[extractor/generic] Simplify upload_date extraction

This commit is contained in:
Sergey M․ 2016-03-18 22:41:16 +06:00
parent 20938f768b
commit 303dcdb995

View File

@ -1259,14 +1259,13 @@ class GenericIE(InfoExtractor):
info_dict = {
'id': video_id,
'title': compat_urllib_parse_unquote(os.path.splitext(url_basename(url))[0]),
'upload_date': unified_strdate(head_response.headers.get('Last-Modified'))
}
# Check for direct link to a video
content_type = head_response.headers.get('Content-Type', '').lower()
m = re.match(r'^(?P<type>audio|video|application(?=/(?:ogg$|(?:vnd\.apple\.|x-)?mpegurl)))/(?P<format_id>[^;\s]+)', content_type)
if m:
upload_date = unified_strdate(
head_response.headers.get('Last-Modified'))
format_id = m.group('format_id')
if format_id.endswith('mpegurl'):
formats = self._extract_m3u8_formats(url, video_id, 'mp4')
@ -1281,7 +1280,6 @@ class GenericIE(InfoExtractor):
info_dict.update({
'direct': True,
'formats': formats,
'upload_date': upload_date,
})
return info_dict
@ -1309,12 +1307,9 @@ class GenericIE(InfoExtractor):
if not is_html(first_bytes):
self._downloader.report_warning(
'URL could be a direct video link, returning it as such.')
upload_date = unified_strdate(
head_response.headers.get('Last-Modified'))
info_dict.update({
'direct': True,
'url': url,
'upload_date': upload_date,
})
return info_dict