djosix / Performance-RNN-PyTorch

Performance RNN - PyTorch

PyTorch implementation of Performance RNN, inspired by Ian Simon and Sageev Oore. "Performance RNN: Generating Music with Expressive Timing and Dynamics." Magenta Blog, 2017. https://magenta.tensorflow.org/performance-rnn.

This model is not implemented in the official way!

Generated Samples

• A sample on C Major Scale [MIDI, MP3]
  • control option: -c '1,0,1,0,1,1,0,1,0,1,0,1;4'
• A sample on C Minor Scale [MIDI, MP3]
  • control option: -c '1,0,1,1,0,1,0,1,1,0,0,1;4'
• A sample on C Major Pentatonic Scale [MIDI, MP3]
  • control option: -c '5,0,4,0,4,1,0,5,0,4,0,1;3'
• A sample on C Minor Pentatonic Scale [MIDI, MP3]
  • control option: -c '5,0,1,4,0,4,0,5,1,0,4,0;3'

Directory Structure

.
├── dataset/
│   ├── midi/
│   │   ├── dataset1/
│   │   │   └── *.mid
│   │   └── dataset2/
│   │       └── *.mid
│   ├── processed/
│   │   └── dataset1/
│   │       └── *.data (preprocess.py)
│   └── scripts/
│       └── *.sh (dataset download scripts)
├── output/
│   └── *.mid (generate.py)
├── save/
│   └── *.sess (train.py)
└── runs/ (tensorboard logdir)

Instructions

• Download datasets

  cd dataset/
  bash scripts/NAME_scraper.sh midi/NAME

• Preprocessing

  # Preprocess all MIDI files under dataset/midi/NAME
  python3 preprocess.py dataset/midi/NAME dataset/processed/NAME

• Training

  # Train on .data files in dataset/processed/MYDATA, and save to save/myModel.sess every 10s
  python3 train.py -s save/myModel.sess -d dataset/processed/MYDATA -i 10
  
  # Or...
  python3 train.py -s save/myModel.sess -d dataset/processed/MYDATA -p hidden_dim=1024
  python3 train.py -s save/myModel.sess -d dataset/processed/MYDATA -b 128 -c 0.3
  python3 train.py -s save/myModel.sess -d dataset/processed/MYDATA -w 100 -S 10

  

• Generating

  # Generate with control sequence from test.data and model from save/test.sess
  python3 generate.py -s save/test.sess -c test.data
  
  # Generate with pitch histogram and note density (C major scale)
  python3 generate.py -s save/test.sess -l 1000 -c '1,0,1,0,1,1,0,1,0,1,0,1;3'
  
  # Or...
  python3 generate.py -s save/test.sess -l 1000 -c ';3' # uniform pitch histogram
  python3 generate.py -s save/test.sess -l 1000 # no control
  
  # Use control sequence from processed data
  python3 generate.py -s save/test.sess -c dataset/processed/some/processed.data

  

  

Pretrained Model

• ecomp.sess
  • default configuration
  • dataset: International Piano-e-Competition, recorded MIDI files
• ecomp_w500.sess
  • window_size: 500
  • control_ratio: 0.7
  • dataset: International Piano-e-Competition, recorded MIDI files

Requirements

• pretty_midi
• numpy
• pytorch >= 0.4
• tensorboardX
• progress