Triangle-GAN

October 30, 2017 · View on GitHub

This is an implemtation for NIPS paper: Triangle Generative Adversarial Networks

1. Experiments Settings:

2. Dataset Format

For domain transfer and classification task, CelebA and MSCOCO dataset need to be in HDF5 format;

For semi-supervised learning tasks, please see here

3. Resources

For domain transfer and classification task: If you want to re-run the CelebA experiment, the feature can be downloaded here: CelebA tag features

2. Basic Model:

Here's is our model:

alt text

The value function for TriGAN model:

alt text The objective of $\Delta$ -GAN is to match the three joint distributions: $p(x, y)$ , $p_x(x, y)$ and $p_y(x, y)$ . If this is achieved, we are ensured that we have learned a bidirectional mapping $p_x(x|y)$ and $p_y(y|x)$ that guarantees the generated fake data pairs $(\hat{x}, y)$ and $(x, \hat{y})$ are indistinguishable from the true data pairs $(x, y)$ . In order to match the joint distributions, an adversarial game is played. Joint pairs are drawn from three distributions: $p(x, y)$ , $p_x(x, y)$ or $p_y(x, y)$ , and two discriminator networks are learned to discriminate among the three, while the two conditional generator networks are trained to fool the discriminators.