CelebA experiments (PredictionOpt only on the generator)

[sanghoon]

Notes

• This work was done only to show some sample outputs. Different random seeds can lead to totally different outcomes. Therefore, we need to investigate outputs from repeated trials to correctly compare two GAN methods.
• For faster training, I used only 50k images from CelebA (resized to be 64x64)
• For these results, prediction methods have been applied only for G.

Large learning rate (0.01)

Vanilla DCGAN

After 2 epochs

After 10 epochs

After 25 epochs

DCGAN w/ prediction

After 2 epochs

After 10 epochs

After 25 epochs

Medium learning rate (0.0001)

Vanilla DCGAN

After 2 epochs

After 10 epochs

After 25 epochs

DCGAN w/ prediction

After 2 epochs

After 10 epochs

After 25 epochs

Small learning rate (1e-5)

Vanilla DCGAN

After 2 epochs

After 10 epochs

After 25 epochs

DCGAN w/ prediction

After 2 epochs

After 10 epochs

After 25 epochs