A seq2seq architecture where a BiLSTM Encoder receives information about the conditioning parameters (i.e., accompanying tracks and musical attributes), while a Transformer-based Decoder with relative global attention produces the generated drum sequences. For further details please read and cite our paper:
Makris D., Guo Z, Kaliakatsos-Papakostas N., Herremans D., “Conditional Drums Generation using Compound Word Representations” to appear EvoMUSART, 2022.
-Python 3.9
-Tensorflow 2.6.0
-music21 6.x or lower
-pretty_midi
-pypianoroll
-numpy
-sklearn
-pre_process.py: Create CP representations using your own MIDI data. However, we offer the pre-processed dataset used in the paper (rar inside the folder). In that case you do not need to run this file.
-post_process.py: Convert the CP data to one-hot encoding streams ready for training.
-model_train.py: Train the model using the hyper-parameters reported in the paper.
-gen_drums.py: Generate conditional drums with your own data. The input MIDI files (./midi_in/) must have two tracks (1st Guitar, 2nd Bass) and with a maximum length of 16 bars. We also offer pre-trained model weights to avoid training here. Place them inside ./aux_files/checkpoints folder.
Random samples from the experimental setup.
If you use this library, please cite the following work: