|
|
# encoding: utf-8 from __future__ import unicode_literals
import re
from .common import InfoExtractor from ..utils import int_or_none
class KontrTubeIE(InfoExtractor): IE_NAME = 'kontrtube' IE_DESC = 'KontrTube.ru - Труба зовёт' _VALID_URL = r'http://(?:www\.)?kontrtube\.ru/videos/(?P<id>\d+)/(?P<display_id>[^/]+)/'
_TEST = { 'url': 'http://www.kontrtube.ru/videos/2678/nad-olimpiyskoy-derevney-v-sochi-podnyat-rossiyskiy-flag/', 'md5': '975a991a4926c9a85f383a736a2e6b80', 'info_dict': { 'id': '2678', 'display_id': 'nad-olimpiyskoy-derevney-v-sochi-podnyat-rossiyskiy-flag', 'ext': 'mp4', 'title': 'Над олимпийской деревней в Сочи поднят российский флаг', 'description': 'md5:80edc4c613d5887ae8ccf1d59432be41', 'thumbnail': 'http://www.kontrtube.ru/contents/videos_screenshots/2000/2678/preview.mp4.jpg', 'duration': 270, } }
def _real_extract(self, url): mobj = re.match(self._VALID_URL, url) video_id = mobj.group('id') display_id = mobj.group('display_id')
webpage = self._download_webpage( url, display_id, 'Downloading page')
video_url = self._html_search_regex( r"video_url\s*:\s*'(.+?)/?',", webpage, 'video URL') thumbnail = self._html_search_regex( r"preview_url\s*:\s*'(.+?)/?',", webpage, 'video thumbnail', fatal=False) title = self._html_search_regex( r'<title>(.+?)</title>', webpage, 'video title') description = self._html_search_meta( 'description', webpage, 'video description')
mobj = re.search( r'<div class="col_2">Длительность: <span>(?P<minutes>\d+)м:(?P<seconds>\d+)с</span></div>', webpage) duration = int(mobj.group('minutes')) * 60 + int(mobj.group('seconds')) if mobj else None
view_count = self._html_search_regex( r'<div class="col_2">Просмотров: <span>(\d+)</span></div>', webpage, 'view count', fatal=False)
comment_count = None comment_str = self._html_search_regex( r'Комментарии: <span>([^<]+)</span>', webpage, 'comment count', fatal=False) if comment_str.startswith('комментариев нет'): comment_count = 0 else: mobj = re.search(r'\d+ из (?P<total>\d+) комментариев', comment_str) if mobj: comment_count = mobj.group('total')
return { 'id': video_id, 'display_id': display_id, 'url': video_url, 'thumbnail': thumbnail, 'title': title, 'description': description, 'duration': duration, 'view_count': int_or_none(view_count), 'comment_count': int_or_none(comment_count), }
|