File size: 4,222 Bytes
76aca0a | 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 | # Sherpa Onnx Punctuation Preprocessor (CharTokenizer)
#
# Model: sherpa-onnx-punct-ct-transformer
# Tokenizer: character-level for Chinese, word-level for English
# Vocab: tokens.json (272727 entries)
# Padding: to 64 tokens
import json
import os
from typing import List, Tuple
import numpy as np
class CharTokenizer:
"""Character/word tokenizer for the sherpa punct CT Transformer model."""
def __init__(self, tokens_path: str, unk_symbol: str = "<unk>"):
if not os.path.exists(tokens_path):
raise FileNotFoundError(f"tokens.json not found: {tokens_path}")
with open(tokens_path, "r", encoding="utf-8") as f:
id2token = json.load(f)
self.id2token = id2token
self.token2id = {tok: idx for idx, tok in enumerate(id2token)}
self.unk_id = self.token2id.get(unk_symbol, 0)
def tokenize(self, text: str) -> List[int]:
"""Split text into tokens and return token IDs.
Chinese characters are segmented individually.
English words are kept as whole tokens.
"""
# Split on whitespace
word_list = text.split()
words = []
for w in word_list:
s = ""
for c in w:
if len(c.encode()) > 1:
# Multi-byte character (Chinese, Japanese, etc.)
if s == "":
s = c
elif len(s[-1].encode()) > 1:
s += c
else:
words.append(s)
s = c
else:
# ASCII character
if s == "":
s = c
elif len(s[-1].encode()) > 1:
words.append(s)
s = c
else:
s += c
if s:
words.append(s)
ids = []
for w in words:
if len(w[0].encode()) > 1:
# Chinese phrase: tokenize each character
for c in w:
ids.append(self.token2id.get(c, self.unk_id))
else:
ids.append(self.token2id.get(w, self.unk_id))
return ids
def tokenize_full(self, text: str) -> List[int]:
"""Tokenize full text without truncation or padding."""
return self.tokenize(text)
def encode(
self, text: str, pad_length: int = 64
) -> Tuple[np.ndarray, int]:
"""Tokenize and pad to fixed length. Truncates if > pad_length.
Returns:
input_array: (1, pad_length) int32 numpy array
original_length: actual token count before padding
"""
ids = self.tokenize(text)
original_len = len(ids)
# Truncate or pad to pad_length
if len(ids) > pad_length:
ids = ids[:pad_length]
original_len = pad_length
padded = np.zeros((1, pad_length), dtype=np.int32)
padded[0, : len(ids)] = ids
return padded, min(original_len, pad_length)
def encode_long(
self, text: str, window_size: int = 64
) -> Tuple[List[np.ndarray], List[int], List[int]]:
"""Tokenize long text into sliding windows for batched inference.
Splits full token sequence into windows of window_size.
Each window is padded to window_size if shorter.
Returns:
windows: list of (1, window_size) int32 arrays
window_token_ids: list of token ID lists per window
window_lens: original token lengths per window (before padding)
"""
ids = self.tokenize(text)
if not ids:
return [], [], []
windows = []
window_token_ids = []
window_lens = []
for start in range(0, len(ids), window_size):
chunk = ids[start:start + window_size]
chunk_len = len(chunk)
padded = np.zeros((1, window_size), dtype=np.int32)
padded[0, :chunk_len] = chunk
windows.append(padded)
window_token_ids.append(chunk)
window_lens.append(chunk_len)
return windows, window_token_ids, window_lens
|