-
Notifications
You must be signed in to change notification settings - Fork 40
/
Copy pathcpmbee_translator.py
202 lines (177 loc) · 7.64 KB
/
cpmbee_translator.py
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
from typing import Dict
from cpm_live.generation.bee import CPMBeeBeamSearch
from cpm_live.models import CPMBeeTorch, CPMBeeConfig
from cpm_live.tokenizers import CPMBeeTokenizer
import torch
import spacy
import re
def is_chinese(ch: str):
if "\u4e00" <= ch <= "\u9fff":
return True
return False
def is_english(ch: str):
return ch.isalpha()
class Translator:
def __init__(self, ckpt_path, batch_size=8):
config = CPMBeeConfig.from_json_file("config/cpm-bee-10b.json")
self.tokenizer = CPMBeeTokenizer()
model = CPMBeeTorch(config=config)
model.load_state_dict(torch.load(ckpt_path))
model.cuda()
self._beam_search = CPMBeeBeamSearch(
model=model,
tokenizer=self.tokenizer,
)
self._batch_size = batch_size
self._nlp_eng = spacy.load("en_core_web_trf")
# self._nlp_chn = spacy.load("zh_core_web_trf")
def _auto_cut(self, text: str):
CUT_TABLE = {
".": 100,
"?": 100,
"!": 100,
"。": 48,
"？": 48,
"！": 48,
}
st = 0
sub_text = []
while st < len(text):
ed = st
while ed + 1 < len(text) and (
text[ed] not in CUT_TABLE or ed < st + CUT_TABLE[text[ed]]
):
ed += 1
sub_text.append(text[st : ed + 1])
st = ed + 1
return sub_text
def _remove_entity(self, nlp: spacy.language.Language, text: str):
doc = nlp(text)
ent_spans = []
for ent in doc.ents:
if ent.label_ in ["PRODUCT"]:
ent_spans.append((ent.start_char, ent.end_char))
sorted(ent_spans, key=lambda x: x[0])
sub_text = []
ent_map = {}
unk_map = {}
p = 0
for ent_s, ent_e in ent_spans:
sub_text.append(self.tokenizer.escape(text[p:ent_s]))
ent = text[ent_s:ent_e]
if ent not in ent_map:
ent_map[ent] = len(ent_map)
unk_map["<unk_{}>".format(ent_map[ent])] = ent
sub_text.append("<unk_{}>".format(ent_map[ent]))
p = ent_e
sub_text.append(self.tokenizer.escape(text[p:]))
return "".join(sub_text), unk_map
def _replace_entity(self, text: str, table: Dict[str, str]):
ret = []
for token in self.tokenizer.tokenize(text):
if token.is_special and token.token in table:
t = token.token
if t.startswith("the "):
t = t[4:]
ret.append(table[token.token])
else:
ret.append(token.token)
return "".join(ret)
def to_chn(self, text: str) -> str:
text, replace_table = self._remove_entity(self._nlp_eng, text)
sub_text = []
for line in text.split("\n"):
sub_text.extend(self._auto_cut(line))
sub_text.append("")
ret = ["\n" for _ in range(len(sub_text))]
curr_batch = []
curr_batch_idx = []
for i, t in enumerate(sub_text):
if len(t) == 0:
ret[i] = "\n"
else:
curr_batch.append(t)
curr_batch_idx.append(i)
if len(curr_batch) >= self._batch_size:
inference_results = self._beam_search.generate(
[{"document": doc, "task": "英翻中", "<ans>": ""} for doc in curr_batch],
max_length=180,
repetition_penalty=1.0,
)
for idx, res in zip(curr_batch_idx, inference_results):
ret[idx] = self._replace_entity(res["<ans>"], replace_table)
curr_batch = []
curr_batch_idx = []
if len(curr_batch) > 0:
inference_results = self._beam_search.generate(
[{"document": doc, "task": "英翻中", "<ans>": ""} for doc in curr_batch],
max_length=180,
repetition_penalty=1.0,
)
for idx, res in zip(curr_batch_idx, inference_results):
ret[idx] = self._replace_entity(res["<ans>"], replace_table)
curr_batch = []
curr_batch_idx = []
return "".join(ret)
def to_eng(self, text: str):
text = self.tokenizer.escape(text)
text = re.sub(r"([^\x00-\x7F])([a-zA-Z])", r"\1 \2", text)
sub_text = []
for line in text.split("\n"):
sub_text.extend(self._auto_cut(line))
sub_text.append("")
ret = ["\n" for _ in range(len(sub_text))]
curr_batch = []
curr_batch_idx = []
for i, t in enumerate(sub_text):
if len(t) == 0:
ret[i] = "\n"
else:
curr_batch.append(t)
curr_batch_idx.append(i)
if len(curr_batch) >= self._batch_size:
inference_results = self._beam_search.generate(
[{"document": doc, "task": "中翻英", "<ans>": ""} for doc in curr_batch],
max_length=180,
repetition_penalty=1.0,
)
for idx, res in zip(curr_batch_idx, inference_results):
ret[idx] = self.tokenizer.unescape(res["<ans>"])
curr_batch = []
curr_batch_idx = []
if len(curr_batch) > 0:
inference_results = self._beam_search.generate(
[{"document": doc, "task": "中翻英", "<ans>": ""} for doc in curr_batch],
max_length=180,
repetition_penalty=1.0,
)
for idx, res in zip(curr_batch_idx, inference_results):
ret[idx] = self.tokenizer.unescape(res["<ans>"])
curr_batch = []
curr_batch_idx = []
is_newline = True
for i in range(len(ret)):
if ret[i] == "\n":
is_newline = True
elif is_newline:
is_newline = False
else:
ret[i] = " " + ret[i]
return "".join(ret)
def main():
translator = Translator("path/to/model")
print(
translator.to_eng(
"""考虑到机器学习模型的“黑盒”本质，模型有可能在不受控的情况下输出包括但不限于虚假信息、错误政治言论、偏见与歧视性话语、对不良行为的煽动与暗示等内容。CPM-Live虽已对相关训练数据进行数据清洗，但仍有可能具有不限于如下所示使用风险。用户使用CPM-Live相关资源前，需明确本节涉及的相关风险，并在使用过程中承担全部风险与责任。
侵犯个人隐私。模型有可能直接或经引导后产生涉及个人隐私的内容。
侵犯内容版权。模型有可能直接或经引导后产生与其他出版物相同、相似的内容。
产生虚假信息。模型有可能直接或经引导后产生不符合事实或客观规律的虚假信息。用户不应故意使用与引导模型制作虚假内容。
产生政治敏感内容。模型有可能直接或经引导后产生与政策、法规等相关的政治敏感内容。
产生偏见与歧视性话语。模型有可能直接或经引导后产生包括但不限于性别、种族等方面的偏见与歧视性话语。
产生对不良行为的煽动与暗示。模型有可能直接或经引导后产生对于违法犯罪等不良行为的煽动与暗示。
产生个体伤害言论。模型有可能直接或经引导后产生对个体进行伤害的言论，如对个人的诋毁、打击言论或鼓励个体进行自我伤害行为的言论等。
"""
)
)
if __name__ == "__main__":
main()