mirror of
https://github.com/yt-dlp/yt-dlp.git
synced 2024-11-07 09:20:36 +00:00
c4da5ff971
Authored by: YuenSzeHong
53 lines
2.3 KiB
Python
53 lines
2.3 KiB
Python
# coding: utf-8
|
|
from __future__ import unicode_literals
|
|
from ..utils import HEADRequest
|
|
from .common import InfoExtractor
|
|
|
|
|
|
class FujiTVFODPlus7IE(InfoExtractor):
|
|
_VALID_URL = r'https?://fod\.fujitv\.co\.jp/title/(?P<sid>[0-9a-z]{4})/(?P<id>[0-9a-z]+)'
|
|
_BASE_URL = 'https://i.fod.fujitv.co.jp/'
|
|
|
|
_TESTS = [{
|
|
'url': 'https://fod.fujitv.co.jp/title/5d40/5d40110076',
|
|
'info_dict': {
|
|
'id': '5d40110076',
|
|
'ext': 'mp4',
|
|
'title': '#1318 『まる子、まぼろしの洋館を見る』の巻',
|
|
'series': 'ちびまる子ちゃん',
|
|
'series_id': '5d40',
|
|
'description': 'md5:b3f51dbfdda162ac4f789e0ff4d65750',
|
|
'thumbnail': 'https://i.fod.fujitv.co.jp/img/program/5d40/episode/5d40110076_a.jpg',
|
|
},
|
|
}]
|
|
|
|
def _real_extract(self, url):
|
|
series_id, video_id = self._match_valid_url(url).groups()
|
|
self._request_webpage(HEADRequest(url), video_id)
|
|
json_info = {}
|
|
token = self._get_cookies(url).get('CT')
|
|
if token:
|
|
json_info = self._download_json('https://fod-sp.fujitv.co.jp/apps/api/episode/detail/?ep_id=%s&is_premium=false' % video_id, video_id, headers={'x-authorization': f'Bearer {token.value}'}, fatal=False)
|
|
else:
|
|
self.report_warning(f'The token cookie is needed to extract video metadata. {self._LOGIN_HINTS["cookies"]}')
|
|
formats, subtitles = [], {}
|
|
src_json = self._download_json(f'{self._BASE_URL}abrjson_v2/tv_android/{video_id}', video_id)
|
|
for src in src_json['video_selector']:
|
|
if not src.get('url'):
|
|
continue
|
|
fmt, subs = self._extract_m3u8_formats_and_subtitles(src['url'], video_id, 'mp4')
|
|
formats.extend(fmt)
|
|
subtitles = self._merge_subtitles(subtitles, subs)
|
|
self._sort_formats(formats, ['tbr'])
|
|
|
|
return {
|
|
'id': video_id,
|
|
'title': json_info.get('ep_title'),
|
|
'series': json_info.get('lu_title'),
|
|
'series_id': series_id,
|
|
'description': json_info.get('ep_description'),
|
|
'formats': formats,
|
|
'subtitles': subtitles,
|
|
'thumbnail': f'{self._BASE_URL}img/program/{series_id}/episode/{video_id}_a.jpg',
|
|
}
|