Awesome

WORK IN PROGRESS ( Not any trained model available yet / Training in progress )

About this project

This is a first implementation of a Colorization Diffusion Based Method. Beware, this implementation comports mistakes but i don't have the time to work on it right now. For a correct implementation ones can look at the following repo. https://github.com/Janspiry/Palette-Image-to-Image-Diffusion-Models

How to train the model

Modify the conf.yml file, set the 'mode' option to 1. Then run the main.py file specifying the path to the config file (absolute or relative) Example : python main.py --config conf.yml

Future of this repo

I have impossibility to train and to test the model implemented due to my lack of computational power. There might be some mistakes in the code, any insight and remark is welcomed

For the validation loop in the training loop, necessity to use/find a more suitable/ an additional metric

Reference

Palette Image_to_Image Diffusion Models https://arxiv.org/pdf/2111.05826v1.pdf

Diffusion Models Beat GANs on Image Synthesis https://arxiv.org/pdf/2105.05233.pdf

The Unet Network script directly comes from the repo of this last : https://github.com/openai/guided-diffusion (with small modifications according to the Palette paper)

A colorization Dataset : https://www.kaggle.com/shravankumar9892/image-colorization ( Palette paper's researchers uses ImageNet )