Take some time to reinforce your learning by undertaking the following exercises:
- What type of GAN would you use to transfer styles on an image?
- What type of GAN would you use to isolate or extract the style?
- Modify the number of critics used in the Wasserstein GAN example and see the effect it has on training.
- Modify the first GAN, the DCGAN, to improve training performance using any technique you learned in this chapter. How did you increase training performance?
- Modify the BatchNormalization momentum parameter and see what effect it has on training.
- Modify a few of the samples by changing the activation from LeakyReLU to another advanced form of activation.
- Modify the Wasserstein GAN example to use your own textures. There is a sample data loader available in the downloaded code sample for the chapter.
- Download one of the other reference GANs from https://github.com/eriklindernoren/Keras-GAN and modify that to use your own dataset.
- Alter the first music generation GAN to use a different corpus.
- Use your own MIDI files to train the second music generation GAN example.
- (BONUS) Which music GAN generated better music? Is it what you expected?
You certainly don't have to work through all these exercises, but give a few a try. Putting this knowledge to practice right away can substantially improve your understanding of the material. Practice does make perfect, after all.