Simpsons GAN
This repo is the implementation of BigGAN and the training target is [Simpsons faces](https://www.kaggle.com/kostastokis/simpsons-faces) in kaggle public dataset. Although the implementation refered to BigGAN, but there are still some modifications in order to have a better fit on Simpsons dataset. And this is still an on going project, since the training process of GAN is a brittle work. But the current results seems already done with 50% of work, which means you can tell the generated images are a little bit Simpsons-ly. I will keep updaing the result/code until the generated images are good to go.
Dependencies
python 3.6
Tensorflow 2.1.0
Model
What original BigGAN do :
- SA-GAN architecture
- Hinge loss
- Conditional batch normalization
- Spectral normalization
- Moving averages of generator's weights
- Orthogonal initialization
- Orthogonal regularization
- Truncation trick
What I do:
- SA-GAN architecture
- Hingeloss
- Spectral normalization
- Orthogonal regularization
I think the conditional batch normalization could improve the results a lot, but unfortunately, the original dataset doesn't have categorical/class label.
Training recipe
Most of the hyper-parameters can refer to project_config.py, which is pretty much the current model setting. The total training epochs is around 1000 epochs, and the steps of each epoch is total_epochs//batch_size. Since the hardware limitation, so the biggest batch size I could use is 64, and the image/generated image resolution is (64,64,3). Both generator and discriminator are using Adam optimizer with learning rate 1e-4 for G and 2e-4 for D, and beta_1 = 0, beta_2 = 0.9
Results
