Mascerade / fast-nst

🎨 Implementation of Fast Neural Style Transfer proposed by Justin Johnson et al. in the paper Perceptual Losses for Real-Time Style Transfer and Super-Resolution

Introduction

Fast Neural Style Transfer is a method proposed by Justin Johnson, Alexandre Alahi, and Li Fei-Fei in their paper Perceptual Losses for Real-Time Style Transfer and Super-Resolution. It is an adaptation of Gatys et. al.'s A Neural Algorithm of Artistic Style where an image is transformed to be in the style of another. Although producing great images, it takes a long time to train and it optimizes on only one image. With Johnson's method, we are able to create a neural network that can convert any image into another style that it was trained on in seconds.

Structure

There is a notebook that has the entire network laid out nicely so that someone could simply look at it and understand how it works. For the Python files, this is the layout:
src - Contains all the different functions necessary functions needed

--> models_architectures - Has the different models used (residual block, VGG19 and the actual transformation network)

Training

In order to train, all you have to do is change the style image within common.py in src and then run train! If you want, you can also adjust the content and style regularizers within training_functions.py in the total_cost function.

To Do

• Save generated images
• Add sample images here
• Add commands to make training/inferencing more user friendly