TRAINING(MDX-Open).md

Leaderboard C (Open)

Final LB Rank : 8th

Final LB Score: 8.537 dB

The final submission is a weighted blending of three models. Training details for each model is given below.

NOTE: All models were trained on 4x 80GB A100 GPUs.

Datasets used:

Training: total 347 songs Public datasets:

• MUSDB18-HQ Train (84)
• MUSDB18-HQ Test (50)
• MedleyDB (songs not already included in MUSDB18-HQ) (74)

Validation: total 16 songs

• MUSDB18-HQ Validation (14)
• MedleyDB (2)

Step 1: Train DWT-Transformer-UNet model

Memory requirements: ~66GB per GPU

CUDA_VISIBLE_DEVICES=0,1,2,3 bash launcher.sh train_dist.py configs/dwt_transformer_unet/mdx/config_mdx_dwt.yaml

Step 2: Train BSRNN model

Memory requirements: ~47GB per GPU

CUDA_VISIBLE_DEVICES=0,1,2,3 bash launcher.sh train_dist.py configs/bsrnn/mdx_labelnoise/config_md_labelnoise_bass.yaml

CUDA_VISIBLE_DEVICES=0,1,2,3 bash launcher.sh train_dist.py configs/bsrnn/mdx_labelnoise/config_md_labelnoise_drums.yaml

CUDA_VISIBLE_DEVICES=0,1,2,3 bash launcher.sh train_dist.py configs/bsrnn/mdx_labelnoise/config_md_labelnoise_other.yaml

CUDA_VISIBLE_DEVICES=0,1,2,3 bash launcher.sh train_dist.py configs/bsrnn/mdx_labelnoise/config_md_labelnoise_vocals.yaml

Step 3: Use HTDemucs model released by the demucs library

https://github.com/facebookresearch/demucs

Trained on 800 songs

Inference settings:

• shifts: 1
• overlap: 0.25

Step 4: Ensemble

The final submission is a weighted blending of the three models. The weights were determined by grid search on the 16 validation songs.

The weights are as follows:

Source	BSRNN	DWT-Transformer-UNet	HTDemucs
bass	0	0.3	0.7
drums	0.3	0.2	0.5
other	0	0.4	0.6
vocals	0.4	0.2	0.4