In this repo I re-implement A Neural Algorithm of Artistic Style from scratch entirly on Google Colab, using Imagenet pre-traing VGG19 network, with a lot of interesting gif demos of the reconstruction process from layers and the style transfer process.
To run the code through colab you just need add my drive folder for style transfer artifacts to your google drive storage since the code references content and style images from this folder, also saves generated images to it to generate videos and demos.
You may make your own directory structure on your drive and change the root_path referenced in the notebook, Also make sure of the root_path in your drive when mounted to Colab.
In summary you may need to
- Set the mounted path
- modify the layer of interset at Precomputing Source Style and Content Representation
- Either Do Style Transfer or Image reconstruction from layer
Modify the layer of interest at Precomputing Source Style and Content Representation then run Recontstruction From Layer Experiment.
Modify the layer of interest at Precomputing Source Style and Content Representation then run Style Transfer Experiment You may also modify the losses and the weighting of content and style loss.
Converting high resolution videos into gifs isn't a simple task, I used gifski it's a great and efficient tool for producing high-quality gifs.