sdujump / triangle-gan

implementation for NIPS paper Triangle Generative Adversarial Networks

Triangle-GAN

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

1. Experiments Settings:

1. Running environment:

tensorflow 1.1.0, python 2.7;

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:

The value function for TriGAN model:

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.

3. Compare with simplified Triple GAN:

figure (a): the joint distribution $p(x,y)$ of real data.

figure (b): $\Delta$-GAN left: the joint distribution $p_x(x,y)$; right: the joint distribution $p_y(x,y)$.

figure (c): Tirple GAN without regularization terms left: the joint distribution $p_x(x,y)$; right: the joint distribution $p_y(x,y)$.

4. Results

1. conditional generated CIFAR images:

2. MSCOCO:

1. generated COCO

2. COCO to attributes to COCO

3. CelebA tasks:

1. faces to attributes to faces

2. face editting