Creating image with WGAN with only 2000 images in traning data

Deep Learning
1

Hi All,
I was trying to create a new image of a Hindu God, trying to see how would a NEW God look like if we create it through machine learning. After scraping from google, wiki and Flickr, I could still collate 2000 pictures and after about 1600 epochs I still can’t get a clear picture.
Anyone who has any suggestions here, please feel free to look at my code and suggest.
https://github.com/Nik2049/GAN_Image_generation

Thanks.

2

Your dataset is simply too small to train a decently performing GAN from scratch I think. I had worked on a model for generating Pokemon using about 2100 images with large amounts of augmentation. The results were pretty similar to yours. Heres’s the link, although it is in pure Tensorflow 1.x(the older ugly version) : https://github.com/Atom-101/PokeGAN

Although the README says I trained for 10 epochs, I gave it around 50.

1 Like
3

@jeremy Can’t we use Imagenet pretrained model as a generator (without a few final layers) here? Apologies for tagging you out of nowhere.