1
mirror of https://github.com/yt-dlp/yt-dlp synced 2024-12-20 18:15:54 +01:00
yt-dlp/youtube_dl/extractor/spike.py

33 lines
1.1 KiB
Python
Raw Normal View History

from __future__ import unicode_literals
2014-03-08 21:06:12 +01:00
import re
from .mtv import MTVServicesInfoExtractor
class SpikeIE(MTVServicesInfoExtractor):
2014-03-08 21:06:12 +01:00
_VALID_URL = r'''(?x)https?://
(www\.spike\.com/(video-clips|episodes)/.+|
m\.spike\.com/videos/video.rbml\?id=(?P<mobile_id>[^&]+))
'''
_TEST = {
'url': 'http://www.spike.com/video-clips/lhtu8m/auction-hunters-can-allen-ride-a-hundred-year-old-motorcycle',
'md5': '1a9265f32b0c375793d6c4ce45255256',
'info_dict': {
'id': 'b9c8221a-4e50-479a-b86d-3333323e38ba',
'ext': 'mp4',
'title': 'Auction Hunters|Can Allen Ride A Hundred Year-Old Motorcycle?',
'description': 'md5:fbed7e82ed5fad493615b3094a9499cb',
},
}
_FEED_URL = 'http://www.spike.com/feeds/mrss/'
_MOBILE_TEMPLATE = 'http://m.spike.com/videos/video.rbml?id=%s'
2014-03-08 21:06:12 +01:00
def _real_extract(self, url):
mobj = re.search(self._VALID_URL, url)
mobile_id = mobj.group('mobile_id')
if mobile_id is not None:
url = 'http://www.spike.com/video-clips/%s' % mobile_id
return super(SpikeIE, self)._real_extract(url)