Neural Style Transfer

This repository contains the code for my implementation of the Neural Style Transfer paper by Gatys et al.

The deep learning model is able to take a content image and render it in the form of a given style image.

Sample

Content Image

Style Image

Rendered Image from model

Setup and Training

See the requirements.txt for dependecies related to Python. After cloning the repository, you can install all these requirements using the make command. It is recommended to create a virtual environment for this purpose. The Makefile will then run the main.py script to start training the model with sample images. The model will run for 20000 iterations before saving the new image in the data folder. From my experiments, this should take around 14.5 minutes on a Tesla P100 GPU. You can train using your own images by specifying the --content and/or --style argument when running the main.py script.

References

• A Neural Algorithm of Artistic Style

Acknowledgements

• Style sample photo is The Starry Night by Vincent van Gogh