Basic Implementation (Study friendly) of InfoGAN in Pytorch

More Information: Original Paper

About this algorithm (in Korean): My Blog Post

• GAN: [Pytorch][Tensorflow]
• DCGAN: [Pytorch][Tensorflow]
• InfoGAN: [Pytorch][Tensorflow]
• Pix2Pix: [Pytorch][Tensorflow]
• DiscoGAN: [Pytorch][Tensorflow]

• Ubuntu 16.04
• Python 3.6 (Anaconda)
• Pytorch 0.2.0
• Torchvision 0.1.9

• Image Size = 28x28x1 MNIST / 64x64x3 CelebA
• Batch Size = 16
• Test Batch Size (Sample Size) = 100 (10x10)
• Learning Rate = 0.0002 for D, 0.001 for G
• Adam_beta1 = 0.5
• z_dim = 62 MNIST / 128 CelebA
• InfoGAN Continuous Code Dimension = 1
• InfoGAN Discrete Code Dimension = 10
• Continuous Code Weight = 0.1~0.5 MNIST / 1.0 CelebA
• Discrete Code Weight = 1
• Epoch = 50 MNIST / 4 CelebA

You can modify hyper-parameter. Look at the parsing part of the code.

Train

python infogan.py

Data Download

chmod +x celebdownload.sh
./celebdownload.sh

Train

python infogan.py --db 'celebA' --image_size 64 --z_dim 128 --continuous_weight 1.0 --num_epochs 4

Will be uploaded

Discrete(Categorical) Code adjusts personality/skin color

Continuous Code adjusts emotion (Smile)