Turning Neural Networks Upside Down

[Syntopia]

Here is a new implementation of Google's inceptionism visualization: https://317070.github.io/LSD/  - including an interactive demo: http://www.twitch.tv/317070

For a great introduction to the Convolutional Neural Networks behind this, try the @karpathy course: http://cs231n.github.io/convolutional-networks/

[Chillheimer]

cool, just read this in the twitch-stream chat:
"benanne: code will be released in a few days! takes some tinkering to get it to run though"


here's a little more info about the real-time render-video i copied from the chat:

 carmethene_tv: @317070 +1, thanks for doing this! any chance of making the software behind this available? it would be fantastic for parties
paradonym: carmethene_tv I'd guess the needed CPU/GPU power is WAY too high for standard desktop systems
benanne: @Carmethene_tv source will be made available in a few days, but it will mainly be useful for experimentation I tihnk
benanne: and yeah, you need a decent GPU and lots of RAM
paradonym: @benanne - does this run on CAD or standard GPUs?
benanne: GTX 980 previously, now it's running on an older GTX 680
benanne: standard gamer GPUs
gwelengu: wow, 32 GB? that's crazy. I'll never be able to run this
ChaozCoder: @paradonym: or maybe it was some kind of bundled GPU i don't know exactly, but nothing like a supercomputer
benanne: you need 4GB of video RAM
ChaozCoder: @gwelengu: he said the machine has 32 GB you probably don't need that much
ChaozCoder: @benanne: ok
benanne: 32GB of CPU RAM but that may be fixed at some point. We suspect some sort of memory leak in the encoder
ChaozCoder: yeah i meant cpu ram not gpu ram
Dustinator2: my rig is a little lacking in the RAM department but my graphics card makes up for it until I can get more
ChaozCoder: memory leaks, that's why i love c# (for non performance critical stuff)
benanne: @Chaozcoder: not much we can do about it though, it's not in the code @317070 wrote but rather in the video encoder it's using. Maybe he'll try with a different one, I don't know
ChaozCoder: @benanne: oh
benanne: he may already have solved it, I don't know, it definitely seems to be more stable today
317070: @Chaozcoder @Benanne I think I solved the problem though. There was some fishy socket-handling by this bot which could have caused the issue.
ChaozCoder: so the bot was causing it, that is hilarious lol
foofoobarfoo: which software/language has been used in implementing this? python/theano..?
ChaozCoder: of all he things
benanne: okay. Still 27GB of memory in use right now
benanne: @Foofoobarfoo yeah, Python, Theano, Lasagne
foofoobarfoo: it's something like a diabolo network? (autoencoders)
benanne: My guess is the code could be made much more memory efficient but he doesn't need to bother because the machine has 32GB anyway
benanne: @Foofoobarfoo: no it's a feedforward neural net run backwards. Technical details here:http://317070.github.io/LSD/
paradonym: Next step: comparing different hardware's dream-images to the exact same thing
wiibrewer: You should train these on larger image sets so they become more defined
carmethene_tv: again, thanks for setting this up, it's a fantastic idea
benanne: @Wiibrewer: yeah, with more data it should work better. We did not train this net ourselves though, we just downloaded the parameters to save time
Dogeapi: was hoping for a link to some code at the github page, but good writeup!
carmethene_tv: so it only recognises what's in the list linked below?
MysteryXi: What is this running on? Isn't this super CPU intensive?
wiibrewer: @Benanne ahh, gotcha. Does it require monitoring to be trained?
wiibrewer: @Benanne meaning does someone have to be there to train it?
benanne: @Wiibrewer: it requires a lot of parameter tuning, but once you've foudn the parameters you just leave it running for a couple of weeks on a GPU
benanne: @Mysteryxi: the neural net is running on a GTX 680 GPU. The CPU only does the frame interpolation and video encoding (which is actually pretty intensive)
wiibrewer: @Benanne awesome, so in theory I could set up and automate this thing myself?
@benanne so this is mainly a RAM-Work? - Could you probably post Hardware occupation? How much load on GPU/CPU/RAM aso?
 benanne: if you have the hardware, sure
* carmethene_tv: @Benanne where should I keep an eye out for the software release?
benanne: $ uptime 18:03:19 up 19:39, 2 users, load average: 5.44, 5.88, 5.92
benanne: hexacore CPU is pretty much maxed out
wiibrewer: @Benanne is 20 gigs of ram, gtx770, quad core cpu strong enough?
benanne: for the GPU we don't have utilization info because it's a gamer GPU. But it's at 73C right now which means it's not maxed out probably
benanne: wiibrewer: if you reduce the resolution a bit it should be. Or it might not even be necessary, I don't know
 benanne: oh yeah, important detail: this is running on a Linux machine. Setting it up on Windows / Mac OS is not impossible but it'll be a lot more tedious
paradonym: this chat nearly convinces me purchasing 32 gb of RAM instead of a GTX 980 ti
MysteryXi: @benanne What distro?
benanne: @Paradonym: get both,980 Ti is sweet too
paradonym: to run this as CPU benchmark
benanne: @Mysteryxi Ubuntu 14.04 LTS





