Code structure

The two main source files are as follows :

• The code for the style transfer algorithm can be found in src/styleTransfer.py.
• ScipyOptimizer.py simply acts as an interface between Keras/Tensorflow and scipy. It is a dependency used by styleTransfer.py.

Moreover :

• The file batchTransfer.py allows us to batch run the style transfer algorithm.
• denoise.py performs total variation denoising on an input image.
• transform_caffe_vgg_normalized_gatys_weights.py loads gatys original normalized weights into a caffe model and transforms them to tensorflow format.
• weight_normalizer.py samples the activation across a directory of images and normalizes VGG19 weights to have unit mean activation across all images, all positions.
• vgg19_loader.py loads a VGG19 model with max pool layers replaced by average pool layers.

Running the code

First, make sure that Tensorflow and Keras are properly installed. Install all the requirements in the requirements.txt file with pip.
We may have missed some requirements, in that case please send a message.

In order to perform style transfer on a singleimage, simply run the following in the terminal : python3 styleTransfer.py -s <pathToStyleImage> -c <pathToContentImage>.
Some other options are also available, type python3 styleTransfer.py --help for details.

If you wish to automate the style transfer process and schedule several runs of the algorithm, modify and run batchTransfer.py accordingly. Details are provided in the surce file.

Some sample images used in the report have been provided in the images folder.