|
|
from __future__ import unicode_literals
import re
from .common import InfoExtractor
from ..compat import ( compat_str, compat_urllib_request, compat_urlparse, ) from ..utils import ( clean_html, int_or_none, parse_iso8601, unescapeHTML, )
class BlipTVIE(InfoExtractor): _VALID_URL = r'https?://(?:\w+\.)?blip\.tv/(?:(?:.+-|rss/flash/)(?P<id>\d+)|((?:play/|api\.swf#)(?P<lookup_id>[\da-zA-Z+_]+)))'
_TESTS = [ { 'url': 'http://blip.tv/cbr/cbr-exclusive-gotham-city-imposters-bats-vs-jokerz-short-3-5796352', 'md5': 'c6934ad0b6acf2bd920720ec888eb812', 'info_dict': { 'id': '5779306', 'ext': 'mov', 'title': 'CBR EXCLUSIVE: "Gotham City Imposters" Bats VS Jokerz Short 3', 'description': 'md5:9bc31f227219cde65e47eeec8d2dc596', 'timestamp': 1323138843, 'upload_date': '20111206', 'uploader': 'cbr', 'uploader_id': '679425', 'duration': 81, } }, { # https://github.com/rg3/youtube-dl/pull/2274 'note': 'Video with subtitles', 'url': 'http://blip.tv/play/h6Uag5OEVgI.html', 'md5': '309f9d25b820b086ca163ffac8031806', 'info_dict': { 'id': '6586561', 'ext': 'mp4', 'title': 'Red vs. Blue Season 11 Episode 1', 'description': 'One-Zero-One', 'timestamp': 1371261608, 'upload_date': '20130615', 'uploader': 'redvsblue', 'uploader_id': '792887', 'duration': 279, } }, { # https://bugzilla.redhat.com/show_bug.cgi?id=967465 'url': 'http://a.blip.tv/api.swf#h6Uag5KbVwI', 'md5': '314e87b1ebe7a48fcbfdd51b791ce5a6', 'info_dict': { 'id': '6573122', 'ext': 'mov', 'upload_date': '20130520', 'description': 'Two hapless space marines argue over what to do when they realize they have an astronomically huge problem on their hands.', 'title': 'Red vs. Blue Season 11 Trailer', 'timestamp': 1369029609, 'uploader': 'redvsblue', 'uploader_id': '792887', } }, { 'url': 'http://blip.tv/play/gbk766dkj4Yn', 'md5': 'fe0a33f022d49399a241e84a8ea8b8e3', 'info_dict': { 'id': '1749452', 'ext': 'mp4', 'upload_date': '20090208', 'description': 'Witness the first appearance of the Nostalgia Critic character, as Doug reviews the movie Transformers.', 'title': 'Nostalgia Critic: Transformers', 'timestamp': 1234068723, 'uploader': 'NostalgiaCritic', 'uploader_id': '246467', } }, { # https://github.com/rg3/youtube-dl/pull/4404 'note': 'Audio only', 'url': 'http://blip.tv/hilarios-productions/weekly-manga-recap-kingdom-7119982', 'md5': '76c0a56f24e769ceaab21fbb6416a351', 'info_dict': { 'id': '7103299', 'ext': 'flv', 'title': 'Weekly Manga Recap: Kingdom', 'description': 'And then Shin breaks the enemy line, and he's all like HWAH! And then he slices a guy and it's all like FWASHING! And... it's really hard to describe the best parts of this series without breaking down into sound effects, okay?', 'timestamp': 1417660321, 'upload_date': '20141204', 'uploader': 'The Rollo T', 'uploader_id': '407429', 'duration': 7251, 'vcodec': 'none', } }, ]
@staticmethod def _extract_url(webpage): mobj = re.search(r'<meta\s[^>]*https?://api\.blip\.tv/\w+/redirect/\w+/(\d+)', webpage) if mobj: return 'http://blip.tv/a/a-' + mobj.group(1) mobj = re.search(r'<(?:iframe|embed|object)\s[^>]*(https?://(?:\w+\.)?blip\.tv/(?:play/|api\.swf#)[a-zA-Z0-9_]+)', webpage) if mobj: return mobj.group(1)
def _real_extract(self, url): mobj = re.match(self._VALID_URL, url) lookup_id = mobj.group('lookup_id')
# See https://github.com/rg3/youtube-dl/issues/857 and # https://github.com/rg3/youtube-dl/issues/4197 if lookup_id: urlh = self._request_webpage( 'http://blip.tv/play/%s' % lookup_id, lookup_id, 'Resolving lookup id') url = compat_urlparse.urlparse(urlh.geturl()) qs = compat_urlparse.parse_qs(url.query) mobj = re.match(self._VALID_URL, qs['file'][0])
video_id = mobj.group('id')
rss = self._download_xml('http://blip.tv/rss/flash/%s' % video_id, video_id, 'Downloading video RSS')
def blip(s): return '{http://blip.tv/dtd/blip/1.0}%s' % s
def media(s): return '{http://search.yahoo.com/mrss/}%s' % s
def itunes(s): return '{http://www.itunes.com/dtds/podcast-1.0.dtd}%s' % s
item = rss.find('channel/item')
