jrmylee / sonata

Sonata

Project for chord recognition using deep RNNs. Designed to be used with the Isophonics Dataset(http://isophonics.net/), but can be used with any dataset with labeled chords corresponding to a time interval.

Note: Datasets do not come with corresponding audio due to copyright concerns.

Setup:

1. Clone Dataset
2. Clone BTC Model
3. add config/config.yaml with relevant paths
4. mkdir checkpoints
5. run train.py
6. pip3 install audiomentations numpy pandas pyrubberband librosa pyyaml mir_eval pretty_midi music21 comet_ml sklearn torch torchvision matplotlib numba==0.48
7. apt-get install ffmpeg

Example config(config/config.yaml):

model: 
  path: #PATH TO MODEL
  feature_size : 108
  timestep : 108
  num_chords : 133
  input_dropout : 0.2
  layer_dropout : 0.2
  attention_dropout : 0.2
  relu_dropout : 0.2
  num_layers : 8
  num_heads : 4
  hidden_size : 128
  total_key_depth : 128
  total_value_depth : 128
  filter_size : 128
  loss : 'ce'
  probs_out : False
  sample_rate: 22050
preprocess:
  data_path: #PATH TO DATASET
  sample_rate: 22050
  hop_size: 2048
  window_size: 10
  song_hz: 22050