jerrylzy / RNN_Jazzy_Haydn

An LSTM RNN for Jazz Music Generation -- Trained on Piano Sonata No. 48, Finale

RNN Jazzy Hadyn

This is an LSTM-based Recurrent Neural Network trained on Piano Sonata No. 48 III Finale.

The training file can be found at kunstderfuge.com.

• All preprocessing files are modified from Ji-Sung Kim's work deepjazz.
• The model is from Coursera.

This is more of an exploration rather than a serious project.

Environment Setup

1. Download and install Anaconda

2. Create and activate environment

   conda update -n base conda
   conda create --name ml numpy scipy h5py jupyter keras
   source activate ml
   pip install --upgrade pip
   pip install music21 matplotlib

3. Install Tensorflow. For example

   pip install --ignore-installed --upgrade https://storage.googleapis.com/tensorflow/mac/cpu/tensorflow-1.6.0-py3-none-any.whl

4. Launch the Jupyter Notebook

   jupyter notebook

5. Run the notebook Jazzy Haydn.ipynb

6. Get the result from output/my_music.midi

Data Preprocessing

Most code is adopted from Ji-Sung Kim's work which is based on Evan Chow's jazzml. However, some parts are tailored specifically for his training data.

• data_utils.py: moved function generate_music to the jupyter notebook. Use variable N_tones for num_classes everywhere.
• grammar.py: return None if either the list measure or chords has a length 0.
• inference_code.py: Use the same N_tones as the one in data_utils.py.
• music_utils.py: moved the one_hot helper function to the jupyter notebook.
• preprocess.py: Multiple changes to the __parse_midi function to make it compatible with our training data.

Model

We adopted the model from Coursera. It is a one-to-many LSTM RNN with 64 memory cells. It uses the output of the previous cell as the input of the current cell.

We train this model with 100 epoches, and generate 50 new values from the network.