Augmentative Generation for Transformer Bach

This repo contains the implementation of the ICML ML4MD Workshop papers Bach or Mock? A Grading Function for Chorales in the Style of J.S. Bach and Incorporating Music Knowledge in Continual Dataset Augmentation for Music Generation.

Setup

A conda environment is provided in environment.yml. Load it with conda env create -f environment.yml. The environment is named tbach and can be activated with conda activate tbach.

Data

To create the folder of Bach chorales in chorales/bach_chorales, run

python scripts/create_bach_dataset.py

Train models

Every time we run main.py, we can do one of three things:

1. load a model with the --load flag
2. train a model with augmentative generation with the --aug_gen flag
3. train a regular model without dataset augmentation with the --base flag

Before training a model, modify the hyperparameters in transformer_bach/bach_decoder_config.py, and set the desired 'savename' and 'description' for the model. The created model will be saved in models/model_id, where model_id combines the provided savename and a timestamp. The given description will be saved in models/model_id/README.txt. To modify the features in the grading function, see the README in Grader/. The default provided are the same features used in our paper.

Then, to train the augmented model, run

python main.py --aug_gen --config=transformer_bach/bach_decoder_config.py

To train the baseline-none model (which is equivalent to training a model without data augmentation), run

python main.py --base --config=transformer_bach/bach_decoder_config.py

On the first run, the dataset will need to be created in data/. Enter y and then index when prompted to create the most general vocabulary possible. After building the dataset, training should start.

Throughout model training, dataset_sizes.csv, grades.csv, and loss.csv will be populated. These files contain information for each epoch about the size of the training dataset, the grades of the generated chorales, and the training and validation losses. To visualize this data, run

python experiments/plot_training.py --model_dir=models/model_id

The plots will be saved in the models/model_id/plots folder.

Generate

Use the --generate flag to load and generate from a model. Generations are saved in the models/model_id/generations folder.

python main.py --load --config=models/model_id/config.py --generate

References

If you use our code for research, please cite our paper(s)!

@inproceedings{fang2020gradingfunction,
    title={Bach or Mock? {A} Grading Function for Chorales in the Style of {J.S. Bach}},
    author={Fang, Alexander and Liu, Alisa and Seetharaman, Prem and Pardo, Bryan},
    booktitle={Machine Learning for Media Discovery (ML4MD) Workshop at the International Conference on Machine Learning (ICML)},
    year={2020}
}

@inproceedings{liu2020auggen,
    title={Incorporating Music Knowledge in Continual Dataset Augmentation for Music Generation},
    author={Liu, Alisa and Fang, Alexander and Hadjeres, Ga{\"e}tan and Seetharaman, Prem and Pardo, Bryan},
    booktitle={Machine Learning for Media Discovery (ML4MD) Workshop at the International Conference on Machine Learning (ICML)},
    year={2020}
}