_bias
self.criterion_pre = MAELoss(normalize_length=length_normalized_loss)
def forward(
self,
speech: torch.Tensor,
speech_lengths: torch.Tensor,
text: torch.Tensor,
text_lengths: torch.Tensor,
) -> Dict[str, Optional[torch.Tensor]]:
"""Frontend + Encoder + Decoder + Calc loss
Args:
speech: (Batch, Length, ...)
speech_lengths: (Batch, )
text: (Batch, Length)
text_lengths: (Batch,)
"""
assert text_lengths.dim() == 1, text_lengths.shape
# Check that batch_size is unified
assert (
speech.shape[0]
== speech_lengths.shape[0]
== text.shape[0]
== text_lengths.shape[0]
), (speech.shape, speech_lengths.shape, text.shape, text_lengths.shape)
# 1. Encoder
encoder_out, encoder_mask = self.encoder(speech, speech_lengths)
encoder_out_lens = encoder_mask.squeeze(1).sum(1)
# 2a. Attention-decoder branch
if self.ctc_weight != 1.0:
loss_att, acc_att, loss_pre = self._calc_att_loss(
encoder_out, encoder_mask, text, text_lengths
)
else:
# loss_att = None
# loss_pre = None
loss_att: torch.Tensor = torch.tensor(0)
loss_pre: torch.Tensor = torch.tensor(0)
# 2b. CTC branch
if self.ctc_weight != 0.0:
loss_ctc = self.ctc(encoder_out, encoder_out_lens, text, text_lengths)
else:
loss_ctc = None
if loss_ctc is None:
loss = loss_att + self.predictor_weight * loss_pre
# elif loss_att is None:
elif loss_att == torch.tensor(0):
loss = loss_ctc
else:
loss = (
self.ctc_weight * loss_ctc
+ (1 - self.ctc_weight) * loss_att
+ self.predictor_weight * loss_pre
)
return {
"loss": loss,
"loss_att": loss_att,
"loss_ctc": loss_ctc,
"loss_pre": loss_pre,
}
def _calc_att_loss(
self,
encoder_out: torch.Tensor,
encoder_mask: torch.Tensor,
ys_pad: torch.Tensor,
ys_pad_lens: torch.Tensor,
) -> Tuple[torch.Tensor, float, torch.Tensor]:
if self.predictor_bias == 1:
_, ys_pad = add_sos_eos(ys_pad, self.sos, self.eos, self.ignore_id)
ys_pad_lens = ys_pad_lens + self.predictor_bias
pre_acoustic_embeds, pre_token_length, _, pre_peak_index = self.predictor(
encoder_out, ys_pad, encoder_mask, ignore_id=self.ignore_id
)
# 1. Forward decoder
decoder_out, _, _ = self.decoder(
encoder_out, encoder_mask, pre_acoustic_embeds, ys_pad_lens
)
# 2. Compute attention loss
loss_att = self.criterion_att(decoder_out, ys_pad)
acc_att = th_accuracy(
decoder_out.view(-1, self.vocab_size),
ys_pad,
ignore_label=self.ignore_id,
)
loss_pre: torch.Tensor = self.criterion_pre(
ys_pad_lens.type_as(pre_token_length), pre_token_length
)
return loss_att, acc_att, loss_pre
def calc_predictor(self, encoder_out, encoder_mask):
encoder_mask = (
~make_pad_mask(encoder_mask, max_len=encoder_out.size(1))[:, None, :]
).to(encoder_out.device)
pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index = self.predictor(
encoder_out, None, encoder_mask, ignore_id=self.ignore_id
)
return pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index
def cal_decoder_with_predictor(
self, encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens
):
decoder_out, _, _ = self.decoder(
encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens
)
decoder_out = torch.log_softmax(decoder_out, dim=-1)
return decoder_out, ys_pad_lens
def recognize(self):
raise NotImplementedError
def paraformer_greedy_search(
self,
speech: torch.Tensor,
speech_lengths: torch.Tensor,
decoding_chunk_size: int = -1,
num_decoding_left_chunks: int = -1,
simulate_streaming: bool = False,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Apply beam search on attention decoder
Args:
speech (torch.Tensor): (batch, max_len, feat_dim)
speech_length (torch.Tensor): (batch, )
decoding_chunk_size (int): decoding chunk for dynamic chunk
trained model.
<0: for decoding, use full chunk.
>0: for decoding, use fixed chunk size as set.
