MNIST Generative Adversarial Networks (PyTorch)

Sam Greydanus. April 2017. MIT License.

About

I use the classic MNIST dataset to achieve ultra-simple GAN results. Think of this repo as a lab where you can get comfortable with GANs before trying them on something more complex (e.g. CIFAR, ImageNet).

• vanilla-gan: 100 lines of code
• cnn-gan: 130 lines of code

Samples

Vanilla discriminator (D) and generator (G) networks

CNN discriminator (D) and vanilla generator (G) network

Dependencies

• All code is written in python 3.6. You will need:
• Numpy
• matplotlib
• PyTorch: much easier to write and debug than TensorFlow!
• Jupyter