Task: pokemon image generation
Dataset: pokemon-image-dataset
Augmentations:
- Resize to 64x64,
- Random horizontal flip
- Color jitter
Training:
- 1000 epochs
- batch_size=128
- Adam: LR = 0.0002 BETA = (0.5, 0.999)
Models:
- DCGAN
- WGAN-GP
Best model: DCGAN with tweaks from ganhacks:
- A modified loss function (for G min (log 1-D) --> max log D)
- Label smoothing
- Adding noise to inputs
- LeakyReLU insead of ReLU
Result:
