|
|
@ -70,3 +70,47 @@ class TwitterCardIE(InfoExtractor): |
|
|
|
'duration': duration, |
|
|
|
'formats': formats, |
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
class TwitterIE(TwitterCardIE): |
|
|
|
_VALID_URL = r'https?://(?:www|m|mobile)?\.?twitter\.com/(?P<id>[^/]+/status/\d+)' |
|
|
|
|
|
|
|
_TESTS = [{ |
|
|
|
'url': 'https://m.twitter.com/thereaIbanksy/status/614301758345490432', |
|
|
|
'md5': '8bbccb487bd7a31349b775915fcd412f', |
|
|
|
'info_dict': { |
|
|
|
'id': '614301758345490432', |
|
|
|
'ext': 'mp4', |
|
|
|
'title': 'thereaIbanksy - This time lapse is so pretty \U0001f60d\U0001f60d', |
|
|
|
'thumbnail': 're:^https?://.*\.jpg', |
|
|
|
'duration': 29.5, |
|
|
|
'description': 'banksy on Twitter: "This time lapse is so pretty \U0001f60d\U0001f60d http://t.co/QB8DDbqiR1"', |
|
|
|
'uploader': 'banksy', |
|
|
|
'uploader_id': 'thereaIbanksy', |
|
|
|
}, |
|
|
|
}] |
|
|
|
|
|
|
|
def _real_extract(self, url): |
|
|
|
id = self._match_id(url) |
|
|
|
username, twid = re.match(r'([^/]+)/status/(\d+)', id).groups() |
|
|
|
name = username |
|
|
|
url = re.sub(r'https?://(m|mobile)\.', 'https://', url) |
|
|
|
webpage = self._download_webpage(url, 'tweet: ' + url) |
|
|
|
description = unescapeHTML(self._search_regex('<title>\s*(.+?)\s*</title>', webpage, 'title')) |
|
|
|
title = description.replace('\n', ' ') |
|
|
|
splitdesc = re.match(r'^(.+?)\s*on Twitter:\s* "(.+?)"$', title) |
|
|
|
if splitdesc: |
|
|
|
name, title = splitdesc.groups() |
|
|
|
title = re.sub(r'\s*https?://[^ ]+', '', title) # strip 'https -_t.co_BJYgOjSeGA' junk from filenames |
|
|
|
card_id = self._search_regex(r'["\']/i/cards/tfw/v1/(\d+)', webpage, '/i/card/...') |
|
|
|
card_url = 'https://twitter.com/i/cards/tfw/v1/' + card_id |
|
|
|
return { |
|
|
|
'_type': 'url_transparent', |
|
|
|
'ie_key': 'TwitterCard', |
|
|
|
'uploader_id': username, |
|
|
|
'uploader': name, |
|
|
|
'url': card_url, |
|
|
|
'webpage_url': url, |
|
|
|
'description': description, |
|
|
|
'title': username + ' - ' + title, |
|
|
|
} |