Calliope: Adversarial Transformer Autoencoder

This repository contains the implementation of Calliope, an Adversarial Transformer Autoencoder for music generation and latent space interpolation. Some examples of the ability of the model can be found here. It is also possible to generate new bars here.

Dataset

In order to create the dataset, it is necessary to choose the desired number of bars n_bars from the config file, the number of songs to generate (early_stop, if it is zero then all the songs found in the raw dataset are converted) and then launch:

python create_bar_dataset.py

Training

The training uses Weight & Bias in order to control the training. Login in W&B with your credential. The training can be started with

python train.py

In order to check the progress of the model on W&B, it is needed a way to convert midi files into wav. It is necessary to download the sound font from here. and check that utilities/midi_to_wav is correctly seeing the file.

Otherwise, the pretrained weights can be downloaded here.

Testing

The model can be tested in different ways. With:

python test.py

are tested the reconstruction ability, the generation ability and the interpolation ability. With:

python test_musegan.py

the same metrics of the MuseGAN model are computed on 20000 generated bars. With:

python test_musicvae.py

the reconstruction accuracy are computed. With:

python create_base.py

the model generates four repetitions of the n_bars generated bars.