mirror of
				https://code.hackerspace.pl/q3k/youtube-dl
				synced 2025-03-16 11:43:02 +00:00 
			
		
		
		
	
		
			
				
	
	
		
			61 lines
		
	
	
		
			2.2 KiB
		
	
	
	
		
			Python
		
	
	
	
	
	
			
		
		
	
	
			61 lines
		
	
	
		
			2.2 KiB
		
	
	
	
		
			Python
		
	
	
	
	
	
| from __future__ import unicode_literals
 | |
| 
 | |
| from .common import InfoExtractor
 | |
| from ..utils import int_or_none
 | |
| 
 | |
| 
 | |
| class TestTubeIE(InfoExtractor):
 | |
|     _VALID_URL = r'https?://testtube\.com/[^/?#]+/(?P<id>[^/?#]+)'
 | |
|     _TESTS = [{
 | |
|         'url': 'https://testtube.com/dnews/5-weird-ways-plants-can-eat-animals?utm_source=FB&utm_medium=DNews&utm_campaign=DNewsSocial',
 | |
|         'info_dict': {
 | |
|             'id': '60163',
 | |
|             'display_id': '5-weird-ways-plants-can-eat-animals',
 | |
|             'duration': 275,
 | |
|             'ext': 'mp4',
 | |
|             'title': '5 Weird Ways Plants Can Eat Animals',
 | |
|             'description': 'Why have some plants evolved to eat meat?',
 | |
|             'thumbnail': 're:^https?://.*\.jpg$',
 | |
|             'uploader': 'DNews',
 | |
|             'uploader_id': 'dnews',
 | |
|         },
 | |
|     }]
 | |
| 
 | |
|     def _real_extract(self, url):
 | |
|         display_id = self._match_id(url)
 | |
| 
 | |
|         webpage = self._download_webpage(url, display_id)
 | |
|         video_id = self._search_regex(
 | |
|             r"player\.loadRevision3Item\('video_id',\s*([0-9]+)\);",
 | |
|             webpage, 'video ID')
 | |
| 
 | |
|         all_info = self._download_json(
 | |
|             'https://testtube.com/api/getPlaylist.json?api_key=ba9c741bce1b9d8e3defcc22193f3651b8867e62&codecs=h264,vp8,theora&video_id=%s' % video_id,
 | |
|             video_id)
 | |
|         info = all_info['items'][0]
 | |
| 
 | |
|         formats = []
 | |
|         for vcodec, fdatas in info['media'].items():
 | |
|             for name, fdata in fdatas.items():
 | |
|                 formats.append({
 | |
|                     'format_id': '%s-%s' % (vcodec, name),
 | |
|                     'url': fdata['url'],
 | |
|                     'vcodec': vcodec,
 | |
|                     'tbr': fdata.get('bitrate'),
 | |
|                 })
 | |
|         self._sort_formats(formats)
 | |
| 
 | |
|         duration = int_or_none(info.get('duration'))
 | |
| 
 | |
|         return {
 | |
|             'id': video_id,
 | |
|             'display_id': display_id,
 | |
|             'title': info['title'],
 | |
|             'description': info.get('summary'),
 | |
|             'thumbnail': info.get('images', {}).get('large'),
 | |
|             'uploader': info.get('show', {}).get('name'),
 | |
|             'uploader_id': info.get('show', {}).get('slug'),
 | |
|             'duration': duration,
 | |
|             'formats': formats,
 | |
|         }
 | 
