mirror of
https://github.com/yt-dlp/yt-dlp.git
synced 2024-10-31 16:27:25 -04:00
82 lines
2.7 KiB
Python
82 lines
2.7 KiB
Python
# -*- coding: utf-8 -*-
|
|
from __future__ import unicode_literals
|
|
|
|
import re
|
|
import json
|
|
|
|
from .common import InfoExtractor
|
|
from ..utils import (
|
|
int_or_none,
|
|
compat_str,
|
|
)
|
|
|
|
|
|
class StreamCZIE(InfoExtractor):
|
|
_VALID_URL = r'https?://(?:www\.)?stream\.cz/.+/(?P<videoid>.+)'
|
|
|
|
_TESTS = [{
|
|
'url': 'http://www.stream.cz/peklonataliri/765767-ecka-pro-deti',
|
|
'md5': '6d3ca61a8d0633c9c542b92fcb936b0c',
|
|
'info_dict': {
|
|
'id': '765767',
|
|
'ext': 'mp4',
|
|
'title': 'Peklo na talíři: Éčka pro děti',
|
|
'description': 'md5:49ace0df986e95e331d0fe239d421519',
|
|
'thumbnail': 'http://im.stream.cz/episode/52961d7e19d423f8f06f0100',
|
|
'duration': 256,
|
|
},
|
|
}, {
|
|
'url': 'http://www.stream.cz/blanik/10002447-tri-roky-pro-mazanka',
|
|
'md5': '246272e753e26bbace7fcd9deca0650c',
|
|
'info_dict': {
|
|
'id': '10002447',
|
|
'ext': 'mp4',
|
|
'title': 'Kancelář Blaník: Tři roky pro Mazánka',
|
|
'description': 'md5:9177695a8b756a0a8ab160de4043b392',
|
|
'thumbnail': 'http://im.stream.cz/episode/537f838c50c11f8d21320000',
|
|
'duration': 368,
|
|
},
|
|
}]
|
|
|
|
def _real_extract(self, url):
|
|
mobj = re.match(self._VALID_URL, url)
|
|
video_id = mobj.group('videoid')
|
|
|
|
webpage = self._download_webpage(url, video_id)
|
|
|
|
data = self._html_search_regex(r'Stream\.Data\.Episode\((.+?)\);', webpage, 'stream data')
|
|
|
|
jsonData = json.loads(data)
|
|
|
|
formats = []
|
|
for video in jsonData['instances']:
|
|
for video_format in video['instances']:
|
|
format_id = video_format['quality']
|
|
|
|
if format_id == '240p':
|
|
quality = 0
|
|
elif format_id == '360p':
|
|
quality = 1
|
|
elif format_id == '480p':
|
|
quality = 2
|
|
elif format_id == '720p':
|
|
quality = 3
|
|
|
|
formats.append({
|
|
'format_id': '%s-%s' % (video_format['type'].split('/')[1], format_id),
|
|
'url': video_format['source'],
|
|
'quality': quality,
|
|
})
|
|
|
|
self._sort_formats(formats)
|
|
|
|
return {
|
|
'id': compat_str(jsonData['episode_id']),
|
|
'title': self._og_search_title(webpage),
|
|
'thumbnail': jsonData['episode_image_original_url'].replace('//', 'http://'),
|
|
'formats': formats,
|
|
'description': self._og_search_description(webpage),
|
|
'duration': int_or_none(jsonData['duration']),
|
|
'view_count': int_or_none(jsonData['stats_total']),
|
|
}
|