aboutsummaryrefslogtreecommitdiff
path: root/devtools/gen_wordpiece_unicode.py
diff options
context:
space:
mode:
Diffstat (limited to 'devtools/gen_wordpiece_unicode.py')
-rwxr-xr-xdevtools/gen_wordpiece_unicode.py139
1 files changed, 139 insertions, 0 deletions
diff --git a/devtools/gen_wordpiece_unicode.py b/devtools/gen_wordpiece_unicode.py
new file mode 100755
index 0000000..2f8642e
--- /dev/null
+++ b/devtools/gen_wordpiece_unicode.py
@@ -0,0 +1,139 @@
+#!/usr/bin/env python3
+"""Generate unicode.h file for libtok: Unicode range tables for the WordPiece
+BasicTokenizer's punctuation and control classification, matching transformers'
+BertTokenizer._is_punctuation and _is_control exactly.
+
+Punctuation is the ASCII set (33-47, 58-64, 91-96, 123-126) plus every codepoint
+whose Unicode category starts with "P". Control is categories Cc, Cf, Co minus
+tab/newline/carriage-return (which route to whitespace). The tables are emitted
+as sorted, merged [lo, hi] ranges for binary search in the C tokenizer.
+
+Run when bumping the target Unicode version (the C side is data-only otherwise):
+
+ uv run python devtools/gen_wordpiece_unicode.py > unicode.h
+"""
+
+import unicodedata
+
+MAX_CP = 0x110000
+
+
+def merge(cps):
+ """Collapse a sorted codepoint list into contiguous [lo, hi] ranges."""
+ ranges = []
+ for cp in cps:
+ if ranges and cp == ranges[-1][1] + 1:
+ ranges[-1][1] = cp
+ else:
+ ranges.append([cp, cp])
+ return ranges
+
+
+def is_punct(cp):
+ if (33 <= cp <= 47) or (58 <= cp <= 64) or (91 <= cp <= 96) or (123 <= cp <= 126):
+ return True
+ return unicodedata.category(chr(cp)).startswith("P")
+
+
+def is_control(cp):
+ if cp in (0x09, 0x0A, 0x0D):
+ return False
+ # Cc/Cf/Co are the assigned control, format and private-use chars. Cn
+ # (unassigned) is intentionally excluded: it never occurs in real text and
+ # would bind the table to a single Unicode version.
+ return unicodedata.category(chr(cp)) in ("Cc", "Cf", "Co")
+
+
+# Hangul syllables decompose by formula (the standard algorithm), so they are
+# excluded from the fold table to keep it small; the C side handles them.
+HANGUL_BASE, HANGUL_COUNT = 0xAC00, 11172
+
+
+def fold(cp):
+ """do_lower_case fold: Python str.lower then NFD then drop combining marks
+ (Mn), matching transformers' lowercase + _run_strip_accents. Returns the
+ output codepoint list (empty when the codepoint is a dropped mark)."""
+ s = chr(cp).lower()
+ return [ord(c) for c in unicodedata.normalize("NFD", s) if unicodedata.category(c) != "Mn"]
+
+
+def emit_fold(entries):
+ out = ["static const wp_fold_t WP_FOLD[] = {"]
+ for cp, f in entries:
+ outs = ", ".join("0x%X" % c for c in (f + [0, 0, 0])[:3])
+ out.append(" {0x%X, {%s}, %d}," % (cp, outs, len(f)))
+ out.append("};")
+ out.append("enum { WP_FOLD_N = (int)(sizeof(WP_FOLD) / sizeof(WP_FOLD[0])) };")
+ return "\n".join(out)
+
+
+def emit(name, ranges):
+ out = [f"static const wp_range_t {name}[] = {{"]
+ line = " "
+ for lo, hi in ranges:
+ s = "{0x%X, 0x%X}, " % (lo, hi)
+ if len(line) + len(s) > 96:
+ out.append(line.rstrip())
+ line = " "
+ line += s
+ if line.strip():
+ out.append(line.rstrip())
+ out.append("};")
+ out.append(f"enum {{ {name}_N = (int)(sizeof({name}) / sizeof({name}[0])) }};")
+ return "\n".join(out)
+
+
+def main():
+ punct = merge([cp for cp in range(MAX_CP) if is_punct(cp)])
+ control = merge([cp for cp in range(MAX_CP) if is_control(cp)])
+ hangul = range(HANGUL_BASE, HANGUL_BASE + HANGUL_COUNT)
+ fold_entries = [
+ (cp, f)
+ for cp in range(MAX_CP)
+ if cp not in hangul
+ for f in [fold(cp)]
+ if f != [cp]
+ ]
+ print(f"""/* Generated by devtools/gen_wordpiece_unicode.py from Python unicodedata
+ * (Unicode {unicodedata.unidata_version}). Do not edit by hand.
+ *
+ * Tables for the BasicTokenizer, matching transformers' BertTokenizer.
+ *
+ * WP_PUNCT_RANGES / WP_CONTROL_RANGES: _is_punctuation and _is_control as sorted
+ * [lo, hi] ranges. Punctuation is the ASCII set (33-47/58-64/91-96/123-126) plus
+ * every category P* codepoint. Control is categories Cc, Cf, Co minus
+ * tab/newline/carriage-return (those route to whitespace). Unassigned (Cn) is
+ * intentionally not control: it never occurs in real text and would bind the
+ * table to one Unicode version.
+ *
+ * WP_FOLD: the do_lower_case fold (Python str.lower then NFD then drop Mn marks),
+ * sorted by codepoint. n is the output length (0 drops a combining mark; up to
+ * 3). Hangul syllables (U+AC00..U+D7A3) decompose by formula on the C side and
+ * are excluded here. */
+#ifndef WORDPIECE_UNICODE_H
+#define WORDPIECE_UNICODE_H
+
+#include <stdint.h>
+
+typedef struct {{
+ uint32_t lo;
+ uint32_t hi;
+}} wp_range_t;
+
+typedef struct {{
+ uint32_t cp;
+ uint32_t out[3];
+ uint8_t n;
+}} wp_fold_t;
+
+{emit("WP_PUNCT_RANGES", punct)}
+
+{emit("WP_CONTROL_RANGES", control)}
+
+{emit_fold(fold_entries)}
+
+#endif /* WORDPIECE_UNICODE_H */""")
+
+
+if __name__ == "__main__":
+ main()