SCNet

This repository is the official implementation of SCNet: Sparse Compression Network for Music Source Separation

---

Training

First, you need to install the requirements.

cd SCNet-main
pip install -r requirements.txt

We use the accelerate package from Hugging Face for multi-gpu training.

accelerate config

You need to modify the dataset path in the /conf/config.yaml. The dataset folder should contain the train and valid parts.

data:
  wav: /path/to/dataset

The training command is as follows. If you do not specify a path, the default path will be used.

accelerate launch -m scnet.train --config_path path/to/config.yaml --save_path path/to/save/checkpoint/

---

Inference

The model checkpoint was trained on the MUSDB dataset. You can download it from the following link:

Download Model Checkpoint

The large version is now available.

SCNet-large
config.yaml

I have performed normalization on the model's input during training, which helps in stabilizing the training process (no code modifications are needed during inference).

python -m scnet.inference --input_dir path/to/test/dir --output_dir path/to/save/result/ --checkpoint_path path/to/checkpoint.th

---

Citing

If you find our work useful in your research, please consider citing:

@misc{tong2024scnet,
      title={SCNet: Sparse Compression Network for Music Source Separation}, 
      author={Weinan Tong and Jiaxu Zhu and Jun Chen and Shiyin Kang and Tao Jiang and Yang Li and Zhiyong Wu and Helen Meng},
      year={2024},
      eprint={2401.13276},
      archivePrefix={arXiv},
      primaryClass={eess.AS}
}