Spaces:
Running
Running
File size: 38,123 Bytes
67c46fd |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 |
"""Beam search module."""
from itertools import chain
import logging
from typing import Any
from typing import Dict
from typing import List
from typing import NamedTuple
from typing import Tuple
from typing import Union
import torch
from funasr_detach.metrics.common import end_detect
from funasr_detach.models.transformer.scorers.scorer_interface import (
PartialScorerInterface,
)
from funasr_detach.models.transformer.scorers.scorer_interface import ScorerInterface
class Hypothesis(NamedTuple):
"""Hypothesis data type."""
yseq: torch.Tensor
score: Union[float, torch.Tensor] = 0
scores: Dict[str, Union[float, torch.Tensor]] = dict()
states: Dict[str, Any] = dict()
def asdict(self) -> dict:
"""Convert data to JSON-friendly dict."""
return self._replace(
yseq=self.yseq.tolist(),
score=float(self.score),
scores={k: float(v) for k, v in self.scores.items()},
)._asdict()
class BeamSearchScama(torch.nn.Module):
"""Beam search implementation."""
def __init__(
self,
scorers: Dict[str, ScorerInterface],
weights: Dict[str, float],
beam_size: int,
vocab_size: int,
sos: int,
eos: int,
token_list: List[str] = None,
pre_beam_ratio: float = 1.5,
pre_beam_score_key: str = None,
):
"""Initialize beam search.
Args:
scorers (dict[str, ScorerInterface]): Dict of decoder modules
e.g., Decoder, CTCPrefixScorer, LM
The scorer will be ignored if it is `None`
weights (dict[str, float]): Dict of weights for each scorers
The scorer will be ignored if its weight is 0
beam_size (int): The number of hypotheses kept during search
vocab_size (int): The number of vocabulary
sos (int): Start of sequence id
eos (int): End of sequence id
token_list (list[str]): List of tokens for debug log
pre_beam_score_key (str): key of scores to perform pre-beam search
pre_beam_ratio (float): beam size in the pre-beam search
will be `int(pre_beam_ratio * beam_size)`
"""
super().__init__()
# set scorers
self.weights = weights
self.scorers = dict()
self.full_scorers = dict()
self.part_scorers = dict()
# this module dict is required for recursive cast
# `self.to(device, dtype)` in `recog.py`
self.nn_dict = torch.nn.ModuleDict()
for k, v in scorers.items():
w = weights.get(k, 0)
if w == 0 or v is None:
continue
assert isinstance(
v, ScorerInterface
), f"{k} ({type(v)}) does not implement ScorerInterface"
self.scorers[k] = v
if isinstance(v, PartialScorerInterface):
self.part_scorers[k] = v
else:
self.full_scorers[k] = v
if isinstance(v, torch.nn.Module):
self.nn_dict[k] = v
# set configurations
self.sos = sos
self.eos = eos
self.token_list = token_list
self.pre_beam_size = int(pre_beam_ratio * beam_size)
self.beam_size = beam_size
self.n_vocab = vocab_size
if (
pre_beam_score_key is not None
and pre_beam_score_key != "full"
and pre_beam_score_key not in self.full_scorers
):
raise KeyError(f"{pre_beam_score_key} is not found in {self.full_scorers}")
self.pre_beam_score_key = pre_beam_score_key
self.do_pre_beam = (
self.pre_beam_score_key is not None
and self.pre_beam_size < self.n_vocab
and len(self.part_scorers) > 0
)
def init_hyp(self, x: torch.Tensor) -> List[Hypothesis]:
"""Get an initial hypothesis data.
Args:
x (torch.Tensor): The encoder output feature
Returns:
Hypothesis: The initial hypothesis.
"""
init_states = dict()
init_scores = dict()
for k, d in self.scorers.items():
init_states[k] = d.init_state(x)
init_scores[k] = 0.0
return [
Hypothesis(
score=0.0,
scores=init_scores,
states=init_states,
yseq=torch.tensor([self.sos], device=x.device),
)
]
@staticmethod
def append_token(xs: torch.Tensor, x: int) -> torch.Tensor:
"""Append new token to prefix tokens.
