This is a repository for an ongoing project of mine where I try to generate believable chord progression continuations from a prompt. The idea is that if I provide the model with, for instance, a few chords like Am7 - Dm, it would output EM7 - Am - a standard logical continuation. The main issue here is, of course, making the model learn the musical context (tonality, harmonic diversity, etc.) and understand the long-term structure of music. To that end, I've developed three models so far: one employing bidirectional LSTMs, one using Perceiver AR (a long-context autoregressive model), and one using GPT2 and treating the chords as text. I've also written a short script that extracts chords from MIDI files to serve as the source of data for the models.
I've written a short script based on the Chorder package that extracts chords from MIDI files. The script is available in the MIDI_to_chords.ipynb notebook. I've used this script and the Lakh MIDI dataset to download and 'chordify' songs of over 250 Soul artists (via a manually compiled list), giving me chords from over 1700 songs. These chords were used as the data for all three models below. The data is available for download as a .txt or as a pickle file.
The code can be found in the chord_progrression_LSTM.ipynb notebook. The results are rather mixed: the model tends to suggest only the most basic chords (I or V), which do make some sense, but do not quite reflect the diversity of chord progressions in real-world music. Making the model's suggestions "hotter" is a hit-or-miss: sometimes they are quite fitting, but sometimes they are as dissonant as they can get. I tried creating my own embeddings for the chords (representing them as constituent pitches), but that provided only marginal improvement to the model's performance.
Better results are achieved with GPT2, thanks to the model's sophisticated architecture (i.e. implementation of attention). The code is available in the chord_progrression_GPT2.ipynb notebook. However, the results are still not quite stellar: progressions seem to get off-key rather quickly. Though, to be fair, some of them are quite pleasant.
My current approach utilizes a brand new architecture - Perceiver AR, developed by the Deep Mind team and implemented in PyTorch by Phil Wang on GitHub. This auto-regressive model is able to attend to very long context windows and generate outputs with long-term coherence. So far I haven't managed to properly fine-tune the model parameters for this specific task and the model is prone to overfitting, but the results are encouraging nonetheless.