Sophie's Secret LJ

(alias the Roost of the Purple Chicken)

Previous Entry Share Next Entry
A vague attempt at documenting my attempt to make neural network art
I like pi!
Neural Network/Deep Learning/Style Transfer art is this sort of thing.

This is mostly for my own reference when I next feel up to poking at the problem some more.

Unfortunately, this stuff is still fairly cutting edge, and all the documentation assumes you're an AI researcher running Linux. I'm sure there'll be simple to use graphics programs aimed at artists eventually, but for now the only things along those lines are websites and apps, which lack the flexibility I want.

So, the first promising looking program I found was Fast Style Transfer in TensorFlow.

Step one: The Machine Intelligence library Tensorflow.

This needs python, which is harder to install on Windows than it should be but I got it in the end.

I missed the Nvidia requirements so my GPU-using install of TensorFlow complained.

This page gave some advice on installing the Nvidia requirements for using the GPU but TensorFlow still wasn't happy.

So I uninstalled TensorFlow and re-installed it to just use the CPU. It worked! Sure, it will, from all accounts, run like molasses but I wanted to make sure everything worked at all before spending any more time debugging TensorFlow.

Step two: Fast Style Transfer

I installed this only to realise I'd missed some dependencies here too. SO MANY DEPENDENCIES. And they should apparently be downloaded as part of a package, but there's multiple options and I got decision fatigue and gave up for the evening.

Something unfortunate I noticed: Fast Style Transfer runs in python, which I have running, but there's a file which, afaict, will only run in Linux. It looks like processing images using a trained neural network doesn't need, but training your own does. Training is also the slow part that needs a GPU.

But training my own neural network was half the point, especially since Fast Style Transfer doesn't have any image processing settings to fiddle with.

Other options:

I did find another TensorFlow based style transfer neural network package, Neural Style that has more settings, but it relies on other people's trained neural networks and the example images don't look as pretty. neural-style-tf looks promising.

This neural-style looks nice but relies on Torch which doesn't run on Windows. On the plus side there's an associated page of similar programs.

I came across this article explaining the science by a guy with a webapp Stylist. I haven't looked into them very deeply.

Deep Learning with GPU on Windows 10 looks promising.

Style App runs on PC or mac and is very simple to use...and none of the filters do what I want. Could be fun for other contexts though.

Anaconda includes a bunch of neccesary sciencey python libraries and brings along it's own version of python.

This entry was originally posted at There are comment count unavailable comments.


Log in

No account? Create an account