This repository contains the codebase for the Image Completion project done for the Computer Vision 2017 Course at IIIT Delhi.

The codebase is maintained by Ambar Pal and Aishwarya Jaiswal

We will first train a GAN on the original data distribution and then use the trained model to perform image inpainting on corrupted images

python train.py --dataset .. --num_train_epochs .. --num_disc_steps .. --num_gen_steps .. --batch_size .. --save_checkpoint_every .. --generate_samples_every .. --flip_alpha ..

The following configurations work best for the various datasets:

• python train.py --dataset 'MNIST' --num_train_epochs 10 --num_disc_steps 1 --num_gen_steps 2 --batch_size 64 --save_checkpoint_every 250 --generate_samples_every 100 --flip_alpha 0.3

• python --dataset 'CIFAR10' --num_train_epochs 50 --num_disc_steps 1 --num_gen_steps 1 --batch_size 64 --save_checkpoint_every 250 --generate_samples_every 100 --flip_alpha 0.2

• python --dataset 'CELEBA' --num_train_epochs 15 --num_disc_steps 1 --num_gen_steps 1 --batch_size 64 --save_checkpoint_every 250 --generate_samples_every 100 --flip_alpha 0.2

• python --dataset 'SVHN' --num_train_epochs 15 --num_disc_steps 1 --num_gen_steps 1 --batch_size 64 --save_checkpoint_every 250 --generate_samples_every 100 --flip_alpha 0.3