Code for unmixing audio signals in four different stems "drums, bass, vocals, others". The code is adapted from "Jukebox: A Generative Model for Music"

Last update: Dec 29, 2022

Related tags

Deep Learning unmix

Overview

Status: Archive (code is provided as-is, no updates expected)

Disclaimer

This code is a based on "Jukebox: A Generative Model for Music" Paper

We adjusted it for our application: Demixing an audio signal into four different stems: drums, basss, vocals, other.

Unmix

Install

Install the conda package manager from https://docs.conda.io/en/latest/miniconda.html

# Required: Sampling
conda create --name unmix python=3.7.5
conda activate unmix
conda install mpi4py=3.0.3 # if this fails, try: pip install mpi4py==3.0.3
conda install pytorch=1.4 torchvision=0.5 cudatoolkit=10.0 -c pytorch
git clone https://github.com/wzaiealmri/unmix.git
cd unmix
pip install -r requirements.txt
pip install -e .

# Required: Training
conda install av=7.0.01 -c conda-forge
pip install ./tensorboardX

# Optional: Apex for faster training with fused_adam
conda install pytorch=1.1 torchvision=0.3 cudatoolkit=10.0 -c pytorch
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./apex

Training

Stage 1: VQVAE

To train the vqvae, run

mpiexec -n {ngpus} python unmix/train.py --hps=vqvae --name=vqvae_drums_b4 --sr=44100 --sample_length=393216 --bs=4 --audio_files_dir="Put the path to the specific stem audio folder" --labels=False --train --aug_shift --aug_blend

Here, {audio_files_dir} is the directory in which you can put the audio files for your stem , and {ngpus} is number of GPU's you want to use to train. The above trains a one-level VQ-VAE with downs_t = (3), and strides_t = (2) meaning we downsample the audio by 2**3 = 8 to get the first level of codes.
Checkpoints are stored in the logs folder. You can monitor the training by running Tensorboard

tensorboard --logdir logs

Stage 2: Encoder

Train encoder

Once the VQ-VAE is trained, we can restore it from its saved checkpoint and train encoder on the learnt codes. To train the encoder, we can run

mpiexec -n {ngpus} python unmix_encoder/train.py --hps=vqvae --name=encoder_drums__b4 --sr=44100 --sample_length=393216 --bs=4 --audio_files_dir="path to the mix dataset" --labels=False --train --aug_shift --aug_blend --encoder=True --channel=_1 --restore_vqvae="path to the specific checkpoint of the vq-vae"

License (Jukebox OpenAI)

Noncommercial Use License

It covers both released code and weights.

Code for unmixing audio signals in four different stems "drums, bass, vocals, others". The code is adapted from "Jukebox: A Generative Model for Music"

Related tags

Overview

Disclaimer

Unmix

Install

Training

Stage 1: VQVAE

Stage 2: Encoder

Train encoder

License (Jukebox OpenAI)

Owner

Wadhah Zai El Amri

Adaptive Pyramid Context Network for Semantic Segmentation (APCNet CVPR'2019)

PyTorch implementation for MINE: Continuous-Depth MPI with Neural Radiance Fields

mbrl-lib is a toolbox for facilitating development of Model-Based Reinforcement Learning algorithms.

Graph WaveNet apdapted for brain connectivity analysis.

This implementation contains the application of GPlearn's symbolic transformer on a commodity futures sector of the financial market.

A super lightweight Lagrangian model for calculating millions of trajectories using ERA5 data

Auto-Encoding Score Distribution Regression for Action Quality Assessment

Learning Representational Invariances for Data-Efficient Action Recognition

Utilizes Pose Estimation to offer sprinters cues based on an image of their running form.

Official implementation of NeurIPS 2021 paper "One Loss for All: Deep Hashing with a Single Cosine Similarity based Learning Objective"

The Rich Get Richer: Disparate Impact of Semi-Supervised Learning

HMLET (Hybrid-Method-of-Linear-and-non-linEar-collaborative-filTering-method)

Reinforcement Learning for Automated Trading

VolumeGAN - 3D-aware Image Synthesis via Learning Structural and Textural Representations

Pre-Training 3D Point Cloud Transformers with Masked Point Modeling

StyleSpace Analysis: Disentangled Controls for StyleGAN Image Generation

CVPR 2021 - Official code repository for the paper: On Self-Contact and Human Pose.

Towers of Babel: Combining Images, Language, and 3D Geometry for Learning Multimodal Vision. ICCV 2021.

Code for Contrastive-Geometry Networks for Generalized 3D Pose Transfer

Using knowledge-informed machine learning on the PRONOSTIA (FEMTO) and IMS bearing data sets. Predict remaining-useful-life (RUL).