sigmoidal nmfd

Sigmoidal NMFD

Code for the paper Vande Veire, Len and De Bie, Tijl and De Boom, Cedric, "Sigmoidal NMFD: Convolutional NMF with Saturating Activations For Drum Loop Decomposition".

Installation

# Download the repositorygit clone https://github.com/aida-ugent/sigmoidal-nmfdcd sigmoidal-nmfd# Install requirements...# ... for conda users:conda create --name sigmoidnmfd --file requirements.txt# ... alternatively:pip install -r requirements.txt

You will also need to install soundfile:

sudo apt-get install libsndfile1

Running sigmoidal NMFD

The script run_nmfd_sigmoid.py applies sigmoidal NMFD to the provided audio file. For example:

python -m scripts.run_nmfd_sigmoid resources/moonkits-hiphop.wav 4 --plot

The original NMFD algorithm can be run as follows:

python -m scripts.run_nmfd_vanilla resources/moonkits-hiphop.wav 4 --plot

The sparse NMFD baseline can be run as follows:

python -m scripts.run_nmfd_sparsity resources/moonkits-hiphop.wav 4 --sparsity 0.1 --plot

Running the experiments from the paper

First, download the ENST dataset.

Then, execute the experiment_nmfdsigmoid_on_enst.py script:

python -m scripts.experiment_nmfdsigmoid_on_enst --dir-enst /path/to/ENST-drums-public --dir-out /home/user/somedirectory --tracklist "resources/tracklists/tracklist_enst_allphrases.csv"

This will automatically crop the ENST phrase files as described in the paper, save them in a new directory, and apply all baselines and the proposed sigmoidal model to all cropped phrases in the dataset. The results are saved in .npz archives (note: this requires about 1 GB of disk space). It will then print out the metric values aggregated over all examples.

The ablation experiments can be executed analogously.
For the ablation experiments on sigmoidal NMFD, including the evaluation of the different optimization strategies:

python -m scripts.experiment_ablation_nmfdsigmoid_on_enst --dir-enst /path/to/ENST-drums-public --dir-out /home/user/somedirectory --tracklist "resources/tracklists/tracklist_enst_allphrases.csv"

For sparse NMFD with an unconstrained warm-up stage:

python -m scripts.experiment_nmfdsparse_with_warmup.py --dir-enst /path/to/ENST-drums-public --dir-out /home/user/somedirectory --tracklist "resources/tracklists/tracklist_enst_allphrases.csv"

Note that parallel processing is supported by adding a --parallel flag in the aforementioned commands for the ablation experiments.

Recreating the initialization templates

The initialization values for the templates W in the NMFD framework can be recreated using thecreate_nmf_drum_templates_from_sample_library.py script.

For example, use these drum samples from Producerspot, as we did for this paper.

Then execute:

python create_nmf_drum_templates_from_sample_library.py --samples-list-file resources/tracklist/templates/samples_kick.csv --output-file resources/templates/kick.npy
python create_nmf_drum_templates_from_sample_library.py --samples-list-file resources/tracklist/templates/samples_snare.csv --output-file resources/templates/snare.npy
python create_nmf_drum_templates_from_sample_library.py --samples-list-file resources/tracklist/templates/samples_hihat.csv --output-file resources/templates/hihat.npy
python create_nmf_drum_templates_from_sample_library.py --samples-list-file resources/tracklist/templates/samples_crash.csv --output-file resources/templates/crash.npy

Copyright information

Copyright 2020 Len Vande Veire.

This code within this repository is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with this program.  If not, see https://www.gnu.org/licenses/.