aboutsummaryrefslogtreecommitdiff
path: root/youtube_dl
diff options
context:
space:
mode:
authorremitamine <remitamine@gmail.com>2016-03-05 12:14:48 +0100
committerremitamine <remitamine@gmail.com>2016-03-05 12:14:48 +0100
commit18b0b23992b5ee4042967b63e366c937484e434e (patch)
treea14cb2e234a1fa341ac277f59b7fe5c84b6108c6 /youtube_dl
parentd1b29d13421fd91093fdf3699aea17d8ba4b636b (diff)
[revision3] add support pages of type embed
Diffstat (limited to 'youtube_dl')
-rw-r--r--youtube_dl/extractor/revision3.py66
1 files changed, 49 insertions, 17 deletions
diff --git a/youtube_dl/extractor/revision3.py b/youtube_dl/extractor/revision3.py
index b1b8800b9..fb2196ad6 100644
--- a/youtube_dl/extractor/revision3.py
+++ b/youtube_dl/extractor/revision3.py
@@ -52,6 +52,20 @@ class Revision3IE(InfoExtractor):
'uploader': 'DNews',
'uploader_id': 'dnews',
},
+ }, {
+ 'url': 'http://testtube.com/tt-editors-picks/the-israel-palestine-conflict-explained-in-ten-min',
+ 'info_dict': {
+ 'id': '73573',
+ 'ext': 'mp4',
+ 'display_id': 'tt-editors-picks/the-israel-palestine-conflict-explained-in-ten-min',
+ 'title': 'The Israel-Palestine Conflict Explained in Ten Minutes',
+ 'description': 'If you\'d like to learn about the struggle between Israelis and Palestinians, this video is a great place to start',
+ 'uploader': 'Editors\' Picks',
+ 'uploader_id': 'tt-editors-picks',
+ 'timestamp': 1453309200,
+ 'upload_date': '20160120',
+ },
+ 'add_ie': ['Youtube'],
}]
_PAGE_DATA_TEMPLATE = 'http://www.%s/apiProxy/ddn/%s?domain=%s'
_API_KEY = 'ba9c741bce1b9d8e3defcc22193f3651b8867e62'
@@ -61,9 +75,38 @@ class Revision3IE(InfoExtractor):
page_info = self._download_json(
self._PAGE_DATA_TEMPLATE % (domain, display_id, domain), display_id)
- if page_info['data']['type'] == 'episode':
- episode_data = page_info['data']
- video_id = compat_str(episode_data['video']['data']['id'])
+ page_data = page_info['data']
+ page_type = page_data['type']
+ if page_type == 'episode' or page_type == 'embed':
+ show_data = page_data['show']['data']
+ video_id = compat_str(page_data['video']['data']['id'])
+
+ preference = qualities(['mini', 'small', 'medium', 'large'])
+ thumbnails = [{
+ 'url': image_url,
+ 'id': image_id,
+ 'preference': preference(image_id)
+ } for image_id, image_url in page_data.get('images', {}).items()]
+
+ info = {
+ 'id': video_id,
+ 'display_id': display_id,
+ 'title': unescapeHTML(page_data['name']),
+ 'description': unescapeHTML(page_data.get('summary')),
+ 'timestamp': parse_iso8601(page_data.get('publishTime'), ' '),
+ 'author': page_data.get('author'),
+ 'uploader': show_data.get('name'),
+ 'uploader_id': show_data.get('slug'),
+ 'thumbnails': thumbnails,
+ }
+
+ if page_type == 'embed':
+ info.update({
+ '_type': 'url_transparent',
+ 'url': page_data['video']['data']['embed'],
+ })
+ return info
+
video_data = self._download_json(
'http://revision3.com/api/getPlaylist.json?api_key=%s&codecs=h264,vp8,theora&video_id=%s' % (self._API_KEY, video_id),
video_id)['items'][0]
@@ -84,26 +127,15 @@ class Revision3IE(InfoExtractor):
})
self._sort_formats(formats)
- preference = qualities(['mini', 'small', 'medium', 'large'])
- thumbnails = [{
- 'url': image_url,
- 'id': image_id,
- 'preference': preference(image_id)
- } for image_id, image_url in video_data.get('images', {}).items()]
-
- return {
- 'id': video_id,
- 'display_id': display_id,
+ info.update({
'title': unescapeHTML(video_data['title']),
'description': unescapeHTML(video_data.get('summary')),
- 'timestamp': parse_iso8601(episode_data.get('publishTime'), ' '),
- 'author': episode_data.get('author'),
'uploader': video_data.get('show', {}).get('name'),
'uploader_id': video_data.get('show', {}).get('slug'),
'duration': int_or_none(video_data.get('duration')),
- 'thumbnails': thumbnails,
'formats': formats,
- }
+ })
+ return info
else:
show_data = page_info['show']['data']
episodes_data = page_info['episodes']['data']