[twitter:amplify] Extract more metadata

pull/2/head
Yen Chi Hsuan 9 years ago
parent acb6e97e6a
commit bdbf4ba40e

@ -243,6 +243,7 @@ class TwitterAmplifyIE(TwitterBaseIE):
'id': '0ba0c3c7-0af3-4c0a-bed5-7efd1ffa2951',
'ext': 'mp4',
'title': 'Twitter Video',
'thumbnail': 're:^https?://.*',
},
}
@ -254,8 +255,35 @@ class TwitterAmplifyIE(TwitterBaseIE):
'twitter:amplify:vmap', webpage, 'vmap url')
video_url = self._get_vmap_video_url(vmap_url, video_id)
thumbnails = []
thumbnail = self._html_search_meta(
'twitter:image:src', webpage, 'thumbnail', fatal=False)
def _find_dimension(target):
w = int_or_none(self._html_search_meta(
'twitter:%s:width' % target, webpage, fatal=False))
h = int_or_none(self._html_search_meta(
'twitter:%s:height' % target, webpage, fatal=False))
return w, h
if thumbnail:
thumbnail_w, thumbnail_h = _find_dimension('image')
thumbnails.append({
'url': thumbnail,
'width': thumbnail_w,
'height': thumbnail_h,
})
video_w, video_h = _find_dimension('player')
formats = [{
'url': video_url,
'width': video_w,
'height': video_h,
}]
return {
'id': video_id,
'title': 'Twitter Video',
'url': video_url,
'formats': formats,
'thumbnails': thumbnails,
}

Loading…
Cancel
Save