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
| import os
import regex
from typing import Iterable, Iterator
from collections import Counter
from collections import deque
import pickle
class tokenizer:
def __init__(self,vocab,merges,special_tokens=None):
self.merges=merges
if special_tokens:
vocab_reverse={byte_token:token_id for token_id,byte_token in vocab.items()}
for special_token in set(special_tokens):
byte_special_token= special_token.encode('utf-8')
if byte_special_token not in vocab_reverse:
vocab[len(vocab)]=byte_special_token
self.vocab=vocab
self.vocab_reverse={byte_token:token_id for token_id,byte_token in self.vocab.items()}
self.pair2id={pair:i for i,pair in enumerate(merges)}
self.special_tokens=special_tokens
def encode(self,text:str)->list[int]:
if self.special_tokens:
sorted_special_tokens=sorted(self.special_tokens,key=len,reverse=True)
sp_tokens_deal_pattern='('+'|'.join(map(regex.escape,sorted_special_tokens))+')'
document=regex.split(sp_tokens_deal_pattern,text)
else:
document=[text]
PAT=r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+"""
pre_tokenize_pattern=regex.compile(PAT)
ids=[]
for doc in document:
if not doc:
continue
byte_doc=doc.encode('utf-8')
if self.special_tokens and doc in self.special_tokens and byte_doc in self.vocab_reverse:
ids.append(self.vocab_reverse[byte_doc])
continue
for match in pre_tokenize_pattern.finditer(doc):
substr=match.group().encode('utf-8')
byte_substr=[bytes([i]) for i in substr]
while(len(byte_substr)>=2):
new_byte_substr=[]
pairs=[(byte_substr[i],byte_substr[i+1]) for i in range(len(byte_substr)-1)]
vaild_pairs=[pair for pair in pairs if pair in self.pair2id]
if not vaild_pairs:
break
first_merge_pair=min(vaild_pairs,key=lambda p:self.pair2id[p])
index=0
while(index<len(byte_substr)):
if index<len(byte_substr)-1 and byte_substr[index]==first_merge_pair[0] and byte_substr[index+1]==first_merge_pair[1]:
new_byte_substr.append(first_merge_pair[0]+first_merge_pair[1])
index+=2
else:
new_byte_substr.append(byte_substr[index])
index+=1
byte_substr=new_byte_substr
ids.extend(self.vocab_reverse[byte] for byte in byte_substr)
return ids
def decode(self,ids)->str:
return b"".join(self.vocab[token_id] for token_id in ids).decode('utf-8',errors='replace')
def encode_iterable(self,iterator:Iterator[str])->Iterator[int]:
for line in iterator:
for token_id in self.encode(line):
yield token_id
@classmethod
def merge_substr(cls,str_tuple,max_pair,new_token_id)->tuple[int,...]:
i=0
new_str_tuple=[]
while i<len(str_tuple):
if i<len(str_tuple)-1 and str_tuple[i]==max_pair[0] and str_tuple[i+1]==max_pair[1]:
new_str_tuple.append(new_token_id)
i+=2
else:
new_str_tuple.append(str_tuple[i])
i+=1
return tuple(new_str_tuple)
@classmethod
def train(cls,
input_path: str | os.PathLike,
vocab_size: int,
special_tokens: list[str],
**kwargs):
with open(input_path,'rb') as f:
text=f.read().decode('utf-8')
if special_tokens:
sorted_special_tokens=sorted(special_tokens,key=len,reverse=True)
sp_tokens_deal_pattern='|'.join(map(regex.escape,sorted_special_tokens))
document=regex.split(sp_tokens_deal_pattern,text)
else:
document=[text]
PAT=r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+"""
pre_tokenize_pattern=regex.compile(PAT)
substr_counter:dict[tuple[int,...],int]=Counter()
for doc in document:
if not doc:
continue
for match in pre_tokenize_pattern.finditer(doc):
substr=match.group()
if substr:
str_tuple=tuple(substr.encode('utf-8'))
substr_counter[str_tuple]+=1
vocab={token:bytes([token]) for token in range(256)}
merges=[]
if special_tokens:
for special_token in special_tokens:
vocab[len(vocab)]=special_token.encode('utf-8')
merge_iters=vocab_size-len(vocab)
for _ in range(merge_iters):
pair_dict={}
for substr,count in substr_counter.items():
for i in range(len(substr)-1):
pair=(substr[i],substr[i+1])
pair_dict[pair]=pair_dict.get(pair,0)+count
if not pair_dict:
break
max_pair=max(pair_dict,key=lambda p:(pair_dict[p],vocab[p[0]],vocab[p[1]]))
new_token_id=len(vocab)
new_substr_counter={}
for substr,count in substr_counter.items():
if max_pair[0] not in substr or max_pair[1] not in substr:
new_substr_counter[substr]=new_substr_counter.get(substr,0)+count
continue
new_substr=tokenizer.merge_substr(substr,max_pair,new_token_id)
new_substr_counter[new_substr]=new_substr_counter.get(new_substr,0)+count
vocab[new_token_id]=vocab[max_pair[0]]+vocab[max_pair[1]]
substr_counter=new_substr_counter
merges.append((vocab[max_pair[0]],vocab[max_pair[1]]))
return vocab,merges
@classmethod
def from_files(cls, vocab_filepath, merges_filepath, special_tokens=None):
with open(vocab_filepath,'rb') as f:
vocab=pickle.load(f)
with open(merges_filepath,'rb') as f:
merges=pickle.load(f)
return cls(vocab,merges,special_tokens)
|
