DeepDream Mandelbrot

[Dinkydau]

This is a Kalles fraktaler render that was "deepdreamed". It is one of the very few times for me to upload something that is not a raw fractal, but still based on a fractal. The colors were also edited. This is a frame of an upcoming zoom video that will be published in 2 editions: the normal edition and a deepdream edition. The relatively low resolution of course has to do with the fact that this is a frame from a zoom video, but the process of deepdreaming also requires a lot of RAM. An extremely high resolution would not be doable on my computer with 16 GB of RAM. I'm not sure how much further I will look into deepdreaming. I might make more.

[kram1032]

I think the concept of deep dreaming would become so much more interesting if we could train such a neural net ourselves.
But training one must take so much longer than having it tell us what it sees.
However, eventually we might train networks to actually inform creative decisions in neat ways.

[3dickulus]

nice
I did this 7680x4320 on a machine with 6G ram + 2G swap partition
http://3dickulus.deviantart.com/art/Medusa-Effect-547596579
using https://github.com/jcjohnson/cnn-vis

[Dinkydau]

Thanks for that. I will have a look.

[Chillheimer]

nice one dinkydau.

hey 3dickulus, whoaaaat, how could you do 7680 with 6gb ram? My absolute max is 2560*1920 with 12gb ram! what is that "swap partition" thing you are talking about? is this just a linux thing, or does this work on windows machines as well?

[quaz0r]

in windows land its called the page file

[Chillheimer]

hmm... any tips how I could use this to achieve larger deepdream-resolutions?
(sorry for threadnapping dinkydau, if there's more response to my questions I'll split the topic)

[GratefulEd]

Hey Dinkydau...you might want to check out Googlenet_places205 if you're not aware of it.

It's available from the GithHub Model Zoo: https://github.com/BVLC/caffe/wiki/Model-Zoo

[Dinkydau]

hmm... any tips how I could use this to achieve larger deepdream-resolutions?
(sorry for threadnapping dinkydau, if there's more response to my questions I'll split the topic)

That's okay because if it works it would be extremely useful. Unfortunately I won't be able to try it for at least a few days because of vacation. (I finished the zoom videos just in time.)

[3dickulus]

hmm... any tips how I could use this to achieve larger deepdream-resolutions?
(sorry for threadnapping dinkydau, if there's more response to my questions I'll split the topic)

indeed, my intent was only to give you, Dinkydau, another option for rendering these crazy images.

@Chilli : I tried all the deep dream code I could get my hands on and CNN-vis is the one that, I think, has the most potential for hacking and the best memory management. It uses CUDA for just about everything via caffe and python, you have to compile caffe with USECUDNN (check the makefile for flags), I think there is a flag in CNN-vis too that needs to be set but going through the trouble of setting up and compiling everything with CUDA options enabled means that 7680x4320 image took 45 min on my quad core xeon @ 2.83Ghz, 6G+2Gswap and a 2G GeForce 760. This setup exploits page swap ? tiling on the GPU because there isn't enough ram (just 2G) to do it all on the GPU and that means optimized memory usage.

one cavaet, you also need to register as a developer to get the cuDNN library code @ NVIDIA cuDNN – GPU Accelerated Deep Learning again well worth the trouble