README.md

Setup

Install Packages

1. python3 -m venv env
2. source env/bin/activate
3. pip install -e .
4. git submodule update --init --recursive
5. pip install -e umx

Download VocalSet Dataset

1. wget https://zenodo.org/record/1442513/files/VocalSet1-2.zip?download=1
2. mv VocalSet.zip?download=1 VocalSet.zip
3. unzip VocalSet.zip

Training

Steps

1. Change Wandb and data root variables in shell_vars.sh and source shell_vars.sh
2. python scripts/train.py +exp=default

Experiments

Training parameters can be configured in cfg/exp/default.yaml. Here are some descriptions

• max_kept_effects={n} max number of Kept effects to apply to each file. Set to -1 to always use all effects (default: -1)
• max_removed_effects={n} max number of Removed effects to apply to each file. Set to -1 to always use all effects (default: -1)
• model={model} architecture to use (see 'Models')
• effects_to_use={effect} Effects to use (see 'Effects') (default: all in the list)
• effects_to_remove={effect} Effects to remove (see 'Effects') (default: all in the list)
• accelerator=null/'gpu' Use GPU (1 device) (default: null)
• render_files=True/False Render files. Disable to skip rendering stage (default: True)
• render_root={path/to/dir}. Root directory to render files to (default: DATASET_ROOT)

Note that "kept effects" are calculated from the difference between effects_to_use and effects_to_remove.

These can also be specified on the command line. Example: python scripts/train.py model=demucs "effects_to_use=[distortion, reverb, chorus]" "effects_to_remove=[distortion]" max_kept_effects=2 max_removed_effects=4 shuffle_kept_effects=False shuffle_removed_effects=True accelerator='gpu' render_root=/scratch/VocalSet'

Printout:

Effect Summary:
Apply kept effects: ['chorus', 'reverb'] (Up to 2, chosen in order) -> Dry
Apply remove effects: ['distortion'] (Up to 4, chosen randomly) -> Wet

Models

• umx
• demucs

Effects

• chorus
• compressor
• distortion
• reverb

Misc.

By default, files are rendered to input_dir / processed / {string_of_effects} / {train|val|test}.