This repo contains the code for my CS4243 project on image inpainting.
At a high level, we scrape 200k images from iNaturalist, create a dataset using torch.data.Dataset
, and then train several GAN models on it.
In particular, we explore:
- Partial and Gated convolutions
- Multitask architectures
- Contrastive formulations of the task
- Discrimination at different scales - specifically, local vs. global scale discrimination
- PixelShuffle and similar upsampling techniques
- Geometric learning via GIN convolutions by formulating the image task as a graph task
You can find most of them in the active_experiments
folder, though it's not complete.
We also explored the use of diffusion models, but did not have the time or bandwidth to.
For a neater presentation of the material, check out our video.
The project was awarded an A+ and the top project for the class. See the digital certificate.
The repo is slightly messy. Will clean it up when I find some time to.