yt-dlp/yt_dlp/extractor/learningonscreen.py

73 lines
2.7 KiB
Python

import functools
import re
from .common import InfoExtractor
from ..utils import (
ExtractorError,
clean_html,
extract_attributes,
join_nonempty,
parse_duration,
unified_timestamp,
)
from ..utils.traversal import find_element, traverse_obj
class LearningOnScreenIE(InfoExtractor):
_VALID_URL = r'https?://learningonscreen\.ac\.uk/ondemand/index\.php/prog/(?P<id>\w+)'
_TESTS = [{
'url': 'https://learningonscreen.ac.uk/ondemand/index.php/prog/005D81B2?bcast=22757013',
'info_dict': {
'id': '005D81B2',
'ext': 'mp4',
'title': 'Planet Earth',
'duration': 3600.0,
'timestamp': 1164567600.0,
'upload_date': '20061126',
'thumbnail': 'https://stream.learningonscreen.ac.uk/trilt-cover-images/005D81B2-Planet-Earth-2006-11-26T190000Z-BBC4.jpg',
},
}]
def _real_initialize(self):
if not self._get_cookies('https://learningonscreen.ac.uk/').get('PHPSESSID-BOB-LIVE'):
self.raise_login_required(method='session_cookies')
def _real_extract(self, url):
video_id = self._match_id(url)
webpage = self._download_webpage(url, video_id)
details = traverse_obj(webpage, (
{find_element(id='programme-details', html=True)}, {
'title': ({find_element(tag='h2')}, {clean_html}),
'timestamp': (
{find_element(cls='broadcast-date')},
{functools.partial(re.match, r'([^<]+)')}, 1, {unified_timestamp}),
'duration': (
{find_element(cls='prog-running-time')}, {clean_html}, {parse_duration}),
}))
title = details.pop('title', None) or traverse_obj(webpage, (
{find_element(id='add-to-existing-playlist', html=True)},
{extract_attributes}, 'data-record-title', {clean_html}))
entries = self._parse_html5_media_entries(
'https://stream.learningonscreen.ac.uk', webpage, video_id, m3u8_id='hls', mpd_id='dash',
_headers={'Origin': 'https://learningonscreen.ac.uk', 'Referer': 'https://learningonscreen.ac.uk/'})
if not entries:
raise ExtractorError('No video found')
if len(entries) > 1:
duration = details.pop('duration', None)
for idx, entry in enumerate(entries, start=1):
entry.update(details)
entry['id'] = join_nonempty(video_id, idx)
entry['title'] = join_nonempty(title, idx)
return self.playlist_result(entries, video_id, title, duration=duration)
return {
**entries[0],
**details,
'id': video_id,
'title': title,
}