This notebook is part of the Udacity Deep Learning Nanodegree and it is based from one of their notebooks!
This repo showcases a notebook where a CycleGAN is defined and trained to read an image from a domain X and transform it so it looks as if it belongs in domain Y. Specifically, we are going to look at a set of images from Yosemite National Park taken during summer and winter and changed between these two domains!
The images do not come with labels but CycleGANs give us a way to learn the mapping between these two domains using an unsupervised approach. After all, CycleGAN is defisned for image-to-image translation and it learns from unpaired training data.
If there are some changes or any ML blasphemies that I committed here and there, feel free to open an issue and point me to the right direction! I'm just learning after all! 😄
This notebook is a proof of concept so it is not meant to yield realistic images. In fact, one shortcoming of this model is that it produces fairly low-resolution images. This is an ongoing area of research and there have been attempts to create higher-resolution formulations that use multi-slace generator models, as described in this paper.
We can see the progress of yielding images from the other domain and how it fares much better as the model is trained through iterations!
| First iterations | After 8000 training iterations |
|---|---|
 |
 |
