Request to Fine-Tune SongGeneration Model for a New Music Genre

[kehinde-elelu]

Hello,

I’m interested in fine-tuning your SongGeneration model on a specific genre of music that is not part of the original training data.

I have the following questions:

Can I fine-tune the current public model (songgeneration_base) using my own genre-specific dataset (Afrobeat)?

What format should my dataset be in (e.g., audio files, Mel-spectrograms, metadata)?

How can I work with the model class to load the songgeneration_base checkpoint and continue training with my dataset?

Are there any scripts or examples in the repository that show how to fine-tune from a pretrained checkpoint?

What are the best practices for adapting the model to a new genre? Should I freeze any layers, or fine-tune the entire model?

I already have the pretrained weights from: https://huggingface.co/tencent/SongGeneration/tree/main/ckpt/songgeneration_base

Appreciate any guidance or tips on how to set up the training pipeline for genre adaptation.

[kehinde-elelu]

@tianyu1412 Can this be done? Or is it a lot of work?

[tencent-ailab]

We will open source the Fine-Tune script, but it involves a lot of work, including data preprocessing, toekenizer, etc. Please give us some time