mirror of
https://github.com/ytdl-org/youtube-dl.git
synced 2024-12-21 03:56:05 +01:00
80 lines
2.7 KiB
Python
80 lines
2.7 KiB
Python
|
# coding: utf-8
|
||
|
from __future__ import unicode_literals
|
||
|
|
||
|
from .common import InfoExtractor
|
||
|
from ..utils import (
|
||
|
determine_ext,
|
||
|
int_or_none,
|
||
|
merge_dicts,
|
||
|
parse_iso8601,
|
||
|
T,
|
||
|
traverse_obj,
|
||
|
txt_or_none,
|
||
|
urljoin,
|
||
|
)
|
||
|
|
||
|
|
||
|
class CaffeineTVIE(InfoExtractor):
|
||
|
_VALID_URL = r'https?://(?:www\.)?caffeine\.tv/[^/]+/video/(?P<id>[0-9a-f-]+)'
|
||
|
_TESTS = [{
|
||
|
'url': 'https://www.caffeine.tv/TsuSurf/video/cffc0a00-e73f-11ec-8080-80017d29f26e',
|
||
|
'info_dict': {
|
||
|
'id': 'cffc0a00-e73f-11ec-8080-80017d29f26e',
|
||
|
'ext': 'mp4',
|
||
|
'title': 'GOOOOD MORNINNNNN #highlights',
|
||
|
'timestamp': 1654702180,
|
||
|
'upload_date': '20220608',
|
||
|
'uploader': 'TsuSurf',
|
||
|
'duration': 3145,
|
||
|
'age_limit': 17,
|
||
|
},
|
||
|
'params': {
|
||
|
'format': 'bestvideo',
|
||
|
},
|
||
|
}]
|
||
|
|
||
|
def _real_extract(self, url):
|
||
|
video_id = self._match_id(url)
|
||
|
json_data = self._download_json(
|
||
|
'https://api.caffeine.tv/social/public/activity/' + video_id,
|
||
|
video_id)
|
||
|
broadcast_info = traverse_obj(json_data, ('broadcast_info', T(dict))) or {}
|
||
|
title = broadcast_info['broadcast_title']
|
||
|
video_url = broadcast_info['video_url']
|
||
|
|
||
|
ext = determine_ext(video_url)
|
||
|
if ext == 'm3u8':
|
||
|
formats = self._extract_m3u8_formats(
|
||
|
video_url, video_id, 'mp4', entry_protocol='m3u8',
|
||
|
fatal=False)
|
||
|
else:
|
||
|
formats = [{'url': video_url}]
|
||
|
self._sort_formats(formats)
|
||
|
|
||
|
return merge_dicts({
|
||
|
'id': video_id,
|
||
|
'title': title,
|
||
|
'formats': formats,
|
||
|
}, traverse_obj(json_data, {
|
||
|
'uploader': ((None, 'user'), 'username'),
|
||
|
}, get_all=False), traverse_obj(json_data, {
|
||
|
'like_count': ('like_count', T(int_or_none)),
|
||
|
'view_count': ('view_count', T(int_or_none)),
|
||
|
'comment_count': ('comment_count', T(int_or_none)),
|
||
|
'tags': ('tags', Ellipsis, T(txt_or_none)),
|
||
|
'is_live': 'is_live',
|
||
|
'uploader': ('user', 'name'),
|
||
|
}), traverse_obj(broadcast_info, {
|
||
|
'duration': ('content_duration', T(int_or_none)),
|
||
|
'timestamp': ('broadcast_start_time', T(parse_iso8601)),
|
||
|
'thumbnail': ('preview_image_path', T(lambda u: urljoin(url, u))),
|
||
|
'age_limit': ('content_rating', T(lambda r: r and {
|
||
|
# assume Apple Store ratings [1]
|
||
|
# 1. https://en.wikipedia.org/wiki/Mobile_software_content_rating_system
|
||
|
'FOUR_PLUS': 0,
|
||
|
'NINE_PLUS': 9,
|
||
|
'TWELVE_PLUS': 12,
|
||
|
'SEVENTEEN_PLUS': 17,
|
||
|
}.get(r, 17))),
|
||
|
}))
|