from .common import InfoExtractor
from ..utils import (
- float_or_none,
- str_to_int,
parse_duration,
+ parse_filesize,
+ str_to_int,
)
'url': 'http://www.snotr.com/video/13708/Drone_flying_through_fireworks',
'info_dict': {
'id': '13708',
- 'ext': 'flv',
+ 'ext': 'mp4',
'title': 'Drone flying through fireworks!',
- 'duration': 247,
- 'filesize_approx': 98566144,
+ 'duration': 248,
+ 'filesize_approx': 40700000,
'description': 'A drone flying through Fourth of July Fireworks',
- }
+ 'thumbnail': 're:^https?://.*\.jpg$',
+ },
+ 'expected_warnings': ['description'],
}, {
'url': 'http://www.snotr.com/video/530/David_Letteman_-_George_W_Bush_Top_10',
'info_dict': {
'id': '530',
- 'ext': 'flv',
+ 'ext': 'mp4',
'title': 'David Letteman - George W. Bush Top 10',
'duration': 126,
- 'filesize_approx': 8912896,
+ 'filesize_approx': 8500000,
'description': 'The top 10 George W. Bush moments, brought to you by David Letterman!',
+ 'thumbnail': 're:^https?://.*\.jpg$',
}
}]
title = self._og_search_title(webpage)
description = self._og_search_description(webpage)
- video_url = 'http://cdn.videos.snotr.com/%s.flv' % video_id
+ info_dict = self._parse_html5_media_entries(
+ url, webpage, video_id, m3u8_entry_protocol='m3u8_native')[0]
view_count = str_to_int(self._html_search_regex(
- r'<p>\n<strong>Views:</strong>\n([\d,\.]+)</p>',
+ r'<p[^>]*>\s*<strong[^>]*>Views:</strong>\s*<span[^>]*>([\d,\.]+)',
webpage, 'view count', fatal=False))
duration = parse_duration(self._html_search_regex(
- r'<p>\n<strong>Length:</strong>\n\s*([0-9:]+).*?</p>',
+ r'<p[^>]*>\s*<strong[^>]*>Length:</strong>\s*<span[^>]*>([\d:]+)',
webpage, 'duration', fatal=False))
- filesize_approx = float_or_none(self._html_search_regex(
- r'<p>\n<strong>Filesize:</strong>\n\s*([0-9.]+)\s*megabyte</p>',
- webpage, 'filesize', fatal=False), invscale=1024 * 1024)
+ filesize_approx = parse_filesize(self._html_search_regex(
+ r'<p[^>]*>\s*<strong[^>]*>Filesize:</strong>\s*<span[^>]*>([^<]+)',
+ webpage, 'filesize', fatal=False))
- return {
+ info_dict.update({
'id': video_id,
'description': description,
'title': title,
- 'url': video_url,
'view_count': view_count,
'duration': duration,
'filesize_approx': filesize_approx,
- }
+ })
+
+ return info_dict