Generating Irish/folk/classical music in ABC format using GPT-2-117M, with good results.
In November 2019, I experimented with training a GPT-2 neural net model to generate folk music in the high-level ABC music text format, following previous work in 2016 which used a char-RNN trained on a ‘The Session’ dataset. A GPT-2 hypothetically can improve on an RNN by better global coherence & copying of patterns, without problems with the hidden-state bottleneck.
I encountered problems with the standard GPT-2 model’s encoding of text which damaged results, but after fixing that, I successfully trained it on n = 205,304 ABC music pieces taken from The Session & ABCnotation.com. The resulting music samples are in my opinion quite pleasant. (A similar model was later retrained by Geerlings & Meroño-2020.)
The ABC folk model & dataset are available for download, and I provide for listening selected music samples as well as medleys of random samples from throughout training.
We followed the ABC folk model with an ABC-MIDI model: a dataset of 453k ABC pieces decompiled from MIDI pieces, which fit into GPT-2-117M with an expanded context window when trained on TPUs. The MIDI pieces are far more diverse and challenging, and GPT-2 underfits and struggles to produce valid samples but when sampling succeeds, it can generate even better musical samples.