A cute multi-layer LSTM network that can perform like a human 🎶 It learns the dynamics of music! The architecture was specifically designed to handle music of different genres.
If you wish to learn more about my findings, then please read my blog post and paper:
Iman Malik, Carl Henrik Ek, "Neural Translation of Musical Style", 2017.
You will need a few things in order to get started.
- Tensorflow
- mido
- pretty_midi
- fluidsynth
I created my own dataset for the model. If you wish to use the Piano Dataset 🎹 for academic purposes, you can download it from here. The Piano Dataset is distributed with a CC-BY 4.0 license. If you use this dataset, please reference this paper:
python main.py -current_run <name-of-session> -bi
Flags:
-load_last : Loads and continues from last epoch.
-load_model: Loads specified model.
-data_dir : Directory of datasets.
-data_set : Dataset name.
-runs_dir : Directory of session files.
-forward_only : For making predictions (not training).
-bi : If you wish to use bi-directional LSTMs. (HIGHLY recommended)
pianoify.ipynb : This was used to ensure the files across the dataset were consistent in their musical properties.
generate_audio.ipynb : This was used to make predicitions using StyleNet and generate the audio.
convert-format.rb : This was used to convert format 1 MIDIs into format 0.
file_util.py : This contains folder/file-handling functions.
midi_util.py : This contains MIDI-handling functions.
model.py : StyleNet's Class.
data_util.py : For shuffling and batching data during training.