Music source sepeartion using stacked hourglass networks

This is the code for the paper 'Music source separation using stacked hourglass networks', ISMIR 2018

Check out the qualitative results here

Usage

Required packages

tensorflow, pysoundfile, librosa, bss_eval (https://github.com/craffel/mir_eval)

Dataset

MIR-1K dataset

DSD 100 dataset

Training

Set the dataset and checkpoint paths at config.py and run

python train_mir_1k.py

for MIR-1K dataset, or

python train_dsd_100.py

for DSD 100 dataset.

Evaluation

Run

python eval_mir_1k.py

for MIR-1K dataset, or

python eval_dsd_100.py

for DSD 100 dataset.

Trained models

These are the checkpoint files for each dataset to reproduce the results on the paper.

MIR-1K DSD 100