[nba] extract all video formats and extract more info

pull/8/head
remitamine 9 years ago
parent 59a9efe85b
commit 8fc226ef99

@ -351,7 +351,10 @@ from .myvideo import MyVideoIE
from .myvidster import MyVidsterIE
from .nationalgeographic import NationalGeographicIE
from .naver import NaverIE
from .nba import NBAIE
from .nba import (
NBAIE,
NBAWatchIE,
)
from .nbc import (
NBCIE,
NBCNewsIE,

@ -2,62 +2,100 @@ from __future__ import unicode_literals
from .common import InfoExtractor
from ..utils import (
remove_end,
parse_duration,
parse_iso8601,
int_or_none,
)
class NBAIE(InfoExtractor):
_VALID_URL = r'https?://(?:watch\.|www\.)?nba\.com/(?:nba/)?video(?P<id>/[^?]*?)/?(?:/index\.html)?(?:\?.*)?$'
class NBABaseIE(InfoExtractor):
def _get_formats(self, video_id):
base_url = 'http://nba.cdn.turner.com/nba/big%s' % video_id
return [{
'url': base_url + '_nba_android_high.mp4',
'width': 480,
'height': 320,
'format_id': '320p',
},{
'url': base_url + '_640x360_664b.mp4',
'width': 640,
'height': 360,
'format_id': '360p',
},{
'url': base_url + '_768x432_1404.mp4',
'width': 768,
'height': 432,
'format_id': '432p',
},{
'url': base_url + '_1280x720.mp4',
'width': 1280,
'height': 720,
'format_id': '720p',
}]
def _real_extract(self, url):
video_id = self._match_id(url)
webpage = self._download_webpage(url, video_id)
ret = self._extract_metadata(webpage, video_id)
ret['id'] = video_id.rpartition('/')[2]
ret['formats'] = self._get_formats(video_id)
return ret
class NBAIE(NBABaseIE):
IE_NAME = 'nba'
_VALID_URL = r'https?://(?:www\.)?nba\.com/(?:nba/)?video(?P<id>/[^?]*?)/?(?:/index\.html)?(?:\?.*)?$'
_TESTS = [{
'url': 'http://www.nba.com/video/games/nets/2012/12/04/0021200253-okc-bkn-recap.nba/index.html',
'md5': 'c0edcfc37607344e2ff8f13c378c88a4',
'md5': '9d902940d2a127af3f7f9d2f3dc79c96',
'info_dict': {
'id': '0021200253-okc-bkn-recap.nba',
'ext': 'mp4',
'title': 'Thunder vs. Nets',
'description': 'Kevin Durant scores 32 points and dishes out six assists as the Thunder beat the Nets in Brooklyn.',
'duration': 181,
'timestamp': 1354680189,
'upload_date': '20121205',
},
}, {
'url': 'http://www.nba.com/video/games/hornets/2014/12/05/0021400276-nyk-cha-play5.nba/',
'only_matching': True,
}, {
}]
def _extract_metadata(self, webpage, video_id):
return {
'title': self._html_search_meta('name', webpage),
'description': self._html_search_meta('description', webpage),
'duration': parse_duration(self._html_search_meta('duration', webpage)),
'thumbnail': self._html_search_meta('thumbnailUrl', webpage),
'timestamp': parse_iso8601(self._html_search_meta('uploadDate', webpage))
}
class NBAWatchIE(NBABaseIE):
IE_NAME = 'nba:watch'
_VALID_URL = r'https?://watch.nba\.com/(?:nba/)?video(?P<id>/[^?]*?)/?(?:/index\.html)?(?:\?.*)?$'
_TESTS = [{
'url': 'http://watch.nba.com/nba/video/channels/playoffs/2015/05/20/0041400301-cle-atl-recap.nba',
'md5': 'b2b39b81cf28615ae0c3360a3f9668c4',
'info_dict': {
'id': '0041400301-cle-atl-recap.nba',
'ext': 'mp4',
'title': 'NBA GAME TIME | Video: Hawks vs. Cavaliers Game 1',
'title': 'Hawks vs. Cavaliers Game 1',
'description': 'md5:8094c3498d35a9bd6b1a8c396a071b4d',
'duration': 228,
},
'params': {
'skip_download': True,
'timestamp': 1432094400,
'upload_date': '20150520',
}
}]
def _real_extract(self, url):
video_id = self._match_id(url)
webpage = self._download_webpage(url, video_id)
video_url = 'http://ht-mobile.cdn.turner.com/nba/big' + video_id + '_nba_1280x720.mp4'
shortened_video_id = video_id.rpartition('/')[2]
title = remove_end(
self._og_search_title(webpage, default=shortened_video_id), ' : NBA.com')
description = self._og_search_description(webpage)
duration_str = self._html_search_meta(
'duration', webpage, 'duration', default=None)
if not duration_str:
duration_str = self._html_search_regex(
r'Duration:</b>\s*(\d+:\d+)', webpage, 'duration', fatal=False)
duration = parse_duration(duration_str)
def _extract_metadata(self, webpage, video_id):
program_id = self._search_regex(r'var\s+programId\s*=\s*(\d+);', webpage, 'program id')
metadata = self._download_json(
'http://smbsolr.cdnak.neulion.com/solr_nbav6/nba/nba/mlt/?wt=json&fl=name,description,image,runtime,releaseDate&q=sequence%3A' + program_id, video_id)['match']['docs'][0]
return {
'id': shortened_video_id,
'url': video_url,
'title': title,
'description': description,
'duration': duration,
'title': metadata['name'],
'description': metadata.get('description'),
'duration': int_or_none(metadata.get('runtime')),
'thumbnail': metadata.get('image'),
'timestamp': parse_iso8601(metadata.get('releaseDate'))
}

Loading…
Cancel
Save