genierator

Me practicing generative adversarial networks.

Running the code

0. Make sure you are using Python 3
1. pip -r install requirements.txt
2. python main.py

The code will train the model and print out the generator's loss as well as the discriminator's accuracy for each epoch. You want both of them to be small: that means that the images generated by the generator are realistic (discriminator classifies generator's images as real). After the training 8 sample pictures by the trained generator will be shown.

TODO

• Use a more interesting dataset
• Do some cool parameter tweaking like this
• Learn a new network model, maybe CGANS
• Use labels when training the discriminator

Gallery of generated examples:

MNIST, 10 epochs, relu

MNIST, 10 epochs, leaky relu

MNIST, 10 epochs, leaky relu, smoothed labels

MNIST, 50 epochs, leaky relu, smoothed labels

Sources:

Understanding Generative Adversarial Networks by Naoki Shibuya (I use his network architecture and training loop)