mirror of
https://github.com/yt-dlp/yt-dlp.git
synced 2024-11-07 09:20:36 +00:00
86e5f3ed2e
Using https://github.com/asottile/pyupgrade 1. `__future__` imports and `coding: utf-8` were removed 2. Files were rewritten with `pyupgrade --py36-plus --keep-percent-format` 3. f-strings were cherry-picked from `pyupgrade --py36-plus` Extractors are left untouched (except removing header) to avoid unnecessary merge conflicts
157 lines
6.4 KiB
Python
157 lines
6.4 KiB
Python
import re
|
|
|
|
from .common import InfoExtractor
|
|
from ..utils import (
|
|
extract_attributes,
|
|
smuggle_url,
|
|
strip_or_none,
|
|
urljoin,
|
|
)
|
|
|
|
|
|
class SkyBaseIE(InfoExtractor):
|
|
BRIGHTCOVE_URL_TEMPLATE = 'http://players.brightcove.net/%s/%s_default/index.html?videoId=%s'
|
|
_SDC_EL_REGEX = r'(?s)(<div[^>]+data-(?:component-name|fn)="sdc-(?:articl|sit)e-video"[^>]*>)'
|
|
|
|
def _process_ooyala_element(self, webpage, sdc_el, url):
|
|
sdc = extract_attributes(sdc_el)
|
|
provider = sdc.get('data-provider')
|
|
if provider == 'ooyala':
|
|
video_id = sdc['data-sdc-video-id']
|
|
video_url = 'ooyala:%s' % video_id
|
|
ie_key = 'Ooyala'
|
|
ooyala_el = self._search_regex(
|
|
r'(<div[^>]+class="[^"]*\bsdc-article-video__media-ooyala\b[^"]*"[^>]+data-video-id="%s"[^>]*>)' % video_id,
|
|
webpage, 'video data', fatal=False)
|
|
if ooyala_el:
|
|
ooyala_attrs = extract_attributes(ooyala_el) or {}
|
|
if ooyala_attrs.get('data-token-required') == 'true':
|
|
token_fetch_url = (self._parse_json(ooyala_attrs.get(
|
|
'data-token-fetch-options', '{}'),
|
|
video_id, fatal=False) or {}).get('url')
|
|
if token_fetch_url:
|
|
embed_token = self._download_json(urljoin(
|
|
url, token_fetch_url), video_id, fatal=False)
|
|
if embed_token:
|
|
video_url = smuggle_url(
|
|
video_url, {'embed_token': embed_token})
|
|
elif provider == 'brightcove':
|
|
video_id = sdc['data-video-id']
|
|
account_id = sdc.get('data-account-id') or '6058004172001'
|
|
player_id = sdc.get('data-player-id') or 'RC9PQUaJ6'
|
|
video_url = self.BRIGHTCOVE_URL_TEMPLATE % (account_id, player_id, video_id)
|
|
ie_key = 'BrightcoveNew'
|
|
|
|
return {
|
|
'_type': 'url_transparent',
|
|
'id': video_id,
|
|
'url': video_url,
|
|
'ie_key': ie_key,
|
|
}
|
|
|
|
def _real_extract(self, url):
|
|
video_id = self._match_id(url)
|
|
webpage = self._download_webpage(url, video_id)
|
|
info = self._process_ooyala_element(webpage, self._search_regex(
|
|
self._SDC_EL_REGEX, webpage, 'sdc element'), url)
|
|
info.update({
|
|
'title': self._og_search_title(webpage),
|
|
'description': strip_or_none(self._og_search_description(webpage)),
|
|
})
|
|
return info
|
|
|
|
|
|
class SkySportsIE(SkyBaseIE):
|
|
IE_NAME = 'sky:sports'
|
|
_VALID_URL = r'https?://(?:www\.)?skysports\.com/watch/video/([^/]+/)*(?P<id>[0-9]+)'
|
|
_TESTS = [{
|
|
'url': 'http://www.skysports.com/watch/video/10328419/bale-its-our-time-to-shine',
|
|
'md5': '77d59166cddc8d3cb7b13e35eaf0f5ec',
|
|
'info_dict': {
|
|
'id': 'o3eWJnNDE6l7kfNO8BOoBlRxXRQ4ANNQ',
|
|
'ext': 'mp4',
|
|
'title': 'Bale: It\'s our time to shine',
|
|
'description': 'md5:e88bda94ae15f7720c5cb467e777bb6d',
|
|
},
|
|
