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
use crate::preprocessing::vocab::sentence_piece_vocab::{SentencePieceModel, Node};
use crate::{Vocab, Tokenizer, MultiThreadedTokenizer};
use crate::preprocessing::tokenizer::base_tokenizer::{Token, TokenRef, Offset, OffsetSize, Mask};
use crate::tokenization_utils::{clean_text, decompose_nfkc, lowercase, is_whitespace, replace_string, split_on_special_tokens};
use crate::preprocessing::tokenizer::tokenization_utils::strip_accents;
use crate::preprocessing::vocab::albert_vocab::AlbertVocab;
pub struct AlbertTokenizer {
model: SentencePieceModel,
vocab: AlbertVocab,
lower_case: bool,
keep_accents: bool,
}
impl AlbertTokenizer {
pub fn from_file(path: &str, lower_case: bool, keep_accents: bool) -> AlbertTokenizer {
let model = SentencePieceModel::from_file(path);
let vocab = AlbertVocab::from_file(path);
AlbertTokenizer { model, vocab, lower_case, keep_accents }
}
pub fn from_existing_vocab_and_model(vocab: AlbertVocab, model: SentencePieceModel,
lower_case: bool, keep_accents: bool) -> AlbertTokenizer {
AlbertTokenizer { model, vocab, lower_case, keep_accents }
}
fn post_process_pieces<'a>(&self, tokens: &'a mut Vec<Token>) -> &'a Vec<Token> {
let mut positions_to_update: Vec<(usize, Vec<Token>)> = vec!();
for (token_idx, token) in tokens.iter().enumerate() {
let mut token_chars = token.text.chars().rev();
if token.text.chars().count() > 1 {
if (token_chars.next().unwrap() == ',') & token_chars.next().unwrap().is_ascii_digit() {
let mut new_token = token.clone();
let last_char = new_token.text.pop().unwrap();
new_token.text = new_token.text.replace('\u{2581}', "");
let updated_tokens = self.model.decode_forward_token_ref(new_token.as_ref());
let updated_tokens = self.model.decode_backward(&updated_tokens);
let mut updated_tokens = self.parse_nodes_to_tokens(updated_tokens);
if (token.text.chars().next().unwrap() != '\u{2581}') &
(updated_tokens[0].text.chars().next().unwrap() == '\u{2581}') {
if updated_tokens[0].text.chars().count() == 1 {
updated_tokens.remove(0);
} else {
let first_char_length = updated_tokens[0].text.chars().next().unwrap().len_utf8();
updated_tokens[0].text = (&updated_tokens[0].text[first_char_length..]).parse().unwrap();
}
}
updated_tokens.push(Token {
text: last_char.to_string(),
offset: Offset { begin: token.offset.end, end: token.offset.end - 1 },
reference_offsets: vec!(*token.reference_offsets.last().unwrap()),
mask: token.mask,
});
positions_to_update.push((token_idx, updated_tokens.clone()));
}
}
};
for (pos, new_tokens) in positions_to_update {
tokens.splice(pos..pos, new_tokens);
}
tokens
}
fn parse_nodes_to_tokens(&self, nodes: Vec<&Node>) -> Vec<Token> {
let mut output: Vec<Token> = Vec::with_capacity(nodes.len() + 1);
let mut is_prev_unknown = false;
for node in nodes {
if is_prev_unknown & (node.index == 0) {
let prev_token = output.last().unwrap();
let mut text = prev_token.text.clone();
text.push_str(node.text);
let mut reference_offsets = prev_token.reference_offsets.clone();
reference_offsets.extend_from_slice(node.reference_offsets);
let consolidated_unknown = Token {
text,
offset: Offset { begin: 0, end: 0 },
reference_offsets,
mask: Default::default(),
};
output.pop();
output.push(consolidated_unknown);
} else {
output.push(Token {
text: node.text.to_owned(),
offset: Offset { begin: 0, end: 0 },
reference_offsets: node.reference_offsets.to_vec(),
