music-vae

Music AutoEncoders (AEs) and Variational AutoEncoders (VAEs)

Setup

Clone the repo:

git clone git@github.com:rshube/music-vae.git

Install the requirements from requirements.txt:

pip install -r requirements.txt

Additionally, one will need to install torch.

Fetch the data

In the top level of the repo, fetch the data sources and generate clips of .wav (this will take a few minutes):

python3 generate_wav_clips.py

The data files are saved in music-vae/data by default. The data save location can be changed by changed the DATA_PATH variable in music-vae/utils/consts.py

Train a model

In the top level of the repo, train and evaluate a model with:

python3 main.py --{optional_args}

The arguments you can add:

Flag	Description	Default
-h	Show options and exit	N/A
--model	The model type to train. Must be one of: SimpleAutoEncoder SimpleVarAutoEncoder ComplexAutoEncoder ComplexVarAutoEncoder SimpleConvAutoEncoder ComplexConvAutoEncoder ImageConvAutoEncoder	SimpleAutoEncoder
--num_epochs	The number of epochs to complete during training	10
--lr	The learning rate	1e-03
--batch_size	The batch_size	32

By default, the model is trained on 10s clips of Lo-Fi music. Running this will save the a plot of the evaluation loss over each epoch of training, the evaluation loss against different types of data sets, and the model weights. The default save location is music-vae/results/{model type}/{date-time}/. This can be changed by changing the save_dir variable in the main function of main.py.