Official repository for Multi-Source Diffusion Models for Simultaneous Music Generation and Separation.

Demo website available here.

The environment for running our code can be installed using conda:

# Install environment
conda env create -f env.yaml

# Activate the environment
conda activate msdm

This step is optional and required only if you are interested in training a model from scratch or reproducing the results from the paper.

Detailed instructions to download our dataset can be found here.

It is possible to download the pretrained models we used in our experiments. These are useful to run the generation, inpainting and separation scripts available in this repository.

Detailed instruction to download our pretrained models can be found here.

⚠️ NOTE:
Before executing the training script, you have to rename the file .env.tmp to .env and change the WANDB_* environment variables to match your personal or institutional account.

You can run the model by invoking the train.py script:

# Activate environment
conda activate msdm

# Run training script
PYTHONPATH=. python train.py exp=train_msdm

The results will be logged into your Weight & Biases account, using the information provided before.

Take a look at the notebook notebook/MSDM-generation.ipynb for generation and inpainting examples.

⚠️ NOTE:
Before running the notevook, download the glorious-star-335 pretrained checkpoint.

The notebook can be executed using the command

jupyter notebook notebooks/MSDM-generation.ipynb

The example found in our demo website were produced using a similar notebook.

⚠️ NOTE:
You need to download the test set and the pretrained checkpoints in order to run the evaluation scripts.

Running the script

# Run evaluation script for MSDM model
PYTHONPATH=. python evaluate.py exp=eval_msdm_dirac

will separate the test set and evaluate its quality using the SI-SNRi metric. After executing the script, the separation results can be found inside of the output/separations/<exp-name> directory. The metrics are reported inside of the metrics.csv file, in the separation folder.

It is possible to use different configuration for the separation script, by changing the exp=<config> input argument. You can choose between the following configurations:

exp=eval_msdm_dirac # < second best performance
exp=eval_msdm_gaussian
exp=eval_weakly_msdm_dirac # < best performance (4x slower)
exp=eval_weakly_msdm_gaussian