Normally it means that my server would be in the process of training a neural network which would get sent to you. Did you specify categories? I see a request that was just empty, which was dropped by the server as looking like a fake request.
How do I specify categories? There's a free-form text input field. Am I supposed to put stuff there? What do I put there? What works?
For that matter, I have multiple empty requests bots (apparently). How do you delete them? I could probably edit stowbotlist.json but I can't imagine that's the intended interface.
Yeah each free form text entry field becomes both a image category and a new folder that images will get sorted into. If you need to clear out the empty bots you can just delete the contents of stowbotlist.json. There will be a cleaner interface eventually but this tool is very new and still a little rough around the edges, especially in those cases where things go wrong. The demo video on the website shows an example of things you can request and how it would work.
Yeah, from the video (videos for text? Really?) you have some categories, but the point is what is the available space of categories?
Do you specify tags you want to filter by? For example, if I wanted to do (random suggestion) SMT IC package identification, how would I do that?
------
From the sound of it, right now, it sounds like you review the requested categories, and manually/semi-manually generate each specific network?
------
I'm basically in a position where it'd be much more valuable to me to have a thing I can point at a directory and have it tell me what it found then something I have to say "sort into these categories".
In general, there's very few cases where I have a dataset where I want to filter by a specific content, and don't already have that metadata. Content discovery would be much more useful.
FWIW, this is a project I'm REALLY interested in. If you go open source, I'd totally get involved.
I have a mirror of *booru image galleries (basically, anime/manga images, but REALLY well tagged: upwards of 20+ tags per image) that I've been occationally meaning to see if I can turn into a google-deep-dream-like nightmare to make the worst hentai possible (it started as the idea for a terrible idea hackathon[1]).
I have a fair bit of experience writing web-scrapers [2], but no idea what I'm doing with neural net stuff. I keep meaning to spend time to figure out how to do anything, but other projects (and work) keep me distracted.
Anyways, if you have any use for 5496863 images (probably mostly hentai) with 196852748 tags, hit me up.
Or you could just run the scraper [3] yourself, but I hope you have ~5+ TB of free disk space.
Thanks for the offer, but I'm not really interested in open sourcing this. I'm familiar with the danbooru set from Gwern's posts about it but I don't really intend to use it for anything.
Yeah, it's kind of specific to a certain weirdo subgroup (ayo~!!).
I've also chatted with Gwern a bit about the project. We had similar interests.
My scraper wound up being very similar to what he implemented, though it was already up and running when he made his first posts (I started my project about 6 months before Gwern).
> Each Stowbot is customized to organize one folder. Creating a Stowbot is easy — just specify which types of image content you want your Stowbot to sort.
Ok, what can they sort on.
Also, is this going to be something I can run locally? If so, I'm super interested.
If it's a web-service only, I have zero interest whatsoever.
-----
~~~Signed up with an e-mail. Sigh:
> We will send out emails to users when the beta is available.
Argh~~~
Just noticed I have a e-mail with a link in it. What?
It seems to imply from the homepage that everything is local, or at least that your images remain local and thus the program runs locally, but I believe the actual training of the neural network occurs on their server, so you have to tell them what things you want to sort by, that might be all.
I'd like to see a bit more information on how that part works and which categories it might struggle with, not everyone just has image folders of common animals.
Yep, the sorting is all local, but the training happens on my server.
The space of "categories of images" is super large, but generally this will work well on things that you can find on the internet with simple keyword searches. It doesn't do very well with facial recognition, but I'll add that eventually since some users have requested it.
20 comments
[ 224 ms ] story [ 1381 ms ] threadI assume you need a fairly extensive training dataset to train a net, and that would limit your ability to provide a net trained for arbitrary tags.
For that matter, I have multiple empty requests bots (apparently). How do you delete them? I could probably edit stowbotlist.json but I can't imagine that's the intended interface.
Do you specify tags you want to filter by? For example, if I wanted to do (random suggestion) SMT IC package identification, how would I do that?
------
From the sound of it, right now, it sounds like you review the requested categories, and manually/semi-manually generate each specific network?
------
I'm basically in a position where it'd be much more valuable to me to have a thing I can point at a directory and have it tell me what it found then something I have to say "sort into these categories".
In general, there's very few cases where I have a dataset where I want to filter by a specific content, and don't already have that metadata. Content discovery would be much more useful.
I have a mirror of *booru image galleries (basically, anime/manga images, but REALLY well tagged: upwards of 20+ tags per image) that I've been occationally meaning to see if I can turn into a google-deep-dream-like nightmare to make the worst hentai possible (it started as the idea for a terrible idea hackathon[1]).
I have a fair bit of experience writing web-scrapers [2], but no idea what I'm doing with neural net stuff. I keep meaning to spend time to figure out how to do anything, but other projects (and work) keep me distracted.
Anyways, if you have any use for 5496863 images (probably mostly hentai) with 196852748 tags, hit me up.
Or you could just run the scraper [3] yourself, but I hope you have ~5+ TB of free disk space.
1: http://www.stupidhackathon.com/ 2: https://github.com/fake-name 3: https://github.com/fake-name/DanbooruScraper
I've also chatted with Gwern a bit about the project. We had similar interests.
My scraper wound up being very similar to what he implemented, though it was already up and running when he made his first posts (I started my project about 6 months before Gwern).
Ok, what can they sort on.
Also, is this going to be something I can run locally? If so, I'm super interested.
If it's a web-service only, I have zero interest whatsoever.
-----
~~~Signed up with an e-mail. Sigh:
> We will send out emails to users when the beta is available.
Argh~~~
Just noticed I have a e-mail with a link in it. What?
I'd like to see a bit more information on how that part works and which categories it might struggle with, not everyone just has image folders of common animals.
The space of "categories of images" is super large, but generally this will work well on things that you can find on the internet with simple keyword searches. It doesn't do very well with facial recognition, but I'll add that eventually since some users have requested it.