Built based on the pytorch fast-neural-style example for artistic style transfer, this repository try to re-implemented Conditional Instance Normalization layers to train transfer network with multiple style images at the same time. What I did here is basically replaced all instance normalization layers in fast-neural-style with conditional instance normalization layers

Conditional Instance Normalization was introduced in A Learned Representation For Artistic Style and implemented with Google Magenta TensorFlow. I first learned this paper while reading Joel Moniz Lasagne and Theano implementation. The creation and usage of the method in this repository is based on my limit knowledge in python, pytorch and neural network.

Please refer to fast-neural-style for more details.

Train

python neural_style.py train --dataset </path/to/train-dataset> --style-image </path/to/style/image> --save-model-dir </path/to/save-model/folder> --epochs 2 --cuda 1 --batch-size 4

• --style-image: the code will grab all files under the path as style images
• --batch-size: number of images fed in each batch does not need to be the same to the number of style images

Stylize

python neural_style.py eval --content-image </path/to/content/image> --model </path/to/saved/model> --output-image </path/to/output/image> --cuda 0 --style-num 19 --style-id 18

• --dataset: path to training dataset, the path should point to a folder containing another folder with all the training images
• --style-num: total number of style images, must be the same as the amount used in training
• --style-id: a number from 0 to style_num - 1, indicating which style to transfer to

COCO 2014 Training images dataset [83K/13GB]

I used 19 style images, most of which are from other great style-transfer-related Github repos I read through:

• pytorch fast-neural-style
• jcjohnson
• hwalsuklee

Model used in the following examples can be found in the pytorch_models folder