MusicGAN

M.Sc. Thesis Project

A data science (Artificial Intelligence) master thesis about music generation with GAN models.

Li, Y., & Linberg, J. (2023). Music Generation with Generative Adversarial Networks (Dissertation).
Thesis paper

Code

Code for training and data visualisation is available in /code.

• Training (with Emotion Constraint), the main training file of the project.
• Training Loss Visualisation, using tensorboard to visualise the training data.
• Generator Inspector, a notebook for generating music and evaluating the model.
• Training Parameters, a file for controlling the model hyperparameters.
• MelSpec, a keras layer implementation for creating the mel-scaled spectrograms, original code is found here: https://github.com/keras-team/keras-io/blob/master/examples/audio/melgan_spectrogram_inversion.py.
• No-longer–used code, kept for the purpose of documentation.

Models

The main models of the project, named A-D, are available in /models/train

• Model A (no PatchGAN)
• Model B (no rhythm loss)
• Model C (no emotion labels)
• Model D (full model)

Generated Examples

Model B

Audio

Model C

Audio

Model D

Audio