edit2: they're updating it right now and are enabling 2word combination. I wonder what the tractor-sloth will look like..

[kram1032]

another blogpost http://googleresearch.blogspot.co.at/2015/07/deepdream-code-example-for-visualizing.html

[Chillheimer]

Wooooooohoooooooo!!!!!!!!!!
Now you can do this by yourself!
http://deepdreams.zainshah.net/
upload an image and let it dream!  

(edit: i guess we'll have to wait a little longer.. site is extremely slow, I've been waiting 10 minutes for the last pic)

[Chillheimer]

soooo cool!

[KRAFTWERK]

  I am in love with this AI!

[KRAFTWERK]

Second test, original image:

[Chillheimer]

me too!
this is so cool! and how much cooler will this become, when improved?
Think of telling the computer what to look for, like in the live-stream linked above..
tell it to search for... fishes, or whatever.. in a picture and it will make it look "fishier"
and then this will be used on movies as well! and in 3d. and on the oculus!
oh my god!

the future is AWESOME!!!!

[KRAFTWERK]

It is really weird 

I wish it would be possible to teach it, to choose which images it should know from the start...

[kram1032]

Those are really pretty
More restricted/targeted versions like in the stream would be great though.
Here's an experiment to try:
Use a high res MSet image to be interpreted.
Apply another step of z^2+c.
Reinterpret
Remap...
Will the image converge to an attractor of both z^2+c and the AI?
(You know,  Orbitmap style)

[cKleinhuis]

people, where are you creating these images ?!?!? the links seem to be broken

[Chillheimer]

people, where are you creating these images ?!?!? the links seem to be broken

still here: http://deepdreams.zainshah.net/
you need lots of patience, it's the first online-page to do this an it's probably totally overwhelmed. So it can take a few minutes until the site opens.
I actually have 6 tabs of it open, trying to dream non-stop..

edt: now it really seems to be completely down.
but I'm sure others will follow. this will get huge!!

here's the results:
http://chillheimer.deviantart.com/gallery/

[ellarien]

What fun!

This one started as a JWildfire image. I get the impression there were a lot of dogs in the initial training set ...

[cKleinhuis]

still here: http://deepdreams.zainshah.net/
you need lots of patience, it's the first online-page to do this an it's probably totally overwhelmed. So it can take a few minutes until the site opens.
I actually have 6 tabs of it open, trying to dream non-stop..

edt: now it really seems to be completely down.
but I'm sure others will follow. this will get huge!!

here's the results:
http://chillheimer.deviantart.com/gallery/

this produces incredible results, in fact training the system so much no one knows what comes out of it

your image with the fractal looks incredible,because the machine uses the predefined streaks to build stuf along it, very astonishing ...

[M Benesi]

you need lots of patience, it's the first online-page to do this an it's probably totally overwhelmed. So it can take a few minutes until the site opens.
I actually have 6 tabs of it open, trying to dream non-stop..

edt: now it really seems to be completely down.

really?