'add_ie': ['Ooyala'],
|
|
}, {
|
|
'url': 'https://www.skysports.com/watch/video/sports/f1/12160544/abu-dhabi-gp-the-notebook',
|
|
'only_matching': True,
|
|
}, {
|
|
'url': 'https://www.skysports.com/watch/video/tv-shows/12118508/rainford-brent-how-ace-programme-helps',
|
|
'only_matching': True,
|
|
}]
|
|
|
|
|
|
class SkyNewsIE(SkyBaseIE):
|
|
IE_NAME = 'sky:news'
|
|
_VALID_URL = r'https?://news\.sky\.com/video/[0-9a-z-]+-(?P<id>[0-9]+)'
|
|
_TEST = {
|
|
'url': 'https://news.sky.com/video/russian-plane-inspected-after-deadly-fire-11712962',
|
|
'md5': '411e8893fd216c75eaf7e4c65d364115',
|
|
'info_dict': {
|
|
'id': 'ref:1ua21xaDE6lCtZDmbYfl8kwsKLooJbNM',
|
|
'ext': 'mp4',
|
|
'title': 'Russian plane inspected after deadly fire',
|
|
'description': 'The Russian Investigative Committee has released video of the wreckage of a passenger plane which caught fire near Moscow.',
|
|
'uploader_id': '6058004172001',
|
|
'timestamp': 1567112345,
|
|
'upload_date': '20190829',
|
|
},
|
|
'add_ie': ['BrightcoveNew'],
|
|
}
|
|
|
|
|
|
class SkyNewsStoryIE(SkyBaseIE):
|
|
IE_NAME = 'sky:news:story'
|
|
_VALID_URL = r'https?://news\.sky\.com/story/[0-9a-z-]+-(?P<id>[0-9]+)'
|
|
_TEST = {
|
|
'url': 'https://news.sky.com/story/budget-2021-chancellor-rishi-sunak-vows-address-will-deliver-strong-economy-fit-for-a-new-age-of-optimism-12445425',
|
|
'info_dict': {
|
|
'id': 'ref:0714acb9-123d-42c8-91b8-5c1bc6c73f20',
|
|
'title': 'md5:e408dd7aad63f31a1817bbe40c7d276f',
|
|
'description': 'md5:a881e12f49212f92be2befe4a09d288a',
|
|
'ext': 'mp4',
|
|
'upload_date': '20211027',
|
|
'timestamp': 1635317494,
|
|
'uploader_id': '6058004172001',
|
|
}
|
|
}
|
|
|
|
def _real_extract(self, url):
|
|
article_id = self._match_id(url)
|
|
webpage = self._download_webpage(url, article_id)
|
|
|
|
entries = [self._process_ooyala_element(webpage, sdc_el, url)
|
|
for sdc_el in re.findall(self._SDC_EL_REGEX, webpage)]
|
|
|
|
return self.playlist_result(
|
|
entries, article_id, self._og_search_title(webpage),
|
|
self._html_search_meta(['og:description', 'description'], webpage))
|
|
|
|
|
|
class SkySportsNewsIE(SkyBaseIE):
|
|
IE_NAME = 'sky:sports:news'
|
|
_VALID_URL = r'https?://(?:www\.)?skysports\.com/([^/]+/)*news/\d+/(?P<id>\d+)'
|
|
_TEST = {
|
|
'url': 'http://www.skysports.com/golf/news/12176/10871916/dustin-johnson-ready-to-conquer-players-championship-at-tpc-sawgrass',
|
|
'info_dict': {
|
|
'id': '10871916',
|
|
'title': 'Dustin Johnson ready to conquer Players Championship at TPC Sawgrass',
|
|
'description': 'Dustin Johnson is confident he can continue his dominant form in 2017 by adding the Players Championship to his list of victories.',
|
|
},
|
|
'playlist_count': 2,
|
|
}
|
|
|
|
def _real_extract(self, url):
|
|
article_id = self._match_id(url)
|
|
webpage = self._download_webpage(url, article_id)
|
|
|
|
entries = []
|
|
for sdc_el in re.findall(self._SDC_EL_REGEX, webpage):
|
|
entries.append(self._process_ooyala_element(webpage, sdc_el, url))
|
|
|
|
return self.playlist_result(
|
|
entries, article_id, self._og_search_title(webpage),
|
|
self._html_search_meta(['og:description', 'description'], webpage))
|