This project serves as the final project of CIS680 Fall 2020: Advanced Topics in Machine Perception at University of Pennsylvania.
Authors: Kun Huang, Zhihao Ruan
This project explores BicycleGAN implementation with modified ResNet generator from CycleGAN, PatchGAN discriminator, and ResNet encoder. The training process is built on the dataset edges2shoes.
The model trained on our PC reaches an FID score of 76.822, and LPIPS of 0.23297. For more information on FID score & LPIPS metric, visit:
- FID score: https://github.com/mseitzer/pytorch-fid
- LPIPS metric: https://github.com/richzhang/PerceptualSimilarity
To train the network on edges2shoes dataset, at the project root folder run:
sh scripts/download_and_train.shTo run inferences as well as generate quantitative evaluation on the trained model, run
sh scripts/infer.shDataset will be automatically detected and downloaded if necessary upon running scripts/download_and_train.sh. If one is interested in investigating the dataset, he/she could visit data/ for more information.
We have trained two different models. One model uses the same discriminator for both cVAE-GAN cycle and cLR-GAN cycle, and the other model uses separate models. For more information please visit checkpoints_archived.
Toward Multimodal Image-to-Image Translation.
Jun-Yan Zhu, Richard Zhang, Deepak Pathak, Trevor Darrell, Alexei A. Efros, Oliver Wang, Eli Shechtman.
UC Berkeley and Adobe Research
In Neural Information Processing Systems, 2017.
GitHub link: https://github.com/junyanz/BicycleGAN