Improvements to A Neural Algorithm of Artistic Style

University of Michigan EECS 445 Artistic Style

Collaborators: Brad Frost (@bfrost2893), Kevin Pitt (@kpittumich15), Nathan Sawicki, Stephen Kovacinski (@Kovacinski), Luke Simonson (@lukesimo)

This project was influenced by the paper A Neural Algorithm of Artistic Style. This was the final project of our EECS 445: Introduction to Machine Learning.

To use the script to make your own awesome artistic photos, you will need to have an NVIDIA GPU with CUDA drivers installed. This code would take hours upon hours when running on the CPU, but we may add this option if necessary.

Installation

git clone https://github.com/bfrost2893/eecs445-project
pip install -r requirements.txt

Make sure that your machine has the CUDA drivers installed. If your machine is configured to work with Theano GPU support, it will work as expected. See this documentation article for support on configuring Theano to work with NVIDIA GPUs.

Usage

This script should work for most major Linux distributions. We implemented using it Ubuntu 14.04 LTS (Trusty). However, it should work for any distribution.

Generate an image using content/content.jpg image and style/picasso paintings with K = 4 nearest neighbors and weights 0.4, 0.3, 0.2, and 0.1 weights on the applied style images.

python art.py --content content/content.jpg --style style/picasso -k 4 -w 0.4,0.3,0.2,0.1

Generate an image using content/content.jpg image and style/picasso.jpg, which only applies one style image to the content with no other options.

python art.py --content content/content.jpg --style style/picasso.jpg

Options

• --content, -c
  • The content image that you wish to apply style to
  • This must end with either '.jpg' or '.png'
• --style, -s
  • The style directory or image that you wish to apply to the content image you specify
  • This must be a relative directory to the script and must not begin with '/', just the name. For example 'content/monet'
• -k
  • The number of style images you wish to apply
  • This will choose the best four style images according to the Euclidean distance between the content and the style images (a simple k-Nearest Neighbors implementation that we used for preprocessing)
• --weights, -w
  • Add custom weights to each neighbor image using a comma separated string
  • Example: -w .6,.4
  • The weights must sum to one and must equal the K value