Args:
xs (torch.Tensor): The prefix token
x (int): The new token to append
Returns:
torch.Tensor: New tensor contains: xs + [x] with xs.dtype and xs.device
"""
x = torch.tensor([x], dtype=xs.dtype, device=xs.device)
return torch.cat((xs, x))
def score_full(
self,
hyp: Hypothesis,
x: torch.Tensor,
x_mask: torch.Tensor = None,
pre_acoustic_embeds: torch.Tensor = None,
) -> Tuple[Dict[str, torch.Tensor], Dict[str, Any]]:
"""Score new hypothesis by `self.full_scorers`.
Args:
hyp (Hypothesis): Hypothesis with prefix tokens to score
x (torch.Tensor): Corresponding input feature
Returns:
Tuple[Dict[str, torch.Tensor], Dict[str, Any]]: Tuple of
score dict of `hyp` that has string keys of `self.full_scorers`
and tensor score values of shape: `(self.n_vocab,)`,
and state dict that has string keys
and state values of `self.full_scorers`
"""
scores = dict()
states = dict()
for k, d in self.full_scorers.items():
scores[k], states[k] = d.score(
hyp.yseq,
hyp.states[k],
x,
x_mask=x_mask,
pre_acoustic_embeds=pre_acoustic_embeds,
)
return scores, states
def score_partial(
self, hyp: Hypothesis, ids: torch.Tensor, x: torch.Tensor
) -> Tuple[Dict[str, torch.Tensor], Dict[str, Any]]:
"""Score new hypothesis by `self.part_scorers`.
Args:
hyp (Hypothesis): Hypothesis with prefix tokens to score
ids (torch.Tensor): 1D tensor of new partial tokens to score
x (torch.Tensor): Corresponding input feature
Returns:
Tuple[Dict[str, torch.Tensor], Dict[str, Any]]: Tuple of
score dict of `hyp` that has string keys of `self.part_scorers`
and tensor score values of shape: `(len(ids),)`,
and state dict that has string keys
and state values of `self.part_scorers`
"""
scores = dict()
states = dict()
for k, d in self.part_scorers.items():
scores[k], states[k] = d.score_partial(hyp.yseq, ids, hyp.states[k], x)
return scores, states
def beam(
self, weighted_scores: torch.Tensor, ids: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Compute topk full token ids and partial token ids.
Args:
weighted_scores (torch.Tensor): The weighted sum scores for each tokens.
Its shape is `(self.n_vocab,)`.
ids (torch.Tensor): The partial token ids to compute topk
Returns:
Tuple[torch.Tensor, torch.Tensor]:
The topk full token ids and partial token ids.
Their shapes are `(self.beam_size,)`
"""
# no pre beam performed
if weighted_scores.size(0) == ids.size(0):
top_ids = weighted_scores.topk(self.beam_size)[1]
return top_ids, top_ids
# mask pruned in pre-beam not to select in topk
tmp = weighted_scores[ids]
weighted_scores[:] = -float("inf")
weighted_scores[ids] = tmp
top_ids = weighted_scores.topk(self.beam_size)[1]
local_ids = weighted_scores[ids].topk(self.beam_size)[1]
return top_ids, local_ids
@staticmethod
def merge_scores(
prev_scores: Dict[str, float],
next_full_scores: Dict[str, torch.Tensor],
full_idx: int,
next_part_scores: Dict[str, torch.Tensor],
part_idx: int,
) -> Dict[str, torch.Tensor]:
"""Merge scores for new hypothesis.
Args:
prev_scores (Dict[str, float]):
The previous hypothesis scores by `self.scorers`
next_full_scores (Dict[str, torch.Tensor]): scores by `self.full_scorers`
full_idx (int): The next token id for `next_full_scores`
next_part_scores (Dict[str, torch.Tensor]):
scores of partial tokens by `self.part_scorers`
part_idx (int): The new token id for `next_part_scores`
Returns:
Dict[str, torch.Tensor]: The new score dict.
Its keys are names of `self.full_scorers` and `self.part_scorers`.
Its values are scalar tensors by the scorers.
