Awesome

InfoGAN

InfoGAN Architecture

Tensorlayer implementation of InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets.

Results

MNIST

Manipulating the First Continuous Latent Code

Changing <img src="https://latex.codecogs.com/svg.latex?c_1" title="c_1" /> will rotate the digits:

Manipulating the Second Continuous Latent Code

Changing <img src="https://latex.codecogs.com/svg.latex?c_2" title="c_2" /> will change the width of the digits:

Manipulating the Discrete Latent Code (Categorical)

Changing <img src="https://latex.codecogs.com/svg.latex?d" title="d" /> will change the type of digits:

Random Generation and Loss Plot

G_loss increases steadily after a sufficient number of iterations, showing the discriminator is getting stronger and stronger and indicating the end of training.

CelebA

Manipulating Discrete Latent Code

Azimuth (pose):

Presence or absence of glasses:

Hair color:

Hair quantity:

Lighting:

Faces

Loss Plot

Azimuth

Random Generation

Chairs

Rotation

Run

MNIST

• Start training using python train.py; this will automatically download the dataset.
• To see the results, execute python test.py and input the number of your saved model.

CelebA

• Set your image folder in config.py.
• Some links for the datasets:
  • CelebA
• Start training.

python train.py

Faces

• Set your data folder in config.py.
• A link for BFM 2009:
  • Basel Face Model. This should be downloaded before generating data.
  • Data is generated using the code in data_generator. Call gen_data in MATLAB.
• Start training using python train.py.
• To see the results, execute python test.py and input the number of your saved model.

Chairs

• Set your image folder in data.py.
• Some links for the datasets:
  • Chairs
• Start training.

python train.py

References

1. InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets
2. Large-scale CelebFaces Attributes (CelebA) Dataset
3. THE MNIST DATABASE of handwritten digits
4. Seeing 3D chairs: exemplar part-based 2D-3D alignment using a large dataset of CAD models

Authors

• 李舒辰 (@lisc55): The experiment on CelebA.
• 王畅 (@wangchang327): The experiment on MNIST.
• 竺沈涵 (@zshCuanNi): The experiment on Faces. Finished by 王畅 (@wangchang327).
• 岳鹏云 (@hswd40): The experiment on Chairs.