You would simply have a vector of 128 for each timestep (I'm not sure how discrete MIDI goes), and any of the 128 instruments turned on would be non-zero. It's

You would simply have a vector of 128 for each timestep (I'm not sure how discrete MIDI goes), and any of the 128 instruments turned on would be non-zero. It's not a dense encoding, by a long shot, but it's quite literal and explicit so you don't have to worry indirectness making it hard to learn (assuming you have a model which can handle such long-range inputs to begin with). Personally, I'd rather try redoing BPE encoding for ABC-MIDI specifically.