0: used for training, it's prohibited here
simulate_streaming (bool): whether do encoder forward in a
streaming fashion
Returns:
torch.Tensor: decoding result, (batch, max_result_len)
"""
assert speech.shape[0] == speech_lengths.shape[0]
assert decoding_chunk_size != 0
device = speech.device
batch_size = speech.shape[0]
# Let's assume B = batch_size and N = beam_size
# 1. Encoder
encoder_out, encoder_mask = self._forward_encoder(
speech,
speech_lengths,
decoding_chunk_size,
num_decoding_left_chunks,
simulate_streaming,
) # (B, maxlen, encoder_dim)
encoder_out_lens = encoder_mask.squeeze(1).sum(1)
# 2. Predictor
predictor_outs = self.calc_predictor(encoder_out, encoder_mask)
pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index = (
predictor_outs[0],
predictor_outs[1],
predictor_outs[2],
predictor_outs[3],
)
pre_token_length = pre_token_length.round().long()
if torch.max(pre_token_length) < 1:
return torch.tensor([]), torch.tensor([])
# 2. Decoder forward
decoder_outs = self.cal_decoder_with_predictor(
encoder_out, encoder_out_lens, pre_acoustic_embeds, pre_token_length
)
decoder_out, ys_pad_lens = decoder_outs[0], decoder_outs[1]
hyps = []
b, n, d = decoder_out.size()
for i in range(b):
x = encoder_out[i, : encoder_out_lens[i], :]
am_scores = decoder_out[i, : pre_token_length[i], :]
yseq = am_scores.argmax(dim=-1)
score = am_scores.max(dim=-1)[0]
score = torch.sum(score, dim=-1)
# pad with mask tokens to ensure compatibility with sos/eos tokens
yseq = torch.tensor(
[self.sos] + yseq.tolist() + [self.eos], device=yseq.device
)
nbest_hyps = [Hypothesis(yseq=yseq, score=score)]
for hyp in nbest_hyps:
assert isinstance(hyp, (Hypothesis)), type(hyp)
# remove sos/eos and get hyps
last_pos = -1
if isinstance(hyp.yseq, list):
token_int = hyp.yseq[1:last_pos]
else:
token_int = hyp.yseq[1:last_pos].tolist()
# remove blank symbol id and unk id, which is assumed to be 0
# and 1
token_int = list(filter(lambda x: x != 0 and x != 1, token_int))
hyps.append(token_int)
return hyps
def paraformer_beam_search(
self,
speech: torch.Tensor,
speech_lengths: torch.Tensor,
beam_search: torch.nn.Module = None,
decoding_chunk_size: int = -1,
num_decoding_left_chunks: int = -1,
simulate_streaming: bool = False,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Apply beam search on attention decoder
Args:
speech (torch.Tensor): (batch, max_len, feat_dim)
speech_lengths (torch.Tensor): (batch, )
beam_search (torch.nn.Moudle): beam search module
decoding_chunk_size (int): decoding chunk for dynamic chunk
trained model.
<0: for decoding, use full chunk.
>0: for decoding, use fixed chunk size as set.
0: used for training, it's prohibited here
simulate_streaming (bool): whether do encoder forward in a
streaming fashion
Returns:
torch.Tensor: decoding result, (batch, max_result_len)
"""
assert speech.shape[0] == speech_lengths.shape[0]
assert decoding_chunk_size != 0
device = speech.device
batch_size = speech.shape[0]
# Let's assume B = batch_size and N = beam_size
# 1. Encoder
encoder_out, encoder_mask = self._forward_encoder(
speech,
speech_lengths,
decoding_chunk_size,
num_decoding_left_chunks,
simulate_streaming,
) # (B, maxlen, encoder_dim)
encoder_out_lens = encoder_mask.squeeze(1).sum(1)
# 2. Predictor
predictor_outs = self.calc_predictor(encoder_out, encoder_mask)
pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index = (
predictor_outs[0],
predictor_outs[1],
predictor_outs[2],
predictor_outs[3],
)
pre_token_length = pre_token_length.round().long()
if torch.max(pre_token_length) < 1:
return torch.tensor([]), torch.tensor([])
# 2. Decoder forward
decoder_outs = self.cal_decoder_with_predictor(
encoder_out, encoder_out_lens, pre_acoustic_embeds, pre_token_length
)
decoder_out, ys_pad_lens = decoder_outs[0], decoder_outs[1]
hyps = []
b, n, d = decoder_out.size()
for i in range(b):
x = encoder_out[i, : encoder_out_lens[i], :]
am_scores = decoder_out[i, : pre_token_length[i], :]
if beam_search is not None:
nbest_hyps = beam_search(x=x, am_scores=am_scores)
nbest_hyps = nbest_hyps[:1]
else:
yseq = am_scores.argmax(dim=-1)
score = am_scores.max(dim=-1)[0]
score = torch.sum(score, dim=-1)
# pad with mask tokens to ensure compatibility with sos/eos
# tokens
yseq = torch.tensor(
[self.sos] + yseq.tolist() + [self.eos], device=yseq.device
)
nbest_hyps = [Hypothesis(yseq=yseq, score=score)]
for hyp in nbest_hyps:
assert isinstance(hyp, (Hypothesis)), type(hyp)
# remove sos/eos and get hyps
last_pos = -1
if isinstance(hyp.yseq, list):
token_int = hyp.yseq[1:last_pos]
else:
token_int = hyp.yseq[1:last_pos].tolist()
# remove blank symbol id and unk id, which is assumed to be 0
# and 1
token_int = list(filter(lambda x: x != 0 and x != 1, token_int))
hyps.append(token_int)
return hyps
|