Image Colourisation

Image Colourisation project for the course Deep Learning at the VUB (2022-2023)

Fabian Denoodt, Ward Gauderis, Pierre Vanvolsem

---

The project is made in the form of a Jupiter Notebook. The notebook can be run on Google Colab, or locally after installing the required packages from requirements.txt with the command pip install -r requirements.txt.

jupyter notebook colourisation.ipynb

This notebook was used to train multiple models, test them ancd visualize the results. But the actual model code is organised in classes in the accompanying Python files.

Overview of the directory structure:

• colourisation.ipynb: The Jupiter Notebook
• dataset.py: The Dataset class
• model.py: The Model class
• euclidean_model.py: The Euclidean Model class with a different loss function
• simplified_model.py: The Simplified Model class with fewer parameters
• requirements.txt: The required packages to run the notebook
• CHECKPOINTS: The directory containing the final checkpoints of the trained models. In total 9 models were trained, but due to a conflict with GitHub's Large File Storage limit, the model weights are now being held hostage by GitHub until we pay for more storage space. However, we do include the last checkpoint of the Simplified final model, which we were able to recover.
• DATA: The directory containing the data used for experimentation. The dataset is not included but is downloaded automatically when running the notebook.
• DATA_FINAL: The directory containing the data used for training the final models. The dataset is not included but is downloaded automatically when running the notebook.
• DOWNLOADER: The directory containing the scripts used to download the data. These are also downloaded automatically when running the notebook.
• GRAPHS: The directory containing all made visualizations.
• PRECALCULATED: The directory containing all precalculated data: the mapping between Q and AB values and multiple distributions over the Q values.