ai-content-maker/.venv/Lib/site-packages/pypinyin/runner.py

145 lines
4.5 KiB
Python

#!/usr/bin/env python
# -*- coding: utf-8 -*-
from __future__ import unicode_literals
from argparse import ArgumentParser
import logging
import sys
import pypinyin
from pypinyin.compat import PY2
style_map = {
'NORMAL': pypinyin.Style.NORMAL,
'zhao': pypinyin.Style.NORMAL,
'TONE': pypinyin.Style.TONE,
'zh4ao': pypinyin.Style.TONE,
'TONE2': pypinyin.Style.TONE2,
'zha4o': pypinyin.Style.TONE2,
'TONE3': pypinyin.Style.TONE3,
'zhao4': pypinyin.Style.TONE3,
'INITIALS': pypinyin.Style.INITIALS,
'zh': pypinyin.Style.INITIALS,
'FIRST_LETTER': pypinyin.Style.FIRST_LETTER,
'z': pypinyin.Style.FIRST_LETTER,
'FINALS': pypinyin.Style.FINALS,
'ao': pypinyin.Style.FINALS,
'FINALS_TONE': pypinyin.Style.FINALS_TONE,
'4ao': pypinyin.Style.FINALS_TONE,
'FINALS_TONE2': pypinyin.Style.FINALS_TONE2,
'a4o': pypinyin.Style.FINALS_TONE2,
'FINALS_TONE3': pypinyin.Style.FINALS_TONE3,
'ao4': pypinyin.Style.FINALS_TONE3,
'BOPOMOFO': pypinyin.Style.BOPOMOFO,
'BOPOMOFO_FIRST': pypinyin.Style.BOPOMOFO_FIRST,
'CYRILLIC': pypinyin.Style.CYRILLIC,
'CYRILLIC_FIRST': pypinyin.Style.CYRILLIC_FIRST,
}
func_map = {
'pinyin': pypinyin.pinyin,
'slug': pypinyin.slug,
}
default_style = 'zh4ao'
class NullWriter(object):
"""数据流黑洞,类似 linux/unix 下 /dev/null 的效果。"""
def write(self, string):
pass
def get_parser():
parser = ArgumentParser(description='convert chinese to pinyin.')
parser.add_argument('-V', '--version', action='version',
version='{0} {1}'.format(
pypinyin.__title__, pypinyin.__version__
))
# 要执行的函数名称
parser.add_argument('-f', '--func',
help='function name (default: "pinyin")',
choices=['pinyin', 'slug'],
default='pinyin')
# 拼音风格
parser.add_argument(
'-s', '--style',
help='pinyin style (default: "{0}")'.format(default_style),
choices=style_map.keys(), default=default_style
)
parser.add_argument('-p', '--separator',
help='slug separator (default: "-")',
default='-')
parser.add_argument('-e', '--errors',
help=('how to handle none-pinyin string'
' (default: "default")'),
choices=['default', 'ignore', 'replace'],
default='default')
# 输出多音字
parser.add_argument('-m', '--heteronym', help='enable heteronym',
action='store_true')
# 要查询的汉字
parser.add_argument('hans', nargs='+', help='chinese string')
return parser
def main():
# 禁用除 CRITICAL 外的日志消息
logging.disable(logging.CRITICAL)
# read hans from stdin
if not sys.stdin.isatty():
pipe_data = sys.stdin.read().strip()
else:
pipe_data = ''
args = sys.argv[1:]
if pipe_data:
args.append(pipe_data)
# 获取命令行选项和参数
parser = get_parser()
options = parser.parse_args(args)
if PY2:
hans = [
han.decode(sys.stdin.encoding or 'utf-8') for han in options.hans
]
else:
hans = options.hans
func = getattr(pypinyin, options.func)
style = style_map[options.style]
heteronym = options.heteronym
separator = options.separator
errors = options.errors
func_kwargs = {
'pinyin': {'heteronym': heteronym, 'errors': errors},
'slug': {'heteronym': heteronym, 'separator': separator,
'errors': errors},
}
if PY2:
kwargs = func_kwargs[func.func_name]
else:
kwargs = func_kwargs[func.__name__]
# 重设标准输出流和标准错误流
# 不输出任何字符,防止污染命令行命令的输出结果
# 其实主要是为了干掉 jieba 内的 print 语句 ;)
sys.stdout = sys.stderr = NullWriter()
results = [func(han, style=style, **kwargs) for han in hans]
# 恢复默认
sys.stdout = sys.__stdout__
sys.stderr = sys.__stderr__
for result in results:
if not result:
print('')
elif result and isinstance(result, (list, tuple)):
if isinstance(result[0], (list, tuple)):
print(' '.join([','.join(s) for s in result]))
else:
print(result)
else:
print(result)
if __name__ == '__main__':
main()