"""
new_scores = dict()
for k, v in next_full_scores.items():
new_scores[k] = prev_scores[k] + v[full_idx]
for k, v in next_part_scores.items():
new_scores[k] = prev_scores[k] + v[part_idx]
return new_scores
def merge_states(self, states: Any, part_states: Any, part_idx: int) -> Any:
"""Merge states for new hypothesis.
Args:
states: states of `self.full_scorers`
part_states: states of `self.part_scorers`
part_idx (int): The new token id for `part_scores`
Returns:
Dict[str, torch.Tensor]: The new score dict.
Its keys are names of `self.full_scorers` and `self.part_scorers`.
Its values are states of the scorers.
"""
new_states = dict()
for k, v in states.items():
new_states[k] = v
for k, d in self.part_scorers.items():
new_states[k] = d.select_state(part_states[k], part_idx)
return new_states
def search(
self,
running_hyps: List[Hypothesis],
x: torch.Tensor,
x_mask: torch.Tensor = None,
pre_acoustic_embeds: torch.Tensor = None,
) -> List[Hypothesis]:
"""Search new tokens for running hypotheses and encoded speech x.
Args:
running_hyps (List[Hypothesis]): Running hypotheses on beam
x (torch.Tensor): Encoded speech feature (T, D)
Returns:
List[Hypotheses]: Best sorted hypotheses
"""
best_hyps = []
part_ids = torch.arange(self.n_vocab, device=x.device) # no pre-beam
for hyp in running_hyps:
# scoring
weighted_scores = torch.zeros(self.n_vocab, dtype=x.dtype, device=x.device)
scores, states = self.score_full(
hyp, x, x_mask=x_mask, pre_acoustic_embeds=pre_acoustic_embeds
)
for k in self.full_scorers:
weighted_scores += self.weights[k] * scores[k]
# partial scoring
if self.do_pre_beam:
pre_beam_scores = (
weighted_scores
if self.pre_beam_score_key == "full"
else scores[self.pre_beam_score_key]
)
part_ids = torch.topk(pre_beam_scores, self.pre_beam_size)[1]
part_scores, part_states = self.score_partial(hyp, part_ids, x)
for k in self.part_scorers:
weighted_scores[part_ids] += self.weights[k] * part_scores[k]
# add previous hyp score
weighted_scores += hyp.score
# update hyps
for j, part_j in zip(*self.beam(weighted_scores, part_ids)):
# will be (2 x beam at most)
best_hyps.append(
Hypothesis(
score=weighted_scores[j],
yseq=self.append_token(hyp.yseq, j),
scores=self.merge_scores(
hyp.scores, scores, j, part_scores, part_j
),
states=self.merge_states(states, part_states, part_j),
)
)
# sort and prune 2 x beam -> beam
best_hyps = sorted(best_hyps, key=lambda x: x.score, reverse=True)[
: min(len(best_hyps), self.beam_size)
]
return best_hyps
def forward(
self,
x: torch.Tensor,
scama_mask: torch.Tensor = None,
pre_acoustic_embeds: torch.Tensor = None,
maxlenratio: float = 0.0,
minlenratio: float = 0.0,
maxlen: int = None,
minlen: int = 0,
) -> List[Hypothesis]:
"""Perform beam search.
Args:
x (torch.Tensor): Encoded speech feature (T, D)
maxlenratio (float): Input length ratio to obtain max output length.
If maxlenratio=0.0 (default), it uses a end-detect function
to automatically find maximum hypothesis lengths
If maxlenratio<0.0, its absolute value is interpreted
as a constant max output length.
minlenratio (float): Input length ratio to obtain min output length.
