musicgen-finetune

Training pipeline for musicgen using a simple dataset format.

How to use:

• Follow instructions in the notebook to load dependencies, assuming you have an environment with Jupyter set up.
• Requires access to a GPU with 8+ GB VRAM for training (preferably > 16 GB), at least 4 GB VRAM for inference.
• In Google Colab: replace input_folder_name to a path as per your mounted Google Drive (e.g. /content/...)

NOTE:

• Longer outputs take progressively more time in inference, especially depending on your hardware.
• The current model is prone to have silent sections in the ouput, highly influenced by the finetune dataset used (e.g. you use starts and ends of songs).

TODO

• Support stereo finetune.
• Optimized training: e.g. LORA

Credits

• Inspired from a notebook from dominiqueGrys.
• Trainer code from a deleted profile/repository.
• Facebook/Meta research for providing the tooling to play with their models.