Demucs-v2-ps-ts

This repository is based on the P9-handin repository created by Mads Lang Laursen. Code is originally based on Demucs V2. The original readme is included in this folder as README-original.md. In addition to the original repository, this version includes new models that have been added to extend its functionality.

Running the code:

To install python dependencies run

pip install -r requirements.txt

This includes all relevant software, except for ffmpeg, sox, and soundstretch which should be installed by:

sudo apt install -y ffmpeg sox soundstretch

Running the training on a single GPU:

python3 -m demucs -b 16 -e 150 --PITSISDR --repeat 1 --audio_channels 1 --wav <train_and_eval_dataset_path> --musdb <test_dataset_path>

Running the training on multiple GPUs:

python3 run.py -b 128 -e 150 --PITSISDR --repeat 1 --audio_channels 1 --wav <train_and_eval_dataset_path> --musdb <test_dataset_path> 

Running the inference:

python3 -m demucs.separate -n <model_name> <mixture_wav_path>

Models

Trained models are available in the Models directory. They are stored as Git LFS objects.

Datasets

Dataset can be shared upon a request.

Singularity & Slurm

The repository contains a demucs.def file for a Singularity container that can be used to run the training. Additionally, scripts for training on with Slurm Workload Managers are found in Slurm directory.