Returns:
list[Hypothesis]: N-best decoding results
"""
if maxlen is None:
# set length bounds
if maxlenratio == 0:
maxlen = x.shape[0]
elif maxlenratio < 0:
maxlen = -1 * int(maxlenratio)
else:
maxlen = max(1, int(maxlenratio * x.size(0)))
minlen = int(minlenratio * x.size(0))
logging.info("decoder input length: " + str(x.shape[0]))
logging.info("max output length: " + str(maxlen))
logging.info("min output length: " + str(minlen))
# main loop of prefix search
running_hyps = self.init_hyp(x)
ended_hyps = []
for i in range(maxlen):
logging.debug("position " + str(i))
mask_enc = None
if scama_mask is not None:
token_num_predictor = scama_mask.size(1)
token_id_slice = min(i, token_num_predictor - 1)
mask_enc = scama_mask[:, token_id_slice : token_id_slice + 1, :]
# if mask_enc.size(1) == 0:
# mask_enc = scama_mask[:, -2:-1, :]
# # mask_enc = torch.zeros_like(mask_enc)
pre_acoustic_embeds_cur = None
if pre_acoustic_embeds is not None:
b, t, d = pre_acoustic_embeds.size()
pad = torch.zeros((b, 1, d), dtype=pre_acoustic_embeds.dtype).to(
device=pre_acoustic_embeds.device
)
pre_acoustic_embeds = torch.cat((pre_acoustic_embeds, pad), dim=1)
token_id_slice = min(i, t)
pre_acoustic_embeds_cur = pre_acoustic_embeds[
:, token_id_slice : token_id_slice + 1, :
]
best = self.search(
running_hyps,
x,
x_mask=mask_enc,
pre_acoustic_embeds=pre_acoustic_embeds_cur,
)
# post process of one iteration
running_hyps = self.post_process(i, maxlen, maxlenratio, best, ended_hyps)
# end detection
if maxlenratio == 0.0 and end_detect([h.asdict() for h in ended_hyps], i):
logging.info(f"end detected at {i}")
break
if len(running_hyps) == 0:
logging.info("no hypothesis. Finish decoding.")
break
else:
logging.debug(f"remained hypotheses: {len(running_hyps)}")
nbest_hyps = sorted(ended_hyps, key=lambda x: x.score, reverse=True)
# check the number of hypotheses reaching to eos
if len(nbest_hyps) == 0:
logging.warning(
"there is no N-best results, perform recognition "
"again with smaller minlenratio."
)
return (
[]
if minlenratio < 0.1
else self.forward(x, maxlenratio, max(0.0, minlenratio - 0.1))
)
# report the best result
for x in nbest_hyps:
yseq = "".join([self.token_list[x] for x in x.yseq])
logging.debug(
"nbest: y: {}, yseq: {}, score: {}".format(x.yseq, yseq, x.score)
)
best = nbest_hyps[0]
for k, v in best.scores.items():
logging.info(
f"{v:6.2f} * {self.weights[k]:3} = {v * self.weights[k]:6.2f} for {k}"
)
logging.info(f"total log probability: {best.score:.2f}")
logging.info(f"normalized log probability: {best.score / len(best.yseq):.2f}")
logging.info(f"total number of ended hypotheses: {len(nbest_hyps)}")
if self.token_list is not None:
logging.info(
"best hypo: "
+ "".join([self.token_list[x] for x in best.yseq[1:-1]])
+ "\n"
)
return nbest_hyps
def post_process(
self,
i: int,
maxlen: int,
maxlenratio: float,
running_hyps: List[Hypothesis],
ended_hyps: List[Hypothesis],
) -> List[Hypothesis]:
"""Perform post-processing of beam search iterations.
Args:
i (int): The length of hypothesis tokens.
maxlen (int): The maximum length of tokens in beam search.
maxlenratio (int): The maximum length ratio in beam search.
running_hyps (List[Hypothesis]): The running hypotheses in beam search.
ended_hyps (List[Hypothesis]): The ended hypotheses in beam search.
Returns:
List[Hypothesis]: The new running hypotheses.
"""
logging.debug(f"the number of running hypotheses: {len(running_hyps)}")
if self.token_list is not None:
logging.debug(
"best hypo: "
+ "".join([self.token_list[x] for x in running_hyps[0].yseq[1:]])
)
# add eos in the final loop to avoid that there are no ended hyps
if i == maxlen - 1:
logging.info("adding <eos> in the last position in the loop")
running_hyps = [
h._replace(yseq=self.append_token(h.yseq, self.eos))
for h in running_hyps
]
# add ended hypotheses to a final list, and removed them from current hypotheses
# (this will be a problem, number of hyps < beam)
remained_hyps = []
for hyp in running_hyps:
if hyp.yseq[-1] == self.eos:
# e.g., Word LM needs to add final <eos> score
for k, d in chain(self.full_scorers.items(), self.part_scorers.items()):
s = d.final_score(hyp.states[k])
hyp.scores[k] += s
hyp = hyp._replace(score=hyp.score + self.weights[k] * s)
ended_hyps.append(hyp)
else:
remained_hyps.append(hyp)
return remained_hyps
class BeamSearchScamaStreaming(torch.nn.Module):
"""Beam search implementation."""
