rafaelcga / demucs-splitter Goto Github PK

Simple Python script to retrieve songs from YouTube using yt-dlp or a given file path; and split them into 6 stems through Hybrid Transformer Demucs. Additionaly, a backing track is created by combining all stems except guitar.

Anaconda installation.

1. Install Demucs in new Anaconda environment.

   • Make sure to initialize Demucs submodule repository:

      git submodule update --init --recursive
     

   • Install dependencies and library:

      cd ./demucs_repo
      conda env update -f environment-cpu.yml  # if you don't have GPUs
      conda env update -f environment-cuda.yml # if you have GPUs
      conda activate demucs
      pip install -e .
     

   • If using NVIDIA GPU, ensure it is recognized by PyTorch:

      import torch
      assert torch.cuda.is_available()

   • If NVIDIA GPU is not recognized:

      conda activate demucs
      conda remove pytorch
     

     • Install latest CUDA bundled PyTorch from https://pytorch.org/.

2. Install yt-dlp in the same environment:

   conda activate demucs
   pip install yt-dlp
   

3. Download ffmpeg binary from https://github.com/yt-dlp/FFmpeg-Builds and place it in the same directory as the script.

Launch split.py from demucs Anaconda environment and input any YouTube URL or audio file path (with a format supported by Torchaudio):

conda activate demucs
python split.py https://www.youtube.com/watch?v=dQw4w9WgXcQ