Drum Pattern Generation with Transformer Inpainting

Submitted for ECS7022P Computational Creativity at Queen Mary University of London

Structure

The main.ipynb file contains the majority of this projects implementation, including an interface developed as a playground for exprimenting with the system, as well as the code used to train it. Three other files contain some class and function definitions, split to clean up the main notebook. These are packaged as a python project (see the pyproject.toml file) and include:

• tokeniser.py - Implementations of a number of tokenisation methods which were experimented with (only DrumSequenceEncoder is used in the final system)
• modules.py - Implementation of the custom transformer layers which include a relative position encoding and associated attention blocks.
• utils.py - Helper methods for processing, displaying and playing MIDI data.

Examples

Below are some example outputs of the system, paired with the input patterns which were used as prompts.

Backbeat

Original: backbeat_original.wav

Subtle Variation: backbeat_minimal.wav

Pocket Beat: backbeat_tight.wav

Lots of shifts: backbeat_shifts.wav

Dense sampling: backbeat_dense.wav

Amen

Original: amen_original

Shifting: amen_shifts

Breakbeats: amen_dense

Be my Baby

Original: be_my_baby_original

Syncopated: be_my_baby_syncopated

Improvised: be_my_baby_improvised