Hands-On Generative Adversarial Networks with Keras

MuseGAN

The Multitrack Sequential Generative Adversarial Network (MuseGAN) was developed by Hao-Wen Dong et al. As the name suggests, MuseGAN is able to synthesize multitrack MIDI files.

MuseGAN has three GAN models for generating multitrack data:

• The jamming model: This consists of multiple generators that independently create a single track each.
• The composer model: This consists of a single generator that creates all the tracks.
• The hybrid model: This consists of multiple generators that are conditioned on global and individual Z vectors:

MuseGAN generator diagrams

In MuseGAN, temporal coherence is obtained by considering the following ...