junekihong/music-generation

Playing around with Keras to try and read midi files.

Juneki Hong ([email protected]) April, 2016

Computer Music Generation

Set up

• run the script

  • ./compile_allegro.sh

  • to compile the allegro codebase (which this project relies on)

  • This should compile the allegro directory and produce two binaries: midi2gro, gro2midi

Prepare Training Data

• Prepare midi files as training data (1st step)

  • ./processMIDIfiles2GRO.sh /path/to/midi/data/*.midi

• Prepare midi files as training data (2nd step)

  • ./processGROfiles2ENCODED.sh gro/*.gro

• Prepare midi files as training data (3rd step)

  • cat encoded/*.encoded > train/train.encoded.dat

Running the Project

• To train a model:

  • python keras/lstm_gro_generation.py train/train.encoded.dat

• To run the model to produce a sample song:

  • python keras/decode_random.py train_model_arch.json train_model_weights.h5 train/train.encoded.dat

• To decode the sample song out to a midi file:

  • python decode.py output.encoded.txt

  • ./gro2midi decoded/output.encoded.decoded output/

  • timidity output/output.encoded.mid