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
use crate::preprocessing::vocab::sentence_piece_vocab::{SentencePieceModel};
use crate::{Vocab, Tokenizer, MultiThreadedTokenizer, XLMRobertaVocab};
use crate::preprocessing::tokenizer::base_tokenizer::{Token, TokenRef, Mask, Offset, OffsetSize};
use crate::tokenization_utils::{clean_text, decompose_nfkc, lowercase, is_whitespace, split_on_special_tokens};
pub struct XLMRobertaTokenizer {
model: SentencePieceModel,
vocab: XLMRobertaVocab,
lower_case: bool,
}
impl XLMRobertaTokenizer {
pub fn from_file(path: &str, lower_case: bool) -> XLMRobertaTokenizer {
let model = SentencePieceModel::from_file(path);
let vocab = XLMRobertaVocab::from_file(path);
XLMRobertaTokenizer { model, vocab, lower_case }
}
pub fn from_existing_vocab_and_model(vocab: XLMRobertaVocab, model: SentencePieceModel, lower_case: bool) -> XLMRobertaTokenizer {
XLMRobertaTokenizer { model, vocab, lower_case }
}
}
impl Tokenizer<XLMRobertaVocab> for XLMRobertaTokenizer {
fn vocab(&self) -> &XLMRobertaVocab { &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 {
clean_text(token, true);
decompose_nfkc(token);
if self.lower_case {
lowercase(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 output: Vec<Token> = self.model.parse_nodes_to_tokens(decoded);
sub_tokens.extend(output)
} else {
sub_tokens.push(token.clone());
}
}
sub_tokens
}
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(XLMRobertaVocab::cls_value()));
output.extend(tokens_1);
output.push(self.vocab.token_to_id(XLMRobertaVocab::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.push(1);
special_tokens_mask.extend(vec![0; length]);
special_tokens_mask.push(1);
token_segment_ids.extend(vec![1; length + 2]);
output.push(self.vocab.token_to_id(XLMRobertaVocab::sep_value()));
output.extend(add_tokens);
output.push(self.vocab.token_to_id(XLMRobertaVocab::sep_value()));
if let Some(add_offsets) = offsets_2 {
offsets.push(None);
offsets.extend(add_offsets);
} else {
offsets.extend(vec![None; length + 2]);
}
if let Some(add_original_offsets) = original_offsets_2 {
original_offsets.push(vec!());
original_offsets.extend(add_original_offsets);
}
offsets.push(None);
original_offsets.push(vec!());
mask.push(Mask::Special);
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)
}
fn convert_tokens_to_string(&self, tokens: Vec<String>) -> String {
tokens.into_iter().map(|v| v.replace('\u{2581}', " ")).collect::<Vec<String>>().join("")
}
}
impl MultiThreadedTokenizer<XLMRobertaVocab> for XLMRobertaTokenizer {}