Use deep learning to generate and harmonize music in the style of Bach
Switch branches/tags
Nothing to show
Clone or download
Latest commit 6656e86 Oct 22, 2016


BachBot is a research project utilizing long short term memory (LSTMs) to generate Bach compositions.


Docker container images are hosted on DockerHub.


docker pull fliang/bachbot:base


docker pull fliang/bachbot:CUDA-7.5

Getting Started

Set up environment

source scripts/activate

Prepare polyphonic Bach chorale corpus and train torch-rnn LSTM.

bachbot chorales prepare_poly && \
	bachbot concatenate_corpus scratch/*.utf && \
	bachbot make_h5 && \
	bachbot train

Sample the trained LSTM and decode output to musicXML.

bachbot sample <path_to_checkpoint> -t <temperature> && \
	bachbot decode decode_chord_constant_t_utf scratch/utf_to_txt.json scratch/sampled_<temperature>.utf


  • source ./scripts/activate.zsh to set up the working environment
    • Pro-tip: Add source ~/bachbot/scripts/activate to your .{bash|zsh}rc
  • To develop on the scripts
     cd ./scripts
     pip install --editable .
    The bachbot shell command will use the entry point defined inside ./scripts/
  • To generate ctags which include system Python libraries
     ctags -R -f ./tags `python -c "from distutils.sysconfig import get_python_lib; print get_python_lib()"`<CR>