def __init__(
self,
scorers: Dict[str, ScorerInterface],
weights: Dict[str, float],
beam_size: int,
vocab_size: int,
sos: int,
eos: int,
token_list: List[str] = None,
pre_beam_ratio: float = 1.5,
pre_beam_score_key: str = None,
):
"""Initialize beam search.
Args:
scorers (dict[str, ScorerInterface]): Dict of decoder modules
e.g., Decoder, CTCPrefixScorer, LM
The scorer will be ignored if it is `None`
weights (dict[str, float]): Dict of weights for each scorers
The scorer will be ignored if its weight is 0
beam_size (int): The number of hypotheses kept during search
vocab_size (int): The number of vocabulary
sos (int): Start of sequence id
eos (int): End of sequence id
token_list (list[str]): List of tokens for debug log
pre_beam_score_key (str): key of scores to perform pre-beam search
pre_beam_ratio (float): beam size in the pre-beam search
will be `int(pre_beam_ratio * beam_size)`
"""
super().__init__()
# set scorers
self.weights = weights
self.scorers = dict()
self.full_scorers = dict()
self.part_scorers = dict()
# this module dict is required for recursive cast
# `self.to(device, dtype)` in `recog.py`
self.nn_dict = torch.nn.ModuleDict()
for k, v in scorers.items():
w = weights.get(k, 0)
if w == 0 or v is None:
continue
assert isinstance(
v, ScorerInterface
), f"{k} ({type(v)}) does not implement ScorerInterface"
self.scorers[k] = v
if isinstance(v, PartialScorerInterface):
self.part_scorers[k] = v
else:
self.full_scorers[k] = v
if isinstance(v, torch.nn.Module):
self.nn_dict[k] = v
# set configurations
self.sos = sos
self.eos = eos
self.token_list = token_list
self.pre_beam_size = int(pre_beam_ratio * beam_size)
self.beam_size = beam_size
self.n_vocab = vocab_size
if (
pre_beam_score_key is not None
and pre_beam_score_key != "full"
and pre_beam_score_key not in self.full_scorers
):
raise KeyError(f"{pre_beam_score_key} is not found in {self.full_scorers}")
self.pre_beam_score_key = pre_beam_score_key
self.do_pre_beam = (
self.pre_beam_score_key is not None
and self.pre_beam_size < self.n_vocab
and len(self.part_scorers) > 0
)
def init_hyp(self, x) -> List[Hypothesis]:
"""Get an initial hypothesis data.
Args:
x (torch.Tensor): The encoder output feature
Returns:
Hypothesis: The initial hypothesis.
"""
init_states = dict()
init_scores = dict()
for k, d in self.scorers.items():
init_states[k] = d.init_state(x)
init_scores[k] = 0.0
return [
Hypothesis(
score=0.0,
scores=init_scores,
states=init_states,
yseq=torch.tensor([self.sos], device=x.device),
)
]
@staticmethod
def append_token(xs: torch.Tensor, x: int) -> torch.Tensor:
"""Append new token to prefix tokens.
Args:
xs (torch.Tensor): The prefix token
x (int): The new token to append
Returns:
torch.Tensor: New tensor contains: xs + [x] with xs.dtype and xs.device
"""
x = torch.tensor([x], dtype=xs.dtype, device=xs.device)
return torch.cat((xs, x))
def score_full(
self,
hyp: Hypothesis,
x: torch.Tensor,
x_mask: torch.Tensor = None,
pre_acoustic_embeds: torch.Tensor = None,
cache: dict = {},
) -> Tuple[Dict[str, torch.Tensor], Dict[str, Any]]:
"""Score new hypothesis by `self.full_scorers`.
