This repositiory is a reimplementation of Zhu, J., et al. "Toward multimodal image-to-image translation.(2017)." arXiv preprint arXiv:1711.11586 (2017) for CIS 680 - Advanced Topics in Machine Perception.
To run, set the project_folder variable and root_dir as required. The directory structure should be as follows:
.
├── datasets
│ ├── edge2shoe # edge2shoe dataset as available on the website
│ └── ...
├── models
├── losses
└── plots
Post setup, simple run all cells sequentially. The model will train and compute the LPIPS and FID score for the trained model.
Inference can be performed by
z_random = torch.rand([1,8]) # generate a random seed vector
# outline is an outline from the original distribution
fake_images = generator(norm(torch.unsqueeze(outline,0).to(device)),z_random.to(device))
``