[clipfish] Modernize

This commit is contained in:
Philipp Hagemeister 2014-03-24 22:30:32 +01:00
parent fb8ae2d438
commit ba40a74666
2 changed files with 17 additions and 22 deletions

View File

@ -1,22 +1,28 @@
from __future__ import unicode_literals
import re
import time
import xml.etree.ElementTree
from .common import InfoExtractor
from ..utils import ExtractorError
from ..utils import (
ExtractorError,
parse_duration,
)
class ClipfishIE(InfoExtractor):
IE_NAME = u'clipfish'
IE_NAME = 'clipfish'
_VALID_URL = r'^https?://(?:www\.)?clipfish\.de/.*?/video/(?P<id>[0-9]+)/'
_TEST = {
u'url': u'http://www.clipfish.de/special/game-trailer/video/3966754/fifa-14-e3-2013-trailer/',
u'file': u'3966754.mp4',
u'md5': u'2521cd644e862936cf2e698206e47385',
u'info_dict': {
u'title': u'FIFA 14 - E3 2013 Trailer',
u'duration': 82,
'url': 'http://www.clipfish.de/special/game-trailer/video/3966754/fifa-14-e3-2013-trailer/',
'md5': '2521cd644e862936cf2e698206e47385',
'info_dict': {
'id': '3966754',
'ext': 'mp4',
'title': 'FIFA 14 - E3 2013 Trailer',
'duration': 82,
},
u'skip': 'Blocked in the US'
}
@ -33,21 +39,10 @@ def _real_extract(self, url):
video_url = doc.find('filename').text
if video_url is None:
xml_bytes = xml.etree.ElementTree.tostring(doc)
raise ExtractorError(u'Cannot find video URL in document %r' %
raise ExtractorError('Cannot find video URL in document %r' %
xml_bytes)
thumbnail = doc.find('imageurl').text
duration_str = doc.find('duration').text
m = re.match(
r'^(?P<hours>[0-9]+):(?P<minutes>[0-9]{2}):(?P<seconds>[0-9]{2}):(?P<ms>[0-9]*)$',
duration_str)
if m:
duration = (
(int(m.group('hours')) * 60 * 60) +
(int(m.group('minutes')) * 60) +
(int(m.group('seconds')))
)
else:
duration = None
duration = parse_duration(doc.find('duration').text)
return {
'id': video_id,

View File

@ -1186,7 +1186,7 @@ def parse_duration(s):
return None
m = re.match(
r'(?:(?:(?P<hours>[0-9]+)[:h])?(?P<mins>[0-9]+)[:m])?(?P<secs>[0-9]+)s?$', s)
r'(?:(?:(?P<hours>[0-9]+)[:h])?(?P<mins>[0-9]+)[:m])?(?P<secs>[0-9]+)s?(?::[0-9]+)?$', s)
if not m:
return None
res = int(m.group('secs'))