1
0
mirror of https://github.com/l1ving/youtube-dl synced 2024-11-27 10:33:23 +08:00
youtube-dl/youtube_dl/extractor/screencast.py
Philipp Hagemeister 1cc79574fc Fix imports and general cleanup
· Import from compat what comes from compat. Yes, some names are available in utils too, but that's an implementation detail.
· Use _match_id consistently whenever possible
· Fix some outdated tests
· Use consistent valid URL (always match the whole protocol, no ^ at start required)
· Use modern test definitions
2014-12-13 12:35:45 +01:00

112 lines
4.1 KiB
Python

# -*- coding: utf-8 -*-
from __future__ import unicode_literals
from .common import InfoExtractor
from ..compat import (
compat_parse_qs,
compat_urllib_request,
)
from ..utils import (
ExtractorError,
)
class ScreencastIE(InfoExtractor):
_VALID_URL = r'https?://www\.screencast\.com/t/(?P<id>[a-zA-Z0-9]+)'
_TESTS = [{
'url': 'http://www.screencast.com/t/3ZEjQXlT',
'md5': '917df1c13798a3e96211dd1561fded83',
'info_dict': {
'id': '3ZEjQXlT',
'ext': 'm4v',
'title': 'Color Measurement with Ocean Optics Spectrometers',
'description': 'md5:240369cde69d8bed61349a199c5fb153',
'thumbnail': 're:^https?://.*\.(?:gif|jpg)$',
}
}, {
'url': 'http://www.screencast.com/t/V2uXehPJa1ZI',
'md5': 'e8e4b375a7660a9e7e35c33973410d34',
'info_dict': {
'id': 'V2uXehPJa1ZI',
'ext': 'mov',
'title': 'The Amadeus Spectrometer',
'description': 're:^In this video, our friends at.*To learn more about Amadeus, visit',
'thumbnail': 're:^https?://.*\.(?:gif|jpg)$',
}
}, {
'url': 'http://www.screencast.com/t/aAB3iowa',
'md5': 'dedb2734ed00c9755761ccaee88527cd',
'info_dict': {
'id': 'aAB3iowa',
'ext': 'mp4',
'title': 'Google Earth Export',
'description': 'Provides a demo of a CommunityViz export to Google Earth, one of the 3D viewing options.',
'thumbnail': 're:^https?://.*\.(?:gif|jpg)$',
}
}, {
'url': 'http://www.screencast.com/t/X3ddTrYh',
'md5': '669ee55ff9c51988b4ebc0877cc8b159',
'info_dict': {
'id': 'X3ddTrYh',
'ext': 'wmv',
'title': 'Toolkit 6 User Group Webinar (2014-03-04) - Default Judgment and First Impression',
'description': 'md5:7b9f393bc92af02326a5c5889639eab0',
'thumbnail': 're:^https?://.*\.(?:gif|jpg)$',
}
},
]
def _real_extract(self, url):
video_id = self._match_id(url)
webpage = self._download_webpage(url, video_id)
video_url = self._html_search_regex(
r'<embed name="Video".*?src="([^"]+)"', webpage,
'QuickTime embed', default=None)
if video_url is None:
flash_vars_s = self._html_search_regex(
r'<param name="flashVars" value="([^"]+)"', webpage, 'flash vars',
default=None)
if not flash_vars_s:
flash_vars_s = self._html_search_regex(
r'<param name="initParams" value="([^"]+)"', webpage, 'flash vars',
default=None)
if flash_vars_s:
flash_vars_s = flash_vars_s.replace(',', '&')
if flash_vars_s:
flash_vars = compat_parse_qs(flash_vars_s)
video_url_raw = compat_urllib_request.quote(
flash_vars['content'][0])
video_url = video_url_raw.replace('http%3A', 'http:')
if video_url is None:
video_meta = self._html_search_meta(
'og:video', webpage, default=None)
if video_meta:
video_url = self._search_regex(
r'src=(.*?)(?:$|&)', video_meta,
'meta tag video URL', default=None)
if video_url is None:
raise ExtractorError('Cannot find video')
title = self._og_search_title(webpage, default=None)
if title is None:
title = self._html_search_regex(
[r'<b>Title:</b> ([^<]*)</div>',
r'class="tabSeperator">></span><span class="tabText">(.*?)<'],
webpage, 'title')
thumbnail = self._og_search_thumbnail(webpage)
description = self._og_search_description(webpage, default=None)
if description is None:
description = self._html_search_meta('description', webpage)
return {
'id': video_id,
'url': video_url,
'title': title,
'description': description,
'thumbnail': thumbnail,
}