video_id = item.find(blip('item_id')).text title = item.find('./title').text description = clean_html(compat_str(item.find(blip('puredescription')).text)) timestamp = parse_iso8601(item.find(blip('datestamp')).text) uploader = item.find(blip('user')).text uploader_id = item.find(blip('userid')).text duration = int(item.find(blip('runtime')).text) media_thumbnail = item.find(media('thumbnail')) thumbnail = media_thumbnail.get('url') if media_thumbnail is not None else item.find(itunes('image')).text categories = [category.text for category in item.findall('category')]
formats = [] subtitles_urls = {}
media_group = item.find(media('group')) for media_content in media_group.findall(media('content')): url = media_content.get('url') role = media_content.get(blip('role')) msg = self._download_webpage( url + '?showplayer=20140425131715&referrer=http://blip.tv&mask=7&skin=flashvars&view=url', video_id, 'Resolving URL for %s' % role) real_url = compat_urlparse.parse_qs(msg.strip())['message'][0]
media_type = media_content.get('type') if media_type == 'text/srt' or url.endswith('.srt'): LANGS = { 'english': 'en', } lang = role.rpartition('-')[-1].strip().lower() langcode = LANGS.get(lang, lang) subtitles_urls[langcode] = url elif media_type.startswith('video/'): formats.append({ 'url': real_url, 'format_id': role, 'format_note': media_type, 'vcodec': media_content.get(blip('vcodec')) or 'none', 'acodec': media_content.get(blip('acodec')), 'filesize': media_content.get('filesize'), 'width': int_or_none(media_content.get('width')), 'height': int_or_none(media_content.get('height')), }) self._check_formats(formats, video_id) self._sort_formats(formats)
subtitles = self.extract_subtitles(video_id, subtitles_urls)
return { 'id': video_id, 'title': title, 'description': description, 'timestamp': timestamp, 'uploader': uploader, 'uploader_id': uploader_id, 'duration': duration, 'thumbnail': thumbnail, 'categories': categories, 'formats': formats, 'subtitles': subtitles, }
def _get_subtitles(self, video_id, subtitles_urls): subtitles = {} for lang, url in subtitles_urls.items(): # For some weird reason, blip.tv serves a video instead of subtitles # when we request with a common UA req = compat_urllib_request.Request(url) req.add_header('User-Agent', 'youtube-dl') subtitles[lang] = [{ # The extension is 'srt' but it's actually an 'ass' file 'ext': 'ass', 'data': self._download_webpage(req, None, note=False), }] return subtitles
class BlipTVUserIE(InfoExtractor): _VALID_URL = r'(?:(?:https?://(?:\w+\.)?blip\.tv/)|bliptvuser:)(?!api\.swf)([^/]+)/*$' _PAGE_SIZE = 12 IE_NAME = 'blip.tv:user' _TEST = { 'url': 'http://blip.tv/actone', 'info_dict': { 'id': 'actone', 'title': 'Act One: The Series', }, 'playlist_count': 5, }
def _real_extract(self, url): mobj = re.match(self._VALID_URL, url) username = mobj.group(1)
page_base = 'http://m.blip.tv/pr/show_get_full_episode_list?users_id=%s&lite=0&esi=1'
page = self._download_webpage(url, username, 'Downloading user page') mobj = re.search(r'data-users-id="([^"]+)"', page) page_base = page_base % mobj.group(1) title = self._og_search_title(page)
# Download video ids using BlipTV Ajax calls. Result size per # query is limited (currently to 12 videos) so we need to query # page by page until there are no video ids - it means we got # all of them.
video_ids = [] pagenum = 1
while True: url = page_base + "&page=" + str(pagenum) page = self._download_webpage( url, username, 'Downloading video ids from page %d' % pagenum)
# Extract video identifiers ids_in_page = []
for mobj in re.finditer(r'href="/([^"]+)"', page): if mobj.group(1) not in ids_in_page: ids_in_page.append(unescapeHTML(mobj.group(1)))
video_ids.extend(ids_in_page)
# A little optimization - if current page is not # "full", ie. does not contain PAGE_SIZE video ids then # we can assume that this page is the last one - there # are no more ids on further pages - no need to query # again.
if len(ids_in_page) < self._PAGE_SIZE: break
pagenum += 1
urls = ['http://blip.tv/%s' % video_id for video_id in video_ids] url_entries = [self.url_result(vurl, 'BlipTV') for vurl in urls] return self.playlist_result( url_entries, playlist_title=title, playlist_id=username)
|