Stylenet

This is a quick implementation of http://arxiv.org/abs/1508.06576

Blog post: http://t-satoshi.blogspot.com/2015/09/a-neural-algorithm-of-artistic-style.html

How the learning goes: https://googledrive.com/host/0B046sNk0DhCDcWZpeHNETWhza3M/top.html

Requirements

chainer (1.3.2) http://chainer.org And pre-tarined caffe VGG model

Warning! Be suer to use chainer 1.3.2!!
This code does not work in chainer 1.6 (I didn't test with from 1.4 to 1.5)

Quick Usage

install Anaconda (https://www.continuum.io/downloads) and then,

pip install chainer==1.3.2
wget http://www.robots.ox.ac.uk/%7Evgg/software/very_deep/caffe/VGG_ILSVRC_19_layers.caffemodel
python style_net.py -g -1 -c kinkaku.jpg -s style.png -d kinkaku

if you want to use GPU (recommended). 0 is GPU ID.

python style_net.py -g 0 -c kinkaku.jpg -s style.png -d kinkaku

Comment

There is a hard coding part. The computation of L_style in forward() This part is coresponding to equation (4) and (5). I did not paramatalized it because it would be too complizated. I think defalt is fine, but if you want, you can easily change it directly. See L_style in forward().

I recommend to make content image a squared size. However, you can use rectangular one, but the output will be forsed to a square. You need to resize agian, ex in python,

from scipy.misc import imread, imresize
img = imread('filename.png')
img = imresize(img,[400,300])

Basic settings is configured, in #Settings part, which is just after the import sentences. 20000 iteraton will be done. Image is saved every 500 iteration.