Code for "Unsupervised Source Separation via Bayesian inference in the latent domain" - Link

The separation is performed on a x64 compressed latent domain. The results can be upsampled via Jukebox upsamplers in order to increment perceptive quality (WIP).

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

conda create --name lqvae-separation python=3.7.5
conda activate lqvae-separation
sudo apt install mpich 
conda install mpi4py==3.0.3
pip install ffmpeg-python==0.2.0
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch 
(if problems try pip install torch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2)
pip install -r requirements.txt
pip install -e .

# for training
pip install av=7.0.1
pip install ./tensorboardX

• Enter inside script/ folder and create the folder checkpoints/ and the folder results/.
• Download the checkpoints contained in this Google Drive folder and put them inside checkpoints/

• Call the following in order to perform bs separations of 3 seconds starting from second shift of the mixture created with the sources in path_1 and path_2. The sources must be WAV files sampled at 22kHz.

  PYTHONPATH=.. python bayesian_inference.py --shift=shift --path_1=path_1 --path_2=path_2 --bs=bs
  

• The default value for bs is 64, and can be handled by an RTX3080 with 16 GB of VRAM. Lower the value if you get CUDA: out of memory.

• The vqvae/vqvae.pyfile of Jukebox has been modified in order to include the linearization loss of the LQ-VAE (it is computed at all levels of the hierarchical VQ-VAE but we only care of the topmost level given that we perform separation there). One can train a new LQ-VAE on custom data (here data/train for train and data/test for test) by running the following from the root of the project

PYTHONPATH=. mpiexec -n 1 python jukebox/train.py --hps=vqvae --sample_length=131072 --bs=8 
--audio_files_dir=data/train/ --labels=False --train --test --aug_shift --aug_blend --name=lq_vae --test_audio_files_dir=data/test

• The trained model uses the vqvae hyperparameters in hparams.py so if you want to change the levels / downsampling factors you have to modify them there.
• The only constraint for training the LQ-VAE is to use an even number for the batch size, given its use of pairs in the loss.
• Given that L_lin enforces the sum operation on the latent domain, you can use the data of both sources together (or any other audio data).
• Checkpoints are save in logs/lq_vae (lq_vae is the name parameter).

• After training the LQ-VAE, train two priors on two different classes by calling

PYTHONPATH=. mpiexec -n 1 python jukebox/train.py --hps=vqvae,small_prior,all_fp16,cpu_ema --name=pior_source
 --audio_files_dir=data/source/train --test_audio_files_dir=data/source/test --labels=False --train --test --aug_shift
  --aug_blend --prior --levels=3 --level=2 --weight_decay=0.01 --save_iters=1000 --min_duration=24 --sample_length=1048576 
  --bs=16 --n_ctx=8192 --sample=True --sample_iters=1000 --restore_vqvae=logs/lq_vae/checkpoint_lq_vae.pth.tar

• Here the data of the source is located in data/source/train and data/source/test and we assume the LQ-VAE has 3 levels (topmost level = 2).
• The Transformer model is defined by the parameters of small_prior in hparams.py and uses a context of n_ctx=8192 codes.
• The checkpoint path of the LQ-VAE trained in the previous step must be passed to --restore_vqvae
• Checkpoints are save in logs/pior_source (pior_source is the name parameter).

• Before separation, the sums between all codes must be computed using the LQ-VAE. This can be done using the codebook_precalc.py in the script folder:

PYTHONPATH=.. python codebook_precalc.py --save_path=checkpoints/codebook_sum_precalc.pt 
--restore_vqvae=../logs/lq_vae/checkpoint_lq_vae.pth.tar` --raw_to_tokens=64 --l_bins=2048
--sample_rate=22050 --alpha=[0.5, 0.5] --downs_t=(2, 2, 2) --commit=1.0 --emb_width=64

• Trained checkpoints can be given to bayesian_inference.py as following:

  PYTHONPATH=.. python bayesian_inference.py --shift=shift --path_1=path_1 --path_2=path_2 --bs=bs --restore_vqvae=checkpoints/checkpoint_step_60001_latent.pth.tar
  --restore_priors 'checkpoints/checkpoint_drums_22050_latent_78_19k.pth.tar' checkpoints/checkpoint_latest.pth.tar' --sum_codebook=checkpoints/codebook_precalc_22050_latent.pt
  

• restore_priors accepts two paths to the first and second prior checkpoints.

• In order to evaluate the pre-trained checkpoints, run bayesian_test.py after you have put the full Slakh drums and bass validation split inside data/bass/validation and data/drums/validation.

• training of upsamplers for increasing the quality of the separations
• better rejection sampling method (maybe use verifiers as in https://arxiv.org/abs/2110.14168)

If you find the code useful for your research, please consider citing

@article{mancusi2021unsupervised,
  title={Unsupervised Source Separation via Bayesian Inference in the Latent Domain},
  author={Mancusi, Michele and Postolache, Emilian and Fumero, Marco and Santilli, Andrea and Cosmo, Luca and Rodol{\`a}, Emanuele},
  journal={arXiv preprint arXiv:2110.05313},
  year={2021}
}

as well as the Jukebox baseline:

• Dhariwal, P., Jun, H., Payne, C., Kim, J. W., Radford, A., & Sutskever, I. (2020). Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341.