Generative-Adversarial-Networks

GANs

This is a collection of implementations of different types of GANs in Keras. Below, you can find a breakdown of which GANs have been implemented and how to run them.

You'll find so far:

• Original GAN
• DCGAN

Original GAN - https://arxiv.org/pdf/1406.2661.pdf

usage: gan.py [-h] [--epochs EPOCHS] [--batch_size BATCH_SIZE]
              [--batch_per_epoch BATCH_PER_EPOCH] [--noise_dim NOISE_DIM]

• --epochs - The number of epochs to train for (400 by default)
• --batch_size - The standard batch size (128 by default)
• --batch_per_epoch - The number of batches to train on per epoch
• --noise_dim - The size of the noise vector (100 by default)

Result from GAN:

DCGAN - https://arxiv.org/pdf/1511.06434.pdf

usage: dcgan.py [-h] [--epochs EPOCHS] [--batch_size BATCH_SIZE]
                [--batch_per_epoch BATCH_PER_EPOCH] [--noise_dim NOISE_DIM]
                [--dataset DATASET] [--load_model LOAD_MODEL] [--latent]

• --epochs - The number of epochs to train for (400 by default)
• --batch_size - The standard batch size (128 by default)
• --batch_per_epoch - The number of batches to train on per epoch
• --noise_dim - The size of the noise vector (100 by default)
• --dataset - The name of the data set
• --load_model - If loading a model for the latent space interpolation, the path to the model's weights
• --latent - The flag for interpolating over the latent space

Result from DCGAN on original 151 Pokemon:

Result from DCGAN on generation 5 Pokemon: