Skip to content

Model

The model consists of two modules and the melodies are generated in a two-stage manner. First, structures are generated by the Structure Generation Net (SGN), which is a convolutional generative adversarial network (GAN). Second, melodies are generated conditioned on the structure by the Melody Generation Net (MGN), which consists of three RNNs.

The architecture of SGN:

Structure Generation Net

The architecture of MGN:

Melody Generation Net