stylegan2-pytorch
stylegan2-pytorch copied to clipboard
lucidrains
The generated samples have very little difference
I'm trying to generate some avatars of NBA players.
After 5000 iterations, I can barely see the silhouette of the player. But the generated samples look the same. Is this normal?
The command I run is
stylegan2_pytorch --data statmuse_NBA --name statmuse --multi-gpus --transparent --batch-size 64 --aug-prob 0.25 --num-train-steps 100000 --no-pl-reg
How many images do you have in your dataset?
1200 images, is it too few?
Not necessarily too few. You could increase --aug-prob, include color in aug_types (next to default translation and cutout), and/or you can use a pre-trained model. I hope this will help a little!
Not necessarily too few. You could increase --aug-prob, include color in aug_types (next to default translation and cutout), and/or you can use a pre-trained model. I hope this will help a little!
@Otje89 Thank you for your suggestion, I will try it.
@chris-aeviator It does matter!
I increased the --aug-prob to 0.7 and collected more images for training.
I also tried transfer learning(from CelebA to my customed dataset)
@FrazierLei cool! Could you share the number of images you have been using to generate these?
@chris-aeviator I removed some images that are actually the same player.
E.g 
Then I collected some avatars of players from other leagues. The total number of images is about 1.5k.
@FrazierLei Nice work! I'm really curious how you went by to do transfer learning with this implementation. Would you mind explaning?
