Introduction: The paper we are going to implement is "Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks". This is an unsupervised learning problem.
Challenges: First of all, it takes really long time to train the model. Every single mistake can lead to loss of 10 hours.
Besides, there is big difference between model of horse2zebra and monet2photo since we have to reserve similar color while transferring the style with CNN. And the solution is to add identity loss, which computes the distance from original painting to the fake painting, to loss function. This is because we desire a same picture after passing monet painting into model of monet2photo and photo2monet. Now our model is perfectly working without making a person look like ghost :)
Insights: Concrete result: Horse to zebra and zebra to horse:
Apple to orange and orange to apple:
Monet to photo and photo to monet:
Plan: In the following days, we’d like to train models of Vangogh, ukiyoe, manga, cezanne2photo so we can compare them together to make our project result more reliable. After that, we will implement our models on videos to transfer its style.