mirror of
https://github.com/l1ving/youtube-dl
synced 2024-11-27 08:15:14 +08:00
66 lines
2.8 KiB
Python
66 lines
2.8 KiB
Python
# coding: utf-8
|
|
from __future__ import unicode_literals
|
|
|
|
from .common import InfoExtractor
|
|
from ..utils import unified_strdate
|
|
|
|
|
|
class ElPaisIE(InfoExtractor):
|
|
_VALID_URL = r'https?://(?:[^.]+\.)?elpais\.com/.*/(?P<id>[^/#?]+)\.html(?:$|[?#])'
|
|
IE_DESC = 'El País'
|
|
|
|
_TESTS = [{
|
|
'url': 'http://blogs.elpais.com/la-voz-de-inaki/2014/02/tiempo-nuevo-recetas-viejas.html',
|
|
'md5': '98406f301f19562170ec071b83433d55',
|
|
'info_dict': {
|
|
'id': 'tiempo-nuevo-recetas-viejas',
|
|
'ext': 'mp4',
|
|
'title': 'Tiempo nuevo, recetas viejas',
|
|
'description': 'De lunes a viernes, a partir de las ocho de la mañana, Iñaki Gabilondo nos cuenta su visión de la actualidad nacional e internacional.',
|
|
'upload_date': '20140206',
|
|
}
|
|
}, {
|
|
'url': 'http://elcomidista.elpais.com/elcomidista/2016/02/24/articulo/1456340311_668921.html#?id_externo_nwl=newsletter_diaria20160303t',
|
|
'md5': '3bd5b09509f3519d7d9e763179b013de',
|
|
'info_dict': {
|
|
'id': '1456340311_668921',
|
|
'ext': 'mp4',
|
|
'title': 'Cómo hacer el mejor café con cafetera italiana',
|
|
'description': 'Que sí, que las cápsulas son cómodas. Pero si le pides algo más a la vida, quizá deberías aprender a usar bien la cafetera italiana. No tienes más que ver este vídeo y seguir sus siete normas básicas.',
|
|
'upload_date': '20160303',
|
|
}
|
|
}]
|
|
|
|
def _real_extract(self, url):
|
|
video_id = self._match_id(url)
|
|
webpage = self._download_webpage(url, video_id)
|
|
|
|
prefix = self._html_search_regex(
|
|
r'var\s+url_cache\s*=\s*"([^"]+)";', webpage, 'URL prefix')
|
|
video_suffix = self._search_regex(
|
|
r"(?:URLMediaFile|urlVideo_\d+)\s*=\s*url_cache\s*\+\s*'([^']+)'", webpage, 'video URL')
|
|
video_url = prefix + video_suffix
|
|
thumbnail_suffix = self._search_regex(
|
|
r"(?:URLMediaStill|urlFotogramaFijo_\d+)\s*=\s*url_cache\s*\+\s*'([^']+)'",
|
|
webpage, 'thumbnail URL', fatal=False)
|
|
thumbnail = (
|
|
None if thumbnail_suffix is None
|
|
else prefix + thumbnail_suffix)
|
|
title = self._html_search_regex(
|
|
(r"tituloVideo\s*=\s*'([^']+)'", webpage, 'title',
|
|
r'<h2 class="entry-header entry-title.*?>(.*?)</h2>'),
|
|
webpage, 'title')
|
|
upload_date = unified_strdate(self._search_regex(
|
|
r'<p class="date-header date-int updated"\s+title="([^"]+)">',
|
|
webpage, 'upload date', default=None) or self._html_search_meta(
|
|
'datePublished', webpage, 'timestamp'))
|
|
|
|
return {
|
|
'id': video_id,
|
|
'url': video_url,
|
|
'title': title,
|
|
'description': self._og_search_description(webpage),
|
|
'thumbnail': thumbnail,
|
|
'upload_date': upload_date,
|
|
}
|