This repository contains the code to reproduce the musical audio source separation presented in
Roma, G., Green, O. & Tremblay, P. A. Improving single-network single-channel separation of musical audio with convolutional layers 14th International Conference on Latent Variable Analysis and Signal Separation (2018)
The original evaluation used the DSD100 dataset via dsdtools. For the 2018 SiSEC campaign it was updated to use the musdb18 dataset.
- numpy
- scipy
- pytorch
- resampy
- musdb
- mir_eval
The two proposed variants are in RGT1 and RGT2, corresponding to the SiSEC submissions. The first uses a soft mask and MSE loss, the second uses negative log likelihood loss with a 2D softmax output to predict a binary mask.
In each directory, configure the dataset path and the analysis path (for temporary analysis files). Then run each script in order as needed:
- python analyze.py
- python train.py
- python predict.py
Also create the folders for results as needed. Using a GPU is advised for the train.py script.
For evaluation, the script eval_track.py will evaluate one track, so it can be run in parallel using e.g. GNU parallel:
parallel -j4 python eval_track.py {} ::: {0..49}
The scripts predict_sisec.py and eval_musdb.py run the evaluation of the Test part of the dataset according to the SiSEC campaign procedure. Since the original approach is single-channel it is run for each channel separately.