Trying the Self-Attention-GAN with dog images

[liku-amare]

Hi, I was able clone this repo on my machine and it successfully worked. But, I wanted to work on dog images generation so I changed the content of the folder CelebA with images of dogs by deleting the original celebrity images. I have about 20k images of dogs, all of different resolution. I tried running the code with this dataset by setting the batch_size to 8 (because of CUDA outof Memory issue) and with some different number of steps (10k, 20k ... ). However, the results weren't even good enough to tell what the object inside the image is. This is the result after 20k steps. The model produces quite descent images for the CelebA dataset even after few hundred steps. What is the problem?