Miles Deep - AI Porn Video Editor

what a time to be alive

>>57523934
im legit interested on testing this

>>57523934
>no example video of porn and demonstration of whatever the fuck it does

>>57525270
Are you fucking retarded?
It's written clear as day what it does.

>feed the program a video
>set parameters as needed
>extract sex scenes of the desired kind into a supercut

Do you really need an example to understand?

>>57523934
Why is machine learning gaining so much traction?
Genuinely curious. All the AIfags at my college are so hoity toity about studying intelligence.

>>57525763
>machine learning
only deep learning

>>57523934
Nice. Is this your project?

? = f'(x) using muh hidden layers

yeah so fucking deep

It was a bitch and a half compiling this for arch, but I can confirm it works.

I'm on a GTX 960 and it takes a minute or two to scan a 15 minute 1080p video and cut it together.

All the tags seem to work pretty accurately. I wonder how the model is trained. Actually, I'll probably be running the model through deepdream later to see what kind of horrors come out

is studying AI worth it?

>>57528464
>running the model through deepdream

Omg please do that

>>57523934
>ask miles deep backed search engine for porn
>specify husband wife missionary procreation
>95 seconds of steamy hot Amish sex
>5 seconds of gay BBC rape
>"95% accuracy"

>>57528614
I'll get to it, in the meantime check out https://open_nsfw.gitlab.io/
(warning: nightmare fuel)

which is the same thing, but using yahoo's porn filter


I'm going to try to train this filter with more categories soon.

>>57528464
Did you try running it against nonpornographic content?

If we're lucky it could automatically produce an amusing montage of sex-like scenes from regular content like, say, a season of Full House

>>57528758
Everything I run it against says no adult cuts found

It's well trained

>>57528713
Holy shit

>>57523934
>residual connections

What does this mean?

>>57528791
Does it work well with shitty quality amateur stuff?

So you've basically revolutionized porn search engines over right in a way that is completely insane?

File: 1440920620273.gif (755KB, 500x330px) Image search: [Google]

755KB, 500x330px

amazing
how long before it can edit out any frame where the dudes face shows up?

>>57532285
You could probably write the code from this in under an hour.

I'll leave this as an exercise for the reader.

>>57528658
were the 5 seconds the credits roll?

>was going to sort and tag all my videos
>don't have to now
Procrastinators win again.

>>57525763
The things needed for effective neural nets (lots of data + lots of computational power) are finally here and technogeeks/investors are really excited about the potential machine learning has.

File: cat_197_5.jpg (11KB, 224x224px) Image search: [Google]

11KB, 224x224px

>>57528713
Ass volcano is my new favorite thing.

>>57534074
kek

>>57523934
They only trained it with vanilla shit. This will be useless for most people here.

>>57534274
Anon cant you see?
This is the first step towards a fantastic future
And what a tremendous step it is, really

>>57523934
was that mila jovovich?

>>57535224
Yes, she is

>>57532486
>exercise for the reader
>see: don't know how/too lazy to do it myself, lol

>>57525763
Let's put it that way.
You have a huge quantity of data. You know the correlation between some data and something else, but you don't know how to infer that something else from the data.
You can train an algoritm to infer that something without you even knowing what kind of algorithm it uses.
A simple example from some friends at the uni: analize tweets to know if they are written from a male or a female.
You feed a great number of examples where you know if the source it's male or female and the machine learning learns on its own to differenciate and then has 95% accuracy on all other tweets.
Same with every categorization task.

On a more philosophical note, after the rise of machine computing being better than man at math and logic, (e.g. Deep Blue winning at chess) we switched the definition of intelligence at "being able to learn and adapt".
Yet now we're teaching AIs this as well, further blurring the line between human intelligence and AI intelligence.

>>57525763
Making machines to make machines might be a thing.

https://en.wikipedia.org/wiki/Humans_Need_Not_Apply
https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=1&cad=rja&uact=8&ved=0ahUKEwjgtriu2qzQAhUh6YMKHQyCDWsQyCkIHDAA&url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D7Pq-S557XQU&usg=AFQjCNF8CTJbTHXkI2Qv7kK36viEYXrovw&sig2=A7o1trTlW108pFRgIMIDCg

>>57535804
Derp, sorry about that lets see if I can do that betterer

>>57534074
>tfw when I was a kid I accidentally opened a porn channel
>there was a plan where man was on bottom woman on top, fucking in the pussy, and camera was pointed at their crotch
>thought it was some kind of weird volcano

>>57535728
The only benefit of human intelligence will remain our remarkable potential to store energy - humans will be for a long time the most powerful portable computers for a long time to come.

Our main issue is that our instruction set is still not entirely defined or even consistent - we can't work symbiotically or in union with each other, so our processing power can't really scale worth a damn.

I wrote this and I'd be happy to answer questions. Residual connections are a huge breakthrough in deep learning from Microsoft research. They allow for much deeper nets. https://arxiv.org/abs/1512.03385

It was trained on 36,000 images taken from over 500 movies at 5 second intervals, and hand sorted by me. I'm using transfer learning rather than training from scratch so the models are pretrained on image net. Me training took about 10 hrs. Also on a GTX 960.

