InfoGAN

This repository contains an implementation of InfoGAN on the MNIST dataset using TensorFlow 2.0.

Requirements

tensorflow==2.11.0
tensorflow-probability==0.19.0
numpy==1.23.4
matplotlib==3.6.2

# if you want to use tfds for celeb_a, svhn, etc datasets
# tensorflow-datasets

Usage

1. Clone the repository:

git clone https://github.com/amir7d0/InfoGAN.git

2. Edit the config.py file to set the training parameters and the dataset to use. Choose dataset from ['mnist', 'fashion_mnist']
3. Run the training script:

python main.py

Directory structure

.
├── infogan
│   ├── config.py
│   ├── datasets.py
│   ├── distributions.py
│   ├── infogan_model_openai.py
│   ├── infogan_model.py
│   ├── models.py
│   └── utils.py
├── main.py
├── README.md
└── results
    └── mnist_results

Files

• config.py: Contains all the configuration parameters for training the model.
• datasets.py: Contains code for loading and preprocessing the dataset.
• distributions.py: Contains the code for the distributions.
• utils.py: Contains the code for Callbacks, sample, and plot functions.
• models.py: Contains the code for the generator, discriminator, and recognition networks.
• infogan_model.py: Contains the code for the InfoGAN class and train_step function.
• infogan_model_openai.py: Contains the code for the InfoGAN class and train_step function.
• train.py: Contains the code for training the model.

In infogan_model.py Continuous Latent code is optimized by minimizing Gaussian negative log likelihood loss (GaussianNLLLoss) which is H(c | G(z,c)) in Eq (4) in InfoGAN paper.

In infogan_model_openai.py Continuous Latent code is optimized by minimizing Gaussian negative log likelihood loss (GaussianNLLLoss) and negative Log Prob which is H(c) + H(c | G(z,c)) in Eq (4) in InfoGAN paper.

Results

MNIST

Settings

• Latent Variables (Dim=74)
  1. Random latent ~ Uniform(-1,1), dim = 62
  2. Discrete Latent Code ~ Categorical(k=10, p=0.1), dim = 10
  3. Continuous Latent Code ~ Uniform(-1,1), dim = 2
• Optimizer
  • generator optimizer = Adam, lr=1e-3, beta1 = 0.5
  • discriminator optimizer = Adam, lr=2e-4, beta1 = 0.5
• Lambda for latent codes: $\lambda_{disc} = 1.0, \lambda_{cont} = 1.0$
• Batch size = 128

Generated Images

Row represents categorical latent code from 0 to 9 and column represents continuous latent code varying from -1 to 1 (left to right).

first continuous variable varying from -1 to 1	second continuous variable varying from -1 to 1

Loss

References

1. X. Chen, Y. Duan, R. Houthooft, J. Schulman, I. Sutskever, P. Abbeel. "Infogan: Interpretable representation learning by information maximizing generative adversarial nets." [arxiv]
2. openai/InfoGAN [repo]
3. lisc55/InfoGAN [repo]