From 1cdb3cc28d4d89a576cc06e5cd8eb80da1f3a3aa Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 26 四月 2024 11:27:39 +0800
Subject: [PATCH] Dev gzf exp (#1665)
---
funasr/datasets/large_datasets/utils/tokenize.py | 52 ++++++++++++++++++++++++++++++++++++----------------
1 files changed, 36 insertions(+), 16 deletions(-)
diff --git a/funasr/datasets/large_datasets/utils/tokenize.py b/funasr/datasets/large_datasets/utils/tokenize.py
index d8ceff2..5a1ddd2 100644
--- a/funasr/datasets/large_datasets/utils/tokenize.py
+++ b/funasr/datasets/large_datasets/utils/tokenize.py
@@ -1,6 +1,8 @@
#!/usr/bin/env python
import re
import numpy as np
+from funasr.datasets.large_datasets.utils.hotword_utils import sample_hotword
+
def forward_segment(text, seg_dict):
word_list = []
@@ -16,39 +18,57 @@
i += len(longest_word)
return word_list
+
def seg_tokenize(txt, seg_dict):
+ pattern = re.compile(r"^[\u4E00-\u9FA50-9]+$")
out_txt = ""
for word in txt:
+ word = word.lower()
if word in seg_dict:
out_txt += seg_dict[word] + " "
else:
- out_txt += "<unk>" + " "
+ if pattern.match(word):
+ for char in word:
+ if char in seg_dict:
+ out_txt += seg_dict[char] + " "
+ else:
+ out_txt += "<unk>" + " "
+ else:
+ out_txt += "<unk>" + " "
return out_txt.strip().split()
-def tokenize(data,
- vocab=None,
- seg_dict=None,
- punc_dict=None,
- bpe_tokenizer=None):
+
+def tokenize(data, vocab=None, seg_dict=None, punc_dict=None, bpe_tokenizer=None, hw_config=None):
assert "text" in data
assert isinstance(vocab, dict)
text = data["text"]
token = []
vad = -2
-
if bpe_tokenizer is not None:
- text = bpe_tokenizer.text2tokens("".join(text))
-
+ text = bpe_tokenizer.text2tokens(" ".join(text))
if seg_dict is not None:
assert isinstance(seg_dict, dict)
- txt = forward_segment("".join(text).lower(), seg_dict)
- text = seg_tokenize(txt, seg_dict)
+ text = seg_tokenize(text, seg_dict)
length = len(text)
+ if "hw_tag" in data:
+ pre_index = None
+ if hw_config["pre_hwlist"] is not None and hw_config["pre_prob"] > 0:
+ # enable preset hotword detect in sampling
+ for hw in hw_config["pre_hwlist"]:
+ hw = " ".join(seg_tokenize(hw, seg_dict))
+ _find = " ".join(text).find(hw)
+ if _find != -1:
+ # _find = text[:_find].count(" ") # bpe sometimes
+ pre_index = [_find, _find + max(hw.count(" "), 1)]
+ break
+ hotword_indxs = sample_hotword(length, **hw_config, pre_index=pre_index)
+ data["hotword_indxs"] = hotword_indxs
+ del data["hw_tag"]
for i in range(length):
x = text[i]
- if i == length-1 and "punc" in data and text[i].startswith("vad:"):
- vad = x[-1][4:]
+ if i == length - 1 and "punc" in data and x.startswith("vad:"):
+ vad = x[4:]
if len(vad) == 0:
vad = -1
else:
@@ -56,7 +76,7 @@
elif x in vocab:
token.append(vocab[x])
else:
- token.append(vocab['<unk>'])
+ token.append(vocab["<unk>"])
if "punc" in data and punc_dict is not None:
punc_token = []
@@ -65,9 +85,9 @@
punc_token.append(punc_dict[punc])
else:
punc_token.append(punc_dict["_"])
- data["punc"] = np.array(punc_token)
+ data["punc"] = np.array(punc_token)
data["text"] = np.array(token)
if vad is not -2:
- data["vad_indexes"]=np.array([vad], dtype=np.int64)
+ data["vad_indexes"] = np.array([vad], dtype=np.int64)
return data
--
Gitblit v1.9.1