mask: Default::default(),
});
}
is_prev_unknown = node.index == 0;
}
output
}
}
impl Tokenizer<AlbertVocab> for AlbertTokenizer {
fn vocab(&self) -> &AlbertVocab { &self.vocab }
fn tokenize_to_tokens(&self, text: TokenRef) -> Vec<Token> {
let mut tokens = split_on_special_tokens(text, &self.vocab)
.into_iter()
.map(|token| token.to_owned())
.collect::<Vec<Token>>();
let mut sub_tokens: Vec<Token> = Vec::new();
for token in tokens.iter_mut() {
if token.mask != Mask::Special && token.mask != Mask::Unknown {
replace_string(token, "``", "\"");
replace_string(token, "\'\'", "\"");
clean_text(token, true);
decompose_nfkc(token);
if self.lower_case {
lowercase(token);
}
if !self.keep_accents {
strip_accents(token);
}
token.text = token.text.replace(|c: char| is_whitespace(&c), "\u{2581}");
if !token.text.starts_with('\u{2581}') {
token.text.insert(0, '\u{2581}');
token.reference_offsets.insert(0, 0);
};
let output = self.model.decode_forward_token_ref(token.as_ref());
let decoded = self.model.decode_backward(&output);
let mut output: Vec<Token> = self.parse_nodes_to_tokens(decoded);
self.post_process_pieces(&mut output);
sub_tokens.extend(output)
} else {
sub_tokens.push(token.clone());
}
}
sub_tokens
}
fn convert_tokens_to_string(&self, tokens: Vec<String>) -> String {
tokens.into_iter().map(|v| v.replace('\u{2581}', " ")).collect::<Vec<String>>().join("")
}
fn build_input_with_special_tokens(&self, tokens_1: Vec<i64>, tokens_2: Option<Vec<i64>>,
offsets_1: Vec<Option<Offset>>, offsets_2: Option<Vec<Option<Offset>>>,
original_offsets_1: Vec<Vec<OffsetSize>>, original_offsets_2: Option<Vec<Vec<OffsetSize>>>,
mask_1: Vec<Mask>, mask_2: Option<Vec<Mask>>) -> (Vec<i64>, Vec<i8>, Vec<i8>, Vec<Option<Offset>>, Vec<Vec<OffsetSize>>, Vec<Mask>) {
let mut output: Vec<i64> = vec!();
let mut token_segment_ids: Vec<i8> = vec!();
let mut special_tokens_mask: Vec<i8> = vec!();
let mut offsets: Vec<Option<Offset>> = vec!();
let mut original_offsets: Vec<Vec<OffsetSize>> = vec!();
let mut mask: Vec<Mask> = vec!();
special_tokens_mask.push(1);
special_tokens_mask.extend(vec![0; tokens_1.len()]);
special_tokens_mask.push(1);
token_segment_ids.extend(vec![0; tokens_1.len() + 2]);
output.push(self.vocab.token_to_id(AlbertVocab::cls_value()));
output.extend(tokens_1);
output.push(self.vocab.token_to_id(AlbertVocab::sep_value()));
offsets.push(None);
offsets.extend(offsets_1);
offsets.push(None);
original_offsets.push(vec!());
original_offsets.extend(original_offsets_1);
original_offsets.push(vec!());
mask.push(Mask::Special);
mask.extend(mask_1);
mask.push(Mask::Special);
if let Some(add_tokens) = tokens_2 {
let length = add_tokens.len();
special_tokens_mask.extend(vec![0; length]);
special_tokens_mask.push(1);
token_segment_ids.extend(vec![1; length + 1]);
output.extend(add_tokens);
output.push(self.vocab.token_to_id(AlbertVocab::sep_value()));
if let Some(add_offsets) = offsets_2 {
offsets.extend(add_offsets);
} else {
offsets.extend(vec![None; length]);
}
if let Some(add_original_offsets) = original_offsets_2 {
original_offsets.extend(add_original_offsets);
}
offsets.push(None);
original_offsets.push(vec!());
if let Some(mask_2) = mask_2 {
mask.extend(mask_2)
} else {
mask.extend(vec![Mask::None; length]);
}
mask.push(Mask::Special);
}
(output, token_segment_ids, special_tokens_mask, offsets, original_offsets, mask)
}
}
impl MultiThreadedTokenizer<AlbertVocab> for AlbertTokenizer {}