Robin Byl, Florence Franchomme, Nédid Ismaili, Adeline Wantiez

This project aims at inpainting missing region(s) of an image based on the remaining pixels using deep learning-based methods. For this, a Generative Adversarial Network (GAN) has been implemented. Its purpose is to generate visually plausible image to fill the missing content. The considered dataset is the Animal Faces-HQ (AFHQ).

To run the code go to Usage section.

Results when feeding the whole image to the discriminator:

Results when feeding the patch to the discriminator:

-dataset
    -afhq
        -cat
        -dog
        -wild
-Model
    Generator.py
    Discriminator.py
-Results
    General_architecture.png
    ResultsExample.png
Final_figures.py
Preprocessing.py
read_data.py
README.md
Training.py
TrainigFunctions.py
useTrainedModels.py

git clone https://github.com/FFrancho99/Project_ML.git

numpy >= 1.23.5
matplotlib >= 3.7.1
torch >= 2.0.0
torchmetrics >= 0.11.4
torchvision >= 0.15.1
uuid >= 1.30

Cuda is required to run the code but if you don't have it, change line 49 of Training.py 'cuda:0' to 'cpu'.
Run Training.py file to create, train, validate and test the model. If you already have a saved model then run useTrainedModels.py.

Dataset: Animal Faces-HQ (AFHQ) from https://www.kaggle.com/datasets/andrewmvd/animal-faces