Args:
hyp (Hypothesis): Hypothesis with prefix tokens to score
x (torch.Tensor): Corresponding input feature
Returns:
Tuple[Dict[str, torch.Tensor], Dict[str, Any]]: Tuple of
score dict of `hyp` that has string keys of `self.full_scorers`
and tensor score values of shape: `(self.n_vocab,)`,
and state dict that has string keys
and state values of `self.full_scorers`
"""
scores = dict()
states = dict()
for k, d in self.full_scorers.items():
scores[k], states[k] = d.score(
hyp.yseq,
hyp.states[k],
x,
x_mask=x_mask,
pre_acoustic_embeds=pre_acoustic_embeds,
cache=cache,
)
return scores, states
def score_partial(
self, hyp: Hypothesis, ids: torch.Tensor, x: torch.Tensor
) -> Tuple[Dict[str, torch.Tensor], Dict[str, Any]]:
"""Score new hypothesis by `self.part_scorers`.
Args:
hyp (Hypothesis): Hypothesis with prefix tokens to score
ids (torch.Tensor): 1D tensor of new partial tokens to score
x (torch.Tensor): Corresponding input feature
Returns:
Tuple[Dict[str, torch.Tensor], Dict[str, Any]]: Tuple of
score dict of `hyp` that has string keys of `self.part_scorers`
and tensor score values of shape: `(len(ids),)`,
and state dict that has string keys
and state values of `self.part_scorers`
"""
scores = dict()
states = dict()
for k, d in self.part_scorers.items():
scores[k], states[k] = d.score_partial(hyp.yseq, ids, hyp.states[k], x)
return scores, states
def beam(
self, weighted_scores: torch.Tensor, ids: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Compute topk full token ids and partial token ids.
Args:
weighted_scores (torch.Tensor): The weighted sum scores for each tokens.
Its shape is `(self.n_vocab,)`.
ids (torch.Tensor): The partial token ids to compute topk
Returns:
Tuple[torch.Tensor, torch.Tensor]:
The topk full token ids and partial token ids.
Their shapes are `(self.beam_size,)`
"""
# no pre beam performed
if weighted_scores.size(0) == ids.size(0):
top_ids = weighted_scores.topk(self.beam_size)[1]
return top_ids, top_ids
# mask pruned in pre-beam not to select in topk
tmp = weighted_scores[ids]
weighted_scores[:] = -float("inf")
weighted_scores[ids] = tmp
top_ids = weighted_scores.topk(self.beam_size)[1]
local_ids = weighted_scores[ids].topk(self.beam_size)[1]
return top_ids, local_ids
@staticmethod
def merge_scores(
prev_scores: Dict[str, float],
next_full_scores: Dict[str, torch.Tensor],
full_idx: int,
next_part_scores: Dict[str, torch.Tensor],
part_idx: int,
) -> Dict[str, torch.Tensor]:
"""Merge scores for new hypothesis.
Args:
prev_scores (Dict[str, float]):
The previous hypothesis scores by `self.scorers`
next_full_scores (Dict[str, torch.Tensor]): scores by `self.full_scorers`
full_idx (int): The next token id for `next_full_scores`
next_part_scores (Dict[str, torch.Tensor]):
scores of partial tokens by `self.part_scorers`
part_idx (int): The new token id for `next_part_scores`
Returns:
Dict[str, torch.Tensor]: The new score dict.
Its keys are names of `self.full_scorers` and `self.part_scorers`.
Its values are scalar tensors by the scorers.
"""
new_scores = dict()
for k, v in next_full_scores.items():
new_scores[k] = prev_scores[k] + v[full_idx]
for k, v in next_part_scores.items():
new_scores[k] = prev_scores[k] + v[part_idx]
return new_scores
def merge_states(self, states: Any, part_states: Any, part_idx: int) -> Any:
"""Merge states for new hypothesis.
Args:
states: states of `self.full_scorers`
part_states: states of `self.part_scorers`
part_idx (int): The new token id for `part_scores`
Returns:
Dict[str, torch.Tensor]: The new score dict.
Its keys are names of `self.full_scorers` and `self.part_scorers`.
Its values are states of the scorers.
