KUIELab-MDX-Net

• presentation slide

0. Environment

• Ubuntu 20.04
• at least four cuda-able GPUs (each >= 2080ti)
• 1.5 TB disk storage for data augmentation
• wandb for logging

Also, you must create .env file by copying .env.sample to set environmental variables.

wandb_api_key=[Your Key] # "xxxxxxxxxxxxxxxxxxxxxxxx"
data_dir=[Your Path] # "/home/ielab/repos/musdbHQ"

• about wandb_api_key
  • we currently only support wandb for logging.
  • for wandb_api_key, visit wandb, go to setting, and then copy your api key
• about data_dir
  • the absolute path where datasets are stored

1. Installation

conda env create -f conda_env_gpu.yaml -n mdx-net
conda activate mdx-net
pip install -r requirements.txt
sudo apt-get install soundstretch

2. Training & Submission

• Leaderboard_A
• Leaderboard_B

3. Leaderboard A vs Leaderboard B

• The main difference between the branch Leaderboard_A and Leaderboard_B is the usage of the test dataset of Musdb18.
  • Leaderboard A does not use test dataset for training: https://github.com/kuielab/mdx-net/blob/Leaderboard_A/configs/experiment/multigpu_default.yaml
  • Leaderboard B uses test dataset for training:

    mdx-net/configs/experiment/multigpu_default.yaml

    Line 16 in b45eff1

    external_datasets:

ACKNOWLEDGEMENT

• This repository is based on Lightning-Hydra Template
• Repository of TFC-TDF-U-Net, our previous ISMIR 2020 paper
• Also, facebook/demucs