contrastive-unpaired-translation VS PyTorch-GAN

For unpaired image-to-image translation, the SOTA is probably Contrastive Unpaired Translation, which is developed by the same group that developed CycleGAN and is kind of the successor algorithm.

Here is link number 1 - Previous text "CUT"

Code for https://arxiv.org/abs/2007.15651 found: https://github.com/taesungp/contrastive-unpaired-translation

there are many image to image translation models you can choose from. A good starting point would be PyTorch-GAN. I guess you're able to pick one by experimenting yourself. Good luck in your interesting work, and report back once you're done.

By the way, if you are interested in GANs, this repository may help you: https://github.com/eriklindernoren/PyTorch-GAN.

If you are familiar with pytorch, I highly recommand to take a look at https://github.com/eriklindernoren/PyTorch-GAN as they have papers and implementations

I also happen to know of a old (but it checks out) mass-port-to-pytorch github: https://github.com/eriklindernoren/PyTorch-GAN . That one is handy for more seeing what kind of models are out there, but they're probably not SOTA.

Hi just an update to whom ever faces a similar issue. I've fixed it.The key change was redefining the encoder loss - where kld is defined without logvar (only mu) and the reconstruction loss is defined with L1 rather than MSE. This was a helpful resource on the matter https://github.com/eriklindernoren/PyTorch-GAN/blob/master/implementations/unit/unit.py