Generating new Pokemon with an implementation of DCGAN in Keras.
This model uses an implementation of DCGAN, DCGAN is a type of generative adversarial network that uses randomly sampled noise to generate images.
The GAN uses 2 networks working against each other, they both improve over the training, yielding better results. A discriminator network tries to detect which images are fakes and which are from our dataset, while the generator tries to produce increasingly realistic fakes to fool the discriminator.
Training a GAN is notoriously tricky, most of my experimentation came from testing different hyper-parameters and network sizes so that the generator and discriminator could be somewhat balanced. In early versions of this project the loss of the discriminator would rapidly approach 0 and the generator never really had a change to improve.
Here are the results I was able to achieve training on google colab for a few hours.
