Music Transformer: Generating Music with Long-Term Structure

• 2019 ICLR, Cheng-Zhi Anna Huang, Google Brain
• Re-producer : Yang-Kichang
• paper link
• paper review

Abstract

1. This Repository is perfectly cometible with pytorch

Contribution

• Domain: Dramatically reduces the memory footprint, allowing it to scale to musical sequences on the order of minutes.
• Algorithm: Reduced space complexity of Transformer from O(N^2D) to O(ND).

Preprocessing

• In this repository using single track method (2nd method in paper.).

• If you want to get implementation of method 1, see here .

• I refered preprocess code from performaceRNN re-built repository..

• Preprocess implementation repository is here.

  

Simple Start ( Repository Setting )

$ git clone https://github.com/jason9693/MusicTransformer-pytorch.git
$ cd MusicTransformer-pytorch
$ git clone https://github.com/jason9693/midi-neural-processor.git
$ mv midi-neural-processor midi_processor

Midi Download

$ sh dataset/script/{ecomp_piano_downloader, midi_world_downloader, ...}.sh

• These shell files are from performaceRNN re-built repository implemented by djosix

Prepare Dataset

$ python preprocess.py {midi_load_dir} {dataset_save_dir}

Trainig

$ python train.py -c {config yml file 1} {config yml file 2} ... -m {model_dir}

Hyper Parameter

• learning rate : 0.0001
• head size : 4
• number of layers : 6
• seqence length : 2048
• embedding dim : 256 (dh = 256 / 4 = 64)
• batch size : 2

Result

• Baseline Transformer ( Green, Gray Color ) vs Music Transformer ( Blue, Red Color )

• Loss

  

• Accuracy

  

Generate Music

$ python generate.py -c {config yml file 1} {config yml file 2} -m {model_dir}

Current generation code

python generate_cmd.py -m train_model\longer_seq -k train-250.pth -i mz_331_2 -l 250

Generated Samples ( Youtube Link )

• click the image.