tried to train neural network to detect my favorite type of porn.
Results: big tits often mistaken for ass. Black hair often mistaken for anything black in area. Model is even 100% sure when there is huge ass black rectangle in photo. There were no fucking black rectangles in training data...
Trained with 1100 training photos and 200 test photos. Got 75% accuracy on training data
is machine learning meme? My models are extremely dumb but my training data are decent.
>My models are extremely dumb
pack it up
How long are you training them?
>Trained with 1100 training photos
ridiculously weak
Also how big is your dataset
What did you use?
>1100 pics
>shit models
Yes machine learning is a meme. The problem is not you being naive. There is nothing stupid about expecting Google tier result with that toy of a setup.
You dumb fucking dipshit.
I use T4 GPU from google. I train for 100 Epochs. Save each model and pick the one with highest accuracy.
how many photos do you suggest? Author of imageAI recommending 1000 photos for good results
I have 5 different categories. For each category I have ~1200 training data and ~200 test data
>
ImageAI
Just use pornhub silly
Use more data, i would take your current dataset and add mirrored versions of images and mildly rotated versions.
Train much longer
Also finetune a pretrained model. You might get better results with that little data
>1100 training photos and 200 test photos
Use more data
Why don't they just call it a quantum network?
>and add mirrored versions of images and mildly rotated versions.
yes i believe ImageAI has such option and I am using it during training.
>Train much longer
how much longer? This is my accuracy progress for each epoch. It stopped increasing in the middle of training
pastebin.com
>how much longer?
See ive never trained image recog
But when i was training GANS i would leave them for over 10 hours
Work on your feature engineering.
i am using T4 which is $2000 GPU. I was told there is no point in more training when accuracy stops increasing
How much more photos? It is very tiresome to manually verify each photo
>Black hair often mistaken for anything black in area.
kek
Convolutional neural networks have spatial invariance you dummy.
porn sites pay people to categorise and verify images, if what you were doing was easy you'd be able to sell the tech for a shit ton
Really? Source.
yes I noticed since I am processing most of content dumps of big tube sites. They have tens thousands of completely miscategorized videos with wrong tags and names. It is sad, they don't deserve their traffic
Post code
Use a pretrained model and only train the last few layers
Use 1cycle policy
Actually just post your dataset and I'll make a better model
How do you pick pretraind model to use?
Something that has been trained on ImageNet and has done well will be fine kaggle.com
No one's fault you suck at feature building.
Deep learning on images shouldn't require this
I can post my dataset, but can you produce resnet model? this is what I would like to use since it should be fast with medium accuracy
I use ImageAI which is library for brainlets
training pretrained model is not yet implemented in current version
Post it if possible, I will give it a shot. I don't know what you mean by "Can you produce resnet", you can download it (pytorch format) at the link I gave
Do you just add extra layers on the end to fine tune? Or leave it and do more epoch of traisning.
How do you guys store training sets, labels, and mangage them?
What about tools for labelling more data?
mega.nz
here is my dataset
I need .h5 format to make it work here. ImageAI supports SqueezeNet, ResNet, InceptionV3 and DenseNet
Basically only train the last 3 layers or something like that
i got image data from yandex image search and manually removed bad images
ok giving it a shot
Getting 14% error rate with pretrained resnet34, will try to unfreeze the entire model and tweak some
12.6% error rate after unfreezing, still not that good :(
Maybe the overlap between the categories is just tough
With such a small dataset you should use data augmentation
>Maybe the overlap between the categories is just tough
>entire human body in every one
uhh nah, it's F I N E
I mean, his mature category has people fucking in it, so I can see the confusion between it
I got to 11.3% error using pytorch / fastai / resnet50 pretrained. That's the best I can do.
I would post an image of some of the mistakes its making but it's a blue board
It doesn't count if you link somewhere else.
I am actually interested only in big tits and black hair category, rest was just random test
Here are the errors it's making. Some of them are reasonable errors in my opinion
imgur.com
Here is the tiny amount of code I used to train the model using fast.ai. (Only other thing I had to do was rename the test folder to valid)
pastebin.com
Here is the course you should take
course.fast.ai
how do people jerk off to this shit? It's like the McDonalds of pornographic material
well thanks for giving it a shot. I will wait for new versions of ImageAI, hoping for better luck training pretrained models. My current results are terrible and discouraged me to learn more about topic
Huh? Using fast.ai I was able to improve your 75% accuracy to 89%. You can learn how to do what I did in about 2-3 hours of lectures on youtube.
