I want to generate 256*256 resolution images in 3 color channels (rgb) using a w-gan with gradient penalty. I have scraped the web and have compiled a dataset of ~2500 images. I used data augmentation (45 degree rotation, flipping) to increase this number to around 10000. Right now I am using a critic model with 6 million parameters and a generator model with 3.6 million parameters and a latent dimension of 128. I have tried training this model but it keeps generating shitty images and has a loss which is constantly in the hundreds. Are these parameter numbers good enough for generating such large images or should I change them?example of an image generated by this model
Asked
Active
Viewed 56 times
