Skip to content

Latest commit

 

History

History
96 lines (60 loc) · 2.41 KB

README.md

File metadata and controls

96 lines (60 loc) · 2.41 KB

Separation Worker

Uses DemucsV4 for separating audio files.

Using docker

tbd

Running locally

Prerequisites

One of the dependencies is audiowaveform, which handles calculating a downsampled representation of an uploaded audio file. For installation refer to their installation guide or if using ubuntu do:

sudo add-apt-repository ppa:chris-needham/ppa
sudo apt-get update
sudo apt-get install audiowaveform

Further, on MacOs install soundstretch/soundtouch with:

brew install sound-touch

For Linux you can install PySoundFile as torchaudio backend later.

First start

Install conda/miniconda if you haven't already.

Using conda environment file

You can use the provided environment.yml file to get started fast.

conda env create -f environment.yml

Manually Installing dependencies

Create a new conda environment with python=3.10 or version of you liking.

conda create --name separation python=3.10

Activate the environment with:

conda activate separation

Install pytorch and dependencies using conda/miniconda:

conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia

A detailed guide how to get started with PyTorch for your environment / system can be found here.

Other needed dependencies:

pip install git+https://github.com/CarlGao4/demucs.git@4.1.0-update
pip install bullmq prisma minio asyncio ffmpeg

Additionally for linux:

pip install PySoundFile

Generating the prisma client

Generate prisma client, api-gateway should've been started by now and a db should exist:

prisma db pull
prisma generate

Then copy .env.template to .env and fill out the needed env vars with your secrets.

Start the worker with python worker.py, you will see whether any GPU is for accelerated computing is available and the model downloading from meta's public file registry.

Once thats done, the worker is ready process jobs.

Configuring demucs models

In worker.py where the SeparationProcessor is instantiated, you can pass the model parameter.

  • htdemucs Demucs V3
  • htdemucs_ft Demucs V4 (Hybrid Transformer)