• 

  Sample image produced by a single layer GAN

Shallow General Adversarial Convolutional Network

This is the code for a shallow GAN using a MNIST dataset. To run the code, simply do:

 python gan.py 

The directory shallow_gan will be created which will hold all results and model parameters, including confusion matrices. All you need will be available in the directory and is easily navigable. Since it's a shallow one layer system, the images obtained are note very clear. A deeper system will produce better results.

Pre-requisites

This code uses the Yann toolbox internally to run, so that needs to be set up properly. Instructions on how to setup the toolbox is provided in the website.

Inspiration

I had the inspiration after watching some of Ian Goodfellow's presentations. After learning about the newest discoveries in machine learning, I was inspired to implement a simple, single layer, shallow GAN myself.

Accomplishments

My implementation of a GAN was successfully able to create an image with handwritten numbers. Clearer images can be obtained with a deeper system.

Future Scope

Using deconvolutional neural network to implement a cifer10 and svhn datasets.

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Built With

• python