TRAINING(CDX-DnR).md

Leaderboard A and B (Used only DnR dataset)

Final LB Rank : 3rd (A), 6th (B)

Final LB Score: 4.144 dB

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

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

Step 1: Data pre-processing

The tr-subset of the DnR dataset was pre-processed to remove silence from the music and speech stems.

Using the following command:

python utitlity_scripts/pre_process_dnr.py

For validation, we use 32 mixtures from the cv-subset. You can prepare the validation data using the following command:

python utitlity_scripts/create_sdx_valid.py

Step 2: 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/cdx/config_cdx_dwt.yaml

Step 3: Train BSRNN model

Memory requirements: ~37GB per GPU

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

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

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

Step 4: Ensemble

The final submission is a weighted blending of the three models. The weights were determined based on the LB performance of individual models.

The weights are as follows:

Source	BSRNN	DWT-Transformer-UNet
music	0.18	0.82
speech	0.18	0.82
sfx	0.82	0.18