Chainer implementation of style transfer using neural network

Implementation of

• A Neural Algorithm of Artistic Style
• Preserving Color in Neural Artistic Style Transfer
• Combining Markov Random Fields and Convolutional Neural Networks for Image Synthesis

Requirements

• Python 2.7
• Chainer 2.0.0
• Cupy 1.0.0
• Pillow

Usage

Download VGG 16 layers caffe model

• Visit https://gist.github.com/ksimonyan/211839e770f7b538e2d8 and download VGG_ILSVRC_16_layers.caffemodel.
• Put downloaded file into this directory.

Convert caffemodel to chainer model

$ python src/create_chainer_model.py

Transfer image style using "A Neural Algorithm of Artistic Style"

$ python src/run.py -c content_image.png -s style_image.png -o out_dir -g 0

Options:

• -c (--content) <file path>: required
  Content image file path
• -s (--style) <file path>: required
  Style image file path
• -o (--out_dir) <directory path>: optional
  Output directory path (default: output)
• -g (--gpu) <GPU device index>: optional
  GPU device index. Negative value indecates CPU (default: -1)
• --w (--width) <integer>: optional
  Image width (default: 256)
• --iter <integer>: optional
  Number of iteration for each iteration (default: 2000)
• --initial_image <string|: optional Initial image of optimization: "random" or "content" (default: random)
• --keep_color: optional Keep color phase if specified
• --match_color_histogram: optional
  Use "Color histogram matching" algorithm in "Preserving Color in Neural Artistic Style Transfer"
• --luminance_only: optional
  Use "Luminance-only" algorithm in "Preserving Color in Neural Artistic Style Transfer"
• --resolution_num <int>: optional
  Number of resolutions (default: 1)
• --save_iter <integer>: optional
  Number of iteration for saving images (default: 100)
• --lr <float>: optional
  Learning rate: "alpha" value of ADAM (default: 10)
• --content_weight <float>: optional
  Weight of content loss (default: 0.005)
• --style_weight <float>: optional
  Weight of style loss (default: 1)
• --tv_weight <float>: optional
  Weight of total variation loss (default: 1e-5)

Transfer image style using Markov Random Fields algorithm

$ python src/run_mrf.py -c content_image.png -s style_image.png -o out_dir -g 0

Options:

• -c (--content) <file path>: required
  Content image file path
• -s (--style) <file path>: required
  Style image file path
• -o (--out_dir) <directory path>: optional
  Output directory path (default: output)
• -g (--gpu) <GPU device index>: optional
  GPU device index. Negative value indecates CPU (default: -1)
• --w (--width) <integer>: optional
  Image width (default: 256)
• --iter <integer>: optional
  Number of iteration for each resolution (default: 100)
• --initial_image <string|: optional Initial image of optimization: "random" or "content" (default: content)
• --keep_color: optional Keep color phase if specified
• --match_color_histogram: optional
  Use "Color histogram matching" algorithm in "Preserving Color in Neural Artistic Style Transfer"
• --luminance_only: optional
  Use "Luminance-only" algorithm in "Preserving Color in Neural Artistic Style Transfer"
• --resolution_num <int>: optional
  Number of resolutions (default: 3)
• --save_iter <integer>: optional
  Number of iteration for saving images (default: 10)
• --lr <float>: optional
  Learning rate: "alpha" value of ADAM (default: 2.0)
• --content_weight <float>: optional
  Weight of content loss (default: 0.2)
• --style_weight <float>: optional
  Weight of style loss (default: 1)
• --tv_weight <float>: optional
  Weight of total variation loss (default: 1e-5)

License

MIT License