So, a little while back i wrote a little simple program to download all images from Jow Forums thread (mainly from /b/ and /e/ because you know...). It became kinda hard for me to sort everything (almost full mega drive), so i decided to build an AI to do it for me. Well, it's not going really that well, i mean.. around 300 samples from each (2 at the time) category isn't really much, but i managed to achieve around 95% accuracy, which is not really that bad. I used tensorflow with tensorboard, and pretty simple 5 layer convnet. I have 5 relu layers, 1 softmax connection layer. I'm using linear regression with "adam" optimizer. I'm not really good with NN, so my hyperparameter are bad at best. I don't really know what to tweak to get better results (bigger image size? Maybe change some numbers in conv_2d function? More layers?), so i'm here to ask for a few advices on what to do. Code is here, feel free to take it, modify or whatever.. pastebin(cx5VyEHd) I can provide datasets and models on request. Results are however somewhat mixed (pic attached, from 2x20 epochs run), i have 6 test pics, 3 from each category, and while sometimes it gets like 5/6 right, sometimes it just fuck it up completely. So, if anybody here do this AI stuff and image recognition, and is willing to help a bit, you're more than welcome to join discussion and i would love your help.
--- I know it's totally useless project and stuff, but i want to get into AI and stuff with something that at least i can use. And since i'm fucked up degenerate and i collected lots of porn throughtout the years.. why not this? It's also cool programming excercise.
AI That Sort Nudes
Other urls found in this thread:
github.com
medium.com
medium.com
github.com
github.com
twitter.com
>6 test pics
Way too small to make any meaningful judgments, you need more data. If you have enough data over all, the just increase the size of your test set (assuming you validation and train are large enough). If they are too small, then you could try augmenting your data, make copies of the samples and do transforms on them like rotating, cropping, changing the colors, etc. You can also use cross validation to test on more of your data without compromising training.
Have sex.
spbp
No u
ok, so i should add more test data and do some more image processing to training images, got it. Thanks.
I have. Sometimes.
Just build a layer on top of Google's CNN for object recognition
Also one thing i noticed (pic), aren't my labels swapped around? If you could take a look at the code.. It seems ok for me, but if i swap them around, predictions are basically 100% right.
Well yes, but actually no. I wanted to learn TF and practice programming a bit. I might do that in future, thanks for advice.
blue board
i want an ai that collects as many lewd images as it can from the internet, places them all in a booru, and uses its ai magic to sort them find the most fapworthy images for each user
Use image segmentation
PROMOTIONS
that's essentially what i'm trying to build...
reasons? i probably will, but how hard it is to implement?
How to get banned 101
Maybe this can help you:
github.com
Don't forget to come back when you finish your project tho
no why? its tech related (AI, NN, image recognition). Just what i'm doing with it is weird, but i'm asking for help with AI itself, not with anything else.
holy fuck someone really did that. i will definately take a look on that, thanks. And I will come back, don't worry, i post here and on /b/ pretty often.
many respec techboy
Current system are implement using ResNet, because imagenet lack NSFW content, you need build one dataset label of tits,pussy and ass.
medium.com
Lol huge porn dataset.
You posted porn on a blue board dude
Its art not porn
resources, yummy. this will be helpful as fuck, i didn't knew NNs can do stuff like that. thanks.
it's not a fucking porn. Its a program, and i need help with it. What's inside the program hopefully doesn't matter, and if, then too bad because that's the thing my program do.
i can't believe someone is actually doing these kind of stuff. many thanks too
also here is some results, did a little tweaking, added more test data and some image processing, and after like 50 epochs, 11/12 images were correctly classified. Thanks anons
First make a classifier that sorts between 2D and 3D then you should have an easier time distinguishing between "tits" and "pussy".
NoFap
Good God vaginas are fucking disgusting.
t. fag
MOOOODS
>he doesn't want to have a family
fag
This might be the best thread I've seen on Jow Forums in a while.
The absurd premise for this interesting topic (which is by no means presented in a shallow way) is just the absolute best.
Posting in legendary thread.
stop denying it, you thought it too
maybe when i was 7
i'm gay and what is this
also i'm gay by the way