rcalland / chainer-infogan

Chainer implementation of InfoGAN

InfoGAN

Chainer implementations of InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets http://arxiv.org/abs/1606.03657.

MNIST Latent Codes

• c₁ ~ Cat(K = 10, p = 0.1)
• c₂ ~ Unif(-1, 1)
• c₃ ~ Unif(-1, 1)

10 categorical and 2 continuous codes are sampled and concatenated with 62 noise variables z and then fed into a generator. Following are sample images created by a generator trained over 100 epochs on the MNIST training dataset consisting of 60000 samples.

c₁ - 10 Categorical

Each row shows 10 random images generated from the same one-hot categorical vector. K = 10 and hence such 10 rows, one for each categorical code. The network learn the latent representations of the 10 different digits with some exceptions such as the digit 1.

c₂ - Continuous, Interpolation

Interpolating over c₂ shows that the network learns to distinguish the width of the digit.

c₃ - Continuous, Interpolation

Interpolating over c₃ shows that the network learns the orientations.

Run

Train

python train.py --out-generator-filename ./trained/generator.model --gpu 0