people suggest to put something on your hands like fingerless gloves. i just stick with my trousers under my hand wrists in bed since that will put extra protection between my tender palms and biting cold aluminum body.
Blog of James Brown
Autonomous Machines & Society.
2023-01-14
2023-01-14
proxy service
recaptcha solving service
free libraries
botright can solve recaptcha, hcaptcha, geetest. it shall be used with playwright. doc
paid services
anycaptcha.com is an AI captcha solving service
2023-01-14
the story from medium is about writing some listeners for quick notifications to get jobs faster.
translate.com is a website for translation. you can receive money via payoneer.
register here as translator.
2023-01-13
this site only unblock “restricted” articles. it does not offer “trending” “user homepage” “most read” “recommendation” things which you can get from official medium website.
map *.medium.com/[URL] to scribe.rip/[URL]
or try to host scribe.rip yourself.
source code hosted on sourcehunt.
alternative scribe instances seems all blocked (in mainland)
2023-01-11
通过在腿的两侧添加两个枕头 成功解决因为天气寒冷被子加厚导致被子压手 没有空间在键盘上面打字的问题 一定程度上也解决了被子压脚的问题
2023-01-11
By looking into the NaturalLanguage Framework on macOS, I find text classifier and embeddings inside.
It will leverage default TTS engine on Windows and macOS, but for linux you must install espeak.
1 | import pyttsx3 |
2023-01-10
1 | jobs | grep -Eo "[0-9][0-9][0-9]+" | xargs -iabc kill -s KILL abc |
2023-01-09
Microsoft powered recommendation system, has now joined the Linux Foundation.
CantoonSegmentation can be used for creating visual effects over static cartoon images.
use grammar or state machine to constraint how llm generate tokens, in order to perform actions and call tools correctly
remove vocal from video and from music downloaded from web, then execute shazam. or only remove vocal from video and execute music search.
voice auto translation by meta
use attention visualization to create pan effects, focus on different parts and illustrate separately
to get response fast, you need to create content fast. sometimes it is important to degrade aspects like resolution and more.
https://github.com/audio-agi/audiosep
changedetection get website updates
自动删评机器人：
根据对类似内容的评论预期（结合大模型）进行删除
根据情感分析删除不利评论
根据视频内容 标题调整封面人物表情
造仿制b站 收集用户点击数据
b站禁用了手机端搜索接口的页面数目，但没有限制电脑端搜索页面数目，估计是为了避免二创素材收集不受影响。
deep danbooru for anime comprehension, scene understanding etc
Do not use removable drives (NAS?) as scratch/data disk for long-term programs. Instead, use internal drives.
If possible, first sync all necessary files from removable drives to internal disk, then run the long-running program from there, or first run from rootfs, then jump to removable drives. When error occurs, check if disk is unmounted, try different methods to reload them and rerun the program.
Another possible cause is insufficient current. Check for similar issues online.
短视频内部模拟不同视频上下翻页效果
如果封面提取关键词进行训练并更换 尝试自己生成配图和文字的方法难以实现 可以考虑更换字体和色彩 文字内容不变
封面文字遵循一定的阅读顺序 从上到下 从左到右 可以设立排版以及合并（用于训练）规则
youtube now can auto generate video chapters, which might be part of text/video summarization.
multinational, multilingual, subtitles, localizations
models for content generation:
one-for-all multi-modal generation
magma: a GPT-style multimodal model that can understand any combination of images and language
视频创作导航 可以用来找自动化创作视频的思路
字由 识别字体 下载免费字体
tools for startup companies, including:
1 | [Website] |
a list of live streaming and content creation tools:
1 | [Content Creation] |
收集大up主的名字 根据提取出来的名词 清除文案中的含有名字的句子
for information cascade, treat recommendation system as your target data, and fit your content into the prediction of “to-view” video of popular videos
分析弹幕关键词是否与大量其他视频标题相重合
B站撞车搬运检测：
油管有带有英文番剧名称的AMV/MAD剪辑
通过B站搜索Weibo Tumblr Tiktok Youtube的视频ID 或者名称，可以找出视频是否被转载
通过搜索其他网站的前缀 比如https://www.youtube.com 可以得到被转载视频的链接 但是感觉数据不是很靠谱 不是很火的那种 要看大流量的还是得爬首页推荐链接 根据话题搜索
video highlights extraction
although you may want to train/extract that manually, it would sure be tedious and not self-updating (unless using reinforcement learning).
