A few resources to further boost your knowledge in this area:
- To gain more intuition about GANs in general, I would suggest you take a look at this blog post.
- If you want to gain intuition about convolution and transpose convolution arithmetic, I would suggest referring to this paper.
- For more advanced techniques on training GANs, you can refer to this paper.
- You might have seen that reduction in loss and the quality of images do not track closely which is unlike other networks that we have seen earlier. To create a GAN that has this relationship that lower loss will result in better images, take a look at Wasserstein GAN. This GAN is based on a different loss function which is known as Wasserstein Loss.
- GAN is a prominent area of research. You can keep track of the new techniques introduced in this category here.
- You are currently using even-sized filters. Instead of using even-sized filters, please use odd-sized filters as the concept of a central or anchor pixel is more defined in the odd-sized filters. In odd-sized kernels, we have the same number of pixels on all sides from the anchor pixel. When using an even-sized filter, since this is not the case, it leads to aliasing errors. You can refer to page 2 of this document for different kinds of aliasing errors. I would recommend using a kernel size of 5 for your network here.
- Please use a stride of 1 in the last layer to avoid checkerboard-like artifacts in the generated images. You can read more about the phenomenon in this blog post.
- You could have implemented one-sided label smoothing for the real loss.
- Also, you can try using squared loss too. It has been found to work better than the cross-entropy loss. You can take a look at this paper where the least-squares loss was introduced. The images generated by Least Squares GAN were found to be of better quality than the classical GANs.