"""
new_states = dict()
for k, v in states.items():
new_states[k] = v
for k, d in self.part_scorers.items():
new_states[k] = d.select_state(part_states[k], part_idx)
return new_states
def search(
self,
running_hyps: List[Hypothesis],
x: torch.Tensor,
x_mask: torch.Tensor = None,
pre_acoustic_embeds: torch.Tensor = None,
cache: dict = {},
) -> List[Hypothesis]:
"""Search new tokens for running hypotheses and encoded speech x.
Args:
running_hyps (List[Hypothesis]): Running hypotheses on beam
x (torch.Tensor): Encoded speech feature (T, D)
Returns:
List[Hypotheses]: Best sorted hypotheses
"""
best_hyps = []
part_ids = torch.arange(self.n_vocab, device=x.device) # no pre-beam
for hyp in running_hyps:
# scoring
weighted_scores = torch.zeros(self.n_vocab, dtype=x.dtype, device=x.device)
scores, states = self.score_full(
hyp,
x,
x_mask=x_mask,
pre_acoustic_embeds=pre_acoustic_embeds,
cache=cache,
)
for k in self.full_scorers:
weighted_scores += self.weights[k] * scores[k]
# partial scoring
if self.do_pre_beam:
pre_beam_scores = (
weighted_scores
if self.pre_beam_score_key == "full"
else scores[self.pre_beam_score_key]
)
part_ids = torch.topk(pre_beam_scores, self.pre_beam_size)[1]
part_scores, part_states = self.score_partial(hyp, part_ids, x)
for k in self.part_scorers:
weighted_scores[part_ids] += self.weights[k] * part_scores[k]
# add previous hyp score
weighted_scores += hyp.score
# update hyps
for j, part_j in zip(*self.beam(weighted_scores, part_ids)):
# will be (2 x beam at most)
best_hyps.append(
Hypothesis(
score=weighted_scores[j],
yseq=self.append_token(hyp.yseq, j),
scores=self.merge_scores(
hyp.scores, scores, j, part_scores, part_j
),
states=self.merge_states(states, part_states, part_j),
)
)
# sort and prune 2 x beam -> beam
best_hyps = sorted(best_hyps, key=lambda x: x.score, reverse=True)[
: min(len(best_hyps), self.beam_size)
]
return best_hyps
def forward(
self,
x: torch.Tensor,
scama_mask: torch.Tensor = None,
pre_acoustic_embeds: torch.Tensor = None,
maxlenratio: float = 0.0,
minlenratio: float = 0.0,
maxlen: int = None,
minlen: int = 0,
cache: dict = {},
) -> List[Hypothesis]:
"""Perform beam search.
Args:
x (torch.Tensor): Encoded speech feature (T, D)
maxlenratio (float): Input length ratio to obtain max output length.
If maxlenratio=0.0 (default), it uses a end-detect function
to automatically find maximum hypothesis lengths
If maxlenratio<0.0, its absolute value is interpreted
as a constant max output length.
minlenratio (float): Input length ratio to obtain min output length.
Returns:
list[Hypothesis]: N-best decoding results
"""
if maxlen is None:
# set length bounds
if maxlenratio == 0:
maxlen = x.shape[0]
elif maxlenratio < 0:
maxlen = -1 * int(maxlenratio)
else:
maxlen = max(1, int(maxlenratio * x.size(0)))
minlen = int(minlenratio * x.size(0))
logging.info("decoder input length: " + str(x.shape[0]))
logging.info("max output length: " + str(maxlen))
logging.info("min output length: " + str(minlen))
# main loop of prefix search
# running_hyps = self.init_hyp(x)
running_hyps = cache["running_hyps"]
ended_hyps = []
for i in range(maxlen):
logging.debug("position " + str(i))
mask_enc = None
# if scama_mask is not None:
# token_num_predictor = scama_mask.size(1)
# token_id_slice = min(i, token_num_predictor-1)
# mask_enc = scama_mask[:, token_id_slice:token_id_slice+1, :]
# # if mask_enc.size(1) == 0:
# # mask_enc = scama_mask[:, -2:-1, :]
# # # mask_enc = torch.zeros_like(mask_enc)
pre_acoustic_embeds_cur = None
if pre_acoustic_embeds is not None:
b, t, d = pre_acoustic_embeds.size()
pad = torch.zeros((b, 1, d), dtype=pre_acoustic_embeds.dtype).to(
device=pre_acoustic_embeds.device
)
pre_acoustic_embeds = torch.cat((pre_acoustic_embeds, pad), dim=1)
token_id_slice = min(i, t)
pre_acoustic_embeds_cur = pre_acoustic_embeds[
:, token_id_slice : token_id_slice + 1, :
]
best = self.search(
running_hyps,
x,
x_mask=mask_enc,
pre_acoustic_embeds=pre_acoustic_embeds_cur,
cache=cache["decoder"],
)
# post process of one iteration
running_hyps = self.post_process(i, maxlen, maxlenratio, best, ended_hyps)
# end detection
if maxlenratio == 0.0 and end_detect([h.asdict() for h in ended_hyps], i):
logging.info(f"end detected at {i}")
break
if len(running_hyps) == 0:
logging.info("no hypothesis. Finish decoding.")