often we determine highlights by sound, visual and voice together. highlights often can be identified without too much context, so it can be chunk based.
bilibili
b站的高能进度条 在油管被叫做”most replayed”
b站有弹幕 所以可以根据弹幕找到精彩片段 VClimax是一个浏览器插件 可以通过弹幕单位时间增长速率，设置相关的阈值，来定位最精彩的内容 (弹幕密度怕还是得要分析) 跳转部分番剧OP 视频搞笑片段精准定位 (怕还得是要机器学习)
bilibili Danmaku Skip is another browser plugin which will identify highlights by analyzing danmaku with parameters like threshold, interval and bias
youtube
youtube’s most played data can be extracted by:
youtube-heatmap (nodejs, using puppeteer (bad!))
youtube operational api‘s (powered by shared API keys and info extractors without key), while apparantly youtube-most-replayed is using this service to retrieve the data from yt.lemonslife.com powered by this library
youtube.js (reverse engineered innertube api) added support for chapters and video heatmap
youtube-dl search youtube video
1 | youtube-dl "ytsearch[optional_result_limit]:[query]" |
record live streaming video, upload video
biliup & biliup-rs (commandline program)
全自动录播、投稿工具，也支持twitch、ytb频道搬运。提供分p上传b站接口
其实直播回放没有什么好看的 很单调 另外b站上传之后可以获得视频预测标签
youtube automation toolkit post same content to multiple platforms: bilibili, douyin, douyu, instagram, reddit, spotify, tiktok, twitch
though the idea is correct by posting original content to multiple platforms to prevent pirating, but the description/title generation is a vital part of the process, which must be done intelligibly (AI or human). as for now the repo is just full of links. if you want tools you click given link.
Download a portion of video
yt-dlp (latest)
check pyjom/tests/download_sections_video_portion_partial_download_youtube_yt_dlp_bilibili/test_bilibili.sh for advanced usage of yt-dlp and more on bilibili parsing.
SINCE YT-DLP IS UPDATED YOU CAN USE --download-sections ARGUMENT FOR YOUTUBE
If you want to download multiple sections of same video, you must specify video output format string via -o
But when using that without “–force-keyframes-at-cuts” (skip re-encoding which can speed up thing but not ensuring quality of video at tail), you better keep margin at tail for 10 seconds (could glitch at last 5 seconds) and head for 5 seconds (maybe head margin is not needed?).
youtube-dl
first acquire download url: youtube-dl [--youtube-skip-dash-manifest] [-f 18] -g "https://www.youtube.com/watch?v=V_f2QkBdbRI" (you need to force the format.)
then use ffmpeg with the url to chop the slice: ffmpeg -ss 00:00:15.00 -i "OUTPUT-OF-FIRST URL" -t 00:00:10.00 -c copy out.mp4
RangeDownloader by A-Soul-Database
A-Soul-Database is a live-streaming replay record database designed for vtubers, organized in some way for easy information retrieval.
RangeDownloader is acting like a server, though sometimes we are not sure how fast(er) it can really be.
