Hi Jow Forums,
I built a binary classifier. It's live at www.isitanime.website
Go play with it.
Also, if anyone wants to help out with some CSS I would be greatly obliged.
Hi Jow Forums
Other urls found in this thread:
api.isitanime.website
cs231n.github.io
pastebin.com
twitter.com
botnet
get a miner on it as well as atleast 16 ads
I knew it
>www.isitanime.website
Can confirm the results are accurate
You're fucking banned faggot.
I think it's just looking for flat colors? I trained it on a bunch of photos of IRL vs a scrape of some boorus
Figures it's the iphone user with the furry gore
>anime website
frogposters >>/out/
????!! fuck you
pepe is not anime while wojak is, I think its looking for combinations of colours
works extremely well
this is the image used, try it for yourself.
hentai my dude
anime.website
OP I posted your thing on /a/, but I think they're too stupid to know how to use it.
The eventual goal is to GAN it so that I can build a filter that makes anything more anime. Bring anime to real life, such is the dream.
This is disappointing. I suppose there wasn't much lain in my training set. If I retrain, I'll grab a bunch of screenshots from some shows and include that in the data
I'm getting a good number of IPs coming in, and I just turned on cloudwatch logs so I can try to count the number of API requests it's making.
Also if you guys want to hit it with some python feel free. Just don't script it through your entire hentai folder, I only get a certain amount (about 130K) of free requests per month. If it goes over that, I can give it more at the cost of it will be a bit slower
#! /usr/bin/python
import requests
url = 'api.isitanime.website
with open('test02.jpg', 'rb') as in_f:
data = in_f.read()
headers = {
'Content-Type': 'image/jpeg',
'accept': 'application/json',
}
r = requests.post(url, headers=headers, data=data)
print(r.text)
can't we have options to say if it's correct? i think that would help improve it, but idk shit about NN.
kek
this. let us confirm or deny if it's actually anime. you can go through and approve or reject each confirmation or denial if you're worried people will lie.
this tripped it up
also:
Top 5 saddest anime moments
WatchMojo.com
I love APIs!
Mind telling us a bit about how you made it? I've wanted to make something like this but don't know where to start
first 0% i've seen
anime mutt
I expected 56%
>moshi moshi, goburino desu
t. original
I'd like to do this at some point. I'm worried about trolls (Jow Forums convinces it that hitler is anime or something) but it would probably be more signal than noise.
I'll be free for a few weeks soon, so maybe that will be my next project. It's not just a "click a few buttons and it happens" thing, but I should be able to get it to work.
Nice
It's a convolutional neural net with keras. It samples the image into a 512x512 square, then dumps it through a bunch of linear algebra until answers come out.
I read cs231n.github.io
One of my favorite things to do is to take a photo of someone and then tell them how anime they are. It's a fun but useless metric
Accurate
>I'd like to do this at some point. I'm worried about trolls (Jow Forums convinces it that hitler is anime or something) but it would probably be more signal than noise.
that's why I said you should have it save the file and you double-check
Image source please
Jow Forums.org
No, it's definitely more than that.
>ruiko
eww
at least the bot has some taste
>not a pipe
then what is it?
an image
where do i learn all this artificial intelligence shit
Looks like Sayori is not anime
That's not an anime, that's an Ohayou.
Thats depressing
I can't see the fucking upload button.
this is bullshit. We all know that Richard D. Wolff is anime.
Does this mean I can find him in real life?
I do not claim to be good at CSS. If you don't believe me, check the comments in the page source.
That's pretty interesting. Thanks for sharing
>I'm worried about trolls (Jow Forums convinces it that hitler is anime or something)
nazi flag is anime
nazi flag is 99.9% anime
post original?
>Does this mean I can find him in real life?
Trump is making anime real
Season 1 was kino
95.4% anime
You trained this on hentai didn't you
every hentai pic I upload is 100%
actually a pretty good result
Safebooru
So it's close to hentai, but not quite NSFW
Potential future training sources:
>+ My anime collection, take a few screenshots per series
>+ Sadpanda image sets?
>+ /c/
>- /co/
>- tumblr
>+- All the stuff you guys have been uploading
I'm not saving the uploads at the moment so rest assured all your shit is wiped from the server as soon as processing is done.
This is a misclassification, but I don't feel so bad about this one
i'm learning about machine learning stuff. could you tell me what ML model you are using to classify images? Is it some kind of pre-trained CNN or did you make it from scratch?
Fuck you, I lost
>www.isitanime.website
neat stuff senpai, but accuracy's a touch off
Convolutional neural net, from scratch but with keras so the code really isn't that bad.
Here's a JSON dump of the model. I'd post source, but that's tied up in a backup tarball I haven't bothered to recover yet
pastebin.com
Yeah I need more training
Is this another random.org?
I see, thanks
As far as ML libraries go, how does keras compare to tensorflow?
I've started with tensorflow in python but it's confusing as fuck
Looking pretty good man
I'll keep fucking around with it for a while
Keras is a higher level abstraction on top of tensorflow (or theano). It figures out most of the stuff that would be boilerplate, so all you really need to do is figure out your hyperparameters. Essentially think tensorflow with a less general, easier API.
If I were going to get good at this, I'd stick with the high level libraries for most of the work, but we're not yet at the point where you can just use those without knowing what's going on on the backend. You should build at least one neural net with straight numpy before you move on to TF, and build at least one with TF before you move on to keras / etc.
The CS people around me seem to prefer PyTorch over TF for some reason. I don't know the difference well enough to make a comparison, but they do more work in this space than I do.
96.6% anime
pretty cool
From the gay screenshots I've seen here, I can conclude that OP is a faggot.
How did you train it?
You should use some hand crafted features — anime usually has large patches of the same color
>I'm worried about trolls (Jow Forums convinces it that hitler is anime or something)
I got you covered senpai!
Tf
Well, shit.
OP image isnt even anime
shit software
>safebooru
i guess its another case of garbage in, garbage out.
pixiv and calarts/tumbrshit dont count
it seems to be focused on sharpness of edges. interface suggestion, add link upload. it should be five minutes for you, and downloading stuff on hdd + uploading it feels just wrong