Symbolic Music Generation with Transformer-GANs

Code for the paper "Symbolic Music Generation with Transformer-GANs" (AAAI 2021)

If you use this code, please cite the paper using the bibtex reference below.

@inproceedings{transformer-gan,
    title={Symbolic Music Generation with Transformer-GANs},
    author={Aashiq Muhamed and Liang Li and Xingjian Shi and Suri Yaddanapudi and Wayne Chi and Dylan Jackson and Rahul Suresh and Zachary C. Lipton and Alexander J. Smola},
    booktitle={35th AAAI Conference on Artificial Intelligence, {AAAI} 2021},
    year={2021},
}

Requirements

• Python 3.6+
• Pytorch
• Transformers

You can install all required Python packages with bash requirements.sh.

Datasets, switching inside data folder

• Downloaded data

bash get_data.sh

• Run music_encoder.py to generate the encoded numpy files
  • Messages stating that pitches are out of range are expected behavior

python3 music_encoder.py --encode_official_maestro \  
    --mode midi_to_npy \  
    --pitch_transpose_lower -3 \  
    --pitch_transpose_upper 3 \  
    --output_folder ./maestro_magenta_s5_t3  

Train and Generate: switching inside model folder

• Train a Transformer XL (No GAN)

python3 -m torch.distributed.launch --nproc_per_node=4 ./train.py \
    --data_dir ../data/maestro_magenta_s5_t3 \
    --cfg ./training_config/experiment_baseline.yml \
    --work_dir exp_dir

• Train a Transformer XL (with GAN)

python3 -m torch.distributed.launch --nproc_per_node=4 ./train.py \
    --data_dir ../data/maestro_magenta_s5_t3 \
    --cfg ./training_config/experiment_spanbert.yml \
    --work_dir exp_dir

• Generate unconditional samples

# generate unconditional samples
python3 generate.py --inference_config inference_config/inference_unconditional.yml

Note, if you are loading an old config.yml file which includes None/" " inside, please change it to a string 'Null' to make sure you can do cfg.merge_from_file.

• Extend music to generate conditional samples

# generate conditional samples
python3 generate.py --inference_config inference_config/inference_conditional.yml

1. Please set condition_len as well as condition_file
2. Change memlen and genlen. memlen=genlen is recommended

Post process for data (convert .txt to .mid)

• Run the following to get midi files from txt files
  • Use --mode to_midi for text file conversions. Use --mode npy_to_midi for numpy file conversions.

python3 ../data/music_encoder.py --input_folder ./Output_Uncondtitionl --output_folder ./Output_Uncondtitionl_MIDI --mode to_midi
python3 ../data/music_encoder.py --input_folder ./Output_Condtitionl --output_folder ./Output_Condtitionl_MIDI --mode to_midi

different methods inside music_encoder

• encoder.to_text(input.mid, output.txt)
• encoder.from_text(input.txt, out.mid)
• encoder.encode_vocab(input.mid) return list of ids
• encoder.decoder_vocab(list(ids)) return out.mid
• encoder.to_text_argumentaion(input.mid, output.txt)