Merge pull request #1342 from whydoubt/fix_mit_26

Fix MIT extractor for Python 2.6
This commit is contained in:
Jaime Marquínez Ferrándiz 2013-08-29 13:42:08 +02:00
commit 54fda45bac
1 changed files with 7 additions and 9 deletions

View File

@ -25,23 +25,21 @@ class TechTVMITIE(InfoExtractor):
def _real_extract(self, url):
mobj = re.match(self._VALID_URL, url)
video_id = mobj.group('id')
webpage = self._download_webpage(
raw_page = self._download_webpage(
'http://techtv.mit.edu/videos/%s' % video_id, video_id)
embed_page = self._download_webpage(
'http://techtv.mit.edu/embeds/%s/' % video_id, video_id,
note=u'Downloading embed page')
clean_page = re.compile(u'<!--.*?-->', re.S).sub(u'', raw_page)
base_url = self._search_regex(r'ipadUrl: \'(.+?cloudfront.net/)',
embed_page, u'base url')
formats_json = self._search_regex(r'bitrates: (\[.+?\])', embed_page,
raw_page, u'base url')
formats_json = self._search_regex(r'bitrates: (\[.+?\])', raw_page,
u'video formats')
formats = json.loads(formats_json)
formats = sorted(formats, key=lambda f: f['bitrate'])
title = get_element_by_id('edit-title', webpage)
description = clean_html(get_element_by_id('edit-description', webpage))
title = get_element_by_id('edit-title', clean_page)
description = clean_html(get_element_by_id('edit-description', clean_page))
thumbnail = self._search_regex(r'playlist:.*?url: \'(.+?)\'',
embed_page, u'thumbnail', flags=re.DOTALL)
raw_page, u'thumbnail', flags=re.DOTALL)
return {'id': video_id,
'title': title,