break
else:
logging.debug(f"remained hypotheses: {len(running_hyps)}")
nbest_hyps = sorted(ended_hyps, key=lambda x: x.score, reverse=True)
# check the number of hypotheses reaching to eos
if len(nbest_hyps) == 0:
logging.warning(
"there is no N-best results, perform recognition "
"again with smaller minlenratio."
)
return (
[]
if minlenratio < 0.1
else self.forward(x, maxlenratio, max(0.0, minlenratio - 0.1))
)
# report the best result
for x in nbest_hyps:
yseq = "".join([self.token_list[x] for x in x.yseq])
logging.debug(
"nbest: y: {}, yseq: {}, score: {}".format(x.yseq, yseq, x.score)
)
best = nbest_hyps[0]
for k, v in best.scores.items():
logging.info(
f"{v:6.2f} * {self.weights[k]:3} = {v * self.weights[k]:6.2f} for {k}"
)
logging.info(f"total log probability: {best.score:.2f}")
logging.info(f"normalized log probability: {best.score / len(best.yseq):.2f}")
logging.info(f"total number of ended hypotheses: {len(nbest_hyps)}")
if self.token_list is not None:
logging.info(
"best hypo: "
+ "".join([self.token_list[x] for x in best.yseq[1:-1]])
+ "\n"
)
return nbest_hyps
def post_process(
self,
i: int,
maxlen: int,
maxlenratio: float,
running_hyps: List[Hypothesis],
ended_hyps: List[Hypothesis],
) -> List[Hypothesis]:
"""Perform post-processing of beam search iterations.
Args:
i (int): The length of hypothesis tokens.
maxlen (int): The maximum length of tokens in beam search.
maxlenratio (int): The maximum length ratio in beam search.
running_hyps (List[Hypothesis]): The running hypotheses in beam search.
ended_hyps (List[Hypothesis]): The ended hypotheses in beam search.
Returns:
List[Hypothesis]: The new running hypotheses.
"""
logging.debug(f"the number of running hypotheses: {len(running_hyps)}")
if self.token_list is not None:
logging.debug(
"best hypo: "
+ "".join([self.token_list[x] for x in running_hyps[0].yseq[1:]])
)
# add eos in the final loop to avoid that there are no ended hyps
if i == maxlen - 1:
logging.info("adding <eos> in the last position in the loop")
running_hyps = [
h._replace(yseq=self.append_token(h.yseq, self.eos))
for h in running_hyps
]
# add ended hypotheses to a final list, and removed them from current hypotheses
# (this will be a problem, number of hyps < beam)
remained_hyps = []
for hyp in running_hyps:
if hyp.yseq[-1] == self.eos:
# e.g., Word LM needs to add final <eos> score
for k, d in chain(self.full_scorers.items(), self.part_scorers.items()):
s = d.final_score(hyp.states[k])
hyp.scores[k] += s
hyp = hyp._replace(score=hyp.score + self.weights[k] * s)
ended_hyps.append(hyp)
else:
remained_hyps.append(hyp)
return remained_hyps
|