UPDATE (2023 29 Oct.):
- Regrettably, due to the policy of the company I am currently working with on this project, open code will likely be possible in January 2024.
UPDATE (2023 6 Oct.):
- YourMT3+ is currently achieving on-par with or clearly better performance (using our own augmentation) than MT3.
- YourMT3+ achieves comparable performance with PerceiverTF on Slakh when our own augmentation (similar to PerceiverTF) is applied.
- Our toolkit comes with easy-to-use auto-downloader/installer for full dataset of MT3 (except for Cerberus, which has overlap with Slakh) + MIR-ST500 for singing.
- Improved data-loading that enables everything on-the-fly by custom batching pipeline (not using auto-batch of PyTorch DataLoader)
- Note-processing and tokenization are implemented dependency-free from
noteseq,prettymidi. - System requirements for fully reproducing result: 500GB disk space, 80G (recommended) GPU memory
We are going to present this work (YourMT3+) as an ISMIR Late-breaking Demo next month in Milano.