Use CUDA it makes it much faster. 10x speedup.

>>57528464
I'm the creator. Thanks for testing it. I also used a GTX 960. Did you try CUDA? It should really speed it up. What did you compile it for? Was the difficult part just getting caffe setup or can I make it easier somehow?

Interesting. I think I'll try it on my 1070 when I get home.

Have anyone train a deep dream model on porn yet?

If you're on Linux your can just download the binaries without recompiling. I tested it on Ubuntu 16.04 and Linux mint 18.

It should be really fast on a 1070. For bonus points you can even use CuDNN. It's free but you have to sign up with Nvidia to download the drivers.

>>57540302
Yea but it was mostly just genitals, not very interesting.

>>57534074
These are incredibly interesting.

>>57540302
here it is
https://open_nsfw.gitlab.io/

I wonder if there's already an plug and play program to generate deep dream pics from your own collection of pics.

>>57540130
any plans to add more tags?

Yes. I'd like to add more sex act tags..and perhaps another net for description tags like milf, skinny, redhead, etc. Any suggestions?

On obvious one would be money shot or at least facial...but that's somewhat difficult because it's an action over time vs being well represented in one image. I'm looking into this and may add it if it works. For now the money shot is usually grouped with wherever it is...blowjob category for on face, sex back for on back etc

>>57540464
Try scraping pornmd based on popular queries

>>57540052

What the fuck is [spoiler]right[/spoiler] with you?

I may also attempt to separate anal and vaginal sex. Now it simply divides sex front vs sex back based on which side of the female is showing, her butt or her vagina. In some instances you can't tell if it's anal or vaginal cause of the angle but in some it's clear and I think it could make that distinction like we do.

>>57540052
Do you have some precision and recall values?
How long did you take to manually sort the 36k images?
What were those 500 movies?
Normal movies with som sex scenes in them or a mixture of normal movies and porns?

File: gan-samples-1.png (607KB, 825x416px) Image search: [Google]

607KB, 825x416px

>>57540485
how long till it can generate new porn on demand?

>>57540464
>>57540503
http://www.pornmd.com/straight/most-popular
Scrape thumbnails from category, scrape thumbnails from query "sex", find what's abnormally common in thumbnails. Adding 'redhead' as a tag wouldn't work, the performer probably has the same hair color throughout.

>>57540573
Good questions. More info is here: https://github.com/ryanjay0/miles-deep/blob/master/README.md

It was only porn videos and all the 'other' category pictures come from the same videos. This isn't like Yahoo's NSFW model that is designed to look at non porn. Porn is its job so it doesn't need to see normal videos.

Making the training set was boring and took a while, but training and refining the models took a lot longer.

>>57539976
what tools do you learned to use and what books did you read in order to be able to do this?

I only have experience with OOP and functional langs

>>57540573
It gets 95% accuracy on my test set which comes from a totally different set of movies with less frames form each movie to reduce dependence. It has 2500 images. You can see the prediction weights for a smaple movie at that link.

>>57540654
Pretty impressive.
I see you mention not using temporal information.
What kind of use did you envision for it?

>>57525763
Easier than ever.

>>57540659

Caffe is a good machine learning framework especially for vision problems:

http://caffe.berkeleyvision.org/

And fine tuning:
http://caffe.berkeleyvision.org/gathered/examples/finetune_flickr_style.html

This is good too:
https://devblogs.nvidia.com/parallelforall/image-segmentation-using-digits-5/

And some general background on CNNs would be good:
http://cs231n.github.io/convolutional-networks/

I have an MS in CS with a focus on machine learning, but I think you could do this without all that schooling.

Also learn python if you don't know it already.

>>57540567
How are the images produced from the neural net?

>>57540744

Some have tried classifying video with different temporal architectures... the easiest is just run a copy of what i have on 10 frames and merge the results at the end (they call that late fusion). There's also early and slow fusion. Slow fusion is best but not really that much better than single frame. It might be better for certain actions though. Heres the paper: http://cs.stanford.edu/people/karpathy/deepvideo/

>>57540810
thanks for the link

>>57540766

Some of it is just hype. This happened in the 80s/90s when neural nets came out and were all the rage. Now again in 2006+ when Hinton showed a better way to train them (aka deep learning). There will be another ai winter. I'd bring a jacket.

>>57540784
Thanks a lot. I heard about Caffe but I didn't know it was a good choice.

>>57540836
I don't think so anon. DeepMind is innovating so much right now. CNN-based reinforcement learning and turing nets have already produced some amazing shit and still has much potential.

>>57540902

That would be great. I hope you're right. But reinforcement learning with neural nets isn't that new. The successes we've seen are largely based on vision problems that were difficult for computers before.

DeepMinds new DNC archictecture is a big leap and so is the one shot learning stuff, but I still think we will see some off the hype dissolve and the resurge before we get to AGI (artifical general intelligence),