Pytorch code for GAN models

This is the pytorch implementation of 3 different GAN models using same convolutional architecture.

• DCGAN (Deep convolutional GAN)
• WGAN-CP (Wasserstein GAN using weight clipping)
• WGAN-GP (Wasserstein GAN using gradient penalty)

Dependecies

The prominent packages are:

• numpy
• scikit-learn
• tensorflow 2.5.0
• pytorch 1.8.1
• torchvision 0.9.1

To install all the dependencies quickly and easily you should use pip

pip install -r requirements.txt

Training

Running training of DCGAN model on Fashion-MNIST dataset:

python main.py --model DCGAN \
               --is_train True \
               --download True \
               --dataroot datasets/fashion-mnist \
               --dataset fashion-mnist \
               --epochs 30 \
               --cuda True \
               --batch_size 64

Running training of WGAN-GP model on CIFAR-10 dataset:

python main.py --model WGAN-GP \
               --is_train True \
               --download True \
               --dataroot datasets/cifar \
               --dataset cifar \
               --generator_iters 40000 \
               --cuda True \
               --batch_size 64

Start tensorboard:

tensorboard --logdir ./logs/

Walk in latent space

Interpolation between a two random latent vector z over 10 random points, shows that generated samples have smooth transitions.

         

Generated examples MNIST, Fashion-MNIST, CIFAR-10

Inception score

About Inception score

         

Useful Resources

• WGAN reddit thread
• Blogpost
• Deconvolution and checkboard Artifacts
• WGAN-CP paper
• WGAN-GP paper
• DCGAN paper
• Working remotely with PyCharm and SSH