File size: 6,427 Bytes
f2878d0 | 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 | //! Lowercase WordPiece, the same normalisation as `WordPiece.encode` in tinydecide.js.
//!
//! Per code point: drop NUL, U+FFFD and control/format characters (tab, newline and CR stay as
//! spaces); space separators become spaces; CJK ideographs become one-character words; everything
//! else is NFD-decomposed, nonspacing marks are dropped, and each remaining character is lowercased.
//! Words split on whitespace (as JavaScript's `\s` sees it) and on punctuation (the ASCII symbol
//! ranges plus general category P*), then each word is split greedily into vocabulary pieces.
use std::collections::HashMap;
use unicode_normalization::UnicodeNormalization;
use unicode_properties::general_category::{GeneralCategory, GeneralCategoryGroup, UnicodeGeneralCategory};
use crate::error::{Error, Result};
/// Token ids plus the byte range of `text` each token came from.
#[derive(Debug, Clone, Default)]
pub struct Encoded {
pub ids: Vec<u32>,
/// `(start, end)` UTF-8 byte offsets into the encoded text.
pub offsets: Vec<(usize, usize)>,
}
#[derive(Debug, Clone)]
pub struct WordPiece {
vocab: HashMap<String, u32>,
unk: u32,
prefix: String,
max_chars: usize,
}
fn is_cjk(cp: u32) -> bool {
(0x4E00..=0x9FFF).contains(&cp)
|| (0x3400..=0x4DBF).contains(&cp)
|| (0x20000..=0x2A6DF).contains(&cp)
|| (0x2A700..=0x2B73F).contains(&cp)
|| (0x2B740..=0x2B81F).contains(&cp)
|| (0x2B820..=0x2CEAF).contains(&cp)
|| (0xF900..=0xFAFF).contains(&cp)
|| (0x2F800..=0x2FA1F).contains(&cp)
}
fn is_punct(ch: char) -> bool {
let c = ch as u32;
if (33..=47).contains(&c) || (58..=64).contains(&c) || (91..=96).contains(&c) || (123..=126).contains(&c) {
return true;
}
ch.general_category_group() == GeneralCategoryGroup::Punctuation
}
/// JavaScript's `\s` (WhiteSpace + LineTerminator).
pub(crate) fn is_js_space(ch: char) -> bool {
matches!(
ch,
'\t' | '\n' | '\u{0B}' | '\u{0C}' | '\r' | ' ' | '\u{A0}' | '\u{1680}' | '\u{2000}'..='\u{200A}'
| '\u{2028}' | '\u{2029}' | '\u{202F}' | '\u{205F}' | '\u{3000}' | '\u{FEFF}'
)
}
type Ch = (char, usize, usize);
impl WordPiece {
/// `vocab[i]` is the piece for id `i` (`None` for unused ids). Later duplicates win, as in the JS.
pub fn new(vocab: &[Option<String>], unk: &str, prefix: &str, max_chars: usize) -> Result<Self> {
let mut map = HashMap::with_capacity(vocab.len());
for (i, p) in vocab.iter().enumerate() {
if let Some(p) = p {
map.insert(p.clone(), i as u32);
}
}
let unk = *map.get(unk).ok_or_else(|| Error::Model(format!("unknown-token piece {unk:?} is not in the vocabulary")))?;
Ok(WordPiece { vocab: map, unk, prefix: prefix.to_string(), max_chars })
}
pub fn unk(&self) -> u32 {
self.unk
}
pub fn encode(&self, text: &str) -> Encoded {
let mut chars: Vec<Ch> = Vec::with_capacity(text.len() + 8);
for (a, ch) in text.char_indices() {
let b = a + ch.len_utf8();
let cp = ch as u32;
let cat = ch.general_category();
if cp == 0
|| cp == 0xFFFD
|| (matches!(cat, GeneralCategory::Control | GeneralCategory::Format) && !matches!(ch, '\t' | '\n' | '\r'))
{
continue;
}
if matches!(ch, ' ' | '\t' | '\n' | '\r') || cat == GeneralCategory::SpaceSeparator {
chars.push((' ', a, b));
continue;
}
if is_cjk(cp) {
chars.push((' ', a, a));
chars.push((ch, a, b));
chars.push((' ', b, b));
continue;
}
for d in std::iter::once(ch).nfd() {
if d.general_category() == GeneralCategory::NonspacingMark {
continue;
}
for l in d.to_lowercase() {
chars.push((l, a, b));
}
}
}
let mut words: Vec<Vec<Ch>> = Vec::new();
let mut cur: Vec<Ch> = Vec::new();
for c in chars {
if c.0 == ' ' || is_js_space(c.0) {
if !cur.is_empty() {
words.push(std::mem::take(&mut cur));
}
continue;
}
if is_punct(c.0) {
if !cur.is_empty() {
words.push(std::mem::take(&mut cur));
}
words.push(vec![c]);
continue;
}
cur.push(c);
}
if !cur.is_empty() {
words.push(cur);
}
let mut out = Encoded::default();
let mut s = String::new();
let mut pieces: Vec<(u32, usize, usize)> = Vec::new();
for w in &words {
let whole = (w[0].1, w[w.len() - 1].2);
if w.len() > self.max_chars {
out.ids.push(self.unk);
out.offsets.push(whole);
continue;
}
pieces.clear();
let mut start = 0;
let mut bad = false;
while start < w.len() {
let mut end = w.len();
let mut got = None;
while start < end {
s.clear();
if start > 0 {
s.push_str(&self.prefix);
}
s.extend(w[start..end].iter().map(|c| c.0));
if let Some(&id) = self.vocab.get(s.as_str()) {
got = Some(id);
break;
}
end -= 1;
}
match got {
Some(id) => {
pieces.push((id, w[start].1, w[end - 1].2));
start = end;
}
None => {
bad = true;
break;
}
}
}
if bad {
out.ids.push(self.unk);
out.offsets.push(whole);
} else {
for &(id, a, b) in &pieces {
out.ids.push(id);
out.offsets.push((a, b));
}
}
}
out
}
}
|