A Baseline for SDX 2023

[ws-choi]

We are Implementing a baseline model for SDX2023.

Currently, a script to train vocals are pused to sdx_label.

• target

  • [x] vocals: python run.py experiment=multigpu_vocals datamodule=moisesdb23_labelnoise_v1
  • [x] drums
  • [ ] bass
  • [ ] other

• datamodule

  • [x] tested on moisesdb23_labelnoise_v1
  • [ ] tested on moisesdb23 bleeding
  • [ ] on-the-fly data augmentation support

• bug

  • [x] multi gpus with ddp not working

Please feel free to make a PR for any items on this todo list.

---

wandb:

• vocals on labelnoise: https://wandb.ai/wschoi/mdx_vocals/runs/hlssqa8a?workspace=user-wschoi

[Ma5onic]

@ws-choi Could you please post/share the .ckpt files from MDX21 Leaderboard B somewhere? I would like to try fine-tuning/transfer-learning, which can't be done with the .onnx from 2021. It would greatly be appreciated.

[ws-choi]

Hi @Ma5onic , I am sorry but we lost the original .ckpt file. I will try to resotre .ckpt using some third party library from .onnx but it might take several months.