ytdl-nightly/youtube_dl/extractor/redtube.py

45 lines
1.2 KiB
Python

import re
from .common import InfoExtractor
class RedTubeIE(InfoExtractor):
_VALID_URL = r'(?:http://)?(?:www\.)?redtube\.com/(?P<id>[0-9]+)'
_TEST = {
u'url': u'http://www.redtube.com/66418',
u'file': u'66418.mp4',
u'md5': u'7b8c22b5e7098a3e1c09709df1126d2d',
u'info_dict': {
u"title": u"Sucked on a toilet",
u"age_limit": 18,
}
}
def _real_extract(self, url):
mobj = re.match(self._VALID_URL, url)
video_id = mobj.group('id')
video_extension = 'mp4'
webpage = self._download_webpage(url, video_id)
self.report_extraction(video_id)
video_url = self._html_search_regex(
r'<source src="(.+?)" type="video/mp4">', webpage, u'video URL')
video_title = self._html_search_regex(
r'<h1 class="videoTitle slidePanelMovable">(.+?)</h1>',
webpage, u'title')
# No self-labeling, but they describe themselves as
# "Home of Videos Porno"
age_limit = 18
return {
'id': video_id,
'url': video_url,
'ext': video_extension,
'title': video_title,
'age_limit': age_limit,
}