A project that trains a LSTM recurrent neural network over a data-set of MIDI files.
Numpy (http://www.numpy.org/)
Tensorflow 0.8 (https://github.com/tensorflow/tensorflow) its best to uses virtual env to do this
virtual env (https://virtualenv.pypa.io/en/stable/)
Python Midi (https://github.com/vishnubob/python-midi.git)
Mingus (https://github.com/bspaans/python-mingus)
Matplotlib (https://github.com/matplotlib/matplotlib)
mkdir data && mkdir models- run
python main.py. This will collect the data, create the chord mapping file in data/nottingham.pickle, and train the model - Run
python rnn_sample.py --config_file new_config_file.configto generate a new MIDI song.
Give it 1-2 hours to train on your local machine, then generate the new song. You don't have to wait for it to finish, just wait until you see the 'saving model' message in terminal. To increase speed you can use a cloud based GPU such as www.fomoro.com
This is an adaptation of an idea from Yoav Zimmerman. The main use of this project is as a proof of concept for a RNN.