mirror of
https://github.com/ytdl-org/youtube-dl.git
synced 2024-11-23 19:35:19 +01:00
118 lines
3.8 KiB
Python
118 lines
3.8 KiB
Python
from __future__ import unicode_literals
|
|
|
|
from .common import InfoExtractor, ExtractorError
|
|
from ..utils import (
|
|
int_or_none,
|
|
float_or_none,
|
|
parse_iso8601,
|
|
)
|
|
|
|
|
|
class VidmeIE(InfoExtractor):
|
|
_VALID_URL = r'https?://vid\.me/(?:e/)?(?P<id>[\da-zA-Z]+)'
|
|
_TESTS = [{
|
|
'url': 'https://vid.me/QNB',
|
|
'md5': 'c62f1156138dc3323902188c5b5a8bd6',
|
|
'info_dict': {
|
|
'id': 'QNB',
|
|
'ext': 'mp4',
|
|
'title': 'Fishing for piranha - the easy way',
|
|
'description': 'source: https://www.facebook.com/photo.php?v=312276045600871',
|
|
'duration': 119.92,
|
|
'timestamp': 1406313244,
|
|
'upload_date': '20140725',
|
|
'thumbnail': 're:^https?://.*\.jpg',
|
|
'view_count': int,
|
|
'like_count': int,
|
|
'comment_count': int,
|
|
},
|
|
}, {
|
|
'url': 'https://vid.me/Gc6M',
|
|
'md5': 'f42d05e7149aeaec5c037b17e5d3dc82',
|
|
'info_dict': {
|
|
'id': 'Gc6M',
|
|
'ext': 'mp4',
|
|
'title': 'O Mere Dil ke chain - Arnav and Khushi VM',
|
|
'duration': 223.72,
|
|
'timestamp': 1441211642,
|
|
'upload_date': '20150902',
|
|
'thumbnail': 're:^https?://.*\.jpg',
|
|
'view_count': int,
|
|
'like_count': int,
|
|
'comment_count': int,
|
|
'comment_count': int,
|
|
},
|
|
'params': {
|
|
'skip_download': True,
|
|
},
|
|
}, {
|
|
# tests uploader field
|
|
'url': 'https://vid.me/4Iib',
|
|
'info_dict': {
|
|
'id': '4Iib',
|
|
'ext': 'mp4',
|
|
'title': 'The Carver',
|
|
'description': 'md5:e9c24870018ae8113be936645b93ba3c',
|
|
'duration': 97.859999999999999,
|
|
'timestamp': 1433203629,
|
|
'upload_date': '20150602',
|
|
'uploader': 'Thomas',
|
|
'thumbnail': 're:^https?://.*\.jpg',
|
|
'view_count': int,
|
|
'like_count': int,
|
|
'comment_count': int,
|
|
},
|
|
'params': {
|
|
'skip_download': True,
|
|
},
|
|
}, {
|
|
# From http://naked-yogi.tumblr.com/post/118312946248/naked-smoking-stretching
|
|
'url': 'https://vid.me/e/Wmur',
|
|
'only_matching': True,
|
|
}]
|
|
|
|
def _real_extract(self, url):
|
|
video_id = self._match_id(url)
|
|
api_url = 'https://api.vid.me/videoByUrl/' + video_id
|
|
data = self._download_json(api_url, video_id)
|
|
|
|
video_data = data.get('video')
|
|
if video_data is None:
|
|
raise ExtractorError('Could not extract the vid.me video data')
|
|
|
|
title = video_data.get('title')
|
|
description = video_data.get('description')
|
|
thumbnail = video_data.get('thumbnail_url')
|
|
timestamp = parse_iso8601(video_data.get('date_created'), ' ')
|
|
duration = float_or_none(video_data.get('duration'))
|
|
view_count = int_or_none(video_data.get('view_count'))
|
|
like_count = int_or_none(video_data.get('likes_count'))
|
|
comment_count = int_or_none(video_data.get('comment_count'))
|
|
|
|
uploader = None
|
|
user_data = video_data.get('user')
|
|
if user_data is not None:
|
|
uploader = user_data.get('username')
|
|
|
|
formats = [{
|
|
'format_id': format['type'],
|
|
'url': format['uri'],
|
|
'width': int_or_none(format['width']),
|
|
'height': int_or_none(format['height']),
|
|
} for format in video_data.get('formats', [])]
|
|
self._sort_formats(formats)
|
|
|
|
return {
|
|
'id': video_id,
|
|
'title': title,
|
|
'description': description,
|
|
'thumbnail': thumbnail,
|
|
'timestamp': timestamp,
|
|
'duration': duration,
|
|
'view_count': view_count,
|
|
'like_count': like_count,
|
|
'comment_count': comment_count,
|
|
'uploader': uploader,
|
|
'formats': formats,
|
|
}
|