|
|
# encoding: utf-8 from __future__ import unicode_literals
import re
from .common import InfoExtractor from ..utils import ( unified_strdate, url_basename, )
class CanalplusIE(InfoExtractor): _VALID_URL = r'https?://(?:www\.canalplus\.fr/.*?/(?P<path>.*)|player\.canalplus\.fr/#/(?P<id>[0-9]+))' _VIDEO_INFO_TEMPLATE = 'http://service.canal-plus.com/video/rest/getVideosLiees/cplus/%s' IE_NAME = 'canalplus.fr'
_TEST = { 'url': 'http://www.canalplus.fr/c-infos-documentaires/pid1830-c-zapping.html?vid=922470', 'md5': '3db39fb48b9685438ecf33a1078023e4', 'info_dict': { 'id': '922470', 'ext': 'flv', 'title': 'Zapping - 26/08/13', 'description': 'Le meilleur de toutes les chaînes, tous les jours.\nEmission du 26 août 2013', 'upload_date': '20130826', }, }
def _real_extract(self, url): mobj = re.match(self._VALID_URL, url) video_id = mobj.groupdict().get('id')
# Beware, some subclasses do not define an id group display_id = url_basename(mobj.group('path'))
if video_id is None: webpage = self._download_webpage(url, display_id) video_id = self._search_regex(r'<canal:player videoId="(\d+)"', webpage, 'video id')
info_url = self._VIDEO_INFO_TEMPLATE % video_id doc = self._download_xml(info_url, video_id, 'Downloading video XML')
video_info = [video for video in doc if video.find('ID').text == video_id][0] media = video_info.find('MEDIA') infos = video_info.find('INFOS')
preferences = ['MOBILE', 'BAS_DEBIT', 'HAUT_DEBIT', 'HD', 'HLS', 'HDS']
formats = [ { 'url': fmt.text + '?hdcore=2.11.3' if fmt.tag == 'HDS' else fmt.text, 'format_id': fmt.tag, 'ext': 'mp4' if fmt.tag == 'HLS' else 'flv', 'preference': preferences.index(fmt.tag) if fmt.tag in preferences else -1, } for fmt in media.find('VIDEOS') if fmt.text ] self._sort_formats(formats)
return { 'id': video_id, 'display_id': display_id, 'title': '%s - %s' % (infos.find('TITRAGE/TITRE').text, infos.find('TITRAGE/SOUS_TITRE').text), 'upload_date': unified_strdate(infos.find('PUBLICATION/DATE').text), 'thumbnail': media.find('IMAGES/GRAND').text, 'description': infos.find('DESCRIPTION').text, 'view_count': int(infos.find('NB_VUES').text), 'like_count': int(infos.find('NB_LIKES').text), 'comment_count': int(infos.find('NB_COMMENTS').text), 'formats': formats, }
|