No idea, it's worse than McDonalds IMO.
Also I would argue that the data is part of the accuracy problem. There are pictures of people fucking that are in the other categories, not "fucking".
>click thread thinking OP is based af
>some lame as porn
fucking why bother
Stealth shill thread.
OP, posting with his bad implementation and bad porn.
user comes in with an alternative, even posts a link to the tutorial.
It's a nice setup, your curiosity is sparked just enough, and your dissatisfaction with the porn choice makes you want to do your own with good shit.
You have all of the resources available to do it.
It's actually a pretty good setup.
You're paranoid
I have 50,000 anime images that I have tagged and manually ranked with an elo score using manual comparisons
good enuf to pump into a off the shelf skiddie ml model?
>Trained with 1100 training photos
>HURRR MACHINE LEARNING IS A MEME
this board is unironically terrible
this desu
Do it
The elo score may be too subjective for a model to pick up on. But post it and I can try it
1100 is fine for transfer learning, you're terrible. the main problem here is the quality of his data
>OMG HOW COULD YOU DO THIS
>IT ISN'T ***REAL*** ML UNTIL YOU USE NINE GORILLION IMAGES
how to learn transfer learning and all this shit
what is wrong with black hair and big tits data those are solid photos. Rest can be discarded as I got your point
I posted the online course I'm taking here
Nothing's wrong with those, check , pretty good for those two imo
lmao how is sex a porn category
Is computer generated porn the future?
As an experiment i reccomend only using fat old men in your imagined and retaching the nural network to identify dad cock
Only as an experiment
It grows bigger as you "train" it
decided to give it another try. Trained densenet with only two categories big tits, black hair model claiming 98% accuracy on test data. But on real data, it is just terrible
>Author of imageAI recommending 1000 photos for good results
For identifying a single boolean value like if a picture does or does not contain a cat. And "good" is best in air quotes even for this.
you're retarded
just a woman with her tits out can be porn, no fucking required
I enjoy the symmetry of your post
The GPU doesn't matter you stupid nigger, add 10,000 more images to your dataset and call back.
Transfer learning can solve harder problems with relatively modest data sets. I can get 94% accuracy classifying between 37 breeds of dogs and cats with only 8,000 images total (robots.ox.ac.uk
I mentioned it just because someone was talking about 10 hours of training, that gpu is beast takes 30 seconds for 1 epoch compared to my PC which trains 15+hours. But I lost faith that more images would fix anything. my 98% accuracy model is completely clueless on real life data see
>talktotransformer.com
Text is generated from this site if you're interested. It uses a demo of a state of the art neural network trained on random internet script.
You probably did too many epochs and overfitted.
Thanks, man. This is some good shit.
Someone post datasets for me to classify
How do you guys do train/test/holdout sets etc, crossfold?
I have hard time evaluating how effective or any improvements.
Even with a hold out set if you keep working towards improving on that over the course of days/weeks could you start overfitting?
>ranked with an elo score
How does one do something like this, exactly?
Asking for a friend.
Using a random 80/20% split for training/validation is simple and does well.
my projects on GitHub, called supercutegrab
it's not the most user-friendly yet
you are overfitting
You can't train a network from scratch, you don't have enough labelled data. Instead do transfer learning. That means take an existing network and use the final hidden layer as features to train a simple model like logistic regression. I did this with tinder and it worked decently.
"My penis is prehensile!" I said, "It can move about and grip things!"
"But it's NOT capable of gripping anything," she countered. "All we've known from science is that I have a penis as a part of me." She gestured at her erect cock. "My 'cock was' is not my 'stool.' My 'stool' is my 'penis.'" I didn't know if she intended that alluding to the fact that a penis is not a flat plate but a "stool" that is "made" from fluid that constantly flows through it. The thought of her talking about a 'coiled' shaft and her own cock being the coiled tube, having an erect penis, I knew that she meant the penis hanging from her body like string, but, it was the way that she was describing that I immediately understood. She could see me in the mirror, so this thought occurred to me: "That's a really strange thing to say, isn't it?" The thought didn't really register until I had to think about it further.
Is random fine or should I make sure that train/test distribution matches.
Just dump training photos from your porn folder, if your porn folder is properly organised that means you should be able to get 1-2 milion training photos and a few hundred thousand test photos really easily
bad model and too few training data.
1k is fucking nothing