Viral videos
Data sources and monitors
Rebang.today
通过B站推荐标签接口可以得到观众的实时需求
提供全站 知乎 微博 IT之家 百度 虎扑 直播吧 少数派 36氪 吾爱破解 天涯 小众软件 反斗限免 哔哩哔哩 抖音 技术期刊 v2ex GitHub的热榜
API：https://api.rebang.today/v1/items?tab=<TAB_NAME>&page=<PAGE_NUM> (potential parameters: sub_tab, date_type (default:now))
知乎有个专门的热榜，地址：https://www.zhihu.com/billboard
| name | tab |
|---|---|
| 全站热榜 | top-all |
| 微博 | |
| V2EX | v2ex |
| 知乎 | zhihu |
| 哔哩哔哩 | bilibili |
| GitHub | github |
| 抖音 | douyin |
| 技术期刊 | journal-tech |
| 虎扑 | hupu |
| 少数派 | sspai |
| 百度 | baidu |
| 36氪 | 36kr |
| 天涯 | tianya |
| 吾爱破解 | 52pojie |
| IT之家 | ithome |
| 全站24小时 | top-daylong |
| 直播吧 | zhibo8 |
| 小众软件 | appinn |
| 反斗限免 | apprcn |
番茄数据
番茄数据提供了从近24小时–近90天B站的最新热门视频，你既可以通过“搜索标题、简介、标签、评论出现关键词”，也可以通过行业分类、播放数、点赞数、投币数、视频时长、观众画像等高级条件，精准定位想要查找领域的热门视频。传播指数的计算方法有待研究。传播指数是根据UP主的粉丝数、视频点赞数、播放数、投币数、分析数等分析出来的综合得分。根据评论热词分析视频观众热点，根据评论用户分析用户画像。
对于各大数据网站 都存在一个收录的接口 如果up主从来没有上过首页 大概率不会被收录 需要手动提交 间接说明大up主是如何被找到的
Image recognizers
Baidu Image recognizer 百度识图
与此相关的识图项目位置：pyjom/tests/search_engine_suggestion_based_qa_bot
可以获取关键字，标签，同样的图，秒懂百科视频，百度百科数据，包含图片的信息，颜值信息
通过把上传接口修改 以及http改成https 现在可以继续使用
位置：pyjom/tests/viral_video_experiments/BaiduSerchImgApi
通过http改成https 修改好了360识图的接口
位置：tests/viral_video_experiments/360ImageSearch
Video collectors
tiktok compilation video generator
collect popular video on tiktok by multiple filters such as hashtags, categories , popularities and search queries
WeiboSpider
需要cookie 收集用户信息 用户粉丝列表 用户关注列表 微博采集 微博评论采集 微博转发采集 基于关键词的微博检索
botTuber: a instagram compilation reposter to youtube
Using instaloader and instalooter, it can download videos from instagram. It merges a series of video and add intro and outro. It only contains one default title starts with “TRY NOT TO LAUGH” in its “auto” mode.
reddit hot videos to youtube
In “TiktokCringe” reddit channel, we are able to get hot posts and video links prefixed by https://v.redd.it (from tiktok to reddit) in json format: https://www.reddit.com/r/TikTokCringe/hot.json?limit=12. This link looks like some API or subscription. Maybe Bilibili and other sites have similar “hot” json urls. The way to extract video links is in atmt.sh. It adds transitions to every video clip.
Video editors
vced
i think it needs to be fine-tuned on large diversive training data.
VCED 可以通过你的文字描述来自动识别视频中相符合的片段进行视频剪辑。该项目基于跨模态搜索与向量检索技术搭建，通过前后端分离的模式，帮助你快速的接触新一代搜索技术。
videofy
it is a self-hosted service, summarize article, get relevant text and image, determine mood to select BGM. try for yourself!
meme video maker
It uses google cloud to select “English” words on image, enable the user to edit the “stage” to show meme step by step.
It requires amazon cloud services and google cloud services.
回声工坊 TRPG Replay Generator
TRPG：桌上角色扮演游戏 有丢骰子（随机元素）的RPG
角色立绘可以是动态的 但是是多个png文件
背景可以设定为 {'black','white','greenscreen'} 中的一个，以建立纯色背景
Has special requirements to media sources. Use .ogg format for BGM. Use .wav format for AFX and voices. Use .png for image. Cannot get background video layers working so you might consider some “green screen” effects.
Use --ExportVideo flag to export video without GUI.
openshot and libopenshot (for python bindings)
Nightcorify is accelerating audio and raising pitch (asetrate strenches the timeline to change sample rate, atempo (not used in here) change the timeline but not the pitch, while aresample changes the sample rate but not timeline), also showing audio wave shape with showwaves.
This library is complex AND WITHOUT PROPER DOC FOR PYTHON thus not recommend for using
keybert and summarization transformer pipeline
Check docs on transformers pipelines for default and fine-tuned task-specific models for each pipeline.
Keybert uses “sentence-transformers”. The author would advise either “all-MiniLM-L6-v2” for English documents or “paraphrase-multilingual-MiniLM-L12-v2” for multi-lingual documents or any other language. Search for “multi” with tag “summarization in huggingface, then you would get huge models. A mT5 model is very large, size upto 2.33GB
Keywords-image pairs can be used for CLIP model training.
watson based video maker
It first downloads wikipedia content from algorithmia, then uses regex to filter out unwanted parts, uses watson AI for sentence cutting, set a limit for max sentences (notice: not summarization), then search image with keywords, finally create video.
