Image Style Transfer Using Convolutional Neural Networks in PyTorch

This is unofficial pytorch implementation of the paper, "Image Style Transfer Using Convolutional Neural Networks" [Gatys+, CVPR2016].

Requirements

• Python 3.5+ (tested with 3.5.4)
• PyTorch 0.2.0+ (tested with 0.3.0.post4)
• TorchVision 0.2.0+ (tested with 0.2.0)
• Numpy 1.11.1+ (tested with 1.13.3)
• Pillow 5.0.0+ (tested with 5.0.0)

Usage

Options

• --content, -c: The path to the content image.
• --style, -s: The path to the style image.
• --epoch, -e: The number of epoch. (Default: 300)
• -content_weight, -c_w: The weight of the content loss. (Default: 1)
• -style_weight, -s_w: The weight of the style loss. (Default: 1000)
• --initialize_noise, -i_n: If you use this option, the transferred image is initialized with white noise. If not, it is initialized with the grayscale content image.
• --cuda: If you have an available GPU, you should use this option.

Examples

With CPU:

python style_transfer.py -c contents/golden_gate.jpg -s styles/kandinsky.jpg

With GPU:

python style_transfer.py -c contents/golden_gate.jpg -s styles/kandinsky.jpg --cuda

Installation

git clone https://github.com/enomotokenji/pytorch-Neural-Style-Transfer.git
cd pytorch-Neural-Style-Transfer

Docker

docker build -t style_transfer .
docker run -it style_transfer

Nvidia Docker

nvidia-docker build -t style_transfer_gpu .
nvidia-docker run style_transfer_gpu

Without Docker

Install PyTorch and dependencies from http://pytorch.org.
We have prepared requirement.txt, but it is preferable to use Anaconda as recommended on http://pytorch.org.

References

• Leon A. Gatys, Alexander S. Ecker and Matthias Bethge. "Image Style Transfer Using Convolutional Neural Networks", in CVPR 2016. [Paper]
• Code is inspired by Neural Transfer with PyTorch.