A simple LSTM-based music generation system.
The system uses a dataset from jukedeck/nottingham-dataset, containing a total of 1034 songs, stored as a compressed file in the dataset folder, which should be extracted into the dataset folder when you need to use it.
Python: 3.7.9
Keras: 2.3.0
tensorflow-gpu: 2.2.0
music21: 6.7.1
PS: Third party libraries can be installed using the pip install command.
This music generation system can be used to generate music in two ways: generate from scratch or based on existed music.
When there are no symbolic music files in the inputs folder, running midi_generator.py will generate multiple MIDI music tracks from zero into the outputs folder, the number generated depends on the parameter MIDI_NUM.
When there are symbolic music files in the inputs folder, after running midi_generator.py, the system will continuation and generate MIDI music to the outputs folder based on this music, the new generated music contains all the notes of the original song (if the original song is not in C major/A minor, it will be transposed).
1. Store all symbolic music files in the dataset folder, if you only have a small amount of data yourself, you can keep the original dataset of this system;
2. Run encoding_module.py, which will give you the vocabulary vocabulary.json of the dataset, and the corpus corpus.bin;
3. Run rnn_model.py, which will give you the model weights weights.hdf5 after training (estimated time here is 2 hours).
After that, you can use midi_generator.py to generate music that matches the musical style of the new dataset.
If you need to finetune the parameters, you can do so in utils.py. It is not recommended to change the parameters in other files.