import re
from .common import InfoExtractor
+from ..compat import compat_str
from ..utils import (
xpath_text,
int_or_none,
tokens = {}
urls = []
formats = []
- rex = re.compile(r'''(?x)
- (?P<width>[0-9]+)x(?P<height>[0-9]+)
- (?:_(?P<bitrate>[0-9]+))?
- ''')
- for video_file in video_data.findall('files/file'):
+ rex = re.compile(
+ r'(?P<width>[0-9]+)x(?P<height>[0-9]+)(?:_(?P<bitrate>[0-9]+))?')
+ # Possible formats locations: files/file, files/groupFiles/files
+ # and maybe others
+ for video_file in video_data.findall('.//file'):
video_url = video_file.text.strip()
if not video_url:
continue
if video_url in urls:
continue
urls.append(video_url)
- format_id = video_file.attrib['bitrate']
+ format_id = video_file.get('bitrate')
if ext == 'smil':
- formats.extend(self._extract_smil_formats(video_url, video_id, fatal=False))
+ formats.extend(self._extract_smil_formats(
+ video_url, video_id, fatal=False))
elif ext == 'm3u8':
m3u8_formats = self._extract_m3u8_formats(
- video_url, video_id, 'mp4', m3u8_id=format_id, fatal=False)
+ video_url, video_id, 'mp4', m3u8_id=format_id or 'hls',
+ fatal=False)
if m3u8_formats:
# Sometimes final URLs inside m3u8 are unsigned, let's fix this
# ourselves
elif ext == 'f4m':
formats.extend(self._extract_f4m_formats(
update_url_query(video_url, {'hdcore': '3.7.0'}),
- video_id, f4m_id=format_id, fatal=False))
+ video_id, f4m_id=format_id or 'hds', fatal=False))
else:
f = {
'format_id': format_id,
'height': int(mobj.group('height')),
'tbr': int_or_none(mobj.group('bitrate')),
})
- elif format_id.isdigit():
- f['tbr'] = int(format_id)
- else:
- mobj = re.match(r'ios_(audio|[0-9]+)$', format_id)
- if mobj:
- if mobj.group(1) == 'audio':
- f.update({
- 'vcodec': 'none',
- 'ext': 'm4a',
- })
- else:
- f['tbr'] = int(mobj.group(1))
+ elif isinstance(format_id, compat_str):
+ if format_id.isdigit():
+ f['tbr'] = int(format_id)
+ else:
+ mobj = re.match(r'ios_(audio|[0-9]+)$', format_id)
+ if mobj:
+ if mobj.group(1) == 'audio':
+ f.update({
+ 'vcodec': 'none',
+ 'ext': 'm4a',
+ })
+ else:
+ f['tbr'] = int(mobj.group(1))
formats.append(f)
self._sort_formats(formats)