Theano/Lasagne based Neural artistic style transfer with Kivy GUI
based on https://github.com/Lasagne/Recipes/tree/master/examples/styletransfer an implementation of the algorithm described in "A Neural Algorithm of Artistic Style" (http://arxiv.org/abs/1508.06576) by Gatys, Ecker and Bethge.
use VGG19 model http://www.robots.ox.ac.uk/~vgg/research/very_deep/
(
wget https://s3.amazonaws.com/lasagne/recipes/pretrained/imagenet/vgg19_normalized.pkl
to get the pickled file)
Several modification are made to speed up the process.
- Doing the transfer in pyramid manner to speed up and seems generating better results.
- Has an option to use ADAM as the optimizer.
- Use shared memory and can precompile theanot functions.
- Has a server and a Kivy GUI client, so you can capture image using a mobile device and the use the GPU on a remote machine.
content and styles Many image files are from https://github.com/jcjohnson/neural-style and some are public domain
CC-By http://agf81.deviantart.com/art/Ivy-Texture-7-326197006?q=gallery%3ACreative-Commons%2F26171837&qo=33 Ivy texture
CC-BY-SA http://maadmann.deviantart.com/art/Fullcolor-437966862?q=gallery%3ACreative-Commons%2F26171826&qo=88 fullcolor
CC-BY https://www.flickr.com/photos/fontplaydotcom/24795772945/in/pool-734017@N25/ color squares
CCB-BY-SA http://oilsoaked.deviantart.com/art/holiday-188505340?q=gallery%3ACreative-Commons%2F26171826&qo=373 Holidy