Note: this is described in detail here: http://danshiebler.com/2016-08-17-musical-tensorflow-part-two-the-rnn-rbm/

Music_RNN_RBM

This repository contains code for generating long sequences of polyphonic music by using an RNN_RBM in TensorFlow.

TLDR:

You can generate music by cloning the directory and running:

python rnn_rbm_generate.py parameter_checkpoints/pretrained.ckpt

This will populate the music_outputs directory with midi files that you can play with an application like GuitarBand.

Training

To train the model, first run the following command to initialize the parameters of the RBM.

python weight_initializations.py

Then, run the following command to train the RNN_RBM model:

python rnn_rbm_train.py <num_epochs>

num_epochs can be any integer. Set it between 50-500, depending on the hyperparameters.

Generation:

The command:

python rnn_rbm_generate.py <path_to_ckpt_file>

will generate music by using the weights stored in the path_to_ckpt_file. You can use the provided file parameter_checkpoints/pretrained.ckpt, or you can use one of the ckpt files that you create. When you run train_rnn_rbm.py, the model creates a epoch_<x>.ckpt file in the parameter_checkpoints directory every couple of epochs.