A Neural Style Transfer implementation

Python, torch

NST principle

Inputs : content image C, style image S

A generated image (tensor G) is created from either a copy of content or from random noise. This tensor is the only trained parameter.
A pretrained VGG model, or a pretrained Resnet151 model is used to extract features from all three tensors. The features are stored at multiple steps during any forward pass in the model, L times

The content and style losses are defined as such :

And the final loss to backpropagate is

Results

Starry Night by Van Gogh, and a town

The Scream, by Edvard Munch and the Windows background image

Tableau I, 1921, by Piet Mondrian and a picture of the London skyline

Emergence of Orange, by Koola Adams and a Tokyo street

Emergence of Orange, by Koola Adams and a city skyline