mirror of
https://github.com/yt-dlp/yt-dlp.git
synced 2024-11-18 06:09:17 +00:00
145 lines
5.5 KiB
Python
145 lines
5.5 KiB
Python
import base64
|
|
import json
|
|
import re
|
|
|
|
from .common import InfoExtractor
|
|
from ..utils import (
|
|
determine_ext,
|
|
int_or_none,
|
|
mimetype2ext,
|
|
qualities,
|
|
traverse_obj,
|
|
try_get,
|
|
url_or_none,
|
|
)
|
|
|
|
|
|
class ImdbIE(InfoExtractor):
|
|
IE_NAME = 'imdb'
|
|
IE_DESC = 'Internet Movie Database trailers'
|
|
_VALID_URL = r'https?://(?:www|m)\.imdb\.com/(?:video|title|list).*?[/-]vi(?P<id>\d+)'
|
|
|
|
_TESTS = [{
|
|
'url': 'http://www.imdb.com/video/imdb/vi2524815897',
|
|
'info_dict': {
|
|
'id': '2524815897',
|
|
'ext': 'mp4',
|
|
'title': 'No. 2',
|
|
'description': 'md5:87bd0bdc61e351f21f20d2d7441cb4e7',
|
|
'duration': 152,
|
|
'thumbnail': r're:^https?://.+\.jpg',
|
|
}
|
|
}, {
|
|
'url': 'https://www.imdb.com/video/vi3516832537',
|
|
'info_dict': {
|
|
'id': '3516832537',
|
|
'ext': 'mp4',
|
|
'title': 'Paul: U.S. Trailer #1',
|
|
'description': 'md5:17fcc4fe11ec29b4399be9d4c5ef126c',
|
|
'duration': 153,
|
|
'thumbnail': r're:^https?://.+\.jpg',
|
|
}
|
|
}, {
|
|
'url': 'http://www.imdb.com/video/_/vi2524815897',
|
|
'only_matching': True,
|
|
}, {
|
|
'url': 'http://www.imdb.com/title/tt1667889/?ref_=ext_shr_eml_vi#lb-vi2524815897',
|
|
'only_matching': True,
|
|
}, {
|
|
'url': 'http://www.imdb.com/title/tt1667889/#lb-vi2524815897',
|
|
'only_matching': True,
|
|
}, {
|
|
'url': 'http://www.imdb.com/videoplayer/vi1562949145',
|
|
'only_matching': True,
|
|
}, {
|
|
'url': 'http://www.imdb.com/title/tt4218696/videoplayer/vi2608641561',
|
|
'only_matching': True,
|
|
}, {
|
|
'url': 'https://www.imdb.com/list/ls009921623/videoplayer/vi260482329',
|
|
'only_matching': True,
|
|
}]
|
|
|
|
def _real_extract(self, url):
|
|
video_id = self._match_id(url)
|
|
webpage = self._download_webpage(f'https://www.imdb.com/video/vi{video_id}', video_id)
|
|
info = self._search_nextjs_data(webpage, video_id)
|
|
video_info = traverse_obj(info, ('props', 'pageProps', 'videoPlaybackData', 'video'), default={})
|
|
title = (traverse_obj(video_info, ('name', 'value'), ('primaryTitle', 'titleText', 'text'))
|
|
or self._html_search_meta(('og:title', 'twitter:title'), webpage, default=None)
|
|
or self._html_extract_title(webpage))
|
|
data = video_info.get('playbackURLs') or try_get(self._download_json(
|
|
'https://www.imdb.com/ve/data/VIDEO_PLAYBACK_DATA', video_id,
|
|
query={
|
|
'key': base64.b64encode(json.dumps({
|
|
'type': 'VIDEO_PLAYER',
|
|
'subType': 'FORCE_LEGACY',
|
|
'id': 'vi%s' % video_id,
|
|
}).encode()).decode(),
|
|
}), lambda x: x[0]['videoLegacyEncodings'])
|
|
quality = qualities(('SD', '480p', '720p', '1080p'))
|
|
formats, subtitles = [], {}
|
|
for encoding in data:
|
|
if not encoding or not isinstance(encoding, dict):
|
|
continue
|
|
video_url = url_or_none(encoding.get('url'))
|
|
if not video_url:
|
|
continue
|
|
ext = mimetype2ext(encoding.get(
|
|
'mimeType')) or determine_ext(video_url)
|
|
if ext == 'm3u8':
|
|
fmts, subs = self._extract_m3u8_formats_and_subtitles(
|
|
video_url, video_id, 'mp4', entry_protocol='m3u8_native',
|
|
preference=1, m3u8_id='hls', fatal=False)
|
|
subtitles = self._merge_subtitles(subtitles, subs)
|
|
formats.extend(fmts)
|
|
continue
|
|
format_id = traverse_obj(encoding, ('displayName', 'value'), 'definition')
|
|
formats.append({
|
|
'format_id': format_id,
|
|
'url': video_url,
|
|
'ext': ext,
|
|
'quality': quality(format_id),
|
|
})
|
|
|
|
return {
|
|
'id': video_id,
|
|
'title': title,
|
|
'alt_title': info.get('videoSubTitle'),
|
|
'formats': formats,
|
|
'description': try_get(video_info, lambda x: x['description']['value']),
|
|
'thumbnail': url_or_none(try_get(video_info, lambda x: x['thumbnail']['url'])),
|
|
'duration': int_or_none(try_get(video_info, lambda x: x['runtime']['value'])),
|
|
'subtitles': subtitles,
|
|
}
|
|
|
|
|
|
class ImdbListIE(InfoExtractor):
|
|
IE_NAME = 'imdb:list'
|
|
IE_DESC = 'Internet Movie Database lists'
|
|
_VALID_URL = r'https?://(?:www\.)?imdb\.com/list/ls(?P<id>\d{9})(?!/videoplayer/vi\d+)'
|
|
_TEST = {
|
|
'url': 'https://www.imdb.com/list/ls009921623/',
|
|
'info_dict': {
|
|
'id': '009921623',
|
|
'title': 'The Bourne Legacy',
|
|
'description': 'A list of trailers, clips, and more from The Bourne Legacy, starring Jeremy Renner and Rachel Weisz.',
|
|
},
|
|
'playlist_count': 8,
|
|
}
|
|
|
|
def _real_extract(self, url):
|
|
list_id = self._match_id(url)
|
|
webpage = self._download_webpage(url, list_id)
|
|
entries = [
|
|
self.url_result('http://www.imdb.com' + m, 'Imdb')
|
|
for m in re.findall(r'href="(/list/ls%s/videoplayer/vi[^"]+)"' % list_id, webpage)]
|
|
|
|
list_title = self._html_search_regex(
|
|
r'<h1[^>]+class="[^"]*header[^"]*"[^>]*>(.*?)</h1>',
|
|
webpage, 'list title')
|
|
list_description = self._html_search_regex(
|
|
r'<div[^>]+class="[^"]*list-description[^"]*"[^>]*><p>(.*?)</p>',
|
|
webpage, 'list description')
|
|
|
|
return self.playlist_result(entries, list_id, list_title, list_description)
|