You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
PaddleSpeech/paddlespeech/server/engine/asr/online/ctc_search.py

209 lines
8.4 KiB

# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import copy
from collections import defaultdict
3 years ago
import paddle
3 years ago
from paddlespeech.cli.log import logger
from paddlespeech.s2t.utils.utility import log_add
__all__ = ['CTCPrefixBeamSearch']
class CTCPrefixBeamSearch:
def __init__(self, config):
"""Implement the ctc prefix beam search
Args:
config (yacs.config.CfgNode): the ctc prefix beam search configuration
"""
self.config = config
# beam size
self.first_beam_size = self.config.beam_size
# TODO(support second beam size)
self.second_beam_size = int(self.first_beam_size * 1.0)
logger.info(
f"first and second beam size: {self.first_beam_size}, {self.second_beam_size}"
)
# state
self.cur_hyps = None
self.hyps = None
self.abs_time_step = 0
self.reset()
def reset(self):
"""Rest the search cache value
"""
self.cur_hyps = None
self.hyps = None
self.abs_time_step = 0
@paddle.no_grad()
def search(self, ctc_probs, device, blank_id=0):
"""ctc prefix beam search method decode a chunk feature
Args:
xs (paddle.Tensor): feature data
ctc_probs (paddle.Tensor): the ctc probability of all the tokens
device (paddle.fluid.core_avx.Place): the feature host device, such as CUDAPlace(0).
blank_id (int, optional): the blank id in the vocab. Defaults to 0.
Returns:
list: the search result
"""
# decode
logger.info("start to ctc prefix search")
assert len(ctc_probs.shape) == 2
batch_size = 1
vocab_size = ctc_probs.shape[1]
first_beam_size = min(self.first_beam_size, vocab_size)
second_beam_size = min(self.second_beam_size, vocab_size)
logger.info(
f"effect first and second beam size: {self.first_beam_size}, {self.second_beam_size}"
)
maxlen = ctc_probs.shape[0]
# cur_hyps: (prefix, (blank_ending_score, none_blank_ending_score))
# 0. blank_ending_score,
# 1. none_blank_ending_score,
# 2. viterbi_blank ending score,
# 3. viterbi_non_blank score,
# 4. current_token_prob,
# 5. times_viterbi_blank, times_b
# 6. times_titerbi_non_blank, times_nb
if self.cur_hyps is None:
self.cur_hyps = [(tuple(), (0.0, -float('inf'), 0.0, 0.0,
-float('inf'), [], []))]
# self.cur_hyps = [(tuple(), (0.0, -float('inf')))]
# 2. CTC beam search step by step
for t in range(0, maxlen):
logp = ctc_probs[t] # (vocab_size,)
# next_hyps = defaultdict(lambda: (-float('inf'), -float('inf')))
next_hyps = defaultdict(
lambda: (-float('inf'), -float('inf'), -float('inf'), -float('inf'), -float('inf'), [], []))
# 2.1 First beam prune: select topk best
# do token passing process
top_k_logp, top_k_index = logp.topk(
first_beam_size) # (first_beam_size,)
for s in top_k_index:
s = s.item()
ps = logp[s].item()
for prefix, (pb, pnb, v_b_s, v_nb_s, cur_token_prob, times_b,
times_nb) in self.cur_hyps:
last = prefix[-1] if len(prefix) > 0 else None
if s == blank_id: # blank
n_pb, n_pnb, n_v_b, n_v_nb, n_cur_token_prob, n_times_b, n_times_nb = next_hyps[
prefix]
n_pb = log_add([n_pb, pb + ps, pnb + ps])
pre_times = times_b if v_b_s > v_nb_s else times_nb
n_times_b = copy.deepcopy(pre_times)
viterbi_score = v_b_s if v_b_s > v_nb_s else v_nb_s
n_v_b = viterbi_score + ps
next_hyps[prefix] = (n_pb, n_pnb, n_v_b, n_v_nb,
n_cur_token_prob, n_times_b,
n_times_nb)
elif s == last:
# Update *ss -> *s;
# case1: *a + a => *a
n_pb, n_pnb, n_v_b, n_v_nb, n_cur_token_prob, n_times_b, n_times_nb = next_hyps[
prefix]
n_pnb = log_add([n_pnb, pnb + ps])
if n_v_nb < v_nb_s + ps:
n_v_nb = v_nb_s + ps
if n_cur_token_prob < ps:
n_cur_token_prob = ps
n_times_nb = copy.deepcopy(times_nb)
n_times_nb[
-1] = self.abs_time_step # 注意,这里要重新使用绝对时间
next_hyps[prefix] = (n_pb, n_pnb, n_v_b, n_v_nb,
n_cur_token_prob, n_times_b,
n_times_nb)
# Update *s-s -> *ss, - is for blank
# Case 2: *aε + a => *aa
n_prefix = prefix + (s, )
n_pb, n_pnb, n_v_b, n_v_nb, n_cur_token_prob, n_times_b, n_times_nb = next_hyps[
n_prefix]
if n_v_nb < v_b_s + ps:
n_v_nb = v_b_s + ps
n_cur_token_prob = ps
n_times_nb = copy.deepcopy(times_b)
n_times_nb.append(self.abs_time_step)
n_pnb = log_add([n_pnb, pb + ps])
next_hyps[n_prefix] = (n_pb, n_pnb, n_v_b, n_v_nb,
n_cur_token_prob, n_times_b,
n_times_nb)
else:
# Case 3: *a + b => *ab, *aε + b => *ab
n_prefix = prefix + (s, )
n_pb, n_pnb, n_v_b, n_v_nb, n_cur_token_prob, n_times_b, n_times_nb = next_hyps[
n_prefix]
viterbi_score = v_b_s if v_b_s > v_nb_s else v_nb_s
pre_times = times_b if v_b_s > v_nb_s else times_nb
if n_v_nb < viterbi_score + ps:
n_v_nb = viterbi_score + ps
n_cur_token_prob = ps
n_times_nb = copy.deepcopy(pre_times)
n_times_nb.append(self.abs_time_step)
n_pnb = log_add([n_pnb, pb + ps, pnb + ps])
next_hyps[n_prefix] = (n_pb, n_pnb, n_v_b, n_v_nb,
n_cur_token_prob, n_times_b,
n_times_nb)
# 2.2 Second beam prune
next_hyps = sorted(
next_hyps.items(),
key=lambda x: log_add([x[1][0], x[1][1]]),
reverse=True)
self.cur_hyps = next_hyps[:second_beam_size]
# 2.3 update the absolute time step
self.abs_time_step += 1
self.hyps = [(y[0], log_add([y[1][0], y[1][1]]), y[1][2], y[1][3],
y[1][4], y[1][5], y[1][6]) for y in self.cur_hyps]
logger.info("ctc prefix search success")
return self.hyps
def get_one_best_hyps(self):
"""Return the one best result
Returns:
list: the one best result, List[str]
"""
return [self.hyps[0][0]]
def get_hyps(self):
"""Return the search hyps
Returns:
list: return the search hyps, List[Tuple[str, float, ...]]
"""
return self.hyps
def finalize_search(self):
"""do nothing in ctc_prefix_beam_search
"""
pass