autodrummer

A text-to-audio model for generating text-conditioned drum beats.

how it works

autodrummer is a transformer model created by converting the MIDI data from Google Magenta's E-GMD dataset to a specialized variant of plaintext upon which to train a transformer. This repo covers everything necessary to convert the raw MIDI data from E-GMD into tokens for the transformer.

installation

1. Download the E-GMD dataset and place it in this repo in a folder named e-gmd
2. git clone https://github.com/angelfaraldo/pymidifile within this repo and replace its quantize.py with this repo's version of quantize.py. you have to run quantize.py from the autodrummer repo directory, not the pymidifile repo directory, or else you'll get import errors.
3. run pip install -r requirements.txt

current workflow (to prepare data for your own finetune)

1. pymidifile/quantize.py: quantize the MIDI data in the e-gmd dataset and convert it to matrices representing (drum_code, hit_time) pairs
2. matrix2text.py: convert said matrices into plaintext
3. df2jsonl.py: convert dataframe of plaintext to jsonl file
4. train model here
5. txt2audio.py or evaluator.py: convert back from plaintext to audio

contact

Jasper on Twitter