Generative Adversarial Network in Keras
I implemented a generative adversarial network (GAN) trained on the MNIST handwritten digits dataset. I used Keras and Tensorflow to implement the generator and discriminator networks.
The results are pretty stunning! They will not fool a human but they do clearly have some resemblance to the dataset’s handwritten digits.
I only used networks with 2-3 hidden layers! A deeper network should deliver more convincing results. Using two convolutional networks should also help too.
This is the results of using convolutional networks! These images were produced by the generator. Wow, way better. Many of these could fool a human.
This is a video of the generator evolving. Before starting to train the network, I create and store a random input vector for the generator. Every 100 training cycles, I forward prop this vector through the generator and store the resulting image. A video of all those images can be found
This is a video of the convolutional generator evolving. Before starting to train the network, I create and store a random input vector for the generator. Every 100 training cycles, I forward prop this vector through the generator and store the resulting image. A video of all those images can be found
It is easiest to run Tensorflow from a virtual enviroment on Linux. Here are instructions on how to setup Tensorflow and the virtual enviroment https://linuxize.com/post/how-to-install-tensorflow-on-ubuntu-18-04/
To activate the virtual enviroment: source venv/bin/activate