In another similar project IMDB (Popular Film/TV series) and Google search trends (as RSS) are included.
Auto-Editor
By passing --edit option, you can remove unwanted parts identified by motion or audio (can be combined). It can import clip with manual “cut-out”. It can export as json.
Pictory
Leveraging 3 millions of tagged video clips and audio, choosing most semantically similar clip to current scene (by extracting keyword -> search images -> compare images to video sources with all embedding things going under the hood (CLIP)), map video word by word to the timeline (to create extractive highlights and remove unwanted words like “um”)
Wisecut
Short videos can attract your viewers and converting them into followers (to view more of your long videos). Make short videos with music, subtitles and facial recognition auto-reframe (detect main speaker). It match the right BGM with the type of content, with audio ducking, which can be achieved with ffmpeg or editly.
It is listed among an AI marketing tools list, which mentions copywriting, social media/email/blog marketing text/content generation (like copy.ai), text to video
Jumpcutter
An audio-slience based video cutter. In jumpcut_file.py it chops audio into chunks and decide if it is slience or not. The core logic is to first compare max volume of each chunk against threshold, then check in neighbors of every chunk if all of them are slient and cut them out. It has audio speed changing methods from audiotsm.
In another implementation, it uses ring buffer by collections.deque and applies VAD (Voice Activity Detetion) by webrtcvad to every chunk of audio.
Gifcurry
Adding text to video, has typing effects, written in haskell. You can add -m flag to export video instead of GIF.
Backgroundremover
A commandline tool powered by torch, removing background from images and video
Moviepy most loved commandline video editor?
There are some cool text effects called “Text with moving letters” (PPT-like), and a dancing video generator based on tempo finder and video loop maker, which can help you adjust video speed according to video period and music bpm. The Star-Wars Text Effect reminds me of easing functions used with page scrolling.
Data collect/analyze
Social media statistics are time series data which should be collected regularly and predictable with time forcasting models.
open-sir
Use sirx over sir.
I think it is hard to use. Many “presumed” parameters are out there. It can fit “reproduction rate” but no individual “alpha” and “beta” values.
In tradirional SIR models, beta is infection rate, gamma is recovery rate. While in open-sir it is different. alpha now is beta, beta now is gamma.
Youtube Viral Video Machine Learning Analysis
Refer to this document for details in data collection and machine learning methods.
Usage:
You can decide whether to copy a video or not when it is posted for only a few days.
Dataset creation:
Monitoring video right at the time it is posted, monitor for a few days, calculate features, then wait for a month or two (it must stablize then), judge the video is viral or not by view counts.
Using multiple machine learning techniques, there are some top features matters the most for viral video forecasting (though you can derive your own by collecting more data (like the follower-view theory if applied), and beware if your video all sucks, you may not get an accurate model out of your data alone):
| Rank | Feature Name | Importance |
|---|---|---|
| 1 | views_acc | 12% |
| 2 | views_1 | 11% |
| 3 | ageRatioReviews_1 | 9% |
| 4 | video_duration | 9% |
| 5 | comments_1 | 5% |
| 6 | channel_uploads | 5% |
| 7 | ageRatioLikes_1 | 4% |
| 8 | comments_acc | 4% |
| 9 | channel_views | 4% |
| 10 | comments_sentiment_compound | 3% |
ViralCaster
TitleParser.py analyses views along with words, getting the most “popular” word or word combinations. It has demo data. It generates “max” “min” “mean” views related to single word or word combinations.
Predictube
peak_detection.py use daily view count to categorize and identify trends. “MonoIncr” might be our desired category.
Video Viralization Tool
It uses relative infection ratio instead of absolute to predict the trend. By “information cascade” it means statistics can be used to predict future view counts. It considers individuals and viewers as nodes. It suggests different relationships between parameters in SIR model and data (likes, shares, comments, new subscribers, subscribers, length, quality, tag keywords, description keywords).
2023-01-05
save this under /lib/systemd/system/neo4j.service
1 | [Unit] |
2023-01-02
robotics
RT-1 robotics transformer and SayCan
VIMA General Robot Manipulation with Multimodal Prompts
multimodal model
magma a GPT-style multimodal model that can understand any combination of images and language
Versatile-Diffusion Text, Images and Variations All in One Diffusion Model