Submission

Submission Summary

• MDX Leaderboard A

  • Submission ID: 216209
  • Submitter: kim_min_seok
  • Final rank: 3rd place
  • Final scores:

    SDR_song	SDR_bass	SDR_drums	SDR_other	SDR_vocals
    6.513	6.707	6.712	4.816	7.816

• MDX Leaderboard B

  • Submission ID: 216211
  • Submitter: kim_min_seok
  • Final rank: 1st place
  • Final scores:

    SDR_song	SDR_bass	SDR_drums	SDR_other	SDR_vocals
    6.581	6.975	6.646	4.962	7.741

Model Summary

• Data
  • All 203 tracks of the Moises dataset was used for training (no validation split)
  • Augmentation
    • Random chunking and mixing sources from different tracks ([1])
• Model
  • A 'multi-source' version of TFC-TDF U-Net[2, 3] with some architectural improvements, including Channel-wise Subband[4]
  • Final submission is an ensemble of 3 models with identical architecture and training procedure but with different random seeds
• Noise-robust Training
  • Leaderboard A: Loss masking
    • Intuitively, data with high training loss is likely to be audio chunks with labelnoise
    • For each training batch, discard (=don't use for weight update) batch elements with higher loss than some quantile
      • ex) only use half of the training batch for each weight update
  • Leaderboard B: Loss masking (along temporal dimension)
    • Compared to labelnoise, data with bleeding seemed to vary less in terms of the amount of noise
    • A more fine-grained masking method performed better (discarding temporal bins with high loss)

[1] S. Uhlich, et al., "Improving music source separation based on deep neural networks through data augmentation and network blending", ICASSP 2017.

[2] W. Choi, et al. "Investigating u-nets with various intermediate blocks for spectrogram-based singing voice separation", ISMIR 2020.

[3] M. Kim, et al. “Kuielab-mdx-net: A two-stream neural network for music demixing”, MDX Workshop at ISMIR 2021.

[4] H. Liu, et al. "Channel-wise Subband Input for Better Voice and Accompaniment Separation on High Resolution Music", INTERSPEECH 2020.

Reproduction

Download mdx_AB.zip, which contains

• pretrained model checkpoints
• config.yaml files (configurations for training and inference)

How to reproduce the submission

1. Create a 'ckpts' folder under my_submission. Unzip the downloaded zip file to 'my_submission/ckpts'.
2. Copy my_submission and requirements.txt to your SDX 2023 Music Demixing Track Starter Kit.
3. Run submit.sh after configuring my_submission/user_config.py
   • for Leaderboard A, set MySeparationModel = A
   • for Leaderboard B, set MySeparationModel = B

How to reproduce the training

• All code needed to reproduce training is in my_submission/src
• See HOW_TO_TRAIN.md