Neural Style Transfer by Gatys et al. Chainer 4.0

Chainer v4.0 implementation of "Image Style Transfer Using Convolutional Neural Networks(2016)" by Gatys et al.

Image Style Transfer Using Convolutional Neural Networks. Gatys, L.A. et al(2016)

The purpose is further research of Gatsy's article. So source code keeps very simple for modification and for that reason, the sophisticated functions of chainer(i.e. updater, trainer) are not used.

Results

Original Image

hourly bell at Kawagoe city Japan.

Style 1

"Starry Night" by Van Gogh.

Result 1

Style 2

"Gardanne" by Paul Cezanne.

Result 2

GPU: GeForce GTX1080i

Elapsed Time: about 12min for generating 600px squared image.

Usage

Environment

• python3.5 (3.0+)
• chainer4.0
• cupy4.0.0
• cuda8.0+
• cuDNN6.0+

Generate Transferred Image

python generate.py -i images/tokinokane600.jpg -s [dir_path]/style.png -o [dir_path for output] -g 0

Shape of images must be square.

Parameters

See source

Acknowlegement

thanks to the authors to which I referred.

• chainer-gogh(https://github.com/pfnet-research/chainer-gogh)